mirror of
https://github.com/norandom/OpenUltraSAST.git
synced 2026-10-16 03:49:26 +00:00
chore(p0): remove the improve lever and the plane improvement loop
Off every active line (PATHFINDER-2026-10-13/03-unified-proposal.md, cuts). Deleted: src/openultrasast/improve/ (1,175 lines), the ousast improve subcommand, plane/tasks/loop.py and loop.yaml with the --loop generation, the memory-snapshot step in memory.seed, pairs.build_pair_signals, and the four improve tests plus the loop cases in other tests. Docs no longer describe the lever. Net: 3,073 lines deleted, 83 added. Targeted tests: 154 passed; 3,436 tests collect. Co-Authored-By: Claude Fable 5.1 <[email protected]>
This commit is contained in:
Marius Ciepluch
co-authored by
Claude Fable 5.1
Claude Fable 5.1
parent
d2010d0693
commit
94e96c6885
@@ -217,8 +217,7 @@ checks, plus a configuration census. A model answers only the residual question.
|
||||
code it comes from and what it misses.
|
||||
- [docs/architecture.md](docs/architecture.md) gives the stage order and the evidence ladder (a
|
||||
state machine, not a label a model may assign). It also covers the central CWE policy, the
|
||||
project score and the self-improving loops (`ousast improve`, with `--memory` for proposals from
|
||||
plane rows).
|
||||
project score.
|
||||
|
||||
Quick rules cover C/C++, Python, JavaScript/TypeScript, Java with Groovy templates, and PHP. Each
|
||||
rule runs only against files of its language.
|
||||
@@ -311,7 +310,6 @@ the fold discipline that keeps learned decisions out of their own training data.
|
||||
uv run ousast benchmark benchmarks/manifests/python-vulnerable.toml --mode quick
|
||||
uv run ousast pairs --slice sast # fire on the vulnerable side, stay silent on the fix
|
||||
uv run python -m openultrasast.pair_gate # CI: local pairs must all pass
|
||||
uv run ousast improve benchmarks/manifests/java-spring-boot-vulnerable.toml --dry-run --memory
|
||||
```
|
||||
|
||||
## Token economics
|
||||
|
||||
+2
-55
@@ -5,7 +5,7 @@ This page covers four things:
|
||||
- how a scan works;
|
||||
- how a finding earns its place;
|
||||
- how severity and the project score are decided;
|
||||
- how the ruleset improves itself.
|
||||
- how scan calibration works.
|
||||
|
||||
State as of 2026-10-02; module paths are under `src/openultrasast/`.
|
||||
|
||||
@@ -141,9 +141,7 @@ deleted. The score gate works like this:
|
||||
|
||||
The `score` stage writes `score.json` and merges it into `manifest.json`.
|
||||
|
||||
## Self-improvement
|
||||
|
||||
### Scoped calibration (every scan)
|
||||
## Scoped calibration (every scan)
|
||||
|
||||
```
|
||||
scan ─▶ verify ─▶ non-accepted findings ─▶ scoped learnings
|
||||
@@ -158,54 +156,3 @@ next scan ─▶ rank ─▶ calibrate (demote rejected scopes) ─▶ findings
|
||||
|
||||
Demotion is scoped and reversible (`tests/test_pipeline_calibration.py`).
|
||||
|
||||
### Rule-level loop: `ousast improve`
|
||||
|
||||
Each round (`improve/evolve.py`) runs these steps:
|
||||
|
||||
1. benchmark with the current ledger;
|
||||
2. per-rule signals (`rule_signals.json`);
|
||||
3. bounded proposals;
|
||||
4. validation (`improve/validator.py`);
|
||||
5. replay smoke;
|
||||
6. re-benchmark;
|
||||
7. a hard gate.
|
||||
|
||||
The levers are rule status (`enabled -> shadow` before `disabled`) and score constants. The
|
||||
loop can never edit pattern text or the authoritative CWE severity. A round is accepted only if
|
||||
all of these hold:
|
||||
|
||||
- recall >= 90%;
|
||||
- FP < 10%;
|
||||
- the project score did not regress;
|
||||
- no matched finding was lost.
|
||||
|
||||
Otherwise the ledger is not written. With `--pair-catalog`, each round is also gated per
|
||||
provenance profile on the catalog's holdout pairs. `--no-pair-gate` skips this check; do not
|
||||
use it for accepted ledgers. The same detection gate runs in CI (`python -m openultrasast.gate`).
|
||||
|
||||
### Proposals from plane memory
|
||||
|
||||
With `--memory [STORE]`, `ousast improve` also reads the rows that plane runs stored per
|
||||
repository and pin (`improve/memory.py`). Two deterministic rules turn these rows into
|
||||
rule-status proposals:
|
||||
|
||||
- a rule whose alerts are repeatedly false across repositories is shadowed;
|
||||
- a shadow rule that keeps hitting agreed, declared vulnerable sites is re-enabled.
|
||||
|
||||
Some rows are dropped before any rule sees them: rows from holdout pairs, from the gated
|
||||
manifest's own cases and from any `--qualify-population`. Each proposal records the rows it
|
||||
came from (`memory_proposals.jsonl` next to the journal). It then goes through the same
|
||||
validator and gate.
|
||||
|
||||
### The loop as a plane Run
|
||||
|
||||
`ousast plane workspaces <population> --validation-set ... --loop` appends the improvement loop
|
||||
to a generated Run. The Run holds `alerts` per case: the quick scan on the vulnerable and fixed
|
||||
pins. Then come the steps `snapshot` (task `memory-snapshot`), `measure`, `propose` and
|
||||
`improve` (`plane/tasks/loop.py`). Each step is model-free with a zero budget. The snapshot
|
||||
applies the same train-on-test guard before any row leaves the store. A recorded live run:
|
||||
`benchmarks/measurements/2026-09-30-harnessx-removal-loop/`.
|
||||
|
||||
When that run was made, quick mode covered PHP with zero rules. So for PHP, `ousast plane
|
||||
alerts-engine` produces a Run's `alerts` from the engine image on the host instead
|
||||
(`benchmarks/measurements/2026-09-30-php-engine-alerts/`).
|
||||
|
||||
+3
-4
@@ -77,8 +77,7 @@ Terms in the table:
|
||||
- **gVisor** is a sandbox runtime that isolates each actor from the node's kernel.
|
||||
- An **actor** is one running Task inside a sandbox.
|
||||
|
||||
Only the reconciler talks to the memory store. A sandboxed task cannot read it. So the reconciler seeds
|
||||
the loop's `memory-snapshot` before the Run starts ([The plane on ax](plane.md)).
|
||||
Only the reconciler talks to the memory store. A sandboxed task cannot read it.
|
||||
|
||||
## A. This host: the local setup (implemented)
|
||||
|
||||
@@ -478,7 +477,7 @@ Each Task manifest under `plane/tasks/` names the same image by digest. The pin
|
||||
`~/.cache/ousast/ax-src/runner-image`. The manifest picks its work with `spec.command`. The runner
|
||||
starts `python -m openultrasast.plane.tasks.<command[0]>` (`runner.py`, `task_module`).
|
||||
|
||||
`repo-facts`, `verify`, `agree`, `features`, `remember`, `alerts`, `loop` and `roles` are modules in
|
||||
`repo-facts`, `verify`, `agree`, `features`, `remember`, `alerts` and `roles` are modules in
|
||||
that image. They are not images of their own. The engine would be the second image (planned).
|
||||
|
||||
So the production answer to "how many containers" is:
|
||||
@@ -594,7 +593,7 @@ The worker is the unit. Every Task's `limits` fit the 1 CPU / 1.5 GiB worker of
|
||||
| `verify` | 250m / 384Mi | 1 CPU / 1Gi | 2 (`va`, `vb`), plus 1 (`vc`) on disputed candidates |
|
||||
| `agree` | 100m / 256Mi | 1 CPU / 512Mi | 1 (`agree`; `final` reuses it after `vc`) |
|
||||
| `features`, `remember` | 250m / 512Mi; 100m / 256Mi | 1 CPU / 1Gi; 1 CPU / 512Mi | optional, for the decision engine and the memory |
|
||||
| `alerts`, `loop` | 250m / 512Mi | 1 CPU / 1280Mi | not in a push Run (loop Runs only) |
|
||||
| `alerts` | 250m / 512Mi | 1 CPU / 1280Mi | not in a push Run |
|
||||
| `roles` | 250m / 384Mi | 1 CPU / 1Gi | not in a push Run (harvest Runs only) |
|
||||
| `engine` (planned) | from a measured run | above 3 GiB (the host run uses `--memory 3g`) | its own pool, not the default one |
|
||||
|
||||
|
||||
@@ -46,8 +46,7 @@ Rules live in `ruleset/<language>/rules.toml` as `[[rule]]` tables. Each has the
|
||||
- optional `framework`/`library`/`fallback_for` tags.
|
||||
|
||||
The fields are defined in `ruleset/store.py` (`ruleset/store.py:12,
|
||||
19-41`). The self-improvement loop may change only `status`, `min_evidence_level` and
|
||||
`precision_estimate`. The pattern is "human-PR-only" (`store.py:23-25`). There are 52 rules:
|
||||
19-41`). Patterns and rule policy change through reviewed commits. There are 52 rules:
|
||||
C 6, Groovy 1, Java 4, JavaScript 12, Python 11, PHP 18, of which 6 PHP rules are `shadow`.
|
||||
Shadow rules are written to `shadow_findings.json` and removed from the report
|
||||
(`cli.py:538-543`).
|
||||
|
||||
+1
-1
@@ -143,5 +143,5 @@ per-family tables are on [decision-engine.md](decision-engine.md).
|
||||
## Removal of the earlier agentic extra
|
||||
|
||||
The earlier agentic extra was removed on 2026-09-30. The removal was checked for byte-identical
|
||||
deterministic outputs: pair replay, quick benchmark, standard scan, `improve --dry-run`. Records:
|
||||
deterministic outputs: pair replay, quick benchmark, standard scan. Records:
|
||||
`benchmarks/measurements/2026-09-30-harnessx-removal-baseline/` and `-equality/`.
|
||||
|
||||
+3
-15
@@ -55,19 +55,10 @@ uv run python -m openultrasast.pair_gate # CI: local pairs must all pass
|
||||
Each case is two isolated trees: the vuln file and the patched file, at the same relative path.
|
||||
|
||||
- `pair_correct` means the harness fired on the vuln side and stayed silent on the fix.
|
||||
- Misses become `miss` signals and fix-side leaks become `fp` signals for `ousast improve`.
|
||||
- Misses become `miss` signals and fix-side leaks become `fp` signals in the pair report.
|
||||
- The catalog and public dataset pointers live in `benchmarks/pairs/`.
|
||||
|
||||
## 6. Self-improve the ruleset from benchmark feedback
|
||||
|
||||
```bash
|
||||
uv run ousast improve benchmarks/manifests/java-spring-boot-vulnerable.toml --dry-run # preview
|
||||
uv run ousast improve benchmarks/manifests/java-spring-boot-vulnerable.toml # apply
|
||||
# Accepted edits land in <target>/.openultrasast/calibration/rule_policy.json,
|
||||
# which the next scan/benchmark of that target loads automatically.
|
||||
```
|
||||
|
||||
## 7. Running the agentic plane
|
||||
## 6. Running the agentic plane
|
||||
|
||||
Model work over whole repositories runs as a Run manifest on ax. Each task gets one isolated
|
||||
actor. Bring-up is in [ops/ax/README.md](ops/ax/README.md).
|
||||
@@ -80,9 +71,6 @@ uv run ousast plane doctor # kind, Agent Substrate
|
||||
uv run ousast plane run plane/runs/validation-46.yaml # per case: repo-facts, verify a/b/c, agree, final, features, remember
|
||||
uv run ousast plane status validation-46 # per-task status and token attribution
|
||||
uv run ousast plane remember validation-46 # re-ingest the run's memory rows (OUSAST_MEMORY)
|
||||
|
||||
# Feed the stored rows back into the ruleset loop (same validator and gate)
|
||||
uv run ousast improve benchmarks/manifests/java-spring-boot-vulnerable.toml --dry-run --memory
|
||||
```
|
||||
|
||||
Each task in the Run carries its own Model and `budget: {usd, calls}`.
|
||||
@@ -92,7 +80,7 @@ Each task in the Run carries its own Model and `budget: {usd, calls}`.
|
||||
- `ousast plane status` prints the token attribution table. It shows calls, prompt, cache-hit and
|
||||
output tokens, and USD per task.
|
||||
|
||||
## 8. Drive it from an MCP client (OpenCode / IDE)
|
||||
## 7. Drive it from an MCP client (OpenCode / IDE)
|
||||
|
||||
MCP (Model Context Protocol) lets an editor or agent call the tool's functions.
|
||||
|
||||
|
||||
+1
-7
@@ -28,9 +28,6 @@ flowchart LR
|
||||
scan --> report["Report: findings with evidence, SARIF, score"]
|
||||
scan -- "alerts on vulnerable and fixed pins" --> plane["Plane on ax: repo-facts, verify a/b/c, agree, features"]
|
||||
plane -- "remember" --> mem[("Memory store: rows, facts, examples, blobs")]
|
||||
mem -- "improve --memory: rule-status proposals" --> gate{"Validator and detection gate"}
|
||||
gate -- "accepted round" --> ledger["Ruleset ledger: rule status"]
|
||||
ledger --> scan
|
||||
labels["Labels from ground truth"] --> mem
|
||||
mem -- "retrieval of similar labelled cases" --> engine["Decision engine: compiled AI classifier"]
|
||||
engine -. "intended, not adopted: BLOCK / ADVISORY" .-> report
|
||||
@@ -41,10 +38,7 @@ flowchart LR
|
||||
- The **plane** runs model work per case: verify passes and roles. It also runs model-free
|
||||
bookkeeping per case: facts, agreement and feature records. Its `remember` task turns a case's
|
||||
artifacts into rows of the [memory store](memory.md).
|
||||
- Memory feeds two consumers:
|
||||
1. `ousast improve --memory` turns stored rows into rule-status proposals. They pass through the
|
||||
same validator and gate as every other edit.
|
||||
2. The decision engine retrieves labelled examples from the same store.
|
||||
- The decision engine retrieves labelled examples from the memory store.
|
||||
- The decision engine's output (dotted line) is the intended path. Nothing user-facing reads it
|
||||
today.
|
||||
|
||||
|
||||
@@ -149,37 +149,6 @@ Calibration **holds** when the slope interval contains 1, the intercept interval
|
||||
**BLOCK** is offered only when calibration holds and a fold reaches the required precision. Otherwise
|
||||
the engine offers **ADVISORY** only.
|
||||
|
||||
## The loop: memory improves decisions without retraining
|
||||
|
||||
```mermaid
|
||||
flowchart LR
|
||||
new["New labels and adjudications"] --> labels["ousast learn labels"]
|
||||
labels --> build["ousast learn memory build, embed"]
|
||||
build --> mem[("Memory store: example rows")]
|
||||
mem --> retr["Retrieved per candidate at decision time"]
|
||||
retr --> prog["Same compiled program"]
|
||||
prog --> dec["Decision"]
|
||||
runs["Plane runs: alert, verdict, coverage rows"] --> mem2[("Memory store: run rows")]
|
||||
mem2 --> imp["ousast improve --memory: M1, M2 proposals"]
|
||||
imp --> gate{"Unchanged validator and gate"}
|
||||
gate -- "accepted round" --> ledger["Rule status in the ledger"]
|
||||
```
|
||||
|
||||
- **Decisions.**
|
||||
- The compiled part of the program is its instruction and demonstrations.
|
||||
- The examples are retrieved from memory for each candidate.
|
||||
- So a new `example` row is eligible for retrieval in the next decision without recompiling. The same
|
||||
evaluation boundary applies.
|
||||
- `ousast learn curve` measures the metric over memory size on a fixed candidate subset.
|
||||
- **Rules.** `ousast improve --memory` reads the run rows (alerts, verdicts, coverage). It proposes
|
||||
rule-status edits only:
|
||||
- shadow a rule that is repeatedly false across repositories;
|
||||
- re-enable a shadow rule that keeps hitting agreed declared sites.
|
||||
|
||||
Some rows are dropped first: rows from holdout pairs, the gated manifest's own cases and any
|
||||
`--qualify-population`. Every proposal goes through the same validator and gate as any other edit
|
||||
(`improve/memory.py`).
|
||||
|
||||
## What the first slice measured
|
||||
|
||||
Record: `benchmarks/measurements/2026-10-01-decision-engine-injection-slice/record.json` (counts
|
||||
|
||||
+1
-14
@@ -25,10 +25,9 @@ replaces itself. The kinds (`KINDS`) and who writes them:
|
||||
| `facts` | `remember` task | the facts entry of a case (points at `facts/<sha256>.json`; reused by memory key) |
|
||||
| `verdict` | `remember` task | per candidate: each verify pass's answer, the final outcome, whether a tie-break ran, cost |
|
||||
| `unit_cost` | `remember` task | per verify pass and file: usd, calls and token usage |
|
||||
| `alert` | `remember` task | a quick-scan or engine alert on the vulnerable or fixed pin (loop Runs) |
|
||||
| `alert` | `remember` task | a quick-scan or engine alert on the vulnerable or fixed pin |
|
||||
| `coverage` | `remember` task | per language and pin: whether any rule could have fired there |
|
||||
| `features` | `remember` task, from the `features` task | one feature record per candidate and family |
|
||||
| `proposal_outcome` | `ousast improve --memory` (not with `--dry-run`), and the loop's `improve` step | what happened to each memory proposal in a round |
|
||||
| `example` | `ousast learn memory build` | a labelled case: label + feature record + excerpt sha (no rationale) |
|
||||
| `experiment` | `ousast learn experiment register` (task `register`) and `run` (task `run`) | a registered manifest's digests and provenance; a run's counts-only summary. Repo `experiments/<id>`, pin the registering commit |
|
||||
| `arm_outcome` | `ousast learn experiment run` | one decision per (unit, arm, replicate): score, verdict, usd, order |
|
||||
@@ -52,8 +51,6 @@ flowchart LR
|
||||
tasks["Plane tasks: facts, verify a/b/c, agree, alerts, features"] --> rem["remember task: memory.jsonl"]
|
||||
rem --> recv["Delivered to the host with the run's artifacts"]
|
||||
recv --> ingest["ingest after ousast plane run, or ousast plane remember RUN"]
|
||||
loop["Loop improve step: proposal_outcome"] --> ingest
|
||||
improve["ousast improve --memory: proposal_outcome"] --> put
|
||||
build["ousast learn memory build: labels + features + excerpts"] --> put["put_rows, put_blob"]
|
||||
embed["ousast learn memory embed"] --> put
|
||||
compile["ousast learn compile, evaluate"] --> put
|
||||
@@ -79,26 +76,16 @@ flowchart LR
|
||||
flowchart LR
|
||||
store[("Memory store")] --> seed["seed before a Run"]
|
||||
seed --> reuse["Facts reused by memory key: repo, pin, candidates, runner image"]
|
||||
seed --> snap["memory-snapshot: rows after the train-on-test guard"]
|
||||
store --> examples["learn: load example rows, excerpts, embeddings"]
|
||||
examples --> retr["Retrieval of similar labelled cases"]
|
||||
store --> responses["Response cache: replay at no cost"]
|
||||
store --> imp["ousast improve --memory: alert, verdict, coverage rows"]
|
||||
imp --> guard["Train-on-test guard"]
|
||||
guard --> rules["Rules M1, M2 to rule-status proposals"]
|
||||
rules --> gate{"Unchanged validator and gate"}
|
||||
```
|
||||
|
||||
- **Seed** (`memory.seed`).
|
||||
- A Run task gets stored facts when its `openultrasast.io/memory-key` matches them. It is then marked done.
|
||||
- A changed pin, candidate set or runner image finds nothing and recomputes.
|
||||
- The loop's `memory-snapshot` is seeded the same way. A sandboxed task cannot read the store.
|
||||
- **Retrieval** (`learn/retrieve.py`): see
|
||||
[Memory and detection](memory-and-detection.md#3-retrieval-of-similar-labelled-cases).
|
||||
- **`improve --memory`** (`improve/memory.py`). Some rows are dropped before any rule sees them:
|
||||
rows from holdout pairs, from the gated manifest's own cases, and from any `--qualify-population`. See
|
||||
[Architecture](architecture.md#proposals-from-plane-memory).
|
||||
|
||||
## Backends: FileStore and the S3 store
|
||||
|
||||
`OUSAST_MEMORY` selects the backend (`open_store`):
|
||||
|
||||
+3
-23
@@ -103,11 +103,10 @@ What each step guarantees:
|
||||
The first resume of a new image can time out while the snapshot is built. So the reconciler retries it
|
||||
for up to `OUSAST_RESUME_TIMEOUT` seconds (900).
|
||||
|
||||
## The Runs: validation and loop
|
||||
## Validation Runs
|
||||
|
||||
`ousast plane workspaces POPULATION --validation-set ...` (`generate.py`) writes one chain per case.
|
||||
`--loop` adds the `alerts` step per case and the loop's singleton steps. Every step is model-free
|
||||
except the verify passes (and `roles` in harvest Runs).
|
||||
Every step is model-free except the verify passes (and `roles` in harvest Runs).
|
||||
|
||||
### Per case
|
||||
|
||||
@@ -122,33 +121,15 @@ flowchart LR
|
||||
vb --> final
|
||||
vc --> final
|
||||
final --> features["features: one record per candidate"]
|
||||
alerts["alerts: quick scan on vulnerable and fixed pins (--loop)"] -.-> features
|
||||
features --> remember["remember: memory.jsonl"]
|
||||
alerts -.-> remember
|
||||
final --> remember
|
||||
```
|
||||
|
||||
### The loop (with `--loop`)
|
||||
|
||||
```mermaid
|
||||
flowchart LR
|
||||
store[("Memory store")] -- "seeded before the Run, after the train-on-test guard" --> snap["memory-snapshot"]
|
||||
rem["remember of every case"] --> measure["loop-measure"]
|
||||
snap --> measure
|
||||
measure --> propose["loop-propose: rules M1 and M2, M3 advisory"]
|
||||
snap -- "index.json" --> propose
|
||||
propose --> improve["loop-improve: one evolve round, unchanged validator and gate"]
|
||||
improve --> outcome["gate.json, journal.json, and rule_policy.json only if accepted"]
|
||||
```
|
||||
|
||||
- **Before the Run**, `ousast plane run` seeds from the store (`memory.seed`).
|
||||
- Each task has an `openultrasast.io/memory-key`: repository, pin, candidates digest and runner image digest.
|
||||
- If the store has facts for that key, the task gets them and is marked done.
|
||||
- The sandboxed `memory-snapshot` task gets the store's rows after the guard.
|
||||
- **After the Run**, it ingests every delivered `memory.jsonl`. `ousast plane remember RUN` repeats
|
||||
this by hand. See [Memory](memory.md).
|
||||
- [Architecture](architecture.md#proposals-from-plane-memory) describes the loop's rules and its guard.
|
||||
Adopting an accepted ledger stays a maintainer commit (`src/openultrasast/plane/tasks/loop.py`).
|
||||
- `ousast plane harvest` writes the decision engine's harvest Runs (verify a/b with agree, and
|
||||
model `roles`).
|
||||
- `ousast plane alerts-engine` produces a Run's `alerts` from the Joern engine image on the host. It
|
||||
@@ -174,8 +155,7 @@ draws it and sizes the pre-push hook at scale.
|
||||
| `agree` | none | a and b agreed or disputed; after pass c on the disputed, 2-of-3 (the step `final`) |
|
||||
| `features` | none | one feature record per candidate for the decision engine |
|
||||
| `remember` | none | the case's artifacts as memory rows |
|
||||
| `alerts` | none | the quick scan on the vulnerable and fixed pins (loop Runs) |
|
||||
| `loop` | none | the steps `snapshot`, `measure`, `propose`, `improve` |
|
||||
| `alerts` | none | the quick scan on the vulnerable and fixed pins |
|
||||
| `roles` | `deepseek-flash` | per-repository source, sink and sanitizer roles without a vocabulary |
|
||||
|
||||
Source: [ax on this host](ops/ax/README.md#tasks).
|
||||
|
||||
@@ -14,10 +14,8 @@ It also lists the controls that keep a scan safe to run in CI and safe to share.
|
||||
- Only `deep` mode executes target-derived code. Its REGRESS stage loads promoted candidates
|
||||
inside the Docker sandbox described below. It runs only when `docker
|
||||
info` succeeds (otherwise the stage is skipped and recorded as `sandbox_unavailable`).
|
||||
- **The ruleset and CWE policy are trusted, governed data.** They change in only two ways:
|
||||
through reviewed commits, or through the bounded self-improvement loop (`ousast improve`). The
|
||||
loop can flip rule *status* and tune score constants. It can never edit pattern text or the
|
||||
authoritative 0–5 severity. See [architecture.md](architecture.md#self-improvement).
|
||||
- **The ruleset and CWE policy are trusted, governed data.** They change
|
||||
through reviewed commits.
|
||||
- **Provider credentials are the operator's.** The core install (PyYAML only) makes no network
|
||||
calls on its own.
|
||||
- A local scan calls a model only in standard or deep mode, and only when a key is available.
|
||||
|
||||
+3
-4
@@ -74,12 +74,11 @@ and to the hosts the bound Model declares (`openultrasast.io/egress-hosts`). The
|
||||
| `agree` | none | a and b agreed or disputed; after pass c on the disputed, 2-of-3 (the step `final`) |
|
||||
| `features` | none | one feature record per candidate for the decision engine |
|
||||
| `remember` | none | the case's artifacts as memory rows (`facts`, `verdict`, `unit_cost`, `alert`, `coverage`, `features`) |
|
||||
| `alerts` | none | the quick scan on the vulnerable and fixed pins (loop Runs) |
|
||||
| `loop` | none | the steps `snapshot`, `measure`, `propose`, `improve`: the improvement loop as a Run |
|
||||
| `alerts` | none | the quick scan on the vulnerable and fixed pins |
|
||||
| `roles` | `deepseek-flash` | per-repository source, sink and sanitizer roles without a vocabulary (decision-engine harvest) |
|
||||
|
||||
Generators: `ousast plane workspaces <population> --validation-set ...` writes the per-case chain
|
||||
(`facts -> va, vb -> agree -> vc -> final -> features -> remember`, plus `alerts` and the loop with `--loop`);
|
||||
(`facts -> va, vb -> agree -> vc -> final -> features -> remember`);
|
||||
`ousast plane harvest --labels ... --plane DIR` writes the decision engine's harvest Runs; `ousast plane
|
||||
alerts-engine` produces a Run's `alerts` from the Joern engine image on the host, for PHP and other languages
|
||||
quick mode does not cover.
|
||||
@@ -98,7 +97,7 @@ is skipped by the index. A facts entry is reused for the same repository, pin, c
|
||||
tested one). Endpoint and credentials come from `.env` or the environment (`S3_ENDPOINT`, `AWS_ACCESS_KEY_ID`,
|
||||
`AWS_SECRET_ACCESS_KEY`), never from a manifest, and are never printed.
|
||||
|
||||
`ousast improve --memory [STORE]` and the `ousast learn` commands (`--memory`) read the same store.
|
||||
The `ousast learn` commands (`--memory`) read the same store.
|
||||
|
||||
## Measured footprint (2026-09-29, idle, two workers)
|
||||
|
||||
|
||||
@@ -1,18 +0,0 @@
|
||||
# loop (harnessx-removal Req 6.3, design section 6): the improvement loop's singletons -- the generator appends the
|
||||
# step (snapshot, measure, propose, improve) to the command. No model, budget {usd: 0, calls: 0}; improve replays the
|
||||
# benchmark gate unchanged and delivers its verdict as gate.json. The image is digest-pinned.
|
||||
apiVersion: ax.io/v1alpha1
|
||||
kind: Task
|
||||
metadata:
|
||||
name: loop
|
||||
atespace: default
|
||||
spec:
|
||||
image: "localhost:5001/ousast-runner@sha256:d2a55293999cc34581ef1f02a1ffb0c2dd8fc0cbe001ac952df81600e868f019"
|
||||
command: ["loop"]
|
||||
resources:
|
||||
requests:
|
||||
cpu: "250m"
|
||||
memory: "512Mi"
|
||||
limits:
|
||||
cpu: "1"
|
||||
memory: "1280Mi"
|
||||
+4
-182
@@ -6,7 +6,6 @@ import logging
|
||||
import os
|
||||
import subprocess
|
||||
import sys
|
||||
import tempfile
|
||||
import time
|
||||
from collections.abc import Mapping, Sequence
|
||||
from dataclasses import asdict, dataclass, replace
|
||||
@@ -39,10 +38,8 @@ from .complexity.ledger import persist_verdicts
|
||||
from .config import ModelLayerConfig, ObligationsConfig, ResolvedConfig, RetiredConfigError, load_config, load_dotenv
|
||||
from .findings import StaticFinding, quick_scan_findings, uncovered_languages, write_findings
|
||||
from .fusion import FusionDecision, fuse_findings
|
||||
from .gate import FALSE_POSITIVE_CEILING, RECALL_FLOOR
|
||||
from .harness import HarnessRuntime, HarnessTraceWriter, write_harness_config
|
||||
from .hunter import run_hunter_pool, write_hunter_trajectories
|
||||
from .improve import RoundOutcome, run_improvement
|
||||
from .learn.cli import add_commands as add_learn_commands
|
||||
from .learn.cli import run as run_learn
|
||||
from .mapping import analyze_entry_points, attach_reachability_hints, ingest_sarif, write_entry_points, write_static_hints
|
||||
@@ -234,55 +231,9 @@ def _main(argv: list[str] | None) -> int:
|
||||
benchmark.add_argument("--mode", choices=("quick", "standard", "deep"), default="quick")
|
||||
benchmark.add_argument("--config", type=Path, default=Path("openultrasast.toml"))
|
||||
|
||||
improve = subparsers.add_parser(
|
||||
"improve",
|
||||
help="run the bounded self-improvement loop against a benchmark manifest and update the loop-owned ruleset ledger",
|
||||
)
|
||||
improve.add_argument("manifest", type=Path)
|
||||
improve.add_argument("--max-rounds", type=int, default=5)
|
||||
improve.add_argument("--recall-floor", type=float, default=RECALL_FLOOR)
|
||||
improve.add_argument("--fp-ceiling", type=float, default=FALSE_POSITIVE_CEILING)
|
||||
improve.add_argument(
|
||||
"--ledger", type=Path, default=None, help="ruleset ledger path (default: <target>/.openultrasast/calibration/rule_policy.json)"
|
||||
)
|
||||
improve.add_argument("--journal", type=Path, default=None, help="improvement journal path (default: alongside the ledger)")
|
||||
improve.add_argument(
|
||||
"--ruleset-dir", type=Path, default=DEFAULT_RULESET_DIR, help="ruleset directory to improve (default: the bundled ruleset)"
|
||||
)
|
||||
improve.add_argument("--dry-run", action="store_true", help="run rounds against a throwaway ledger; never touch the target's ledger")
|
||||
improve.add_argument(
|
||||
"--pair-catalog",
|
||||
type=Path,
|
||||
default=DEFAULT_CATALOG,
|
||||
help="pair catalog whose holdout split gates every round per provenance profile",
|
||||
)
|
||||
improve.add_argument(
|
||||
"--no-pair-gate", action="store_true", help="skip the per-profile holdout clause (faster; not for accepted ledgers)"
|
||||
)
|
||||
improve.add_argument(
|
||||
"--profile-tolerance", type=float, default=0.0, help="allowed per-profile drop in pair_correct/Youden before rejecting"
|
||||
)
|
||||
improve.add_argument("--min-holdout-pairs", type=int, default=5, help="profiles with fewer holdout pairs are reported, not gated")
|
||||
improve.add_argument(
|
||||
"--memory",
|
||||
nargs="?",
|
||||
const="",
|
||||
default=None,
|
||||
metavar="STORE",
|
||||
help="also propose rule-status edits from the plane memory store (a directory, file:// or s3:// URL; "
|
||||
"no value: OUSAST_MEMORY or the default store); off by default",
|
||||
)
|
||||
improve.add_argument(
|
||||
"--qualify-population",
|
||||
action="append",
|
||||
default=[],
|
||||
metavar="NAME",
|
||||
help="population a memory proposal will be qualified on; its rows never propose (repeatable)",
|
||||
)
|
||||
|
||||
pairs = subparsers.add_parser(
|
||||
"pairs",
|
||||
help="scan isolated vuln-vs-fixed pairs (TP on vuln, silent on fix) and emit improve-loop signals",
|
||||
help="scan isolated vuln-vs-fixed pairs (TP on vuln, silent on fix) and report misses and fixed-side false positives",
|
||||
)
|
||||
pairs.add_argument("--catalog", type=Path, default=DEFAULT_CATALOG)
|
||||
pairs.add_argument("--slice", choices=SLICE_NAMES, default="all")
|
||||
@@ -392,17 +343,6 @@ def _main(argv: list[str] | None) -> int:
|
||||
plane_ws.add_argument(
|
||||
"--runner-image", type=Path, help="with --validation-set: re-pin the task templates to the image in this file (ops/ax/up.sh)"
|
||||
)
|
||||
plane_ws.add_argument(
|
||||
"--loop", action="store_true",
|
||||
help="with --validation-set: append the improvement loop (alerts per case, then snapshot, measure, propose, improve)",
|
||||
) # fmt: skip
|
||||
plane_ws.add_argument("--loop-commit", help="with --loop: the commit under test the gate runs at (default: git rev-parse HEAD)")
|
||||
plane_ws.add_argument("--loop-repo", help="with --loop: this repository's URL (default: git remote get-url origin)")
|
||||
plane_ws.add_argument("--loop-manifest", help="with --loop: the benchmark manifest the proposals are gated on")
|
||||
plane_ws.add_argument(
|
||||
"--qualify-population", action="append", default=[], metavar="NAME", help="with --loop: a population whose rows never propose"
|
||||
)
|
||||
|
||||
raw = list(argv if argv is not None else sys.argv[1:])
|
||||
if _advisory_hook(raw):
|
||||
# Git hook mode (stdin) without an explicit blocking opt-in: a check that cannot run must
|
||||
@@ -430,23 +370,6 @@ def _main(argv: list[str] | None) -> int:
|
||||
return _scan(args.path, args.config, args.mode, args.fail_on)
|
||||
if args.command == "benchmark":
|
||||
return _benchmark(args.manifest, args.config, args.mode)
|
||||
if args.command == "improve":
|
||||
return _improve(
|
||||
args.manifest,
|
||||
max_rounds=args.max_rounds,
|
||||
recall_floor=args.recall_floor,
|
||||
fp_ceiling=args.fp_ceiling,
|
||||
ledger=args.ledger,
|
||||
journal=args.journal,
|
||||
ruleset_dir=args.ruleset_dir,
|
||||
dry_run=args.dry_run,
|
||||
pair_catalog=None if args.no_pair_gate else args.pair_catalog,
|
||||
profile_tolerance=args.profile_tolerance,
|
||||
min_holdout_pairs=args.min_holdout_pairs,
|
||||
memory=args.memory,
|
||||
qualify_populations=tuple(args.qualify_population),
|
||||
guard_catalog=args.pair_catalog,
|
||||
)
|
||||
if args.command == "pairs":
|
||||
return _pairs(
|
||||
args.catalog,
|
||||
@@ -1103,75 +1026,6 @@ def _benchmark(manifest_path: Path, config_path: Path, mode: str) -> int:
|
||||
return 0
|
||||
|
||||
|
||||
def _improve(
|
||||
manifest_path: Path,
|
||||
*,
|
||||
max_rounds: int,
|
||||
recall_floor: float,
|
||||
fp_ceiling: float,
|
||||
ledger: Path | None,
|
||||
journal: Path | None,
|
||||
ruleset_dir: Path,
|
||||
dry_run: bool,
|
||||
pair_catalog: Path | None = None,
|
||||
profile_tolerance: float = 0.0,
|
||||
min_holdout_pairs: int = 5,
|
||||
memory: str | None = None,
|
||||
qualify_populations: tuple[str, ...] = (),
|
||||
guard_catalog: Path | None = None,
|
||||
) -> int:
|
||||
if not manifest_path.exists() or not manifest_path.is_file():
|
||||
raise SystemExit(f"benchmark manifest is not a file: {manifest_path}")
|
||||
pair_cases: tuple[object, ...] = ()
|
||||
if pair_catalog is not None:
|
||||
if not pair_catalog.is_file():
|
||||
raise SystemExit(f"pair catalog is not a file: {pair_catalog} (pass --no-pair-gate to skip the per-profile holdout clause)")
|
||||
pair_cases = select_split(load_pair_catalog(pair_catalog), "holdout")
|
||||
manifest = load_benchmark_manifest(manifest_path)
|
||||
target = resolve_benchmark_source(manifest_path, manifest)
|
||||
policy = load_policy()
|
||||
session = _memory_session(memory, manifest, target, ruleset_dir, qualify_populations, guard_catalog) if memory is not None else None
|
||||
|
||||
# The loop writes its accepted ledger where `scan`/`benchmark` read it, so a
|
||||
# subsequent scan of this target automatically picks up the improved ruleset.
|
||||
default_ledger = target / CALIBRATION_DIR / "rule_policy.json"
|
||||
with tempfile.TemporaryDirectory(prefix="ousast-improve-") as scratch:
|
||||
if dry_run:
|
||||
ledger_path = Path(scratch) / "rule_policy.json"
|
||||
journal_path = Path(scratch) / "improve_journal.json"
|
||||
else:
|
||||
ledger_path = ledger or default_ledger
|
||||
journal_path = journal or ledger_path.with_name("improve_journal.json")
|
||||
|
||||
outcomes = run_improvement(
|
||||
target,
|
||||
manifest,
|
||||
ledger_path=ledger_path,
|
||||
journal_path=journal_path,
|
||||
ruleset_dir=ruleset_dir,
|
||||
policy=policy,
|
||||
max_rounds=max_rounds,
|
||||
recall_floor=recall_floor,
|
||||
fp_ceiling=fp_ceiling,
|
||||
pair_cases=pair_cases,
|
||||
profile_tolerance=profile_tolerance,
|
||||
min_holdout_pairs=min_holdout_pairs,
|
||||
proposals=session.proposals(ledger_path, journal_path) if session else (),
|
||||
)
|
||||
_print_improve_outcomes(outcomes, manifest_path, target, ledger_path, dry_run=dry_run)
|
||||
if session is not None:
|
||||
session.finish(outcomes, journal_path, dry_run=dry_run)
|
||||
return 0
|
||||
|
||||
|
||||
def _memory_session(
|
||||
memory: str, manifest: object, target: Path, ruleset_dir: Path, populations: tuple[str, ...], catalog: Path | None
|
||||
) -> Any:
|
||||
from .improve.memory import MemorySession # lazy: the default `improve` path never touches the store
|
||||
|
||||
return MemorySession.open(memory, manifest, target, ruleset_dir, populations, catalog)
|
||||
|
||||
|
||||
def _pairs(
|
||||
catalog: Path,
|
||||
slice_name: str,
|
||||
@@ -1473,29 +1327,6 @@ def _check_obligations(
|
||||
return statics, payload, cited
|
||||
|
||||
|
||||
def _print_improve_outcomes(outcomes: list[RoundOutcome], manifest_path: Path, target: Path, ledger_path: Path, *, dry_run: bool) -> None:
|
||||
print(f"manifest={manifest_path}")
|
||||
print(f"target={target}")
|
||||
accepted = [o for o in outcomes if o.accepted]
|
||||
for outcome in outcomes:
|
||||
edits = ", ".join(f"{e.rule_id}:{e.from_status}->{e.to_status}" for e in outcome.edits) or "-"
|
||||
print(
|
||||
f"round {outcome.round}: {outcome.reason} | "
|
||||
f"recall {outcome.recall_before:.2%}->{outcome.recall_after:.2%} "
|
||||
f"fp {outcome.fp_before:.2%}->{outcome.fp_after:.2%} "
|
||||
f"score {outcome.score_before}->{outcome.score_after} | edits: {edits}"
|
||||
)
|
||||
if outcome.profile_regressions or outcome.profiles_under_minimum:
|
||||
print(f" profiles: regressed={outcome.profile_regressions or '-'} under_minimum={outcome.profiles_under_minimum or '-'}")
|
||||
print(f"rounds={len(outcomes)} accepted={len(accepted)}")
|
||||
if dry_run:
|
||||
print("dry_run=true (no ledger or mechanism store written)")
|
||||
elif accepted:
|
||||
print(f"ledger={ledger_path}")
|
||||
else:
|
||||
print("ledger=unchanged (no round accepted)")
|
||||
|
||||
|
||||
def _load_scan_findings(run_dir: Path) -> list[StaticFinding]:
|
||||
return load_findings(run_dir / "findings.json")
|
||||
|
||||
@@ -1644,7 +1475,7 @@ def _plane_memory(
|
||||
assert run_manifest is not None
|
||||
seeded = memory.seed(run_manifest, store)
|
||||
if seeded:
|
||||
print(f"memory {store.describe()}: seeded {len(seeded)} tasks (reused facts, loop snapshot): {', '.join(seeded)}")
|
||||
print(f"memory {store.describe()}: seeded {len(seeded)} tasks (reused facts): {', '.join(seeded)}")
|
||||
return 0
|
||||
annotations: dict[str, str] = {}
|
||||
if run_manifest is not None:
|
||||
@@ -1665,7 +1496,7 @@ def _plane_memory(
|
||||
|
||||
|
||||
def _plane_increment(args: argparse.Namespace, atespace: str) -> int:
|
||||
from .plane.generate import LOOP_MANIFEST, Loop, increment
|
||||
from .plane.generate import increment
|
||||
|
||||
if args.candidates is None or args.triage is None:
|
||||
print("--validation-set needs --candidates and --triage", file=sys.stderr)
|
||||
@@ -1679,18 +1510,9 @@ def _plane_increment(args: argparse.Namespace, atespace: str) -> int:
|
||||
)
|
||||
if args.runner_image is not None:
|
||||
command += f" --runner-image {str(args.runner_image).replace(home, '~', 1)}"
|
||||
loop = None
|
||||
if args.loop:
|
||||
git = lambda *a: subprocess.run(["git", *a], capture_output=True, text=True, check=True).stdout.strip() # noqa: E731
|
||||
loop = Loop(
|
||||
repo=args.loop_repo or git("remote", "get-url", "origin"), commit=args.loop_commit or git("rev-parse", "HEAD"),
|
||||
manifest=args.loop_manifest or LOOP_MANIFEST, populations=tuple(args.qualify_population),
|
||||
) # fmt: skip
|
||||
command += f" --loop --loop-commit {loop.commit} --loop-repo {loop.repo} --loop-manifest {loop.manifest}"
|
||||
command += "".join(f" --qualify-population {p}" for p in loop.populations)
|
||||
written = increment(
|
||||
args.population, args.validation_set, args.candidates, args.triage, plane=args.plane, run_name=args.run, command=command,
|
||||
repos=args.repos, runner_image=args.runner_image, loop=loop, atespace=atespace,
|
||||
repos=args.repos, runner_image=args.runner_image, atespace=atespace,
|
||||
) # fmt: skip
|
||||
print(f"{len(written)} manifests written under {args.plane} (Run {args.run})")
|
||||
return 0
|
||||
|
||||
@@ -1,32 +0,0 @@
|
||||
"""Bounded, gated self-improvement loop over the ruleset/policy governance data."""
|
||||
|
||||
from ..pairs import build_pair_signals
|
||||
from .evolve import RoundOutcome, build_rule_signals, propose_status_edits, run_improvement, run_round
|
||||
from .journal import load_journal, reverted_edit_keys
|
||||
from .validator import (
|
||||
VALID_LEVERS,
|
||||
EvolveBounds,
|
||||
EvolveValidator,
|
||||
PolicyConstantEdit,
|
||||
RuleStatusEdit,
|
||||
StrictValidationError,
|
||||
edits_to_ledger,
|
||||
)
|
||||
|
||||
__all__ = [
|
||||
"EvolveBounds",
|
||||
"EvolveValidator",
|
||||
"PolicyConstantEdit",
|
||||
"RoundOutcome",
|
||||
"RuleStatusEdit",
|
||||
"StrictValidationError",
|
||||
"VALID_LEVERS",
|
||||
"build_pair_signals",
|
||||
"build_rule_signals",
|
||||
"edits_to_ledger",
|
||||
"load_journal",
|
||||
"propose_status_edits",
|
||||
"reverted_edit_keys",
|
||||
"run_improvement",
|
||||
"run_round",
|
||||
]
|
||||
@@ -1,416 +0,0 @@
|
||||
"""Deterministic self-improvement loop (tasks 7.4, 7.5, 7.6).
|
||||
|
||||
One round: benchmark with the current ledger -> per-rule signals -> propose bounded
|
||||
edits (auto-shadow precision-draggers) -> validate -> replay smoke -> re-benchmark ->
|
||||
hard acceptance gate (recall >= floor AND FP < ceiling AND project score not regressed
|
||||
AND no matched-finding regression) -> accept (persist the ledger) or revert byte-for-byte.
|
||||
|
||||
The edit *proposers* are the benchmark's per-rule signal and, opt-in, the plane's memory store
|
||||
(:mod:`.memory`, ``run_round(proposals=...)``); both go through this same validator and gate.
|
||||
The project score is the optimization reward; the recall/FP gate is a separate hard
|
||||
feasibility constraint that is never folded into the reward.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import tempfile
|
||||
from collections.abc import Mapping, Sequence
|
||||
from dataclasses import dataclass, field
|
||||
from pathlib import Path
|
||||
|
||||
from ..benchmark import BenchmarkManifest, BenchmarkResult, BenchmarkRun, evaluate_benchmark
|
||||
from ..findings import StaticFinding, quick_scan_findings
|
||||
from ..mapping import analyze_entry_points, attach_reachability_hints
|
||||
from ..policy import CwePolicy, load_policy
|
||||
from ..preprocess import preprocess_repository
|
||||
from ..rank import rank_targets
|
||||
from ..ruleset import DEFAULT_RULESET_DIR, PatternRule, load_ruleset, read_rule_ledger, write_rule_ledger
|
||||
from ..scoring import build_score_artifact
|
||||
from .journal import append_round, load_journal, next_round_index, reverted_edit_keys
|
||||
from .memory import MemoryProposal, applicable, proposal_id_of, write_sidecar
|
||||
from .validator import EvolveValidator, RuleStatusEdit, StrictValidationError, edits_to_ledger
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class RoundOutcome:
|
||||
round: int
|
||||
accepted: bool
|
||||
reason: str
|
||||
edits: list[RuleStatusEdit] = field(default_factory=list)
|
||||
recall_before: float = 0.0
|
||||
recall_after: float = 0.0
|
||||
fp_before: float = 0.0
|
||||
fp_after: float = 0.0
|
||||
score_before: int = 0
|
||||
score_after: int = 0
|
||||
matched_before: int = 0
|
||||
matched_after: int = 0
|
||||
profile_regressions: list[str] = field(default_factory=list)
|
||||
profiles_under_minimum: list[str] = field(default_factory=list)
|
||||
per_profile_before: dict[str, dict[str, float]] = field(default_factory=dict)
|
||||
per_profile_after: dict[str, dict[str, float]] = field(default_factory=dict)
|
||||
# learning-harness Req 5.2: the candidates this round declined to learn from, by pair name.
|
||||
degradations: list[dict[str, object]] = field(default_factory=list)
|
||||
|
||||
|
||||
MAX_PROFILES = 4
|
||||
|
||||
|
||||
def evaluate_profiles(
|
||||
pair_cases: Sequence[object],
|
||||
rules: tuple[PatternRule, ...],
|
||||
) -> dict[str, dict[str, float]]:
|
||||
"""Pair-corpus metrics per provenance profile under ``rules`` (holdout split, no inventory sidecar).
|
||||
|
||||
Only ``seeded`` and ``reviewed`` pairs gate (Req 9.3); ``advisory`` and ``title`` pairs are scored by ``pairs``, never here.
|
||||
|
||||
Deliberately split-blind: this *measures* a ruleset, it never learns from a pair, so every pair is scored and the
|
||||
teacher rule has nothing to say here. What the rule governs is the store these rules were built from
|
||||
(`semantic.seed`, `semantic.loo`), and a filter here would only shrink the evidence (learning-harness Req 5.1).
|
||||
"""
|
||||
from ..pairs import GATING_TIERS, PairCase, evaluate_catalog, select_tier, select_vendored
|
||||
|
||||
# Gates score vendored pairs only (Req 10.4); pointer pairs are for `pairs --pointers`, never for the improve gate.
|
||||
cases = list(select_tier(select_vendored([case for case in pair_cases if isinstance(case, PairCase)]), GATING_TIERS))
|
||||
if not cases:
|
||||
return {}
|
||||
result = evaluate_catalog(cases, ruleset=rules, inventory_sidecar=False)
|
||||
return {
|
||||
profile: {
|
||||
"pairs": float(metrics.pairs),
|
||||
"pair_correct": float(metrics.pair_correct),
|
||||
"pair_pass_rate": metrics.pair_pass_rate,
|
||||
"youden": metrics.youden,
|
||||
}
|
||||
for profile, metrics in result.per_profile.items()
|
||||
}
|
||||
|
||||
|
||||
def profile_regressions(
|
||||
before: dict[str, dict[str, float]],
|
||||
after: dict[str, dict[str, float]],
|
||||
*,
|
||||
tolerance: float = 0.0,
|
||||
min_pairs: int = 5,
|
||||
) -> tuple[list[str], list[str]]:
|
||||
"""(regressed profiles, profiles under the minimum).
|
||||
|
||||
A profile regresses when its pair pass rate or Youden drops by more than ``tolerance``
|
||||
(both are rates in [-1, 1], so one tolerance applies to both).
|
||||
"""
|
||||
regressed: list[str] = []
|
||||
small: list[str] = []
|
||||
for profile in sorted(set(before) | set(after)):
|
||||
old = before.get(profile, {"pairs": 0.0, "pair_correct": 0.0, "pair_pass_rate": 0.0, "youden": 0.0})
|
||||
new = after.get(profile, old)
|
||||
if old["pairs"] < min_pairs:
|
||||
small.append(profile)
|
||||
continue
|
||||
old_rate = old.get("pair_pass_rate", old["pair_correct"] / old["pairs"] if old["pairs"] else 0.0)
|
||||
new_rate = new.get("pair_pass_rate", new["pair_correct"] / new["pairs"] if new["pairs"] else 0.0)
|
||||
if new_rate < old_rate - tolerance or new["youden"] < old["youden"] - tolerance:
|
||||
regressed.append(profile)
|
||||
return regressed, small
|
||||
|
||||
|
||||
def build_rule_signals(result: BenchmarkResult) -> list[dict[str, object]]:
|
||||
"""Convert a benchmark result into per-rule loop signals (recall: miss; precision: fp)."""
|
||||
signals: list[dict[str, object]] = []
|
||||
for miss in result.misses:
|
||||
signals.append({"rule_id": miss.rule_id or "", "signal": "miss", "cwe": miss.cwe, "path": miss.path})
|
||||
for fp in result.false_positives:
|
||||
signals.append({"rule_id": fp.rule_id, "signal": "fp", "path": fp.path, "line": fp.line})
|
||||
return sorted(signals, key=lambda item: (str(item.get("signal")), str(item.get("rule_id")), str(item.get("path"))))
|
||||
|
||||
|
||||
def propose_status_edits(
|
||||
per_rule: Mapping[str, dict[str, int]],
|
||||
ruleset_by_id: Mapping[str, PatternRule],
|
||||
current_ledger: Mapping[str, dict[str, object]],
|
||||
blocked_keys: set[str],
|
||||
) -> list[RuleStatusEdit]:
|
||||
"""Auto-shadow precision-draggers: enabled rules with false positives -> shadow (never delete)."""
|
||||
edits: list[RuleStatusEdit] = []
|
||||
for rule_id, counts in sorted(per_rule.items()):
|
||||
if counts.get("false_positives", 0) <= 0:
|
||||
continue
|
||||
rule = ruleset_by_id.get(rule_id)
|
||||
if rule is None:
|
||||
continue
|
||||
current_status = str(current_ledger.get(rule_id, {}).get("status", rule.status))
|
||||
if current_status != "enabled":
|
||||
continue # already staged or disabled
|
||||
edit = RuleStatusEdit(rule_id=rule_id, from_status="enabled", to_status="shadow", rationale="")
|
||||
if edit.key() in blocked_keys:
|
||||
continue # novelty gate: previously reverted without a new rationale
|
||||
edits.append(edit)
|
||||
return edits
|
||||
|
||||
|
||||
def run_round(
|
||||
target: Path,
|
||||
manifest: BenchmarkManifest,
|
||||
*,
|
||||
ledger_path: Path,
|
||||
journal_path: Path,
|
||||
ruleset_dir: Path = DEFAULT_RULESET_DIR,
|
||||
policy: dict[str, CwePolicy] | None = None,
|
||||
recall_floor: float = 0.9,
|
||||
fp_ceiling: float = 0.1,
|
||||
pair_cases: Sequence[object] | None = None,
|
||||
profile_tolerance: float = 0.0,
|
||||
min_holdout_pairs: int = 5,
|
||||
proposals: Sequence[MemoryProposal] = (),
|
||||
) -> RoundOutcome:
|
||||
policy = policy if policy is not None else load_policy()
|
||||
rounds = load_journal(journal_path)
|
||||
round_index = next_round_index(rounds)
|
||||
blocked = reverted_edit_keys(rounds)
|
||||
current_ledger = read_rule_ledger(ledger_path)
|
||||
|
||||
before_findings, before_rules = _scan(target, ruleset_dir, current_ledger, policy)
|
||||
before = _evaluate(target, manifest, before_findings, before_rules, policy)
|
||||
(ledger_path.parent / "rule_signals.json").parent.mkdir(parents=True, exist_ok=True)
|
||||
_write_signals(ledger_path.parent / "rule_signals.json", build_rule_signals(before.result))
|
||||
|
||||
ruleset_by_id = {rule.rule_id: rule for rule in before_rules}
|
||||
edits = propose_status_edits(before.result.metrics.per_rule, ruleset_by_id, current_ledger, blocked)
|
||||
if proposals:
|
||||
# harnessx-removal Req 6.2: memory proposals join after the benchmark's edits; the first edit per key wins.
|
||||
taken = {edit.key() for edit in edits}
|
||||
for proposal in applicable(proposals, ruleset_by_id, current_ledger, rounds):
|
||||
if proposal.edit.key() not in taken:
|
||||
taken.add(proposal.edit.key())
|
||||
edits.append(proposal.edit)
|
||||
write_sidecar(journal_path, round_index, [p for p in proposals if p.edit in edits])
|
||||
refusals: list[dict[str, object]] = []
|
||||
if not edits:
|
||||
return _outcome(round_index, accepted=False, reason="no_proposals", edits=[], before=before, after=before, degradations=refusals)
|
||||
|
||||
validator = EvolveValidator()
|
||||
try:
|
||||
for edit in edits:
|
||||
validator.validate(edit, ruleset_by_id, policy)
|
||||
except StrictValidationError as exc:
|
||||
_record(journal_path, round_index, edits, "rejected", before, before, str(exc))
|
||||
return _outcome(
|
||||
round_index, accepted=False, reason=f"validation_failed: {exc}", edits=edits, before=before, after=before, degradations=refusals
|
||||
)
|
||||
|
||||
candidate = edits_to_ledger(edits, current_ledger)
|
||||
try:
|
||||
_load_with(ruleset_dir, candidate) # replay smoke gate: candidate ruleset must boot
|
||||
except Exception as exc: # noqa: BLE001 — any boot failure rejects the round
|
||||
_record(journal_path, round_index, edits, "rejected", before, before, f"replay_failed: {exc}")
|
||||
return _outcome(
|
||||
round_index, accepted=False, reason="replay_failed", edits=edits, before=before, after=before, degradations=refusals
|
||||
)
|
||||
|
||||
after_findings, after_rules = _scan(target, ruleset_dir, candidate, policy)
|
||||
after = _evaluate(target, manifest, after_findings, after_rules, policy)
|
||||
|
||||
accepted = (
|
||||
after.recall >= recall_floor and after.fp_rate < fp_ceiling and after.score >= before.score and after.matched >= before.matched
|
||||
)
|
||||
profiles_before: dict[str, dict[str, float]] = {}
|
||||
profiles_after: dict[str, dict[str, float]] = {}
|
||||
regressed: list[str] = []
|
||||
small: list[str] = []
|
||||
if accepted and pair_cases:
|
||||
# pair-corpus-honesty Req 7.2: a change may not help one provenance profile by hurting another.
|
||||
profiles_before = evaluate_profiles(pair_cases, before_rules)
|
||||
profiles_after = evaluate_profiles(pair_cases, after_rules)
|
||||
if len(profiles_before) > MAX_PROFILES:
|
||||
raise ValueError(f"at most {MAX_PROFILES} provenance profiles are supported; got {sorted(profiles_before)}")
|
||||
regressed, small = profile_regressions(profiles_before, profiles_after, tolerance=profile_tolerance, min_pairs=min_holdout_pairs)
|
||||
if regressed:
|
||||
accepted = False
|
||||
if accepted:
|
||||
if edits:
|
||||
write_rule_ledger(ledger_path, candidate)
|
||||
reason = "accepted"
|
||||
elif regressed:
|
||||
reason = f"profile_regression:{regressed[0]}" # every regressed profile is listed in profile_regressions
|
||||
else:
|
||||
reason = "reverted" # ledger_path untouched -> byte-for-byte revert
|
||||
_record(
|
||||
journal_path,
|
||||
round_index,
|
||||
edits,
|
||||
"accepted" if accepted else "reverted",
|
||||
before,
|
||||
after,
|
||||
reason,
|
||||
regressed,
|
||||
small,
|
||||
)
|
||||
outcome = _outcome(round_index, accepted=accepted, reason=reason, edits=edits, before=before, after=after, degradations=refusals)
|
||||
return RoundOutcome(
|
||||
**{
|
||||
**outcome.__dict__,
|
||||
"profile_regressions": regressed,
|
||||
"profiles_under_minimum": small,
|
||||
"per_profile_before": profiles_before,
|
||||
"per_profile_after": profiles_after,
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
def run_improvement(
|
||||
target: Path,
|
||||
manifest: BenchmarkManifest,
|
||||
*,
|
||||
ledger_path: Path,
|
||||
journal_path: Path,
|
||||
ruleset_dir: Path = DEFAULT_RULESET_DIR,
|
||||
policy: dict[str, CwePolicy] | None = None,
|
||||
max_rounds: int = 5,
|
||||
recall_floor: float = 0.9,
|
||||
fp_ceiling: float = 0.1,
|
||||
pair_cases: Sequence[object] | None = None,
|
||||
profile_tolerance: float = 0.0,
|
||||
min_holdout_pairs: int = 5,
|
||||
proposals: Sequence[MemoryProposal] = (),
|
||||
) -> list[RoundOutcome]:
|
||||
"""Run improvement rounds until convergence (no new proposals) or ``max_rounds``."""
|
||||
outcomes: list[RoundOutcome] = []
|
||||
for _ in range(max_rounds):
|
||||
outcome = run_round(
|
||||
target,
|
||||
manifest,
|
||||
ledger_path=ledger_path,
|
||||
journal_path=journal_path,
|
||||
ruleset_dir=ruleset_dir,
|
||||
policy=policy,
|
||||
recall_floor=recall_floor,
|
||||
fp_ceiling=fp_ceiling,
|
||||
pair_cases=pair_cases,
|
||||
profile_tolerance=profile_tolerance,
|
||||
min_holdout_pairs=min_holdout_pairs,
|
||||
proposals=proposals,
|
||||
)
|
||||
outcomes.append(outcome)
|
||||
if outcome.reason == "no_proposals":
|
||||
break
|
||||
return outcomes
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class _Eval:
|
||||
result: BenchmarkResult
|
||||
recall: float
|
||||
fp_rate: float
|
||||
score: int
|
||||
matched: int
|
||||
|
||||
|
||||
def _scan(
|
||||
target: Path, ruleset_dir: Path, ledger: Mapping[str, dict[str, object]], policy: dict[str, CwePolicy]
|
||||
) -> tuple[list[StaticFinding], tuple[PatternRule, ...]]:
|
||||
rules = _load_with(ruleset_dir, ledger)
|
||||
_, targets = preprocess_repository(target)
|
||||
targets = attach_reachability_hints(targets, analyze_entry_points(target, targets))
|
||||
rankings = rank_targets(targets)
|
||||
findings = quick_scan_findings(target, targets, rankings, rules, policy)
|
||||
reported = [finding for finding in findings if finding.status != "shadow"]
|
||||
return reported, rules
|
||||
|
||||
|
||||
def _evaluate(
|
||||
target: Path, manifest: BenchmarkManifest, findings: list[StaticFinding], rules: tuple[PatternRule, ...], policy: dict[str, CwePolicy]
|
||||
) -> _Eval:
|
||||
run = BenchmarkRun(benchmark_run_id="improve", root=target, manifest=manifest)
|
||||
result = evaluate_benchmark(run=run, mode="quick", findings=findings, scan_id=None, scan_run_dir=None)
|
||||
metrics = result.metrics
|
||||
cwe_by_rule = {rule.rule_id: rule.cwe for rule in rules}
|
||||
rule_cwe_by_id = {f.finding_id: cwe_by_rule.get(f.finding_id.split(":", 1)[0], "") for f in findings}
|
||||
score = build_score_artifact(findings, rule_cwe_by_id, policy).project_score
|
||||
recall = metrics.recall if metrics.recall is not None else 1.0
|
||||
fp_rate = metrics.false_positive_findings_total / metrics.actual_findings_total if metrics.actual_findings_total else 0.0
|
||||
return _Eval(result=result, recall=recall, fp_rate=fp_rate, score=score, matched=metrics.matched_findings_total)
|
||||
|
||||
|
||||
def _load_with(ruleset_dir: Path, ledger: Mapping[str, dict[str, object]]) -> tuple[PatternRule, ...]:
|
||||
import json
|
||||
|
||||
with tempfile.NamedTemporaryFile("w", suffix=".json", delete=False) as handle:
|
||||
handle.write(json.dumps(ledger))
|
||||
temp_path = Path(handle.name)
|
||||
try:
|
||||
return load_ruleset(ruleset_dir, temp_path)
|
||||
finally:
|
||||
temp_path.unlink(missing_ok=True)
|
||||
|
||||
|
||||
def _write_signals(path: Path, signals: list[dict[str, object]]) -> None:
|
||||
import json
|
||||
|
||||
path.write_text(json.dumps(signals, indent=2, sort_keys=True) + "\n")
|
||||
|
||||
|
||||
def _outcome(
|
||||
round_index: int,
|
||||
*,
|
||||
accepted: bool,
|
||||
reason: str,
|
||||
edits: list[RuleStatusEdit],
|
||||
before: _Eval,
|
||||
after: _Eval,
|
||||
degradations: list[dict[str, object]] | None = None,
|
||||
) -> RoundOutcome:
|
||||
return RoundOutcome(
|
||||
round=round_index,
|
||||
accepted=accepted,
|
||||
reason=reason,
|
||||
edits=edits,
|
||||
degradations=list(degradations or []),
|
||||
recall_before=before.recall,
|
||||
recall_after=after.recall,
|
||||
fp_before=before.fp_rate,
|
||||
fp_after=after.fp_rate,
|
||||
score_before=before.score,
|
||||
score_after=after.score,
|
||||
matched_before=before.matched,
|
||||
matched_after=after.matched,
|
||||
)
|
||||
|
||||
|
||||
def _record(
|
||||
journal_path: Path,
|
||||
round_index: int,
|
||||
edits: list[RuleStatusEdit],
|
||||
outcome: str,
|
||||
before: _Eval,
|
||||
after: _Eval,
|
||||
reason: str = "",
|
||||
profile_regressions: list[str] | None = None,
|
||||
profiles_under_minimum: list[str] | None = None,
|
||||
) -> None:
|
||||
append_round(
|
||||
journal_path,
|
||||
{
|
||||
"round": round_index,
|
||||
"outcome": outcome,
|
||||
"reason": reason or outcome,
|
||||
"profile_regressions": list(profile_regressions or []),
|
||||
"profiles_under_minimum": list(profiles_under_minimum or []),
|
||||
"edits": [_journal_edit(e) for e in edits],
|
||||
"recall_before": round(before.recall, 4),
|
||||
"recall_after": round(after.recall, 4),
|
||||
"fp_before": round(before.fp_rate, 4),
|
||||
"fp_after": round(after.fp_rate, 4),
|
||||
"score_before": before.score,
|
||||
"score_after": after.score,
|
||||
},
|
||||
)
|
||||
|
||||
|
||||
def _journal_edit(edit: RuleStatusEdit) -> dict[str, object]:
|
||||
entry: dict[str, object] = {
|
||||
"key": edit.key(), "lever": edit.lever, "rule_id": edit.rule_id, "from": edit.from_status, "to": edit.to_status,
|
||||
"rationale": edit.rationale,
|
||||
} # fmt: skip
|
||||
proposal_id = proposal_id_of(edit)
|
||||
if proposal_id is not None:
|
||||
entry["evidence"] = proposal_id # only a memory edit carries it; every other entry is unchanged byte for byte
|
||||
return entry
|
||||
@@ -1,53 +0,0 @@
|
||||
"""Improvement journal: per-round audit + novelty memory (task 7.3, 7.9).
|
||||
|
||||
Persistent JSON record of every self-improvement round so each rule/policy edit is
|
||||
attributable and reversible, and so a previously-reverted edit is not re-proposed
|
||||
without a new rationale (the novelty gate).
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
def load_journal(path: Path) -> list[dict[str, object]]:
|
||||
if not path.exists():
|
||||
return []
|
||||
payload = json.loads(path.read_text())
|
||||
return payload if isinstance(payload, list) else []
|
||||
|
||||
|
||||
def write_journal(path: Path, rounds: list[dict[str, object]]) -> None:
|
||||
path.parent.mkdir(parents=True, exist_ok=True)
|
||||
path.write_text(json.dumps(rounds, indent=2, sort_keys=True) + "\n")
|
||||
|
||||
|
||||
def reverted_edit_keys(rounds: list[dict[str, object]]) -> set[str]:
|
||||
"""Edit keys from reverted rounds — blocked from re-proposal unless a new rationale appears."""
|
||||
keys: set[str] = set()
|
||||
for entry in rounds:
|
||||
if entry.get("outcome") != "reverted":
|
||||
continue
|
||||
for edit in _edits(entry):
|
||||
key = edit.get("key")
|
||||
# Rule edits are blocked only without a new rationale; mechanism edits are blocked outright (the loop
|
||||
# always writes a rationale for them, and a reverted store edit must not be re-proposed every round).
|
||||
if isinstance(key, str) and (not edit.get("rationale") or edit.get("lever") == "mechanisms"):
|
||||
keys.add(key)
|
||||
return keys
|
||||
|
||||
|
||||
def next_round_index(rounds: list[dict[str, object]]) -> int:
|
||||
return len(rounds) + 1
|
||||
|
||||
|
||||
def append_round(path: Path, entry: dict[str, object]) -> None:
|
||||
rounds = load_journal(path)
|
||||
rounds.append(entry)
|
||||
write_journal(path, rounds)
|
||||
|
||||
|
||||
def _edits(entry: dict[str, object]) -> list[dict[str, object]]:
|
||||
edits = entry.get("edits")
|
||||
return [edit for edit in edits if isinstance(edit, dict)] if isinstance(edits, list) else []
|
||||
@@ -1,549 +0,0 @@
|
||||
"""Rule-status proposals from the plane's memory store (harnessx-removal Req 6.2/6.4, design section 5).
|
||||
|
||||
The store (:mod:`..plane.memory`) holds what plane runs learned per repository and pin. Two deterministic rules turn
|
||||
its ``alert`` and ``verdict`` rows into :class:`RuleStatusEdit` proposals that :func:`.evolve.run_round` validates
|
||||
and gates exactly like the benchmark's own edits; a third reads dispute rates as advisory routing input only.
|
||||
|
||||
Join key: an ``alert`` and a ``verdict`` match when repo, pin and ``path::function`` are equal. A ``disputed``
|
||||
final counts as no evidence in either direction.
|
||||
|
||||
- **M1, repeated false alerts** (``enabled -> shadow``): at least :data:`M1_MIN_FALSE` alerts of the rule are false
|
||||
-- at a ``rejected`` candidate, or on a ``fixed`` pin inside the declared fix range (the alert row says
|
||||
``in_fix_range: true``, or its ``path::function`` is a ``site_match`` site of the same repository) -- across at
|
||||
least :data:`M1_MIN_REPOS` repositories, and no alert of the rule sits at an ``agreed`` or ``site_match``
|
||||
candidate.
|
||||
- **M2, missed declared sites** (``shadow -> enabled``): the shadow rule's alerts hit at least :data:`M2_MIN_HITS`
|
||||
``agreed`` candidates with ``site_match`` true across at least :data:`M2_MIN_REPOS` repositories, no enabled rule
|
||||
alerts at those candidates, the rule has no ``fixed``-pin alert at the same function, and at most
|
||||
:data:`M2_MAX_REJECTED` alert at a ``rejected`` candidate.
|
||||
- **M3, disputed families**: families with at least :data:`M3_MIN_CANDIDATES` candidates and a dispute rate of at
|
||||
least :data:`M3_MIN_DISPUTE_RATE`. Advisory (``signals.json``); never an edit, never shown to the validator.
|
||||
|
||||
**Coverage**: M1 and M2 skip every repository and pin whose ``coverage`` rows record the rule's language only as
|
||||
``none`` (quick mode had no rule for it and no engine alerts were produced): a rule cannot be demoted or promoted
|
||||
on evidence that could not exist. The skipped alerts are counted per ``repo@pin`` under ``coverage_none`` in the
|
||||
proposal's ``excluded``.
|
||||
|
||||
**Train-on-test guard** (:class:`Guard`), applied before any rule sees a row: a row is dropped when its population
|
||||
is one the proposal will be qualified on, its (repo, pin) is a case of the gated benchmark manifest, or its
|
||||
repository is a holdout pair's repository. The pair catalog records a holdout pair's *fix* commit, not the
|
||||
vulnerable pin (its parent), so the holdout clause matches by repository, a superset of (repo, pin). This is the
|
||||
rule the closed-loop leak broke, where a holdout pair taught the shape that then "recovered" it.
|
||||
|
||||
**Provenance**: each proposal's rationale is ``memory:<proposal_id>``, ``proposal_id`` being the sha256 of (rule,
|
||||
edit key, sorted evidence row ids, thresholds). The sidecar ``memory_proposals.jsonl`` next to the journal carries
|
||||
the evidence rows (id, repo, pin, run, task, object key and version -- the bucket's object version id, a file's content hash),
|
||||
the counts that met each threshold, what the guard excluded and the store's index digest.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import hashlib
|
||||
import json
|
||||
from collections.abc import Iterable, Mapping, Sequence
|
||||
from dataclasses import dataclass, field
|
||||
from datetime import UTC, datetime
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from ..plane.memory import MemoryStore, Record, repo_key, row_id, rows_digest
|
||||
from ..ruleset import PatternRule
|
||||
from .validator import RuleStatusEdit
|
||||
|
||||
M1_MIN_FALSE = 3
|
||||
M1_MIN_REPOS = 2
|
||||
M2_MIN_HITS = 2
|
||||
M2_MIN_REPOS = 2
|
||||
M2_MAX_REJECTED = 1
|
||||
M3_MIN_CANDIDATES = 5
|
||||
M3_MIN_DISPUTE_RATE = 0.3
|
||||
THRESHOLDS: dict[str, dict[str, float]] = {
|
||||
"M1": {"min_false_alerts": M1_MIN_FALSE, "min_repos": M1_MIN_REPOS, "max_agreed_hits": 0},
|
||||
"M2": {"min_site_hits": M2_MIN_HITS, "min_repos": M2_MIN_REPOS, "max_rejected": M2_MAX_REJECTED, "max_fixed_pin_same_function": 0},
|
||||
"M3": {"min_candidates": M3_MIN_CANDIDATES, "min_dispute_rate": M3_MIN_DISPUTE_RATE},
|
||||
}
|
||||
RATIONALE_PREFIX = "memory:"
|
||||
SIDECAR = "memory_proposals.jsonl"
|
||||
SIGNALS = "signals.json"
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class MemoryProposal:
|
||||
edit: RuleStatusEdit
|
||||
provenance: dict[str, Any]
|
||||
|
||||
@property
|
||||
def proposal_id(self) -> str:
|
||||
return str(self.provenance["proposal_id"])
|
||||
|
||||
|
||||
def proposal_id_of(edit: object) -> str | None:
|
||||
"""The proposal id a memory edit carries in its rationale; None for any other edit."""
|
||||
rationale = getattr(edit, "rationale", "")
|
||||
return rationale[len(RATIONALE_PREFIX) :] if isinstance(rationale, str) and rationale.startswith(RATIONALE_PREFIX) else None
|
||||
|
||||
|
||||
# --- the train-on-test guard ---------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _tail(repo: str) -> str:
|
||||
"""``owner/name`` of a repository, lower-cased: the pair catalog names repositories without a host."""
|
||||
return "/".join(repo_key(repo).lower().split("/")[-2:])
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Guard:
|
||||
populations: frozenset[str] = frozenset()
|
||||
gated_cases: frozenset[tuple[str, str]] = frozenset() # (repo_key, pin) of the benchmark being gated
|
||||
holdout_repos: frozenset[str] = frozenset() # owner/name of every holdout pair
|
||||
|
||||
@classmethod
|
||||
def build(
|
||||
cls, populations: Iterable[str] = (), gated_cases: Iterable[tuple[str, str]] = (), holdout_pairs: Iterable[object] = ()
|
||||
) -> Guard:
|
||||
repos = {_tail(str(getattr(p, "repo", ""))) for p in holdout_pairs if getattr(p, "repo", "")}
|
||||
return cls(frozenset(populations), frozenset((repo_key(r), p) for r, p in gated_cases), frozenset(repos))
|
||||
|
||||
def reason(self, row: Mapping[str, Any]) -> str | None:
|
||||
if row.get("population") in self.populations:
|
||||
return "population"
|
||||
if (row.get("repo"), row.get("pin")) in self.gated_cases:
|
||||
return "gated_case"
|
||||
if _tail(str(row.get("repo"))) in self.holdout_repos:
|
||||
return "holdout_pair"
|
||||
return None
|
||||
|
||||
def apply(self, records: Iterable[Record]) -> tuple[list[Record], dict[str, Any]]:
|
||||
"""(kept records, what was excluded: row count, and counts per population / repo@pin per clause)."""
|
||||
kept: list[Record] = []
|
||||
excluded: dict[str, Any] = {"rows": 0, "population": {}, "gated_case": {}, "holdout_pair": {}}
|
||||
for record in records:
|
||||
why = self.reason(record.row)
|
||||
if why is None:
|
||||
kept.append(record)
|
||||
continue
|
||||
excluded["rows"] += 1
|
||||
label = str(record.row.get("population")) if why == "population" else f"{record.row.get('repo')}@{record.row.get('pin')}"
|
||||
excluded[why][label] = excluded[why].get(label, 0) + 1
|
||||
for clause in ("population", "gated_case", "holdout_pair"):
|
||||
excluded[clause] = dict(sorted(excluded[clause].items()))
|
||||
return kept, excluded
|
||||
|
||||
|
||||
# --- the rules -----------------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _cand(row: Mapping[str, Any]) -> str:
|
||||
return f"{row.get('path')}::{row.get('function')}"
|
||||
|
||||
|
||||
def _status(rule_id: str, ruleset_by_id: Mapping[str, PatternRule], ledger: Mapping[str, dict[str, object]]) -> str | None:
|
||||
rule = ruleset_by_id.get(rule_id)
|
||||
return None if rule is None else str(ledger.get(rule_id, {}).get("status", rule.status))
|
||||
|
||||
|
||||
@dataclass
|
||||
class _View:
|
||||
verdicts: dict[tuple[str, str, str], Record] = field(default_factory=dict) # (repo, pin, candidate)
|
||||
sites: set[tuple[str, str]] = field(default_factory=set) # (repo, candidate) with site_match true
|
||||
alerts: dict[str, list[Record]] = field(default_factory=dict) # rule_id -> alerts
|
||||
coverage: dict[tuple[str, str, str], set[str]] = field(default_factory=dict) # (repo, pin, language) -> recorded kinds
|
||||
|
||||
@classmethod
|
||||
def of(cls, records: Sequence[Record]) -> _View:
|
||||
"""One alert per (repo, pin, rule, path, line, function, pin role): every loop run re-alerts an unchanged pin under a
|
||||
new run name, and a repeat is the same evidence, not more of it."""
|
||||
view = cls()
|
||||
seen: set[tuple[object, ...]] = set()
|
||||
for rec in records:
|
||||
row = rec.row
|
||||
if row["kind"] == "coverage" and row.get("language"):
|
||||
view.coverage.setdefault((row["repo"], row["pin"], str(row["language"])), set()).add(str(row.get("coverage")))
|
||||
elif row["kind"] == "verdict":
|
||||
view.verdicts[(row["repo"], row["pin"], str(row.get("candidate")))] = rec
|
||||
if row.get("site_match") is True:
|
||||
view.sites.add((row["repo"], str(row.get("candidate"))))
|
||||
elif row["kind"] == "alert" and row.get("rule_id"):
|
||||
place = (row["repo"], row["pin"], row["rule_id"], *(row.get(k) for k in ("path", "line", "function", "pin_role")))
|
||||
if place not in seen:
|
||||
seen.add(place)
|
||||
view.alerts.setdefault(str(row["rule_id"]), []).append(rec)
|
||||
return view
|
||||
|
||||
def verdict(self, alert: Mapping[str, Any]) -> Record | None:
|
||||
return self.verdicts.get((alert["repo"], alert["pin"], _cand(alert)))
|
||||
|
||||
def fixed_in_range(self, alert: Mapping[str, Any]) -> bool:
|
||||
if alert.get("pin_role") != "fixed":
|
||||
return False
|
||||
flag = alert.get("in_fix_range")
|
||||
return bool(flag) if flag is not None else (alert["repo"], _cand(alert)) in self.sites
|
||||
|
||||
def blind(self, repo: str, pin: str, languages: Iterable[str]) -> bool:
|
||||
"""True when a language of the rule was recorded at this repository and pin with coverage ``none`` only: no
|
||||
rule could fire there, so what the rule did or did not say there is no evidence either way."""
|
||||
return any(self.coverage.get((repo, pin, language)) == {"none"} for language in languages)
|
||||
|
||||
def seeing(self, languages_by_rule: Mapping[str, Sequence[str]]) -> tuple[_View, dict[str, int]]:
|
||||
"""This view without the alerts at a repository and pin blind to their rule's language, and how many were
|
||||
dropped per ``repo@pin``."""
|
||||
kept: dict[str, list[Record]] = {}
|
||||
dropped: dict[str, int] = {}
|
||||
for rule_id, alerts in self.alerts.items():
|
||||
languages = languages_by_rule.get(rule_id, ())
|
||||
for alert in alerts:
|
||||
if self.blind(alert.row["repo"], alert.row["pin"], languages):
|
||||
label = f"{alert.row['repo']}@{alert.row['pin']}"
|
||||
dropped[label] = dropped.get(label, 0) + 1
|
||||
else:
|
||||
kept.setdefault(rule_id, []).append(alert)
|
||||
return _View(self.verdicts, self.sites, kept, self.coverage), dict(sorted(dropped.items()))
|
||||
|
||||
|
||||
def _m1(rule_id: str, view: _View) -> tuple[list[Record], dict[str, int]] | None:
|
||||
evidence: list[Record] = []
|
||||
repos: set[str] = set()
|
||||
false_alerts = 0
|
||||
for alert in view.alerts.get(rule_id, []):
|
||||
verdict = view.verdict(alert.row)
|
||||
final = verdict.row.get("final") if verdict else None
|
||||
if final == "agreed" or (verdict is not None and verdict.row.get("site_match") is True):
|
||||
return None # the rule caught something the plane agreed on: never demote it on this evidence
|
||||
if final == "rejected" or view.fixed_in_range(alert.row):
|
||||
false_alerts += 1
|
||||
repos.add(alert.row["repo"])
|
||||
evidence.extend([alert, verdict] if verdict is not None and final == "rejected" else [alert])
|
||||
if false_alerts < M1_MIN_FALSE or len(repos) < M1_MIN_REPOS:
|
||||
return None
|
||||
return evidence, {"false_alerts": false_alerts, "repos": len(repos), "agreed_hits": 0}
|
||||
|
||||
|
||||
def _m2(rule_id: str, view: _View, enabled: set[str]) -> tuple[list[Record], dict[str, int]] | None:
|
||||
alerts = view.alerts.get(rule_id, [])
|
||||
hits: dict[tuple[str, str, str], tuple[Record, Record]] = {}
|
||||
rejected = 0
|
||||
for alert in alerts:
|
||||
verdict = view.verdict(alert.row)
|
||||
if verdict is None or alert.row.get("pin_role") == "fixed":
|
||||
continue
|
||||
if verdict.row.get("final") == "rejected":
|
||||
rejected += 1
|
||||
elif verdict.row.get("final") == "agreed" and verdict.row.get("site_match") is True:
|
||||
hits[(alert.row["repo"], alert.row["pin"], _cand(alert.row))] = (alert, verdict)
|
||||
if not hits:
|
||||
return None
|
||||
caught = {
|
||||
(a.row["repo"], a.row["pin"], _cand(a.row)) for rid in enabled if rid != rule_id for a in view.alerts.get(rid, [])
|
||||
} # fmt: skip
|
||||
missed = {k: v for k, v in hits.items() if k not in caught}
|
||||
fixed_same = sum(
|
||||
1 for a in alerts if a.row.get("pin_role") == "fixed" and any((a.row["repo"], _cand(a.row)) == (r, c) for r, _, c in missed)
|
||||
)
|
||||
repos = {repo for repo, _, _ in missed}
|
||||
if len(missed) < M2_MIN_HITS or len(repos) < M2_MIN_REPOS or fixed_same > 0 or rejected > M2_MAX_REJECTED:
|
||||
return None
|
||||
evidence = [rec for key in sorted(missed) for rec in missed[key]]
|
||||
return evidence, {"site_hits": len(missed), "repos": len(repos), "rejected": rejected, "fixed_pin_same_function": 0}
|
||||
|
||||
|
||||
def _evidence(records: Sequence[Record]) -> list[dict[str, Any]]:
|
||||
unique = {rec.row["id"]: rec for rec in records}
|
||||
return [
|
||||
{"id": i, "kind": r.row["kind"], "repo": r.row["repo"], "pin": r.row["pin"], "run": r.row["run"], "task": r.row["task"],
|
||||
"key": r.key, "version": r.version}
|
||||
for i, r in sorted(unique.items())
|
||||
] # fmt: skip
|
||||
|
||||
|
||||
def evidence_digest(evidence_ids: Iterable[str]) -> str:
|
||||
return hashlib.sha256("\n".join(sorted(evidence_ids)).encode("utf-8")).hexdigest()
|
||||
|
||||
|
||||
def _proposal(
|
||||
rule: str, edit_key: str, rule_id: str, records: Sequence[Record], counts: dict[str, int], excluded: dict[str, Any], index: str
|
||||
) -> dict[str, Any]:
|
||||
evidence = _evidence(records)
|
||||
ids = [e["id"] for e in evidence]
|
||||
seed = json.dumps([rule, edit_key, ids, THRESHOLDS[rule]], sort_keys=True, separators=(",", ":"))
|
||||
return {
|
||||
"proposal_id": hashlib.sha256(seed.encode("utf-8")).hexdigest(),
|
||||
"rule": rule,
|
||||
"edit_key": edit_key,
|
||||
"rule_id": rule_id,
|
||||
"evidence": evidence,
|
||||
"evidence_digest": evidence_digest(ids),
|
||||
"counts": counts,
|
||||
"thresholds": THRESHOLDS[rule],
|
||||
"excluded": excluded,
|
||||
"index_digest": index,
|
||||
}
|
||||
|
||||
|
||||
def propose_from_memory(
|
||||
rows: Sequence[Record],
|
||||
ruleset_by_id: Mapping[str, PatternRule],
|
||||
current_ledger: Mapping[str, dict[str, object]],
|
||||
journal_rounds: Sequence[Mapping[str, object]],
|
||||
*,
|
||||
excluded: Guard,
|
||||
index_digest: str = "",
|
||||
) -> list[MemoryProposal]:
|
||||
"""M1 and M2 over the rows the guard keeps, minus proposals already reverted on the same evidence."""
|
||||
kept, dropped = excluded.apply(rows)
|
||||
view, blind = _View.of(kept).seeing({rid: rule.languages for rid, rule in ruleset_by_id.items()})
|
||||
if blind:
|
||||
dropped = {**dropped, "coverage_none": blind}
|
||||
statuses = {rid: _status(rid, ruleset_by_id, current_ledger) for rid in ruleset_by_id}
|
||||
enabled = {rid for rid, status in statuses.items() if status == "enabled"}
|
||||
proposals: list[MemoryProposal] = []
|
||||
for rule_id in sorted(view.alerts):
|
||||
status = statuses.get(rule_id)
|
||||
if status == "enabled" and (m1 := _m1(rule_id, view)):
|
||||
name, found, to_status = "M1", m1, "shadow"
|
||||
elif status == "shadow" and (m2 := _m2(rule_id, view, enabled)):
|
||||
name, found, to_status = "M2", m2, "enabled"
|
||||
else:
|
||||
continue # unknown to the ruleset (the loop cannot add rules), disabled, or no rule fired
|
||||
key = RuleStatusEdit(rule_id, status, to_status).key()
|
||||
provenance = _proposal(name, key, rule_id, found[0], found[1], dropped, index_digest)
|
||||
edit = RuleStatusEdit(rule_id, status, to_status, rationale=RATIONALE_PREFIX + provenance["proposal_id"])
|
||||
proposals.append(MemoryProposal(edit, provenance))
|
||||
return applicable(proposals, ruleset_by_id, current_ledger, journal_rounds)
|
||||
|
||||
|
||||
def applicable(
|
||||
proposals: Sequence[MemoryProposal],
|
||||
ruleset_by_id: Mapping[str, PatternRule],
|
||||
current_ledger: Mapping[str, dict[str, object]],
|
||||
journal_rounds: Sequence[Mapping[str, object]],
|
||||
) -> list[MemoryProposal]:
|
||||
"""Proposals whose rule still has the ``from`` status and whose evidence no reverted round already tried."""
|
||||
tried: set[tuple[str, str]] = set()
|
||||
for entry in journal_rounds:
|
||||
edits = entry.get("edits")
|
||||
if entry.get("outcome") == "reverted" and isinstance(edits, list):
|
||||
tried |= {(str(e.get("key")), str(e.get("evidence"))) for e in edits if isinstance(e, dict) and e.get("evidence")}
|
||||
return [
|
||||
p for p in proposals
|
||||
if _status(p.edit.rule_id, ruleset_by_id, current_ledger) == p.edit.from_status and (p.edit.key(), p.proposal_id) not in tried
|
||||
] # fmt: skip
|
||||
|
||||
|
||||
def disputed_families(rows: Sequence[Record], guard: Guard) -> list[dict[str, Any]]:
|
||||
"""M3: families disputed often enough to route differently. Advisory only: never an edit."""
|
||||
kept, _ = guard.apply(rows)
|
||||
per: dict[str, list[int]] = {}
|
||||
for rec in kept:
|
||||
if rec.row["kind"] == "verdict" and rec.row.get("family"):
|
||||
tally = per.setdefault(str(rec.row["family"]), [0, 0])
|
||||
tally[0] += 1
|
||||
tally[1] += rec.row.get("final") == "disputed"
|
||||
return [
|
||||
{"family": fam, "candidates": n, "disputed": d, "dispute_rate": round(d / n, 4), "advice": "schedule pass c up front"}
|
||||
for fam, (n, d) in sorted(per.items())
|
||||
if n >= M3_MIN_CANDIDATES and d / n >= M3_MIN_DISPUTE_RATE
|
||||
]
|
||||
|
||||
|
||||
# --- the store, the sidecar and the outcomes -----------------------------------------------------------------------
|
||||
|
||||
|
||||
def index_digest(store: MemoryStore) -> str:
|
||||
return rows_digest(store.index())
|
||||
|
||||
|
||||
def write_sidecar(journal_path: Path, round_index: int, proposals: Sequence[MemoryProposal]) -> None:
|
||||
"""Append each proposal once (by id) to ``memory_proposals.jsonl`` next to the journal."""
|
||||
if not proposals:
|
||||
return
|
||||
path = journal_path.with_name(SIDECAR)
|
||||
seen = {json.loads(line)["proposal_id"] for line in path.read_text().splitlines() if line.strip()} if path.is_file() else set()
|
||||
new = [p for p in proposals if p.proposal_id not in seen]
|
||||
if not new:
|
||||
return
|
||||
path.parent.mkdir(parents=True, exist_ok=True)
|
||||
with path.open("a") as sink:
|
||||
for p in new:
|
||||
sink.write(json.dumps({**p.provenance, "round": round_index}, sort_keys=True, separators=(",", ":")) + "\n")
|
||||
|
||||
|
||||
def read_sidecar(journal_path: Path) -> dict[str, dict[str, Any]]:
|
||||
path = journal_path.with_name(SIDECAR)
|
||||
if not path.is_file():
|
||||
return {}
|
||||
return {row["proposal_id"]: row for row in (json.loads(line) for line in path.read_text().splitlines() if line.strip())}
|
||||
|
||||
|
||||
def write_signals(directory: Path, signals: Sequence[Mapping[str, Any]]) -> Path:
|
||||
path = directory / SIGNALS
|
||||
path.parent.mkdir(parents=True, exist_ok=True)
|
||||
path.write_text(json.dumps({"advisory": True, "disputed_families": list(signals), "thresholds": THRESHOLDS["M3"]}, indent=2) + "\n")
|
||||
return path
|
||||
|
||||
|
||||
def outcome_of(reason: str, accepted: bool) -> str | None:
|
||||
if accepted:
|
||||
return "accepted"
|
||||
if reason == "no_proposals":
|
||||
return None
|
||||
if reason.startswith(("validation_failed", "replay_failed")):
|
||||
return "rejected"
|
||||
return "reverted"
|
||||
|
||||
|
||||
def outcome_rows(
|
||||
edits: Sequence[RuleStatusEdit],
|
||||
outcome: str,
|
||||
round_index: int,
|
||||
reason: str,
|
||||
proposals: Mapping[str, MemoryProposal],
|
||||
gate_run: str,
|
||||
) -> list[dict[str, Any]]:
|
||||
"""One ``proposal_outcome`` row per (repository, pin) of each memory edit's evidence."""
|
||||
rows: list[dict[str, Any]] = []
|
||||
for edit in edits:
|
||||
pid = proposal_id_of(edit)
|
||||
proposal = proposals.get(pid or "")
|
||||
if proposal is None:
|
||||
continue
|
||||
by_place: dict[tuple[str, str], dict[str, Any]] = {}
|
||||
for ev in proposal.provenance["evidence"]:
|
||||
by_place.setdefault((ev["repo"], ev["pin"]), ev)
|
||||
for (repo, pin), _ in sorted(by_place.items()):
|
||||
task = f"round-{round_index}"
|
||||
rows.append(
|
||||
{
|
||||
"id": row_id("proposal_outcome", gate_run, task, f"{pid}:{repo}:{pin}"), "kind": "proposal_outcome",
|
||||
"repo": repo, "pin": pin, "run": gate_run, "task": task, "population": "improve-loop", "split": "loop",
|
||||
"image": "host", "edit_key": edit.key(), "proposal_id": pid,
|
||||
"evidence_digest": proposal.provenance["evidence_digest"], "outcome": outcome, "reason": reason,
|
||||
"gate_run": gate_run,
|
||||
}
|
||||
) # fmt: skip
|
||||
return rows
|
||||
|
||||
|
||||
def gate_run_name(manifest_name: str) -> str:
|
||||
return f"improve:{manifest_name}:{datetime.now(UTC).strftime('%Y%m%dT%H%M%SZ')}"
|
||||
|
||||
|
||||
# --- the `ousast improve --memory` session -------------------------------------------------------------------------
|
||||
|
||||
|
||||
def benchmark_cases(manifest: object) -> frozenset[tuple[str, str]]:
|
||||
"""(repo, pin) of the gated benchmark's cases. Local sources (the only kind `resolve_benchmark_source` accepts)
|
||||
name no repository pin, so they exclude nothing here; a pinned source would list its case."""
|
||||
source = getattr(manifest, "source", None)
|
||||
repo, pin = getattr(source, "repo", ""), getattr(source, "pin", "")
|
||||
return frozenset({(repo_key(repo), pin)}) if repo and pin else frozenset()
|
||||
|
||||
|
||||
def store_spec(value: str) -> str | None:
|
||||
"""``--memory`` value -> ``open_store`` spec: empty uses OUSAST_MEMORY or the default, a bare path is a FileStore."""
|
||||
if not value:
|
||||
return None
|
||||
return value if "://" in value else f"file://{Path(value).resolve()}"
|
||||
|
||||
|
||||
@dataclass
|
||||
class MemorySession:
|
||||
"""One `improve --memory` invocation: the store read once, proposals per round, outcomes written back."""
|
||||
|
||||
store: MemoryStore
|
||||
guard: Guard
|
||||
records: list[Record]
|
||||
ruleset_dir: Path
|
||||
gate_run: str
|
||||
made: dict[str, MemoryProposal] = field(default_factory=dict)
|
||||
|
||||
@classmethod
|
||||
def open(
|
||||
cls, spec: str, manifest: object, target: Path, ruleset_dir: Path, populations: Sequence[str], catalog: Path | None
|
||||
) -> MemorySession:
|
||||
from ..pairs import load_pair_catalog, select_split
|
||||
from ..plane.memory import FileStore, open_store
|
||||
|
||||
store = open_store(store_spec(spec))
|
||||
if isinstance(store, FileStore) and not (store.root / "index.jsonl").is_file():
|
||||
raise SystemExit(f"memory store {store.describe()} has no index.jsonl: an unread store would propose nothing")
|
||||
if catalog is None or not catalog.is_file():
|
||||
raise SystemExit(f"--memory needs the pair catalog for the holdout guard: {catalog} is not a file")
|
||||
holdout = select_split(load_pair_catalog(catalog), "holdout")
|
||||
guard = Guard.build(populations, benchmark_cases(manifest), holdout)
|
||||
records = store.rows()
|
||||
print(
|
||||
f"memory={store.describe()} index_entries={len(store.index())} rows={len(records)} "
|
||||
f"guard: populations={sorted(guard.populations) or '-'} gated_cases={len(guard.gated_cases)} "
|
||||
f"holdout_repos={len(guard.holdout_repos)}"
|
||||
)
|
||||
return cls(store, guard, records, ruleset_dir, gate_run_name(str(getattr(manifest, "name", "manifest"))))
|
||||
|
||||
def proposals(self, ledger_path: Path, journal_path: Path) -> list[MemoryProposal]:
|
||||
from ..ruleset import load_ruleset, read_rule_ledger
|
||||
from .journal import load_journal
|
||||
|
||||
rules = load_ruleset(self.ruleset_dir, ledger_path if ledger_path.is_file() else None)
|
||||
found = propose_from_memory(
|
||||
self.records,
|
||||
{r.rule_id: r for r in rules},
|
||||
read_rule_ledger(ledger_path),
|
||||
load_journal(journal_path),
|
||||
excluded=self.guard,
|
||||
index_digest=index_digest(self.store),
|
||||
)
|
||||
kept, excluded = self.guard.apply(self.records)
|
||||
kinds: dict[str, int] = {}
|
||||
for rec in kept:
|
||||
kinds[rec.row["kind"]] = kinds.get(rec.row["kind"], 0) + 1
|
||||
print(f"memory_rows kept={len(kept)} by_kind={json.dumps(dict(sorted(kinds.items())))}")
|
||||
print(f"memory_excluded={json.dumps(excluded, sort_keys=True)}")
|
||||
for p in found:
|
||||
print(f"memory_proposal {p.edit.key()} rule={p.provenance['rule']} id={p.proposal_id} evidence={len(p.provenance['evidence'])}")
|
||||
print(f"memory_proposals={len(found)}")
|
||||
self.made.update({p.proposal_id: p for p in found})
|
||||
return found
|
||||
|
||||
def finish(self, outcomes: Sequence[Any], journal_path: Path, *, dry_run: bool) -> None:
|
||||
signals = disputed_families(self.records, self.guard)
|
||||
path = write_signals(journal_path.parent, signals)
|
||||
print(f"memory_signals disputed_families={[s['family'] for s in signals] or '-'} ({path.name}, advisory)")
|
||||
rows: list[dict[str, Any]] = []
|
||||
for o in outcomes:
|
||||
outcome = outcome_of(o.reason, o.accepted)
|
||||
if outcome is not None:
|
||||
rows += outcome_rows(o.edits, outcome, o.round, o.reason, self.made, self.gate_run)
|
||||
if dry_run:
|
||||
print(f"memory_outcomes={len(rows)} (dry run: not written to the store)")
|
||||
return
|
||||
if rows:
|
||||
self.store.put_rows(rows)
|
||||
print(f"memory_outcomes={len(rows)} written to {self.store.describe()}")
|
||||
|
||||
|
||||
__all__ = [
|
||||
"Guard",
|
||||
"MemorySession",
|
||||
"benchmark_cases",
|
||||
"store_spec",
|
||||
"M1_MIN_FALSE",
|
||||
"M1_MIN_REPOS",
|
||||
"M2_MAX_REJECTED",
|
||||
"M2_MIN_HITS",
|
||||
"M2_MIN_REPOS",
|
||||
"M3_MIN_CANDIDATES",
|
||||
"M3_MIN_DISPUTE_RATE",
|
||||
"MemoryProposal",
|
||||
"THRESHOLDS",
|
||||
"applicable",
|
||||
"disputed_families",
|
||||
"evidence_digest",
|
||||
"gate_run_name",
|
||||
"index_digest",
|
||||
"outcome_of",
|
||||
"outcome_rows",
|
||||
"propose_from_memory",
|
||||
"proposal_id_of",
|
||||
"read_sidecar",
|
||||
"write_sidecar",
|
||||
"write_signals",
|
||||
]
|
||||
@@ -1,125 +0,0 @@
|
||||
"""Bounded-lever edit model + validator for the self-improvement loop (tasks 7.1, 7.2).
|
||||
|
||||
The loop may only adjust governance *data* through two bounded levers and never the
|
||||
two authorities a human/upstream owns: detection **pattern text** and the 0-5 CWE
|
||||
**severity**. Edits are typed so pattern/severity changes are structurally
|
||||
impossible; the validator additionally enforces status/bounds/staging/resolution.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from collections.abc import Mapping
|
||||
from dataclasses import dataclass
|
||||
|
||||
from ..policy import CwePolicy
|
||||
from ..ruleset import PatternRule
|
||||
from ..ruleset.store import VALID_STATUS
|
||||
|
||||
# The only levers the loop may pull (mirrors the design's journal._VALID_LEVERS).
|
||||
VALID_LEVERS = frozenset({"rule", "policy", "mechanisms"})
|
||||
_IDENTIFIER = re.compile(r"^[A-Za-z_$][\w$]*$")
|
||||
MECHANISM_ACTIONS = frozenset({"admit", "retract"})
|
||||
# Policy constants the loop may tune; the 0-5 severity is intentionally absent.
|
||||
TUNABLE_POLICY_CONSTANTS = frozenset({"K", "MIN_SCORE"})
|
||||
|
||||
|
||||
class StrictValidationError(ValueError):
|
||||
"""Raised when a proposed loop edit violates the bounded-lever contract."""
|
||||
|
||||
def __init__(self, message: str, *, kind: str = "rule_change") -> None:
|
||||
super().__init__(message)
|
||||
self.kind = kind
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class RuleStatusEdit:
|
||||
"""A `rule`-lever edit: change a rule's status / evidence floor / precision estimate."""
|
||||
|
||||
rule_id: str
|
||||
from_status: str
|
||||
to_status: str
|
||||
rationale: str = ""
|
||||
lever: str = "rule"
|
||||
|
||||
def key(self) -> str:
|
||||
return f"rule:{self.rule_id}:{self.from_status}->{self.to_status}"
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class PolicyConstantEdit:
|
||||
"""A `policy`-lever edit: tune a score constant (K / MIN_SCORE). Never severity."""
|
||||
|
||||
name: str
|
||||
from_value: float
|
||||
to_value: float
|
||||
rationale: str = ""
|
||||
lever: str = "policy"
|
||||
|
||||
def key(self) -> str:
|
||||
return f"policy:{self.name}:{self.from_value}->{self.to_value}"
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class EvolveBounds:
|
||||
k_range: tuple[float, float] = (20.0, 200.0)
|
||||
min_score_range: tuple[int, int] = (0, 100)
|
||||
|
||||
|
||||
class EvolveValidator:
|
||||
"""Enforces the bounded-lever safety contract; raises StrictValidationError on any breach."""
|
||||
|
||||
def __init__(self, bounds: EvolveBounds | None = None) -> None:
|
||||
self._bounds = bounds or EvolveBounds()
|
||||
|
||||
def validate(self, edit: object, ruleset_by_id: Mapping[str, PatternRule], policy: Mapping[str, CwePolicy]) -> None:
|
||||
if isinstance(edit, RuleStatusEdit):
|
||||
self._validate_rule(edit, ruleset_by_id, policy)
|
||||
elif isinstance(edit, PolicyConstantEdit):
|
||||
self._validate_policy(edit)
|
||||
else:
|
||||
raise StrictValidationError(f"unknown edit type {type(edit).__name__}", kind="rule_change")
|
||||
|
||||
def _validate_rule(self, edit: RuleStatusEdit, ruleset_by_id: Mapping[str, PatternRule], policy: Mapping[str, CwePolicy]) -> None:
|
||||
if edit.to_status not in VALID_STATUS:
|
||||
raise StrictValidationError(f"invalid status {edit.to_status!r} for {edit.rule_id}")
|
||||
rule = ruleset_by_id.get(edit.rule_id)
|
||||
if rule is None:
|
||||
raise StrictValidationError(f"edit targets unknown rule_id {edit.rule_id!r} (loop cannot add rules)")
|
||||
if edit.from_status == "enabled" and edit.to_status == "disabled":
|
||||
raise StrictValidationError(f"{edit.rule_id}: enabled->disabled jump is forbidden; stage via shadow first")
|
||||
if rule.cwe not in policy:
|
||||
raise StrictValidationError(f"{edit.rule_id}: CWE {rule.cwe} does not resolve in the policy")
|
||||
|
||||
def _validate_policy(self, edit: PolicyConstantEdit) -> None:
|
||||
if edit.name not in TUNABLE_POLICY_CONSTANTS:
|
||||
raise StrictValidationError(
|
||||
f"policy constant {edit.name!r} is not loop-tunable (severity is upstream-owned)", kind="policy_change"
|
||||
)
|
||||
if edit.name == "K" and not (self._bounds.k_range[0] <= edit.to_value <= self._bounds.k_range[1]):
|
||||
raise StrictValidationError(f"K={edit.to_value} out of bounds {self._bounds.k_range}", kind="policy_change")
|
||||
if edit.name == "MIN_SCORE" and not (self._bounds.min_score_range[0] <= edit.to_value <= self._bounds.min_score_range[1]):
|
||||
raise StrictValidationError(f"MIN_SCORE={edit.to_value} out of bounds {self._bounds.min_score_range}", kind="policy_change")
|
||||
|
||||
|
||||
def edits_to_ledger(edits: list[RuleStatusEdit], base: Mapping[str, dict[str, object]]) -> dict[str, dict[str, object]]:
|
||||
"""Overlay rule-status edits onto a copy of the base ledger."""
|
||||
ledger: dict[str, dict[str, object]] = {key: dict(value) for key, value in base.items()}
|
||||
for edit in edits:
|
||||
entry = dict(ledger.get(edit.rule_id, {}))
|
||||
entry["status"] = edit.to_status
|
||||
ledger[edit.rule_id] = entry
|
||||
return ledger
|
||||
|
||||
|
||||
__all__ = [
|
||||
"EvolveBounds",
|
||||
"EvolveValidator",
|
||||
"MECHANISM_ACTIONS",
|
||||
"PolicyConstantEdit",
|
||||
"RuleStatusEdit",
|
||||
"StrictValidationError",
|
||||
"TUNABLE_POLICY_CONSTANTS",
|
||||
"VALID_LEVERS",
|
||||
"edits_to_ledger",
|
||||
]
|
||||
@@ -1,8 +1,8 @@
|
||||
"""Vuln-vs-fixed pair corpus: TP on the vuln tree, silent on the fix tree.
|
||||
|
||||
Isolated trees keep the two sides from contaminating each other (a `safe.py`
|
||||
sitting next to `app.py` cannot measure false positives). Pair outcomes feed
|
||||
the improvement loop as miss (vuln) and fp (fix) signals.
|
||||
sitting next to `app.py` cannot measure false positives). Pair outcomes report
|
||||
misses on vulnerable code and false positives on fixed code.
|
||||
|
||||
pair-corpus-honesty: the overlay scorer counts ``coverage`` records with a
|
||||
source as detections and as leaks, matches by function when a label names one,
|
||||
@@ -61,7 +61,7 @@ FunctionRanges = dict[str, tuple[tuple[str, int, int], ...]]
|
||||
SLICE_NAMES = ("all", "local", "github", "sast", "vfc", "vibe-py", "vfc-js", "agent-vfc", "owasp")
|
||||
PROVENANCES = frozenset({"human", "agent", "mixed", "synthetic"})
|
||||
SPLITS = frozenset({"train", "holdout"})
|
||||
# Req 9: how a label was established. Only seeded and reviewed pairs steer the improve loop.
|
||||
# Req 9: how a label was established.
|
||||
REVIEW_TIERS = frozenset({"seeded", "advisory", "title", "reviewed"})
|
||||
GATING_TIERS = frozenset({"seeded", "reviewed"})
|
||||
# Req 10: pointer pairs are harvested into a cache outside the repository, never vendored.
|
||||
@@ -442,7 +442,7 @@ def evaluate_catalog(
|
||||
inventory_sidecar: bool = True,
|
||||
pointers: bool | None = None,
|
||||
) -> PairEvalResult:
|
||||
"""Score every case. ``ruleset`` lets the improve loop score a candidate ledger; ``inventory_sidecar=False`` skips the second scan.
|
||||
"""Score every case. ``ruleset`` selects the rules to score; ``inventory_sidecar=False`` skips the second scan.
|
||||
|
||||
Pointer pairs (``vendored = false``) are harvested into the cache when ``pointers`` (or OPENULTRASAST_PAIRS_NETWORK=1) allows the
|
||||
network, otherwise skipped with a ``pointer_pair_skipped`` degradation (Req 10.2, 10.3).
|
||||
@@ -497,14 +497,14 @@ def evaluate_catalog(
|
||||
|
||||
|
||||
def build_pair_signals(outcomes: Sequence[PairOutcome], *, split: str | None = None) -> list[dict[str, object]]:
|
||||
"""Same miss/fp vocabulary as ``build_rule_signals``, tagged with the pair name, profile, and mechanisms.
|
||||
"""Miss and false-positive signals, tagged with the pair name, profile, and mechanisms.
|
||||
|
||||
``split`` restricts the signals to one split, so a learning path teaches from the train rows only (Req 5.1).
|
||||
"""
|
||||
signals: list[dict[str, object]] = []
|
||||
for outcome in outcomes:
|
||||
if outcome.known_limit is not None:
|
||||
continue # unachievable by construction; the loop must not chase it
|
||||
continue # excluded from the achievable metrics
|
||||
if split is not None and outcome.split != split:
|
||||
continue
|
||||
common: dict[str, object] = {
|
||||
@@ -653,7 +653,7 @@ def select_slice(cases: Sequence[PairCase], slice_name: str | None) -> tuple[Pai
|
||||
|
||||
|
||||
def select_tier(cases: Sequence[PairCase], tiers: Iterable[str]) -> tuple[PairCase, ...]:
|
||||
"""Pairs whose review tier is in ``tiers`` (the improve loop passes GATING_TIERS)."""
|
||||
"""Pairs whose review tier is in ``tiers``."""
|
||||
wanted = frozenset(tiers)
|
||||
return tuple(case for case in cases if case.review_tier in wanted)
|
||||
|
||||
|
||||
@@ -20,13 +20,6 @@ under ``--plane`` (default ``plane/``), everything a Run over the set needs besi
|
||||
with budgets, annotated with the population and split (``openultrasast.io/population``/``split``: the files'
|
||||
stems) the memory rows carry.
|
||||
|
||||
``--loop`` (harnessx-removal design section 6) appends the improvement loop: per case ``alerts`` (quick-mode rules
|
||||
on the vulnerable and the fixed pin, which gets its Workspace ``<case>-fixed``; ``case.json`` then also carries the
|
||||
fix's new-side ranges) feeding ``remember``, and once ``memory-snapshot`` (the store's rows after the train-on-test
|
||||
guard, written on the host by ``memory.seed``) -> ``loop-measure`` -> ``loop-propose`` -> ``loop-improve``, the last
|
||||
two bound to this repository at the commit under test. All of them model-free with budget ``{usd: 0, calls: 0}``;
|
||||
the gate's verdict is ``<run>/loop-improve/gate.json``, and adopting an accepted ledger stays a maintainer commit.
|
||||
|
||||
Each ``repo-facts`` Task carries ``openultrasast.io/memory-key`` (repository, pin, candidates digest, runner image
|
||||
digest): the key under which stored facts are reused instead of recomputed (``plane/memory.py``, ``seed``).
|
||||
|
||||
@@ -67,12 +60,12 @@ import yaml
|
||||
from ..model.endpoint import price_of
|
||||
from .budget import cost_of
|
||||
from .manifests import GIT_COMMITS_ANNOTATION, RUN_API_VERSION, Task, load_manifests
|
||||
from .memory import MEMORY_KEY_ANNOTATION, POPULATION_ANNOTATION, SNAPSHOT_ANNOTATION, SPLIT_ANNOTATION, memory_key
|
||||
from .memory import MEMORY_KEY_ANNOTATION, POPULATION_ANNOTATION, SPLIT_ANNOTATION, memory_key
|
||||
from .reconciler import _ax_name, _doc
|
||||
from .tasks.repo_facts import _digest
|
||||
from .tasks.verify import MAX_STEPS, units_of
|
||||
|
||||
__all__ = ["CaseInputs", "Loop", "fix_ranges", "increment", "read_cases", "render", "repin_templates"]
|
||||
__all__ = ["CaseInputs", "fix_ranges", "increment", "read_cases", "render", "repin_templates"]
|
||||
|
||||
REFERENCE_USD_PER_CANDIDATE = 0.022 # two passes: verifier-batched-check-2026-09-29.json cost.usd_per_candidate_two_passes
|
||||
REFERENCE_MODEL = "deepseek-flash" # verify_batched.py's default (DEFAULT_DETECTOR_MODEL) in the batched check
|
||||
@@ -86,12 +79,7 @@ CALLS_PER_HUNT = MAX_STEPS + 2 # every tool step, the answer, one empty-content
|
||||
ATESPACE = "default" # the default when no profile is given; ``increment`` takes ``profile.atespace``
|
||||
CASE_PATH = "/workspace/case"
|
||||
INPUTS_PATH = "/workspace/inputs"
|
||||
FIXED_PATH = "/workspace/fixed"
|
||||
PROJECT_PATH = "/workspace/project"
|
||||
TEMPLATES = ("repo-facts", "verify", "agree", "remember", "alerts", "loop", "features", "roles")
|
||||
FIXED_RANGES_BASIS = "git diff -U0 <vulnerable> <fixed>, new side: the fix's lines at the fixed pin"
|
||||
LOOP_MANIFEST = "benchmarks/manifests/java-spring-boot-vulnerable.toml" # the equality baseline's `improve` manifest
|
||||
LOOP_CATALOG = "benchmarks/pairs/catalog.toml"
|
||||
TEMPLATES = ("repo-facts", "verify", "agree", "remember", "alerts", "features", "roles")
|
||||
|
||||
Candidate = tuple[str, str, int]
|
||||
|
||||
@@ -105,20 +93,6 @@ class CaseInputs:
|
||||
recorded_usd: float
|
||||
triage_usd: float = 0.0
|
||||
ranges: Mapping[str, Sequence[tuple[int, int]]] | None = None
|
||||
fixed_ranges: Mapping[str, Sequence[tuple[int, int]]] | None = None
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Loop:
|
||||
"""The improvement loop appended to a Run (``--loop``, design section 6): ``repo`` at ``commit`` is this
|
||||
repository at the commit under test, where ``manifest`` and ``catalog`` (paths inside it) gate the proposals;
|
||||
rows of ``populations`` never propose (the train-on-test guard)."""
|
||||
|
||||
repo: str
|
||||
commit: str
|
||||
manifest: str = LOOP_MANIFEST
|
||||
catalog: str = LOOP_CATALOG
|
||||
populations: tuple[str, ...] = ()
|
||||
|
||||
|
||||
def fix_ranges(repo: Path, vulnerable: str, fixed: str, side: str = "old") -> dict[str, list[tuple[int, int]]]:
|
||||
@@ -165,11 +139,10 @@ def _scan_rows(triage_dir: Path, case_id: str) -> list[dict[str, Any]]:
|
||||
|
||||
|
||||
def read_cases(
|
||||
validation_set: Path, candidates_dir: Path, triage_dir: Path, population: Path, repos: Path | None = None, *, fixed_side: bool = False
|
||||
validation_set: Path, candidates_dir: Path, triage_dir: Path, population: Path, repos: Path | None = None
|
||||
) -> list[CaseInputs]:
|
||||
"""The set's cases in id order, each with its candidates, the triage-kept subset, the recorded cost and its
|
||||
triage share, and -- with ``repos`` (one clone per case id) -- the fix's old-side ranges (and, with
|
||||
``fixed_side``, its new-side ranges, which the loop's `alerts` task marks fixed-pin alerts by)."""
|
||||
triage share, and -- with ``repos`` (one clone per case id) -- the fix's old-side ranges."""
|
||||
subset: dict[str, list[list[str]]] = json.loads(validation_set.read_text(encoding="utf-8"))
|
||||
records = {c["id"]: c for c in tomllib.loads(population.read_text(encoding="utf-8"))["case"]}
|
||||
cases: list[CaseInputs] = []
|
||||
@@ -199,11 +172,6 @@ def read_cases(
|
||||
recorded_usd=round(sum(float(r.get("usd") or 0.0) for r in rows), 6),
|
||||
triage_usd=_triage_usd(rows),
|
||||
ranges=None if repos is None else fix_ranges(repos / case_id, records[case_id]["vulnerable"], records[case_id]["fixed"]),
|
||||
fixed_ranges=(
|
||||
fix_ranges(repos / case_id, records[case_id]["vulnerable"], records[case_id]["fixed"], side="new")
|
||||
if repos is not None and fixed_side
|
||||
else None
|
||||
),
|
||||
)
|
||||
)
|
||||
return cases
|
||||
@@ -226,9 +194,6 @@ def _case_record(item: CaseInputs) -> dict[str, Any]:
|
||||
if item.ranges is not None:
|
||||
record["ranges"] = {path: [list(span) for span in spans] for path, spans in item.ranges.items()}
|
||||
record["ranges_basis"] = RANGES_BASIS
|
||||
if item.fixed_ranges is not None:
|
||||
record["fixed_ranges"] = {path: [list(span) for span in spans] for path, spans in item.fixed_ranges.items()}
|
||||
record["fixed_ranges_basis"] = FIXED_RANGES_BASIS
|
||||
record["sites_in_set"] = [s for s in sites if s in in_set]
|
||||
record["cost"] = {
|
||||
"candidates_before_triage": len(item.candidates), "recorded_usd": item.recorded_usd,
|
||||
@@ -304,95 +269,38 @@ def _free() -> dict[str, int]:
|
||||
return {"usd": 0, "calls": 0} # model-free: budget.py refuses any call, so a stray one fails loudly
|
||||
|
||||
|
||||
def _chain_inputs(case_id: str, alerts: bool) -> dict[str, str]:
|
||||
"""Every artifact of a case's chain -- facts, the three passes with their summaries, the final agree -- and, in a
|
||||
loop, its alerts with their coverage."""
|
||||
def _chain_inputs(case_id: str) -> dict[str, str]:
|
||||
"""Every artifact of a case's chain: facts, the three passes with their summaries, and the final agreement."""
|
||||
inputs = {"facts": f"{case_id}-facts/facts.json", "facts_summary": f"{case_id}-facts/summary.json"}
|
||||
for label in ("a", "b", "c"):
|
||||
inputs.update({f"pass_{label}": f"{case_id}-v{label}/units.jsonl", f"summary_{label}": f"{case_id}-v{label}/summary.json"})
|
||||
inputs["agreed"] = f"{case_id}-final/agreed.json"
|
||||
if alerts:
|
||||
inputs["alerts"] = f"{case_id}-alerts/alerts.jsonl"
|
||||
inputs["alerts_summary"] = f"{case_id}-alerts/summary.json" # the coverage: whether a zero could mean clean
|
||||
return inputs
|
||||
|
||||
|
||||
def _features(item: CaseInputs, templates: Mapping[str, Task], alerts: bool) -> tuple[dict, dict]:
|
||||
def _features(item: CaseInputs, templates: Mapping[str, Task]) -> tuple[dict, dict]:
|
||||
"""(Task, Run entry) of the case's `features` (learned-decision-engine design section 1): model-free, bound to the
|
||||
vulnerable Workspace for the functions' spans and entry-point distance; the chain's artifacts as input."""
|
||||
case_id = item.case["id"]
|
||||
env = {"OUSAST_WORKSPACE_DIR": f"{CASE_PATH}/repo"}
|
||||
task = _task(templates["features"], f"features-{case_id}", env, [(_ax_name(case_id, "vulnerable"), CASE_PATH)])
|
||||
inputs = _chain_inputs(case_id, alerts)
|
||||
inputs = _chain_inputs(case_id)
|
||||
entry = {"name": f"{case_id}-features", "task": f"features-{case_id}", "inputs": inputs, "outputs": ["features.jsonl", "summary.json"]}
|
||||
return task, {**entry, "budget": _free()}
|
||||
|
||||
|
||||
def _remember(item: CaseInputs, templates: Mapping[str, Task], population: str, split: str, alerts: bool) -> tuple[dict, dict]:
|
||||
def _remember(item: CaseInputs, templates: Mapping[str, Task], population: str, split: str) -> tuple[dict, dict]:
|
||||
"""(Task, Run entry) of the case's `remember` (design section 4): bound to the vulnerable Workspace for the
|
||||
repository and pin its rows carry; every artifact of the case's chain (and, in a loop, its alerts) and the case's
|
||||
repository and pin its rows carry; every artifact of the case's chain and the case's
|
||||
feature records as input."""
|
||||
case_id = item.case["id"]
|
||||
env = {"OUSAST_POPULATION": population or "unknown", "OUSAST_SPLIT": split or "unknown"}
|
||||
task = _task(templates["remember"], f"remember-{case_id}", env, [(_ax_name(case_id, "vulnerable"), CASE_PATH)])
|
||||
inputs = {**_chain_inputs(case_id, alerts), "features": f"{case_id}-features/features.jsonl"}
|
||||
inputs = {**_chain_inputs(case_id), "features": f"{case_id}-features/features.jsonl"}
|
||||
entry = {"name": f"{case_id}-remember", "task": f"remember-{case_id}", "inputs": inputs, "outputs": ["memory.jsonl"], "budget": _free()}
|
||||
return task, entry
|
||||
|
||||
|
||||
def _alerts(item: CaseInputs, templates: Mapping[str, Task], atespace: str = ATESPACE) -> tuple[dict, dict, dict]:
|
||||
"""(fixed-pin Workspace, Task, Run entry) of the case's `alerts`: quick-mode rules on both pins."""
|
||||
case = item.case
|
||||
git = [{"name": "repo", "repo": case["repo"], "dir": "repo", "depth": 1}]
|
||||
fixed = _doc("Workspace", _ax_name(case["id"], "fixed"), {"git": git}, atespace)
|
||||
fixed["metadata"]["annotations"] = {GIT_COMMITS_ANNOTATION: f"repo={case['fixed']}"}
|
||||
bindings = [(_ax_name(case["id"], "vulnerable"), CASE_PATH), (fixed["metadata"]["name"], FIXED_PATH)]
|
||||
bindings.append((_ax_name(case["id"], "inputs"), INPUTS_PATH))
|
||||
env = {
|
||||
"OUSAST_WORKSPACE_DIR": f"{CASE_PATH}/repo", "OUSAST_FIXED_DIR": f"{FIXED_PATH}/repo",
|
||||
"OUSAST_INPUT_CASE": f"{INPUTS_PATH}/case.json", "OUSAST_VULNERABLE_PIN": str(case["vulnerable"]),
|
||||
"OUSAST_FIXED_PIN": str(case["fixed"]),
|
||||
} # fmt: skip
|
||||
task = _task(templates["alerts"], f"alerts-{case['id']}", env, bindings)
|
||||
entry = {"name": f"{case['id']}-alerts", "task": f"alerts-{case['id']}", "outputs": ["alerts.jsonl", "summary.json"], "budget": _free()}
|
||||
entry["serialize"] = "engine"
|
||||
return fixed, task, entry
|
||||
|
||||
|
||||
def _loop(
|
||||
cases: Sequence[CaseInputs], templates: Mapping[str, Task], loop: Loop, atespace: str = ATESPACE
|
||||
) -> tuple[dict, list[dict], list[dict]]:
|
||||
"""(project Workspace, Tasks, Run entries) of the loop's singletons: snapshot -> measure -> propose -> improve."""
|
||||
git = [{"name": "repo", "repo": loop.repo, "dir": "repo", "depth": 1}]
|
||||
project = _doc("Workspace", _ax_name("openultrasast", loop.commit[:12]), {"git": git}, atespace)
|
||||
project["metadata"]["annotations"] = {GIT_COMMITS_ANNOTATION: f"repo={loop.commit}"}
|
||||
bound = [(project["metadata"]["name"], PROJECT_PATH)]
|
||||
env = {
|
||||
"OUSAST_PROJECT_DIR": f"{PROJECT_PATH}/repo", "OUSAST_LOOP_MANIFEST": loop.manifest, "OUSAST_LOOP_CATALOG": loop.catalog,
|
||||
"OUSAST_QUALIFY_POPULATIONS": json.dumps(sorted(loop.populations)),
|
||||
} # fmt: skip
|
||||
snapshot = _task(templates["loop"], "memory-snapshot", {}, [], ["snapshot"])
|
||||
guard = {"catalog": loop.catalog, "manifest": loop.manifest, "populations": sorted(loop.populations)}
|
||||
snapshot["metadata"].setdefault("annotations", {})[SNAPSHOT_ANNOTATION] = json.dumps(guard, sort_keys=True, separators=(",", ":"))
|
||||
tasks = [
|
||||
snapshot, _task(templates["loop"], "loop-measure", {}, [], ["measure"]),
|
||||
_task(templates["loop"], "loop-propose", env, bound, ["propose"]),
|
||||
_task(templates["loop"], "loop-improve", env, bound, ["improve"]),
|
||||
] # fmt: skip
|
||||
remembered = {f"remember_{c.case['id']}": f"{c.case['id']}-remember/memory.jsonl" for c in cases}
|
||||
entries = [
|
||||
{"name": "memory-snapshot", "task": "memory-snapshot", "outputs": ["rows.jsonl", "index.json"], "budget": _free()},
|
||||
{"name": "loop-measure", "task": "loop-measure", "inputs": {**remembered, "snapshot": "memory-snapshot/rows.jsonl"},
|
||||
"outputs": ["measure.json", "rows.jsonl"], "budget": _free()},
|
||||
{"name": "loop-propose", "task": "loop-propose",
|
||||
"inputs": {"measure": "loop-measure/measure.json", "rows": "loop-measure/rows.jsonl", "index": "memory-snapshot/index.json"},
|
||||
"outputs": ["proposals.jsonl", "memory_proposals.jsonl", "signals.json"], "budget": _free()},
|
||||
{"name": "loop-improve", "task": "loop-improve", "inputs": {"proposals": "loop-propose/proposals.jsonl"},
|
||||
"outputs": ["gate.json", "journal.json"], "budget": _free(), "serialize": "engine"},
|
||||
] # fmt: skip
|
||||
return project, tasks, entries
|
||||
|
||||
|
||||
def render(
|
||||
cases: Sequence[CaseInputs],
|
||||
templates: Mapping[str, Task],
|
||||
@@ -401,13 +309,11 @@ def render(
|
||||
*,
|
||||
population: str = "",
|
||||
split: str = "",
|
||||
loop: Loop | None = None,
|
||||
atespace: str = ATESPACE,
|
||||
) -> dict[str, str]:
|
||||
"""Relative path under the plane root -> file text, for every generated manifest. ``population`` and ``split``
|
||||
(the population file's and the set's stems) become the Run's annotations, which the memory rows carry. Each
|
||||
case ends in `features` -> `remember`; ``loop`` adds per case `alerts` (and the fixed-pin Workspace), and once the
|
||||
snapshot, `loop-measure`, `loop-propose` and `loop-improve` (design section 6), all model-free."""
|
||||
case ends in `features` -> `remember`."""
|
||||
header = f"# generated by `{command}`; do not edit, regenerate.\n"
|
||||
out: dict[str, str] = {}
|
||||
tasks: list[dict[str, Any]] = []
|
||||
@@ -454,22 +360,12 @@ def render(
|
||||
tasks.sort(key=lambda doc: 1 if doc["metadata"]["name"].startswith("verify-c-") else 0) # stable: the old Tasks first
|
||||
entries.extend(tiebreak) # appended: the first Run's entries stay byte-identical, so a rerun keeps their state
|
||||
for item in cases: # appended after the tie-break for the same reason
|
||||
if loop is not None:
|
||||
fixed, alerts_task, alerts_entry = _alerts(item, templates, atespace)
|
||||
out[f"workspaces/{fixed['metadata']['name']}.yaml"] = _dump(header, [fixed])
|
||||
tasks.append(alerts_task)
|
||||
entries.append(alerts_entry)
|
||||
features_task, features_entry = _features(item, templates, loop is not None)
|
||||
features_task, features_entry = _features(item, templates)
|
||||
tasks.append(features_task)
|
||||
entries.append(features_entry)
|
||||
remember_task, remember_entry = _remember(item, templates, population, split, loop is not None)
|
||||
remember_task, remember_entry = _remember(item, templates, population, split)
|
||||
tasks.append(remember_task)
|
||||
entries.append(remember_entry)
|
||||
if loop is not None:
|
||||
project, loop_tasks, loop_entries = _loop(cases, templates, loop, atespace)
|
||||
out[f"workspaces/{project['metadata']['name']}.yaml"] = _dump(header, [project])
|
||||
tasks.extend(loop_tasks)
|
||||
entries.extend(loop_entries)
|
||||
out[f"tasks/{run_name}.yaml"] = _dump(header, tasks)
|
||||
annotations = {k: v for k, v in ((POPULATION_ANNOTATION, population), (SPLIT_ANNOTATION, split)) if v}
|
||||
metadata: dict[str, Any] = {"name": run_name, **({"annotations": annotations} if annotations else {})}
|
||||
@@ -519,24 +415,21 @@ def increment(
|
||||
command: str,
|
||||
repos: Path | None = None,
|
||||
runner_image: Path | None = None,
|
||||
loop: Loop | None = None,
|
||||
atespace: str | None = None,
|
||||
) -> list[Path]:
|
||||
"""Write the generated manifests under ``plane`` from the templates in ``plane/tasks`` (re-pinned first when
|
||||
``runner_image`` is given), with the improvement loop when ``loop`` is given; returns the paths. Workspaces go
|
||||
``runner_image`` is given); returns the paths. Workspaces go
|
||||
to ``atespace`` (default: the profile's)."""
|
||||
if atespace is None:
|
||||
from .profile import load_profile
|
||||
|
||||
atespace = load_profile().atespace
|
||||
if loop is not None and repos is None:
|
||||
raise ValueError("--loop needs --repos: the alerts task marks fixed-pin alerts by the fix's new-side ranges")
|
||||
if runner_image is not None:
|
||||
repin_templates(plane, runner_image)
|
||||
loaded = load_manifests([plane / "tasks" / f"{name}.yaml" for name in TEMPLATES])
|
||||
cases = read_cases(validation_set, candidates_dir, triage_dir, population, repos, fixed_side=loop is not None)
|
||||
cases = read_cases(validation_set, candidates_dir, triage_dir, population, repos)
|
||||
files = render(
|
||||
cases, loaded.tasks, run_name, command, population=population.stem, split=validation_set.stem, loop=loop, atespace=atespace
|
||||
cases, loaded.tasks, run_name, command, population=population.stem, split=validation_set.stem, atespace=atespace
|
||||
)
|
||||
written: list[Path] = []
|
||||
for relative, text in files.items():
|
||||
|
||||
@@ -2,7 +2,7 @@
|
||||
|
||||
One store, keyed by repository and pin, holds what plane runs learned: repository facts by content hash, and rows
|
||||
of the kinds in ``KINDS`` (``facts``, ``verdict``, ``unit_cost``, ``alert``, ``coverage``, ``features``, ...) that
|
||||
``remember`` (:mod:`.tasks.remember`) derives from a run's delivered artifacts, plus the loop's and the decision
|
||||
``remember`` (:mod:`.tasks.remember`) derives from a run's delivered artifacts, plus the decision
|
||||
engine's own. Layout, identical for both backends::
|
||||
|
||||
index.jsonl one row per ingested (run, task): sha256 of its rows, row count
|
||||
@@ -34,10 +34,6 @@ Fact reuse: a facts entry is valid for (repo, pin, candidates digest, runner ima
|
||||
writes as the Task annotation ``openultrasast.io/memory-key``. :func:`seed` runs before a Run: for every entry whose
|
||||
Task carries a key with stored facts and no state yet, it writes ``<run>/<task>/facts.json`` and a ``done``
|
||||
summary and marks the task done under the run lock (``reconciler.mark_done``), so the reconciler skips it.
|
||||
|
||||
The loop's ``memory-snapshot`` (a Task annotated ``openultrasast.io/memory-snapshot`` with the train-on-test guard's
|
||||
parameters) is seeded the same way: a sandboxed task cannot read this store, so :func:`seed` writes the store's rows
|
||||
after the guard into ``<run>/memory-snapshot/`` (:func:`.tasks.loop.write_snapshot`) and marks it done.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
@@ -62,7 +58,6 @@ log = logging.getLogger(__name__)
|
||||
MEMORY_KEY_ANNOTATION = "openultrasast.io/memory-key"
|
||||
POPULATION_ANNOTATION = "openultrasast.io/population"
|
||||
SPLIT_ANNOTATION = "openultrasast.io/split"
|
||||
SNAPSHOT_ANNOTATION = "openultrasast.io/memory-snapshot" # the loop's guard: {"catalog", "manifest", "populations"}
|
||||
# The decision engine's kinds (learned-decision-engine, Data Models): `features` from the `features` task through
|
||||
# `remember`; `label`, `decision`, `experiment`, `arm_outcome` and `experiment_result` for the later tasks; `example`
|
||||
# (a labelled case of the local memory: label + feature record + excerpt sha) from `ousast learn memory build`.
|
||||
@@ -984,8 +979,7 @@ def ingest(run_dir: Path, store: MemoryStore | None = None) -> list[IngestResult
|
||||
def seed(run_manifest: Path, store: MemoryStore | None = None, *, results_root: Path | None = None) -> list[str]:
|
||||
"""Reuse stored facts for a Run's tasks whose ``openultrasast.io/memory-key`` matches (Req 6.1): each such task
|
||||
without a directory or a state gets ``facts.json`` and a ``done`` summary and is marked done under the run
|
||||
lock. Returns the seeded task names. A changed pin, candidate set or runner image finds nothing and recomputes.
|
||||
A ``memory-snapshot`` Task gets the store's rows after the guard its annotation names, marked done likewise."""
|
||||
lock. Returns the seeded task names. A changed pin, candidate set or runner image finds nothing and recomputes."""
|
||||
from . import reconciler
|
||||
|
||||
run_spec, manifests = reconciler.load_run(Path(run_manifest))
|
||||
@@ -995,16 +989,8 @@ def seed(run_manifest: Path, store: MemoryStore | None = None, *, results_root:
|
||||
seeded: list[str] = []
|
||||
for entry in run_spec.tasks:
|
||||
annotations = manifests.tasks[entry.task].metadata.annotations
|
||||
text, snapshot = annotations.get(MEMORY_KEY_ANNOTATION), annotations.get(SNAPSHOT_ANNOTATION)
|
||||
if not (text or snapshot) or (base / entry.name).exists() or state.get(entry.name, {}).get("status", "pending") != "pending":
|
||||
continue
|
||||
if snapshot:
|
||||
from .tasks.loop import write_snapshot # lazy: the guard reads the pair catalog and the benchmark manifest
|
||||
|
||||
info = write_snapshot(base / entry.name, store, json.loads(snapshot))
|
||||
reused = {"snapshot": store.describe(), **{k: info[k] for k in ("rows", "kept", "index_digest")}}
|
||||
reconciler.mark_done(run_spec.metadata.name, entry.name, reused=reused, results_root=results_root)
|
||||
seeded.append(entry.name)
|
||||
text = annotations.get(MEMORY_KEY_ANNOTATION)
|
||||
if not text or (base / entry.name).exists() or state.get(entry.name, {}).get("status", "pending") != "pending":
|
||||
continue
|
||||
key = parse_memory_key(str(text))
|
||||
where = {"candidates_digest": key["candidates"], "image": key["image"]}
|
||||
@@ -1037,7 +1023,6 @@ __all__ = [
|
||||
"POPULATION_ANNOTATION",
|
||||
"QUERYABLE",
|
||||
"QUERY_FIELDS",
|
||||
"SNAPSHOT_ANNOTATION",
|
||||
"SPLIT_ANNOTATION",
|
||||
"FileStore",
|
||||
"IngestResult",
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
"""`alerts` (harnessx-removal Requirement 6.3, design section 6): what the deterministic rules say about a case.
|
||||
|
||||
No model, budget ``{usd: 0, calls: 0}``, one per case in a ``--loop`` Run, serialized like ``repo-facts`` (a full-tree
|
||||
No model, budget ``{usd: 0, calls: 0}``, serialized like ``repo-facts`` (a full-tree
|
||||
scan per pin). It runs the quick-mode scan -- ``preprocess_repository``, the entry-point reachability hints,
|
||||
``rank_targets`` and ``quick_scan_findings`` over the image's bundled ruleset, enabled and shadow rules alike, exactly
|
||||
what ``ousast scan --mode quick`` runs -- on the case's vulnerable pin and on its fixed pin, and writes one row per
|
||||
@@ -12,8 +12,7 @@ what ``ousast scan --mode quick`` runs -- on the case's vulnerable pin and on it
|
||||
candidates' ``path::function``; ``<global>`` outside any function); a language those patterns do not cover keeps the
|
||||
scan's own reachability attribution. ``in_fix_range`` says whether the line lies inside the fix: the old-side ranges
|
||||
of ``case.json`` on the vulnerable pin, the new-side ``fixed_ranges`` on the fixed pin (null when ``case.json`` has
|
||||
none). A fixed-pin alert inside the fix is a false alert on code the fix wrote: rule M1 of ``improve/memory.py``
|
||||
counts it.
|
||||
none). A fixed-pin alert inside the fix is a false alert on code the fix wrote.
|
||||
|
||||
**Coverage** (an honest zero): quick mode has rules only for the languages its ruleset declares
|
||||
(:func:`quick_languages`, derived from the bundled ruleset directory, never listed by hand). ``summary.json`` records
|
||||
|
||||
@@ -1,302 +0,0 @@
|
||||
"""The improvement loop as plane tasks (harnessx-removal Requirement 6.3, design section 6).
|
||||
|
||||
``command: ["loop", <step>]``; every step is model-free with budget ``{usd: 0, calls: 0}`` (``budget.py`` refuses any
|
||||
call, so a stray one fails loudly), and each writes ``summary.json`` under the task contract. Records travel as
|
||||
``{"key", "version", "row"}`` lines: the store object a row came from and its version (the bucket's object version id, a file's
|
||||
content hash), which a proposal's provenance cites (Req 6.4).
|
||||
|
||||
- ``snapshot``: never runs in the sandbox, which cannot read the host's store. ``ousast plane run`` seeds it before
|
||||
the Run (``memory.seed`` -> :func:`write_snapshot`): the store's rows after the train-on-test guard to
|
||||
``rows.jsonl``, and ``index.json`` (store, index digest, row counts, what the guard excluded). Run in a sandbox it
|
||||
fails, naming the seed that did not happen.
|
||||
- ``measure``: every case's ``remember`` rows (inputs ``remember_<case>``) plus the snapshot -> ``measure.json``,
|
||||
the increment's metrics for this run, the store and both (declared sites agreed, usd per candidate, dispute rates
|
||||
per family, alerts per rule and pin) -- and ``rows.jsonl``, the union by row id that ``propose`` reads.
|
||||
- ``propose``: ``improve/memory.py``'s rules M1/M2 over those rows after the guard (``OUSAST_QUALIFY_POPULATIONS``, the
|
||||
gated manifest's cases, the pair catalog's holdout pairs), against the benchmark target's ledger and journal in
|
||||
this repository at the commit under test (``OUSAST_PROJECT_DIR``) -> ``proposals.jsonl`` (edit + provenance),
|
||||
``memory_proposals.jsonl`` (the sidecar) and ``signals.json`` (M3, advisory).
|
||||
- ``improve``: one ``evolve.run_round`` with those proposals -- the unchanged validator and gate, the pair catalog's
|
||||
holdout clause included -- on copies of the target's ledger and journal -> ``gate.json`` (the ``RoundOutcome``
|
||||
and which edits came from memory), ``journal.json``, the sidecar, and ``rule_policy.json`` only when the round
|
||||
was accepted; ``memory.jsonl`` holds the ``proposal_outcome`` rows the host ingests after the Run. Adopting an
|
||||
accepted ledger stays a maintainer commit.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import hashlib
|
||||
import json
|
||||
import os
|
||||
import shutil
|
||||
import sys
|
||||
import tempfile
|
||||
import traceback
|
||||
from collections.abc import Callable, Iterable, Mapping, Sequence
|
||||
from dataclasses import asdict
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from ..memory import REMEMBER_OUTPUT, MemoryStore, Record, validate_row
|
||||
|
||||
INPUT_PREFIX = "OUSAST_INPUT_REMEMBER_"
|
||||
CALIBRATION = Path(".openultrasast") / "calibration" # cli.CALIBRATION_DIR: where `improve` keeps a target's ledger
|
||||
ADOPTION = "adoption is a maintainer commit: copy rule_policy.json to {ledger} (README, Rule-level loop)"
|
||||
|
||||
|
||||
# --- records -------------------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _record_line(record: Record) -> str:
|
||||
return json.dumps({"key": record.key, "version": record.version, "row": record.row}, sort_keys=True, separators=(",", ":")) + "\n"
|
||||
|
||||
|
||||
def read_records(path: Path) -> list[Record]:
|
||||
records: list[Record] = []
|
||||
for number, raw in enumerate(path.read_text(encoding="utf-8").splitlines(), 1):
|
||||
if raw.strip():
|
||||
item = json.loads(raw)
|
||||
records.append(Record(validate_row(item["row"], f"{path}:{number}"), str(item["key"]), item.get("version")))
|
||||
return records
|
||||
|
||||
|
||||
def write_records(path: Path, records: Iterable[Record]) -> None:
|
||||
path.write_text("".join(_record_line(r) for r in sorted(records, key=lambda r: str(r.row["id"]))), encoding="utf-8")
|
||||
|
||||
|
||||
def _guard(params: Mapping[str, Any], root: Path) -> Any:
|
||||
from ...benchmark import load_benchmark_manifest
|
||||
from ...improve.memory import Guard, benchmark_cases
|
||||
from ...pairs import load_pair_catalog, select_split
|
||||
|
||||
manifest = load_benchmark_manifest(root / str(params["manifest"]))
|
||||
holdout = select_split(load_pair_catalog(Path(str(params["catalog"]))), "holdout")
|
||||
return Guard.build([str(p) for p in params.get("populations") or []], benchmark_cases(manifest), holdout)
|
||||
|
||||
|
||||
def write_snapshot(directory: Path, store: MemoryStore, params: Mapping[str, Any], root: Path | None = None) -> dict[str, Any]:
|
||||
"""The store's rows after the guard ``params`` names (``catalog``, ``manifest``, ``populations``; paths relative to
|
||||
``root``, the working directory) into ``directory``: ``rows.jsonl``, ``index.json`` and a ``done`` summary."""
|
||||
from ...improve.memory import index_digest
|
||||
|
||||
here = Path.cwd()
|
||||
try:
|
||||
os.chdir(root or here) # the pair catalog names its files relative to the repository root
|
||||
guard = _guard(params, Path.cwd())
|
||||
finally:
|
||||
os.chdir(here)
|
||||
directory.mkdir(parents=True, exist_ok=False) # after the guard loaded: a failed load leaves no half-seeded task
|
||||
records = store.rows()
|
||||
kept, excluded = guard.apply(records)
|
||||
write_records(directory / "rows.jsonl", kept)
|
||||
info = {
|
||||
"store": store.describe(), "index_entries": len(store.index()), "index_digest": index_digest(store), "rows": len(records),
|
||||
"kept": len(kept), "excluded": excluded, "guard": dict(params),
|
||||
} # fmt: skip
|
||||
(directory / "index.json").write_text(json.dumps(info, indent=2, sort_keys=True) + "\n", encoding="utf-8")
|
||||
summary = {"status": "done", "units_done": 1, "units_total": 1, "usd": None, "calls": 0, "usage": {}, "model": None, "reused": info}
|
||||
(directory / "summary.json").write_text(json.dumps(summary, indent=2, sort_keys=True) + "\n", encoding="utf-8")
|
||||
return info
|
||||
|
||||
|
||||
# --- measure -------------------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def metrics(rows: Sequence[Mapping[str, Any]]) -> dict[str, Any]:
|
||||
"""The increment's metrics over verdict, unit-cost and alert rows (a disputed final is neither agreed nor not)."""
|
||||
verdicts = [r for r in rows if r["kind"] == "verdict"]
|
||||
finals = {f: sum(1 for v in verdicts if v.get("final") == f) for f in ("agreed", "rejected", "disputed")}
|
||||
costs = [r.get("usd") for r in rows if r["kind"] == "unit_cost"]
|
||||
usd = None if any(c is None for c in costs) else round(sum(float(c or 0) for c in costs), 6)
|
||||
families: dict[str, list[int]] = {}
|
||||
for v in verdicts:
|
||||
tally = families.setdefault(str(v.get("family")), [0, 0])
|
||||
tally[0] += 1
|
||||
tally[1] += v.get("final") == "disputed"
|
||||
alerts: dict[str, dict[str, int]] = {}
|
||||
for a in (r for r in rows if r["kind"] == "alert"):
|
||||
role = "fixed" if a.get("pin_role") == "fixed" else "vulnerable"
|
||||
tally_a = alerts.setdefault(str(a.get("rule_id")), {"vulnerable": 0, "fixed": 0, "fixed_in_fix_range": 0})
|
||||
tally_a[role] += 1
|
||||
tally_a["fixed_in_fix_range"] += role == "fixed" and a.get("in_fix_range") is True
|
||||
coverage: dict[str, dict[str, int]] = {} # language -> pins per coverage kind: a `none` pin's zero alerts mean nothing
|
||||
for c in (r for r in rows if r["kind"] == "coverage"):
|
||||
tally_c = coverage.setdefault(str(c.get("language")), {"quick": 0, "engine": 0, "none": 0})
|
||||
tally_c[str(c.get("coverage"))] = tally_c.get(str(c.get("coverage")), 0) + 1
|
||||
return {
|
||||
"cases": len({(v["repo"], v["pin"]) for v in verdicts}),
|
||||
"candidates": len(verdicts),
|
||||
**finals,
|
||||
"tiebreaks": sum(1 for v in verdicts if v.get("tiebreak")),
|
||||
"declared_sites_matched": sum(1 for v in verdicts if v.get("site_match") is True),
|
||||
"declared_sites_agreed": sum(1 for v in verdicts if v.get("site_match") is True and v.get("final") == "agreed"),
|
||||
"usd": usd,
|
||||
"usd_per_candidate": None if usd is None or not verdicts else round(usd / len(verdicts), 6),
|
||||
"dispute_rate": {f: {"candidates": n, "disputed": d, "rate": round(d / n, 4)} for f, (n, d) in sorted(families.items())},
|
||||
"alerts": dict(sorted(alerts.items())),
|
||||
"coverage": dict(sorted(coverage.items())),
|
||||
}
|
||||
|
||||
|
||||
def measure(env: Mapping[str, str], output: Path) -> dict[str, Any]:
|
||||
run = env.get("OUSAST_RUN") or "unknown"
|
||||
current: list[Record] = []
|
||||
read: dict[str, int] = {}
|
||||
for name in sorted(k for k in env if k.startswith(INPUT_PREFIX)):
|
||||
path = Path(env[name])
|
||||
data = path.read_bytes()
|
||||
read[name.removeprefix(INPUT_PREFIX).lower()] = len(data)
|
||||
version = hashlib.sha256(data).hexdigest()
|
||||
for number, raw in enumerate(data.decode("utf-8").splitlines(), 1):
|
||||
if raw.strip():
|
||||
row = validate_row(json.loads(raw), f"{path}:{number}")
|
||||
current.append(Record(row, f"runs/{run}/{row['task']}/{REMEMBER_OUTPUT}", version))
|
||||
if not current:
|
||||
raise ValueError(f"read no remember rows from {len(read)} inputs ({read}): an unread run is not an empty one")
|
||||
snapshot = read_records(Path(env["OUSAST_INPUT_SNAPSHOT"])) if env.get("OUSAST_INPUT_SNAPSHOT") else []
|
||||
merged = {str(r.row["id"]): r for r in snapshot}
|
||||
merged.update({str(r.row["id"]): r for r in current}) # this run's row wins over its earlier ingest
|
||||
write_records(output / "rows.jsonl", merged.values())
|
||||
report = {
|
||||
"run": run, "inputs_bytes": read, "rows": {"run": len(current), "snapshot": len(snapshot), "union": len(merged)},
|
||||
"run_metrics": metrics([r.row for r in current]), "store_metrics": metrics([r.row for r in snapshot]),
|
||||
"union_metrics": metrics([r.row for r in merged.values()]),
|
||||
} # fmt: skip
|
||||
(output / "measure.json").write_text(json.dumps(report, indent=2, sort_keys=True) + "\n", encoding="utf-8")
|
||||
return {"rows": report["rows"]}
|
||||
|
||||
|
||||
# --- propose and improve -------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _project(env: Mapping[str, str]) -> tuple[Any, Path, Path, Path]:
|
||||
"""(manifest, target, ledger, journal) in the repository at the commit under test; the working directory
|
||||
becomes the repository root, where the pair catalog's relative paths resolve."""
|
||||
from ...benchmark import load_benchmark_manifest, resolve_benchmark_source
|
||||
|
||||
project = Path(env.get("OUSAST_PROJECT_DIR") or "")
|
||||
if not env.get("OUSAST_PROJECT_DIR") or not (project / str(env.get("OUSAST_LOOP_MANIFEST"))).is_file():
|
||||
raise ValueError(f"OUSAST_PROJECT_DIR {project} holds no benchmark manifest {env.get('OUSAST_LOOP_MANIFEST')}")
|
||||
os.chdir(project)
|
||||
manifest_path = Path(str(env["OUSAST_LOOP_MANIFEST"]))
|
||||
manifest = load_benchmark_manifest(manifest_path)
|
||||
target = resolve_benchmark_source(manifest_path, manifest)
|
||||
ledger = target / CALIBRATION / "rule_policy.json"
|
||||
return manifest, target, ledger, ledger.with_name("improve_journal.json")
|
||||
|
||||
|
||||
def _params(env: Mapping[str, str]) -> dict[str, Any]:
|
||||
populations = json.loads(env.get("OUSAST_QUALIFY_POPULATIONS") or "[]")
|
||||
return {"catalog": env.get("OUSAST_LOOP_CATALOG"), "manifest": env.get("OUSAST_LOOP_MANIFEST"), "populations": populations}
|
||||
|
||||
|
||||
def propose(env: Mapping[str, str], output: Path) -> dict[str, Any]:
|
||||
from ...improve.journal import load_journal, next_round_index
|
||||
from ...improve.memory import disputed_families, propose_from_memory, write_sidecar, write_signals
|
||||
from ...ruleset import DEFAULT_RULESET_DIR, load_ruleset, read_rule_ledger
|
||||
|
||||
records = read_records(Path(env["OUSAST_INPUT_ROWS"]))
|
||||
index = json.loads(Path(env["OUSAST_INPUT_INDEX"]).read_text(encoding="utf-8"))
|
||||
_, _, ledger, journal = _project(env)
|
||||
guard = _guard(_params(env), Path.cwd())
|
||||
rules = load_ruleset(DEFAULT_RULESET_DIR, ledger if ledger.is_file() else None)
|
||||
rounds = load_journal(journal)
|
||||
found = propose_from_memory(
|
||||
records, {r.rule_id: r for r in rules}, read_rule_ledger(ledger), rounds, excluded=guard, index_digest=str(index["index_digest"])
|
||||
)
|
||||
lines = [json.dumps({"edit": asdict(p.edit), "provenance": p.provenance}, sort_keys=True) + "\n" for p in found]
|
||||
(output / "proposals.jsonl").write_text("".join(lines), encoding="utf-8")
|
||||
sidecar = output / "journal.json"
|
||||
write_sidecar(sidecar, next_round_index(rounds), found)
|
||||
sidecar.with_name("memory_proposals.jsonl").touch() # present when nothing was proposed, as a declared output
|
||||
write_signals(output, disputed_families(records, guard))
|
||||
kept, excluded = guard.apply(records)
|
||||
return {"proposals": len(found), "rows": len(records), "kept": len(kept), "excluded": excluded["rows"]}
|
||||
|
||||
|
||||
def improve(env: Mapping[str, str], output: Path) -> dict[str, Any]:
|
||||
from ...gate import FALSE_POSITIVE_CEILING, RECALL_FLOOR
|
||||
from ...improve.evolve import run_round
|
||||
from ...improve.memory import MemoryProposal, outcome_of, outcome_rows, proposal_id_of
|
||||
from ...improve.validator import RuleStatusEdit
|
||||
from ...pairs import load_pair_catalog, select_split
|
||||
from ...policy import load_policy
|
||||
from ...ruleset import DEFAULT_RULESET_DIR
|
||||
|
||||
path = Path(env["OUSAST_INPUT_PROPOSALS"])
|
||||
items = [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines() if line.strip()]
|
||||
proposals = [MemoryProposal(RuleStatusEdit(**item["edit"]), item["provenance"]) for item in items]
|
||||
manifest, target, ledger, journal = _project(env)
|
||||
pairs = select_split(load_pair_catalog(Path(str(env["OUSAST_LOOP_CATALOG"]))), "holdout")
|
||||
with tempfile.TemporaryDirectory(prefix="ousast-loop-") as scratch:
|
||||
work_ledger = Path(scratch) / "rule_policy.json"
|
||||
if ledger.is_file():
|
||||
shutil.copyfile(ledger, work_ledger)
|
||||
if journal.is_file():
|
||||
shutil.copyfile(journal, output / "journal.json")
|
||||
outcome = run_round(
|
||||
target, manifest, ledger_path=work_ledger, journal_path=output / "journal.json", ruleset_dir=DEFAULT_RULESET_DIR,
|
||||
policy=load_policy(), recall_floor=RECALL_FLOOR, fp_ceiling=FALSE_POSITIVE_CEILING, pair_cases=pairs, proposals=proposals,
|
||||
) # fmt: skip
|
||||
if outcome.accepted and work_ledger.is_file():
|
||||
shutil.copyfile(work_ledger, output / "rule_policy.json")
|
||||
(output / "journal.json").touch()
|
||||
(output / "memory_proposals.jsonl").touch()
|
||||
memory_edits = [e.key() for e in outcome.edits if proposal_id_of(e)]
|
||||
verdict = {
|
||||
"accepted": outcome.accepted, "reason": outcome.reason, "round": outcome.round, "proposals": len(proposals),
|
||||
"memory_edits": memory_edits, "benchmark_edits": [e.key() for e in outcome.edits if not proposal_id_of(e)],
|
||||
"manifest": env.get("OUSAST_LOOP_MANIFEST"), "outcome": asdict(outcome),
|
||||
"adoption": ADOPTION.format(ledger=ledger.as_posix()) if outcome.accepted else None,
|
||||
} # fmt: skip
|
||||
(output / "gate.json").write_text(json.dumps(verdict, indent=2, sort_keys=True, default=str) + "\n", encoding="utf-8")
|
||||
status = outcome_of(outcome.reason, outcome.accepted)
|
||||
by_id = {p.proposal_id: p for p in proposals}
|
||||
gate_run = f"plane:{env.get('OUSAST_RUN') or 'unknown'}"
|
||||
rows = outcome_rows(outcome.edits, status, outcome.round, outcome.reason, by_id, gate_run) if status else []
|
||||
(output / REMEMBER_OUTPUT).write_text("".join(json.dumps(r, sort_keys=True) + "\n" for r in rows), encoding="utf-8")
|
||||
return {"accepted": outcome.accepted, "reason": outcome.reason, "edits": len(outcome.edits), "memory_edits": len(memory_edits)}
|
||||
|
||||
|
||||
def snapshot(env: Mapping[str, str], output: Path) -> dict[str, Any]:
|
||||
raise RuntimeError(
|
||||
"memory-snapshot ran in the sandbox: `ousast plane run` writes it on the host from the memory store before the Run "
|
||||
"starts (memory.seed), so the seed did not happen; rerun through `ousast plane run`"
|
||||
)
|
||||
|
||||
|
||||
STEPS: dict[str, Callable[[Mapping[str, str], Path], dict[str, Any]]] = {
|
||||
"snapshot": snapshot, "measure": measure, "propose": propose, "improve": improve,
|
||||
} # fmt: skip
|
||||
|
||||
|
||||
def main(argv: Sequence[str] | None = None, environ: Mapping[str, str] | None = None) -> int:
|
||||
args = list(sys.argv[1:] if argv is None else argv)
|
||||
env = environ if environ is not None else os.environ
|
||||
output = Path(env["OUSAST_OUTPUT_DIR"]) if env.get("OUSAST_OUTPUT_DIR") else None
|
||||
summary: dict[str, Any] = {"status": "failed", "units_done": 0, "units_total": 1, "usd": 0, "calls": 0, "usage": {}, "model": None}
|
||||
here = Path.cwd()
|
||||
try:
|
||||
if output is None:
|
||||
raise ValueError("OUSAST_OUTPUT_DIR is not set")
|
||||
if len(args) != 1 or args[0] not in STEPS:
|
||||
raise ValueError(f"expected one step of {sorted(STEPS)}, got {args}")
|
||||
output.mkdir(parents=True, exist_ok=True)
|
||||
summary.update(step=args[0], **STEPS[args[0]](env, output))
|
||||
summary.update(status="done", units_done=1)
|
||||
except Exception: # noqa: BLE001 -- a crash is `failed` with its traceback
|
||||
summary["reason"] = traceback.format_exc()[-2000:]
|
||||
finally:
|
||||
os.chdir(here)
|
||||
if output is not None:
|
||||
(output / "summary.json").write_text(json.dumps(summary, indent=2, sort_keys=True, default=str) + "\n", encoding="utf-8")
|
||||
shown = {k: summary.get(k) for k in ("status", "step", "proposals", "accepted", "reason", "rows")}
|
||||
print(json.dumps(shown, default=str), file=sys.stderr)
|
||||
return 0 if summary["status"] == "done" else 2
|
||||
|
||||
|
||||
__all__ = ["main", "measure", "metrics", "propose", "improve", "read_records", "write_records", "write_snapshot"]
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -20,8 +20,8 @@ class RulesetError(ValueError):
|
||||
class PatternRule:
|
||||
"""A detection rule expressed as governed data (no rule-local severity).
|
||||
|
||||
``pattern`` is human-PR-only; the self-improvement loop may only change
|
||||
``status``/``min_evidence_level``/``precision_estimate`` via the ledger.
|
||||
Patterns change through reviewed commits. The ledger overlays
|
||||
``status``, ``min_evidence_level`` and ``precision_estimate``.
|
||||
"""
|
||||
|
||||
rule_id: str
|
||||
@@ -42,10 +42,10 @@ class PatternRule:
|
||||
|
||||
|
||||
def load_ruleset(directory: Path = DEFAULT_RULESET_DIR, ledger: Path | None = None, *, priors: Priors = "all") -> tuple[PatternRule, ...]:
|
||||
"""Load every ``*.toml`` rule file under ``directory`` and apply the loop ledger.
|
||||
"""Load every ``*.toml`` rule file under ``directory`` and apply the rule policy ledger.
|
||||
|
||||
Rules are returned sorted by ``rule_id`` for determinism. The optional ledger
|
||||
(``rule_policy.json``) overlays loop-owned ``status``/``min_evidence_level``/
|
||||
(``rule_policy.json``) overlays ``status``/``min_evidence_level``/
|
||||
``precision_estimate`` per ``rule_id``.
|
||||
|
||||
``priors`` keeps or drops the rules tagged ``framework``/``library``: ``"all"`` (the default, today's rules),
|
||||
@@ -103,12 +103,12 @@ def write_ruleset(path: Path, rules: Iterable[PatternRule]) -> None:
|
||||
|
||||
|
||||
def read_rule_ledger(path: Path) -> dict[str, dict[str, object]]:
|
||||
"""Read the loop-owned ``rule_policy.json`` overlay (status/floors per rule_id)."""
|
||||
"""Read the ``rule_policy.json`` overlay (status/floors per rule_id)."""
|
||||
return _load_ledger(path)
|
||||
|
||||
|
||||
def write_rule_ledger(path: Path, entries: dict[str, dict[str, object]]) -> None:
|
||||
"""Write the loop-owned ``rule_policy.json`` overlay (used by the evolve loop)."""
|
||||
"""Write the ``rule_policy.json`` overlay."""
|
||||
path.parent.mkdir(parents=True, exist_ok=True)
|
||||
path.write_text(json.dumps(entries, indent=2, sort_keys=True) + "\n")
|
||||
|
||||
|
||||
@@ -1,69 +0,0 @@
|
||||
"""`ousast improve` CLI: drives the self-improvement loop and the loop-owned ledger."""
|
||||
|
||||
from pathlib import Path
|
||||
|
||||
from openultrasast.cli import main
|
||||
from openultrasast.ruleset import PatternRule, read_rule_ledger, write_ruleset
|
||||
|
||||
|
||||
def _rule(rule_id: str, pattern: str) -> PatternRule:
|
||||
return PatternRule(rule_id=rule_id, title=rule_id, languages=("python",), cwe="CWE-95", tags=("injection",), pattern=pattern)
|
||||
|
||||
|
||||
def _expected_block(rule_id: str) -> str:
|
||||
return (
|
||||
'[[expected]]\ncwe = "CWE-95"\nclass = "code injection"\npath = "app.py"\n'
|
||||
f'line = 2\nrule_id = "{rule_id}"\nsink = "eval"\nevidence = "eval executes input"\n'
|
||||
)
|
||||
|
||||
|
||||
def _scaffold(tmp_path: Path) -> tuple[Path, Path]:
|
||||
repo = tmp_path / "repo"
|
||||
repo.mkdir()
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\ndef g():\n print('hello')\n")
|
||||
ruleset_dir = tmp_path / "ruleset"
|
||||
write_ruleset(ruleset_dir / "python" / "rules.toml", [_rule("good-eval", r"\beval\s*\("), _rule("noisy-print", r"\bprint\s*\(")])
|
||||
manifest = tmp_path / "bench.toml"
|
||||
manifest.write_text('name = "t"\nlanguage = "python_web"\n\n[source]\npath = "repo"\n\n' + _expected_block("good-eval"))
|
||||
return repo, ruleset_dir
|
||||
|
||||
|
||||
def test_improve_accepts_and_writes_loop_owned_ledger(tmp_path: Path) -> None:
|
||||
repo, ruleset_dir = _scaffold(tmp_path)
|
||||
manifest = tmp_path / "bench.toml"
|
||||
|
||||
exit_code = main(["improve", str(manifest), "--ruleset-dir", str(ruleset_dir)])
|
||||
|
||||
assert exit_code == 0
|
||||
# The ledger lands exactly where `scan`/`benchmark` read it for this target.
|
||||
ledger = repo / ".openultrasast" / "calibration" / "rule_policy.json"
|
||||
assert ledger.exists()
|
||||
assert read_rule_ledger(ledger)["noisy-print"]["status"] == "shadow"
|
||||
journal = repo / ".openultrasast" / "calibration" / "improve_journal.json"
|
||||
assert journal.exists()
|
||||
|
||||
|
||||
def test_improve_dry_run_never_touches_the_targets_ledger(tmp_path: Path) -> None:
|
||||
repo, ruleset_dir = _scaffold(tmp_path)
|
||||
manifest = tmp_path / "bench.toml"
|
||||
|
||||
exit_code = main(["improve", str(manifest), "--ruleset-dir", str(ruleset_dir), "--dry-run"])
|
||||
|
||||
assert exit_code == 0
|
||||
assert not (repo / ".openultrasast" / "calibration" / "rule_policy.json").exists()
|
||||
|
||||
|
||||
def test_improve_rejects_missing_manifest(tmp_path: Path) -> None:
|
||||
import pytest
|
||||
|
||||
with pytest.raises(SystemExit):
|
||||
main(["improve", str(tmp_path / "nope.toml")])
|
||||
|
||||
|
||||
def test_improve_fails_loud_when_pair_catalog_is_missing(tmp_path: Path) -> None:
|
||||
import pytest
|
||||
|
||||
_, ruleset_dir = _scaffold(tmp_path)
|
||||
manifest = tmp_path / "bench.toml"
|
||||
with pytest.raises(SystemExit, match="pair catalog"):
|
||||
main(["improve", str(manifest), "--ruleset-dir", str(ruleset_dir), "--dry-run", "--pair-catalog", str(tmp_path / "missing.toml")])
|
||||
@@ -132,8 +132,6 @@ GOVERNANCE_SCORING_BENCHMARK_PLANES = (
|
||||
"openultrasast.scoring.project_score",
|
||||
"openultrasast.ruleset",
|
||||
"openultrasast.ruleset.store",
|
||||
"openultrasast.improve",
|
||||
"openultrasast.improve.evolve",
|
||||
"openultrasast.fusion",
|
||||
"openultrasast.skills",
|
||||
"openultrasast.redaction",
|
||||
|
||||
@@ -1,324 +0,0 @@
|
||||
"""Self-improvement loop: bounded levers, validator, gates, accept/revert (Phase 3 task 7)."""
|
||||
|
||||
import json
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
|
||||
from openultrasast.benchmark import load_benchmark_manifest
|
||||
from openultrasast.improve import (
|
||||
EvolveValidator,
|
||||
PolicyConstantEdit,
|
||||
RuleStatusEdit,
|
||||
StrictValidationError,
|
||||
build_rule_signals,
|
||||
propose_status_edits,
|
||||
run_improvement,
|
||||
run_round,
|
||||
)
|
||||
from openultrasast.policy import load_policy
|
||||
from openultrasast.ruleset import PatternRule, read_rule_ledger, write_ruleset
|
||||
|
||||
POLICY = load_policy()
|
||||
|
||||
|
||||
def _rule(rule_id: str, pattern: str, cwe: str = "CWE-95") -> PatternRule:
|
||||
return PatternRule(rule_id=rule_id, title=rule_id, languages=("python",), cwe=cwe, tags=("injection",), pattern=pattern)
|
||||
|
||||
|
||||
def _expected(rule_id: str) -> str:
|
||||
return (
|
||||
'[[expected]]\ncwe = "CWE-95"\nclass = "code injection"\npath = "app.py"\n'
|
||||
f'line = 2\nrule_id = "{rule_id}"\nsink = "eval"\nevidence = "eval executes input"\n'
|
||||
)
|
||||
|
||||
|
||||
# ---- validator (7.1, 7.2) ---------------------------------------------------
|
||||
|
||||
|
||||
def test_validator_rejects_enabled_to_disabled_jump() -> None:
|
||||
ruleset = {"r": _rule("r", r"\beval\s*\(")}
|
||||
with pytest.raises(StrictValidationError, match="shadow"):
|
||||
EvolveValidator().validate(RuleStatusEdit("r", "enabled", "disabled"), ruleset, POLICY)
|
||||
|
||||
|
||||
def test_validator_accepts_enabled_to_shadow() -> None:
|
||||
ruleset = {"r": _rule("r", r"\beval\s*\(")}
|
||||
EvolveValidator().validate(RuleStatusEdit("r", "enabled", "shadow"), ruleset, POLICY) # no raise
|
||||
|
||||
|
||||
def test_validator_rejects_unknown_rule_and_bad_status() -> None:
|
||||
ruleset = {"r": _rule("r", r"\beval\s*\(")}
|
||||
with pytest.raises(StrictValidationError, match="unknown rule_id"):
|
||||
EvolveValidator().validate(RuleStatusEdit("nope", "enabled", "shadow"), ruleset, POLICY)
|
||||
with pytest.raises(StrictValidationError):
|
||||
EvolveValidator().validate(RuleStatusEdit("r", "enabled", "bogus"), ruleset, POLICY)
|
||||
|
||||
|
||||
def test_validator_never_tunes_severity_and_bounds_constants() -> None:
|
||||
with pytest.raises(StrictValidationError, match="severity is upstream-owned"):
|
||||
EvolveValidator().validate(PolicyConstantEdit("severity", 3, 5), {}, POLICY)
|
||||
with pytest.raises(StrictValidationError):
|
||||
EvolveValidator().validate(PolicyConstantEdit("K", 60, 5000), {}, POLICY) # out of bounds
|
||||
EvolveValidator().validate(PolicyConstantEdit("K", 60, 90), {}, POLICY) # in bounds, no raise
|
||||
|
||||
|
||||
# ---- proposer + signals (7.4, 7.6) -----------------------------------------
|
||||
|
||||
|
||||
def test_propose_status_edits_auto_shadows_precision_draggers_only() -> None:
|
||||
ruleset = {"good": _rule("good", r"\beval\s*\("), "noisy": _rule("noisy", r"\bprint\s*\(")}
|
||||
per_rule = {"good": {"matched": 1, "missed": 0, "false_positives": 0}, "noisy": {"matched": 0, "missed": 0, "false_positives": 2}}
|
||||
|
||||
edits = propose_status_edits(per_rule, ruleset, current_ledger={}, blocked_keys=set())
|
||||
assert [(e.rule_id, e.to_status) for e in edits] == [("noisy", "shadow")]
|
||||
# novelty gate: a previously-reverted edit is not re-proposed.
|
||||
assert propose_status_edits(per_rule, ruleset, {}, {edits[0].key()}) == []
|
||||
# already shadow -> not re-proposed.
|
||||
assert propose_status_edits(per_rule, ruleset, {"noisy": {"status": "shadow"}}, set()) == []
|
||||
|
||||
|
||||
# ---- end-to-end loop (7.5) --------------------------------------------------
|
||||
|
||||
|
||||
def _manifest(tmp_path: Path, expected_block: str) -> tuple[Path, object]:
|
||||
manifest_path = tmp_path / "bench.toml"
|
||||
manifest_path.write_text('name = "t"\nlanguage = "python_web"\n\n[source]\npath = "repo"\n\n' + expected_block)
|
||||
return manifest_path, load_benchmark_manifest(manifest_path)
|
||||
|
||||
|
||||
def _ruleset_dir(tmp_path: Path, rules: list[PatternRule]) -> Path:
|
||||
rd = tmp_path / "ruleset" / "python"
|
||||
write_ruleset(rd / "rules.toml", rules)
|
||||
return tmp_path / "ruleset"
|
||||
|
||||
|
||||
def test_round_auto_shadows_noisy_rule_and_accepts(tmp_path: Path) -> None:
|
||||
repo = tmp_path / "repo"
|
||||
repo.mkdir()
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\ndef g():\n print('hello')\n")
|
||||
ruleset_dir = _ruleset_dir(tmp_path, [_rule("good-eval", r"\beval\s*\("), _rule("noisy-print", r"\bprint\s*\(")])
|
||||
_, manifest = _manifest(tmp_path, _expected("good-eval"))
|
||||
ledger = tmp_path / "rule_policy.json"
|
||||
journal = tmp_path / "journal.json"
|
||||
|
||||
outcome = run_round(repo, manifest, ledger_path=ledger, journal_path=journal, ruleset_dir=ruleset_dir)
|
||||
|
||||
assert outcome.accepted and outcome.reason == "accepted"
|
||||
assert outcome.fp_before > 0.0 and outcome.fp_after == 0.0
|
||||
assert outcome.recall_after == 1.0 # the true positive is preserved
|
||||
assert read_rule_ledger(ledger)["noisy-print"]["status"] == "shadow"
|
||||
assert json.loads(journal.read_text())[0]["outcome"] == "accepted"
|
||||
|
||||
|
||||
def test_round_reverts_byte_for_byte_when_shadowing_would_drop_recall(tmp_path: Path) -> None:
|
||||
repo = tmp_path / "repo"
|
||||
repo.mkdir()
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\ndef g(y):\n return eval(y)\n")
|
||||
ruleset_dir = _ruleset_dir(tmp_path, [_rule("dual-eval", r"\beval\s*\(")])
|
||||
_, manifest = _manifest(tmp_path, _expected("dual-eval"))
|
||||
ledger = tmp_path / "rule_policy.json"
|
||||
journal = tmp_path / "journal.json"
|
||||
|
||||
outcome = run_round(repo, manifest, ledger_path=ledger, journal_path=journal, ruleset_dir=ruleset_dir)
|
||||
|
||||
assert not outcome.accepted and outcome.reason == "reverted"
|
||||
assert outcome.score_after > outcome.score_before # score improved but the recall gate blocks it
|
||||
assert not ledger.exists() # byte-for-byte revert: the ledger was never written
|
||||
assert json.loads(journal.read_text())[0]["outcome"] == "reverted"
|
||||
|
||||
|
||||
def test_run_improvement_converges_after_accepting(tmp_path: Path) -> None:
|
||||
repo = tmp_path / "repo"
|
||||
repo.mkdir()
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\ndef g():\n print('hello')\n")
|
||||
ruleset_dir = _ruleset_dir(tmp_path, [_rule("good-eval", r"\beval\s*\("), _rule("noisy-print", r"\bprint\s*\(")])
|
||||
_, manifest = _manifest(tmp_path, _expected("good-eval"))
|
||||
|
||||
outcomes = run_improvement(
|
||||
repo,
|
||||
manifest,
|
||||
ledger_path=tmp_path / "rule_policy.json",
|
||||
journal_path=tmp_path / "journal.json",
|
||||
ruleset_dir=ruleset_dir,
|
||||
max_rounds=4,
|
||||
)
|
||||
|
||||
assert outcomes[0].accepted # round 1 shadows the noisy rule
|
||||
assert outcomes[-1].reason == "no_proposals" # converges (nothing left to improve)
|
||||
|
||||
|
||||
def test_build_rule_signals_separates_miss_and_fp() -> None:
|
||||
from openultrasast.benchmark import BenchmarkFalsePositive, BenchmarkMetrics, BenchmarkMiss, BenchmarkResult
|
||||
|
||||
result = BenchmarkResult(
|
||||
benchmark_run_id="r",
|
||||
benchmark_name="n",
|
||||
language="python_web",
|
||||
mode="quick",
|
||||
scan_id=None,
|
||||
scan_run_dir=None,
|
||||
metrics=BenchmarkMetrics(1, 1, 0, 1, 1, 0.0, 0.0, None, {}, {}, {}),
|
||||
misses=[BenchmarkMiss("CWE-89", "sqli", "a.py", "ev", "no match", rule_id="sqli")],
|
||||
false_positives=[BenchmarkFalsePositive("noisy:a.py:1", "a.py", 1, "fp", rule_id="noisy")],
|
||||
baseline_deltas=[],
|
||||
calibration_records=[],
|
||||
rule_recommendations=[],
|
||||
)
|
||||
signals = build_rule_signals(result)
|
||||
assert {s["signal"] for s in signals} == {"miss", "fp"}
|
||||
assert {s["rule_id"] for s in signals} == {"sqli", "noisy"}
|
||||
|
||||
|
||||
# ---- pair-corpus-honesty: per-profile holdout clause (Req 7.2-7.4) ----------
|
||||
|
||||
|
||||
def _pair(
|
||||
tmp_path: Path, name: str, vuln_src: str, fixed_src: str, rule_id: str, sink: str, *, provenance: str, review_tier: str = "reviewed"
|
||||
):
|
||||
from openultrasast.benchmark import ExpectedFinding
|
||||
from openultrasast.pairs import PairCase
|
||||
|
||||
vuln = tmp_path / f"{name}-vuln.py"
|
||||
fixed = tmp_path / f"{name}-fixed.py"
|
||||
vuln.write_text(vuln_src)
|
||||
fixed.write_text(fixed_src)
|
||||
return PairCase(
|
||||
name=name,
|
||||
slice="github",
|
||||
language="python",
|
||||
origin="test",
|
||||
vuln_file=vuln,
|
||||
fixed_file=fixed,
|
||||
relpath="app.py",
|
||||
expected=(
|
||||
ExpectedFinding(
|
||||
cwe="CWE-95",
|
||||
vulnerability_class="x",
|
||||
path="app.py",
|
||||
evidence="",
|
||||
rule_id=rule_id,
|
||||
sink=sink,
|
||||
mechanism="source_reaches_sink",
|
||||
),
|
||||
),
|
||||
min_recall=1.0,
|
||||
fix_policy="silent",
|
||||
provenance=provenance,
|
||||
split="holdout",
|
||||
review_tier=review_tier,
|
||||
reviewer="test" if review_tier == "reviewed" else "",
|
||||
)
|
||||
|
||||
|
||||
def test_round_rejects_when_a_profile_regresses_on_holdout(tmp_path: Path) -> None:
|
||||
repo = tmp_path / "repo"
|
||||
repo.mkdir()
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\ndef g():\n print('hello')\n")
|
||||
ruleset_dir = _ruleset_dir(tmp_path, [_rule("good-eval", r"\beval\s*\("), _rule("noisy-print", r"\bprint\s*\(")])
|
||||
_, manifest = _manifest(tmp_path, _expected("good-eval"))
|
||||
ledger = tmp_path / "rule_policy.json"
|
||||
journal = tmp_path / "journal.json"
|
||||
# The agent profile's holdout pair is only detected by the rule the round wants to shadow.
|
||||
agent_pairs = [
|
||||
_pair(tmp_path, f"agent{i}", "def h(y):\n print(y)\n", "def h(y):\n return y\n", "noisy-print", "print", provenance="agent")
|
||||
for i in range(5)
|
||||
]
|
||||
outcome = run_round(repo, manifest, ledger_path=ledger, journal_path=journal, ruleset_dir=ruleset_dir, pair_cases=agent_pairs)
|
||||
assert not outcome.accepted
|
||||
assert outcome.reason == "profile_regression:agent"
|
||||
assert outcome.profile_regressions == ["agent"]
|
||||
assert outcome.per_profile_before["agent"]["pair_correct"] == 5.0 and outcome.per_profile_after["agent"]["pair_correct"] == 0.0
|
||||
assert not ledger.exists()
|
||||
assert json.loads(journal.read_text())[0]["profile_regressions"] == ["agent"]
|
||||
|
||||
|
||||
def test_round_accepts_when_profiles_hold_and_reports_small_profiles(tmp_path: Path) -> None:
|
||||
repo = tmp_path / "repo"
|
||||
repo.mkdir()
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\ndef g():\n print('hello')\n")
|
||||
ruleset_dir = _ruleset_dir(tmp_path, [_rule("good-eval", r"\beval\s*\("), _rule("noisy-print", r"\bprint\s*\(")])
|
||||
_, manifest = _manifest(tmp_path, _expected("good-eval"))
|
||||
ledger = tmp_path / "rule_policy.json"
|
||||
journal = tmp_path / "journal.json"
|
||||
human_pairs = [
|
||||
_pair(tmp_path, f"h{i}", "def h(y):\n return eval(y)\n", "def h(y):\n return y\n", "good-eval", "eval", provenance="human")
|
||||
for i in range(5)
|
||||
]
|
||||
tiny_agent = [
|
||||
_pair(tmp_path, "a0", "def h(y):\n print(y)\n", "def h(y):\n return y\n", "noisy-print", "print", provenance="agent")
|
||||
]
|
||||
outcome = run_round(
|
||||
repo, manifest, ledger_path=ledger, journal_path=journal, ruleset_dir=ruleset_dir, pair_cases=human_pairs + tiny_agent
|
||||
)
|
||||
assert outcome.accepted and outcome.reason == "accepted"
|
||||
assert outcome.profile_regressions == []
|
||||
assert outcome.profiles_under_minimum == ["agent"] # one pair: reported, not gated
|
||||
assert ledger.exists()
|
||||
|
||||
|
||||
def test_profile_gate_ignores_title_and_advisory_tiers(tmp_path: Path) -> None:
|
||||
"""Req 9.3: only seeded and reviewed pairs gate; a regression confined to title pairs never rejects a round."""
|
||||
repo = tmp_path / "repo"
|
||||
repo.mkdir()
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\ndef g():\n print('hello')\n")
|
||||
ruleset_dir = _ruleset_dir(tmp_path, [_rule("good-eval", r"\beval\s*\("), _rule("noisy-print", r"\bprint\s*\(")])
|
||||
_, manifest = _manifest(tmp_path, _expected("good-eval"))
|
||||
ledger = tmp_path / "rule_policy.json"
|
||||
journal = tmp_path / "journal.json"
|
||||
title_pairs = [
|
||||
_pair(
|
||||
tmp_path,
|
||||
f"t{i}",
|
||||
"def h(y):\n print(y)\n",
|
||||
"def h(y):\n return y\n",
|
||||
"noisy-print",
|
||||
"print",
|
||||
provenance="agent",
|
||||
review_tier="title",
|
||||
)
|
||||
for i in range(5)
|
||||
]
|
||||
outcome = run_round(repo, manifest, ledger_path=ledger, journal_path=journal, ruleset_dir=ruleset_dir, pair_cases=title_pairs)
|
||||
assert outcome.accepted and outcome.reason == "accepted"
|
||||
assert outcome.profile_regressions == [] and "agent" not in outcome.per_profile_before
|
||||
ledger.unlink()
|
||||
journal.unlink()
|
||||
reviewed_pairs = [
|
||||
_pair(tmp_path, f"r{i}", "def h(y):\n print(y)\n", "def h(y):\n return y\n", "noisy-print", "print", provenance="agent")
|
||||
for i in range(5)
|
||||
]
|
||||
outcome = run_round(
|
||||
repo, manifest, ledger_path=ledger, journal_path=journal, ruleset_dir=ruleset_dir, pair_cases=reviewed_pairs + title_pairs
|
||||
)
|
||||
assert not outcome.accepted and outcome.reason == "profile_regression:agent"
|
||||
assert outcome.per_profile_before["agent"]["pairs"] == 5.0 # the five title pairs are scored elsewhere, never here
|
||||
|
||||
|
||||
def test_profile_gate_scores_vendored_pairs_only(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
|
||||
"""Req 10.4 / design: gates call select_vendored; a pointer pair never reaches the improve gate, cached or not."""
|
||||
from openultrasast.improve.evolve import evaluate_profiles
|
||||
from openultrasast.pairs import PairCase
|
||||
from openultrasast.ruleset import load_ruleset
|
||||
|
||||
monkeypatch.setenv("OPENULTRASAST_PAIRS_NETWORK", "1")
|
||||
vendored = _pair(
|
||||
tmp_path, "v", "def h(y):\n return eval(y)\n", "def h(y):\n return y\n", "python-unsafe-eval", "eval", provenance="human"
|
||||
)
|
||||
pointer_src = tmp_path / "cache" / "github" / "p"
|
||||
pointer_src.mkdir(parents=True)
|
||||
(pointer_src / "vuln.py").write_text("def h(y):\n return eval(y)\n")
|
||||
(pointer_src / "fixed.py").write_text("def h(y):\n return y\n")
|
||||
pointer = PairCase(
|
||||
**{
|
||||
**vendored.__dict__,
|
||||
"name": "p",
|
||||
"provenance": "agent",
|
||||
"vendored": False,
|
||||
"vuln_file": pointer_src / "vuln.py",
|
||||
"fixed_file": pointer_src / "fixed.py",
|
||||
"recipe": (("repo", "o/r"), ("parent", "a"), ("commit", "b"), ("path", "app.py"), ("mode", "name")),
|
||||
}
|
||||
)
|
||||
profiles = evaluate_profiles([vendored, pointer], load_ruleset())
|
||||
assert set(profiles) == {"human"} # the agent pointer pair is scored by `pairs --pointers`, never by the gate
|
||||
@@ -1,330 +0,0 @@
|
||||
"""Proposals from plane memory (harnessx-removal task 3, design section 5): rules M1/M2, the advisory M3 signal,
|
||||
the train-on-test guard, provenance, and the unchanged default `improve`."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import pytest
|
||||
|
||||
from openultrasast.benchmark import load_benchmark_manifest
|
||||
from openultrasast.cli import main
|
||||
from openultrasast.improve import EvolveValidator, RuleStatusEdit, run_round
|
||||
from openultrasast.improve.memory import (
|
||||
SIDECAR,
|
||||
Guard,
|
||||
MemoryProposal,
|
||||
disputed_families,
|
||||
outcome_of,
|
||||
outcome_rows,
|
||||
propose_from_memory,
|
||||
read_sidecar,
|
||||
)
|
||||
from openultrasast.pairs import PairCase
|
||||
from openultrasast.plane.memory import FileStore, Record, row_id
|
||||
from openultrasast.ruleset import PatternRule, read_rule_ledger, write_ruleset
|
||||
|
||||
P = {name: ch * 40 for name, ch in (("a", "a"), ("b", "b"), ("c", "c"), ("f", "f"))}
|
||||
NO_GUARD = Guard()
|
||||
|
||||
|
||||
def _rule(rule_id: str, pattern: str = r"\bexec\s*\(", status: str = "enabled") -> PatternRule:
|
||||
return PatternRule(
|
||||
rule_id=rule_id, title=rule_id, languages=("python",), cwe="CWE-95", tags=("injection",), pattern=pattern, status=status
|
||||
)
|
||||
|
||||
|
||||
def _row(kind: str, repo: str, pin: str, subject: str, population: str = "dev", **fields: Any) -> dict[str, Any]:
|
||||
return {
|
||||
"id": row_id(kind, "run-1", f"{repo}-task", subject), "kind": kind, "repo": f"github.com/o/{repo}", "pin": pin,
|
||||
"run": "run-1", "task": f"{repo}-task", "population": population, "split": "validation", "image": "sha256:x", **fields,
|
||||
} # fmt: skip
|
||||
|
||||
|
||||
def verdict(repo: str, pin: str, fn: str, final: str, site_match: bool | None = None, **kw: Any) -> dict[str, Any]:
|
||||
cand = f"app.py::{fn}"
|
||||
return _row("verdict", repo, pin, cand, candidate=cand, family="injection", final=final, site_match=site_match, **kw)
|
||||
|
||||
|
||||
def alert(rule: str, repo: str, pin: str, fn: str, role: str = "vulnerable", status: str = "enabled", **kw: Any) -> dict[str, Any]:
|
||||
fields = {"rule_id": rule, "rule_status": status, "path": "app.py", "line": 3, "function": fn, "pin_role": role}
|
||||
return _row("alert", repo, pin, f"{rule}:{fn}:{role}", **fields, **kw)
|
||||
|
||||
|
||||
def recs(rows: list[dict[str, Any]]) -> list[Record]:
|
||||
return [Record(r, f"repos/{r['repo']}/{r['pin']}.jsonl", None) for r in rows]
|
||||
|
||||
|
||||
def propose(rows: list[dict[str, Any]], rules: list[PatternRule], guard: Guard = NO_GUARD, ledger: Any = None) -> list[MemoryProposal]:
|
||||
return propose_from_memory(recs(rows), {r.rule_id: r for r in rules}, ledger or {}, [], excluded=guard)
|
||||
|
||||
|
||||
def false_alerts(rule: str = "noisy", repos: tuple[str, ...] = ("r1", "r1", "r2"), population: str = "dev") -> list[dict[str, Any]]:
|
||||
rows: list[dict[str, Any]] = []
|
||||
for i, repo in enumerate(repos):
|
||||
rows.append(verdict(repo, P["a"], f"f{i}", "rejected", population=population))
|
||||
rows.append(alert(rule, repo, P["a"], f"f{i}", population=population))
|
||||
return rows
|
||||
|
||||
|
||||
# ---- M1: repeated false alerts -------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_m1_demotes_a_rule_with_false_alerts_across_repositories() -> None:
|
||||
[proposal] = propose(false_alerts(), [_rule("noisy")])
|
||||
assert (proposal.edit.from_status, proposal.edit.to_status) == ("enabled", "shadow")
|
||||
assert proposal.edit.rationale == f"memory:{proposal.proposal_id}"
|
||||
assert proposal.provenance["rule"] == "M1"
|
||||
assert proposal.provenance["counts"] == {"false_alerts": 3, "repos": 2, "agreed_hits": 0}
|
||||
assert len(proposal.provenance["evidence"]) == 6 # three alerts and the three rejected verdicts they sit at
|
||||
|
||||
|
||||
@pytest.mark.parametrize("repos", [("r1", "r2"), ("r1", "r1", "r1")], ids=["two-alerts", "one-repository"])
|
||||
def test_m1_stays_silent_below_threshold(repos: tuple[str, ...]) -> None:
|
||||
assert propose(false_alerts(repos=repos), [_rule("noisy")]) == []
|
||||
|
||||
|
||||
def test_m1_never_demotes_a_rule_with_an_agreed_hit() -> None:
|
||||
rows = false_alerts() + [verdict("r3", P["b"], "hit", "agreed"), alert("noisy", "r3", P["b"], "hit")]
|
||||
assert propose(rows, [_rule("noisy")]) == []
|
||||
site = false_alerts() + [verdict("r3", P["b"], "site", "rejected", site_match=True), alert("noisy", "r3", P["b"], "site")]
|
||||
assert propose(site, [_rule("noisy")]) == []
|
||||
|
||||
|
||||
def test_m1_counts_fixed_pin_alerts_in_the_fix_range_and_ignores_disputes() -> None:
|
||||
rows = [
|
||||
verdict("r1", P["a"], "sink", "agreed", site_match=True),
|
||||
verdict("r2", P["b"], "sink", "agreed", site_match=True),
|
||||
alert("noisy", "r1", P["f"], "sink", role="fixed"), # the declared site's function on the fixed pin
|
||||
alert("noisy", "r2", P["f"], "sink", role="fixed"),
|
||||
alert("noisy", "r2", P["c"], "other", role="fixed", in_fix_range=True),
|
||||
verdict("r1", P["a"], "maybe", "disputed"),
|
||||
alert("noisy", "r1", P["a"], "maybe"), # disputed: evidence in neither direction
|
||||
]
|
||||
[proposal] = propose(rows, [_rule("noisy")])
|
||||
assert proposal.provenance["counts"]["false_alerts"] == 3
|
||||
|
||||
|
||||
def test_m1_skips_rules_the_ruleset_does_not_know_or_that_are_not_enabled() -> None:
|
||||
assert propose(false_alerts(rule="ghost"), [_rule("noisy")]) == []
|
||||
assert propose(false_alerts(), [_rule("noisy")], ledger={"noisy": {"status": "shadow"}}) == []
|
||||
|
||||
|
||||
# ---- M2: missed declared sites -------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def site_hits(rule: str = "quiet") -> list[dict[str, Any]]:
|
||||
return [
|
||||
verdict("r1", P["a"], "s1", "agreed", site_match=True),
|
||||
verdict("r2", P["b"], "s2", "agreed", site_match=True),
|
||||
alert(rule, "r1", P["a"], "s1", status="shadow"),
|
||||
alert(rule, "r2", P["b"], "s2", status="shadow"),
|
||||
]
|
||||
|
||||
|
||||
def test_m2_promotes_a_shadow_rule_that_hits_declared_sites_nothing_else_catches() -> None:
|
||||
[proposal] = propose(site_hits(), [_rule("quiet", status="shadow"), _rule("other")])
|
||||
assert (proposal.edit.from_status, proposal.edit.to_status, proposal.provenance["rule"]) == ("shadow", "enabled", "M2")
|
||||
assert proposal.provenance["counts"]["site_hits"] == 2
|
||||
|
||||
|
||||
def test_m2_stays_silent_below_threshold_or_when_already_caught() -> None:
|
||||
rules = [_rule("quiet", status="shadow"), _rule("other")]
|
||||
assert propose(site_hits()[:3], rules) == [] # one hit
|
||||
assert propose([*site_hits(), alert("other", "r1", P["a"], "s1")], rules) == [] # an enabled rule catches s1
|
||||
|
||||
|
||||
def test_m2_stays_silent_on_a_fixed_pin_alert_or_rejected_candidates() -> None:
|
||||
rules = [_rule("quiet", status="shadow")]
|
||||
assert propose([*site_hits(), alert("quiet", "r1", P["f"], "s1", role="fixed", status="shadow")], rules) == []
|
||||
rejected = [
|
||||
verdict("r3", P["c"], "x1", "rejected"),
|
||||
verdict("r3", P["c"], "x2", "rejected"),
|
||||
alert("quiet", "r3", P["c"], "x1", status="shadow"),
|
||||
alert("quiet", "r3", P["c"], "x2", status="shadow"),
|
||||
]
|
||||
assert propose([*site_hits(), *rejected[::2]], rules) != [] # one rejected alert is tolerated
|
||||
assert propose([*site_hits(), *rejected], rules) == []
|
||||
|
||||
|
||||
# ---- coverage: no rule could fire, so no evidence either way -------------------------------------------------------
|
||||
|
||||
|
||||
def covered(repo: str, pin: str, kind: str, language: str = "python", role: str = "vulnerable") -> dict[str, Any]:
|
||||
return _row("coverage", repo, pin, f"{language}:{role}:{kind}", language=language, coverage=kind, files=3, pin_role=role)
|
||||
|
||||
|
||||
def test_m1_skips_a_repository_and_pin_whose_coverage_for_the_rules_language_is_none() -> None:
|
||||
blind = [*false_alerts(), covered("r2", P["a"], "none")]
|
||||
assert propose(blind, [_rule("noisy")]) == [], "r2's alert could not be evidence: one repository is left"
|
||||
wider = [*false_alerts(repos=("r1", "r1", "r2", "r3")), covered("r3", P["a"], "none")]
|
||||
[proposal] = propose(wider, [_rule("noisy")])
|
||||
assert proposal.provenance["counts"] == {"false_alerts": 3, "repos": 2, "agreed_hits": 0}
|
||||
assert proposal.provenance["excluded"]["coverage_none"] == {f"github.com/o/r3@{P['a']}": 1}
|
||||
|
||||
|
||||
def test_coverage_none_of_another_language_or_beside_engine_coverage_is_no_skip() -> None:
|
||||
other = [*false_alerts(), covered("r2", P["a"], "none", language="php")]
|
||||
assert len(propose(other, [_rule("noisy")])) == 1, "the rule's language (python) was covered there"
|
||||
engine = [*false_alerts(), covered("r2", P["a"], "none"), covered("r2", P["a"], "engine")]
|
||||
assert len(propose(engine, [_rule("noisy")])) == 1, "a later run covered the pin: its alerts are evidence"
|
||||
|
||||
|
||||
def test_m2_skips_a_blind_repository_and_pin() -> None:
|
||||
rules = [_rule("quiet", status="shadow"), _rule("other")]
|
||||
assert propose([*site_hits(), covered("r2", P["b"], "none")], rules) == []
|
||||
|
||||
|
||||
# ---- M3: advisory only ---------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_disputed_families_are_advisory_and_never_reach_the_validator(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
|
||||
rows = [verdict("r1", P["a"], f"d{i}", "disputed" if i < 2 else "agreed") for i in range(5)]
|
||||
[signal] = disputed_families(recs(rows), NO_GUARD)
|
||||
assert signal["family"] == "injection" and signal["dispute_rate"] == 0.4
|
||||
assert disputed_families(recs(rows[1:]), NO_GUARD) == [] # four candidates: under the minimum
|
||||
assert propose(rows, [_rule("noisy")]) == [] # a dispute is no proposal
|
||||
|
||||
seen: list[object] = []
|
||||
original = EvolveValidator.validate
|
||||
monkeypatch.setattr(EvolveValidator, "validate", lambda self, edit, *a: (seen.append(edit), original(self, edit, *a))[1])
|
||||
repo, rules_dir, manifest = _scaffold(tmp_path)
|
||||
[proposal] = propose(false_alerts(), [_rule("noisy")])
|
||||
run_round(
|
||||
repo, manifest, ledger_path=tmp_path / "l.json", journal_path=tmp_path / "j.json", ruleset_dir=rules_dir, proposals=[proposal]
|
||||
)
|
||||
assert seen and all(isinstance(edit, RuleStatusEdit) for edit in seen)
|
||||
|
||||
|
||||
# ---- the train-on-test guard ---------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_guard_drops_the_qualification_population_and_reports_it() -> None:
|
||||
rows = false_alerts(population="independent-v2/validation") + false_alerts(repos=("r4",), population="dev")
|
||||
guard = Guard.build(populations=["independent-v2/validation"])
|
||||
kept, excluded = guard.apply(recs(rows))
|
||||
assert len(kept) == 2 and excluded["rows"] == 6 and excluded["population"] == {"independent-v2/validation": 6}
|
||||
assert propose(rows, [_rule("noisy")], guard) == [] # the only evidence is excluded: no proposal
|
||||
assert propose(rows, [_rule("noisy")]) != [] # the same rows without the guard would propose
|
||||
|
||||
|
||||
def test_guard_drops_holdout_pair_repositories_and_gated_cases() -> None:
|
||||
holdout = PairCase(
|
||||
name="h", slice="s", language="python", origin="o", vuln_file=Path("v"), fixed_file=Path("f"), relpath="app.py",
|
||||
expected=(), min_recall=1.0, fix_policy="silent", repo="O/R1", split="holdout",
|
||||
) # fmt: skip
|
||||
guard = Guard.build(gated_cases=[("https://github.com/o/r2", P["a"])], holdout_pairs=[holdout])
|
||||
kept, excluded = guard.apply(recs(false_alerts()))
|
||||
assert kept == []
|
||||
assert excluded["holdout_pair"] == {f"github.com/o/r1@{P['a']}": 4}
|
||||
assert excluded["gated_case"] == {f"github.com/o/r2@{P['a']}": 2}
|
||||
[proposal] = propose(false_alerts(repos=("r1", "r2", "r3", "r3", "r4")), [_rule("noisy")], guard)
|
||||
assert {e["repo"] for e in proposal.provenance["evidence"]} == {"github.com/o/r3", "github.com/o/r4"}
|
||||
assert proposal.provenance["excluded"]["rows"] == 4
|
||||
|
||||
|
||||
# ---- provenance, the round, and outcomes ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _scaffold(tmp_path: Path) -> tuple[Path, Path, Any]:
|
||||
repo = tmp_path / "repo"
|
||||
repo.mkdir()
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\n")
|
||||
rules_dir = tmp_path / "ruleset"
|
||||
write_ruleset(rules_dir / "python" / "rules.toml", [_rule("good-eval", r"\beval\s*\("), _rule("noisy")])
|
||||
manifest = tmp_path / "bench.toml"
|
||||
manifest.write_text(
|
||||
'name = "t"\nlanguage = "python_web"\n\n[source]\npath = "repo"\n\n[[expected]]\ncwe = "CWE-95"\n'
|
||||
'class = "code injection"\npath = "app.py"\nline = 2\nrule_id = "good-eval"\nsink = "eval"\nevidence = "eval"\n'
|
||||
)
|
||||
return repo, rules_dir, load_benchmark_manifest(manifest)
|
||||
|
||||
|
||||
def test_provenance_round_trips_journal_sidecar_store(tmp_path: Path) -> None:
|
||||
store = FileStore(tmp_path / "memory")
|
||||
store.ingest_rows("run-1", "t", false_alerts())
|
||||
records = store.rows()
|
||||
[proposal] = propose_from_memory(records, {"noisy": _rule("noisy")}, {}, [], excluded=NO_GUARD, index_digest="idx")
|
||||
repo, rules_dir, manifest = _scaffold(tmp_path)
|
||||
journal, ledger = tmp_path / "cal" / "journal.json", tmp_path / "cal" / "rule_policy.json"
|
||||
|
||||
outcome = run_round(repo, manifest, ledger_path=ledger, journal_path=journal, ruleset_dir=rules_dir, proposals=[proposal])
|
||||
|
||||
assert outcome.accepted and read_rule_ledger(ledger)["noisy"]["status"] == "shadow"
|
||||
[entry] = json.loads(journal.read_text())[0]["edits"]
|
||||
assert entry["evidence"] == proposal.proposal_id and entry["rationale"] == f"memory:{proposal.proposal_id}"
|
||||
side = read_sidecar(journal)[entry["evidence"]]
|
||||
assert side["edit_key"] == entry["key"] and side["index_digest"] == "idx" and side["round"] == 1
|
||||
by_id = {rec.row["id"]: rec for rec in records}
|
||||
for ev in side["evidence"]: # every cited row is in the store, at the object and version recorded
|
||||
assert by_id[ev["id"]].key == ev["key"] and by_id[ev["id"]].version == ev["version"]
|
||||
assert ev["version"].startswith("sha256:")
|
||||
|
||||
rows = outcome_rows(
|
||||
outcome.edits, outcome_of(outcome.reason, outcome.accepted) or "", 1, outcome.reason, {proposal.proposal_id: proposal}, "g"
|
||||
)
|
||||
store.put_rows(rows)
|
||||
back = store.rows(kind="proposal_outcome")
|
||||
assert {r.row["outcome"] for r in back} == {"accepted"} and {r.row["proposal_id"] for r in back} == {proposal.proposal_id}
|
||||
assert len(back) == 2 # one per (repository, pin) of the evidence
|
||||
|
||||
again = run_round(repo, manifest, ledger_path=ledger, journal_path=journal, ruleset_dir=rules_dir, proposals=[proposal])
|
||||
assert again.reason == "no_proposals" # the rule is shadow now: the proposal no longer applies
|
||||
|
||||
|
||||
def test_a_reverted_memory_proposal_is_not_retried_on_the_same_evidence(tmp_path: Path) -> None:
|
||||
repo, rules_dir, manifest = _scaffold(tmp_path)
|
||||
rows = [verdict(r, P["a"], f"f{i}", "rejected") for i, r in enumerate(("r1", "r1", "r2"))]
|
||||
rows += [alert("good-eval", r, P["a"], f"f{i}") for i, r in enumerate(("r1", "r1", "r2"))]
|
||||
[proposal] = propose(rows, [_rule("good-eval", r"\beval\s*\(")])
|
||||
journal = tmp_path / "j.json"
|
||||
first = run_round(repo, manifest, ledger_path=tmp_path / "l.json", journal_path=journal, ruleset_dir=rules_dir, proposals=[proposal])
|
||||
assert first.reason == "reverted" and outcome_of(first.reason, first.accepted) == "reverted" # shadowing drops recall
|
||||
second = run_round(repo, manifest, ledger_path=tmp_path / "l.json", journal_path=journal, ruleset_dir=rules_dir, proposals=[proposal])
|
||||
assert second.reason == "no_proposals"
|
||||
assert outcome_of("validation_failed: x", False) == "rejected" and outcome_of("no_proposals", False) is None
|
||||
|
||||
|
||||
def test_default_round_journal_has_no_evidence_and_no_sidecar(tmp_path: Path) -> None:
|
||||
repo, rules_dir, manifest = _scaffold(tmp_path)
|
||||
(repo / "app.py").write_text("def f(x):\n return eval(x)\ndef g(y):\n exec(y)\n")
|
||||
journal = tmp_path / "j.json"
|
||||
outcome = run_round(repo, manifest, ledger_path=tmp_path / "l.json", journal_path=journal, ruleset_dir=rules_dir)
|
||||
assert outcome.edits # the benchmark's own edit
|
||||
[entry] = json.loads(journal.read_text())[0]["edits"]
|
||||
assert set(entry) == {"key", "lever", "rule_id", "from", "to", "rationale"}
|
||||
assert not (tmp_path / SIDECAR).exists()
|
||||
|
||||
|
||||
# ---- the CLI -------------------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_default_improve_output_matches_the_recorded_baseline(capsys: pytest.CaptureFixture[str]) -> None:
|
||||
root = Path(__file__).resolve().parents[1]
|
||||
baseline = root / "benchmarks/measurements/2026-09-30-harnessx-removal-baseline/improve/stdout.txt"
|
||||
manifest = "benchmarks/manifests/java-spring-boot-vulnerable.toml"
|
||||
import os
|
||||
|
||||
cwd = Path.cwd()
|
||||
os.chdir(root)
|
||||
try:
|
||||
assert main(["improve", manifest, "--dry-run", "--no-pair-gate"]) == 0
|
||||
finally:
|
||||
os.chdir(cwd)
|
||||
assert capsys.readouterr().out.replace(str(root), "<ROOT>") == baseline.read_text()
|
||||
|
||||
|
||||
def test_improve_memory_with_an_empty_store_has_no_proposals(tmp_path: Path, capsys: pytest.CaptureFixture[str]) -> None:
|
||||
repo, rules_dir, _ = _scaffold(tmp_path)
|
||||
store = tmp_path / "memory"
|
||||
store.mkdir()
|
||||
(store / "index.jsonl").write_text("")
|
||||
catalog = Path(__file__).resolve().parents[1] / "benchmarks/pairs/catalog.toml"
|
||||
args = ["improve", str(tmp_path / "bench.toml"), "--ruleset-dir", str(rules_dir), "--dry-run", "--pair-catalog", str(catalog)]
|
||||
assert main([*args, "--no-pair-gate", "--memory", str(store)]) == 0
|
||||
out = capsys.readouterr().out
|
||||
assert "rows=0" in out and "memory_proposals=0" in out and "round 1: no_proposals" in out
|
||||
with pytest.raises(SystemExit, match="no index.jsonl"):
|
||||
main([*args, "--memory", str(tmp_path / "absent")])
|
||||
@@ -13,12 +13,12 @@ from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import pytest
|
||||
from test_plane_loop import LOOP, _templates, _write
|
||||
import yaml
|
||||
|
||||
from openultrasast import cli
|
||||
from openultrasast.plane import engine_alerts
|
||||
from openultrasast.plane.engine_alerts import EngineAlertsError, alerts_engine, engine_rows, parse_site, read_record
|
||||
from openultrasast.plane.generate import CaseInputs, render
|
||||
from openultrasast.plane.manifests import RUN_API_VERSION
|
||||
|
||||
FIXTURE = Path(__file__).parent / "fixtures" / "plane-engine"
|
||||
CASE = "php-probe"
|
||||
@@ -44,17 +44,29 @@ def world(tmp_path: Path) -> dict[str, Any]:
|
||||
(repo / "app" / "index.php").write_text(text)
|
||||
_git(repo, "-c", "user.name=t", "-c", "[email protected]", "commit", "-qam", "fixed")
|
||||
fixed = _git(repo, "rev-parse", "HEAD")
|
||||
record = {"id": CASE, "repo": "https://example.com/reserved/probe", "family": "injection", "vulnerable": vulnerable, "fixed": fixed}
|
||||
candidate = ("app/index.php", "<global>", 3)
|
||||
case = CaseInputs(
|
||||
record, (candidate,), (candidate,), {}, 0.0, ranges={"app/index.php": [(3, 3)]}, fixed_ranges={"app/index.php": [(3, 3)]}
|
||||
)
|
||||
files = render([case], _templates(), "validation-46", "t", population="p", split="s", loop=LOOP)
|
||||
plane = tmp_path / "plane"
|
||||
for name in ("tasks", "workspaces", "runs"):
|
||||
(plane / name).mkdir(parents=True)
|
||||
task = yaml.safe_load((Path(__file__).resolve().parents[1] / "plane/tasks/alerts.yaml").read_text())
|
||||
task["metadata"]["name"] = f"{CASE}-alerts"
|
||||
task["spec"]["env"] = [{"name": f"OUSAST_{role.upper()}_PIN", "value": pin} for role, pin in
|
||||
(("vulnerable", vulnerable), ("fixed", fixed))]
|
||||
task["spec"]["workspaces"] = [{"name": "case", "path": "/workspace/case"}]
|
||||
(plane / "tasks/alerts.yaml").write_text(yaml.safe_dump(task))
|
||||
case = {"family": "injection", "ranges": {"app/index.php": [[3, 3]]}, "fixed_ranges": {"app/index.php": [[3, 3]]}}
|
||||
workspace = {"apiVersion": task["apiVersion"], "kind": "Workspace", "metadata": {"name": "case"}, "spec": {
|
||||
"git": [{"name": "repo", "repo": "https://example.com/reserved/probe", "dir": "repo"}],
|
||||
"files": [{"path": "case.json", "content": json.dumps(case)}],
|
||||
}}
|
||||
(plane / "workspaces/case.yaml").write_text(yaml.safe_dump(workspace))
|
||||
run = plane / "runs/validation-46.yaml"
|
||||
run.write_text(yaml.safe_dump({"apiVersion": RUN_API_VERSION, "kind": "Run", "metadata": {"name": "validation-46"},
|
||||
"spec": {"tasks": [{"name": f"{CASE}-alerts", "task": f"{CASE}-alerts",
|
||||
"outputs": ["alerts.jsonl", "summary.json"], "budget": {"usd": 0, "calls": 0}}]}}))
|
||||
source = tmp_path / "frozen"
|
||||
(source / "src" / "openultrasast").mkdir(parents=True)
|
||||
(source / "benchmarks" / "push").mkdir(parents=True)
|
||||
(source / "benchmarks" / "push" / "finding_dump.py").write_text("")
|
||||
run = _write(tmp_path / "plane", files)
|
||||
return {"run": run, "source": source, "cache": tmp_path / "cache", "pins": (vulnerable, fixed), "results": tmp_path / "results"}
|
||||
|
||||
|
||||
|
||||
@@ -1,313 +0,0 @@
|
||||
"""The improvement loop as a plane Run (harnessx-removal Req 6.3, design section 6).
|
||||
|
||||
The whole graph runs through the reconciler and the fake ax of ``test_plane_reconciler``, whose ``exec`` mode runs the
|
||||
real task modules on the inputs the receiver serves: a recorded case's chain (validation-46's lmdeploy case from
|
||||
``tests/fixtures/plane-remember``, already done) -> ``alerts`` -> ``remember`` -> ``loop-measure`` -> ``loop-propose``
|
||||
-> ``loop-improve``, with ``memory-snapshot`` seeded on the host from a crafted store.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import shutil
|
||||
import subprocess
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import pytest
|
||||
import yaml
|
||||
from test_plane_reconciler import Fake, fake # noqa: F401 -- `fake` is the fixture
|
||||
|
||||
import openultrasast
|
||||
from openultrasast import cli
|
||||
from openultrasast.plane import reconciler
|
||||
from openultrasast.plane.generate import TEMPLATES, CaseInputs, Loop, fix_ranges, render
|
||||
from openultrasast.plane.manifests import load_manifests
|
||||
from openultrasast.plane.memory import SNAPSHOT_ANNOTATION, FileStore, row_id
|
||||
from openultrasast.plane.tasks import alerts
|
||||
|
||||
FIXTURE = Path(__file__).parent / "fixtures" / "plane-remember" / "validation-46"
|
||||
PLANE = Path(__file__).resolve().parents[1] / "plane"
|
||||
SRC = str(Path(openultrasast.__file__).resolve().parents[1])
|
||||
CASE = "lmdeploy-media-url-ssrf"
|
||||
VULNERABLE = "71648d239c978452720a6d6692fe914efc8a7b94"
|
||||
FIXED = "ab" * 20
|
||||
OTHER = "12" * 20
|
||||
EVAL = "python-unsafe-eval" # a bundled, enabled rule
|
||||
CHAIN = ("facts", "va", "vb", "agree", "vc", "final")
|
||||
LOOP_TASKS = ("memory-snapshot", "loop-measure", "loop-propose", "loop-improve")
|
||||
|
||||
|
||||
def _case(**ranges: Any) -> CaseInputs:
|
||||
record = {
|
||||
"id": CASE, "repo": "https://example.com/reserved/lmdeploy", "family": "untrusted_destination", "vulnerable": VULNERABLE,
|
||||
"fixed": FIXED, "sites": ["lmdeploy/vl/media/connection.py::_load_http_url"],
|
||||
} # fmt: skip
|
||||
candidate = ("lmdeploy/vl/media/connection.py", "_load_http_url", 40)
|
||||
return CaseInputs(record, (candidate,), (candidate,), {}, 0.02, **ranges)
|
||||
|
||||
|
||||
def _templates() -> dict:
|
||||
return load_manifests([PLANE / "tasks" / f"{name}.yaml" for name in TEMPLATES]).tasks
|
||||
|
||||
|
||||
LOOP = Loop(repo="https://example.invalid/openultrasast.git", commit="c" * 40, manifest="bench.toml", catalog="catalog.toml")
|
||||
|
||||
|
||||
def _write(plane: Path, files: dict[str, str]) -> Path:
|
||||
shutil.copytree(PLANE / "models", plane / "models")
|
||||
for relative, text in files.items():
|
||||
(plane / relative).parent.mkdir(parents=True, exist_ok=True)
|
||||
(plane / relative).write_text(text)
|
||||
return plane / "runs" / "validation-46.yaml"
|
||||
|
||||
|
||||
# --- alerts --------------------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_alerts_attribute_functions_and_mark_the_fix_on_both_pins(tmp_path: Path) -> None:
|
||||
vulnerable, fixed = tmp_path / "v", tmp_path / "f"
|
||||
(vulnerable / "pkg").mkdir(parents=True)
|
||||
(fixed / "pkg").mkdir(parents=True)
|
||||
(vulnerable / "pkg" / "run.py").write_text("import os\n\ndef handle(x):\n y = x\n return eval(y)\n")
|
||||
(fixed / "pkg" / "run.py").write_text("import os\n\ndef handle(x):\n return eval(x, {})\n\nprint(eval('1'))\n")
|
||||
case = tmp_path / "case.json"
|
||||
case.write_text(json.dumps({"ranges": {"pkg/run.py": [[4, 5]]}, "fixed_ranges": {"pkg/run.py": [[4, 4]]}}))
|
||||
out = tmp_path / "out"
|
||||
env = {
|
||||
"OUSAST_OUTPUT_DIR": str(out), "OUSAST_WORKSPACE_DIR": str(vulnerable), "OUSAST_FIXED_DIR": str(fixed),
|
||||
"OUSAST_INPUT_CASE": str(case), "OUSAST_VULNERABLE_PIN": VULNERABLE, "OUSAST_FIXED_PIN": FIXED,
|
||||
} # fmt: skip
|
||||
assert alerts.main(env) == 0
|
||||
rows = [json.loads(line) for line in (out / "alerts.jsonl").read_text().splitlines()]
|
||||
got = [(r["rule_id"], r["pin_role"], r["pin"], r["line"], r["function"], r["in_fix_range"]) for r in rows if r["rule_id"] == EVAL]
|
||||
assert got == [
|
||||
(EVAL, "vulnerable", VULNERABLE, 5, "handle", True),
|
||||
(EVAL, "fixed", FIXED, 4, "handle", True),
|
||||
(EVAL, "fixed", FIXED, 6, "<global>", False),
|
||||
]
|
||||
summary = json.loads((out / "summary.json").read_text())
|
||||
assert (summary["status"], summary["units_done"], summary["usd"], summary["calls"], summary["model"]) == ("done", 2, 0, 0, None)
|
||||
assert summary["read"]["vulnerable"]["files"] == 1 and summary["read"]["vulnerable"]["bytes"] > 0, "the scan read its tree"
|
||||
assert alerts.in_range(None, "pkg/run.py", 4) is None, "no ranges: unknown, not outside"
|
||||
|
||||
empty = tmp_path / "empty"
|
||||
empty.mkdir()
|
||||
assert alerts.main({**env, "OUSAST_FIXED_DIR": str(empty)}) == 2
|
||||
assert "an unread tree is not a clean one" in json.loads((out / "summary.json").read_text())["reason"]
|
||||
|
||||
|
||||
def test_quick_coverage_is_derived_from_the_ruleset_directory(tmp_path: Path) -> None:
|
||||
shipped = alerts.quick_languages()
|
||||
assert {"python", "javascript", "typescript", "java", "c", "groovy", "php"} <= shipped, "PHP has quick rules"
|
||||
assert "php" in alerts.engine_languages(), "the engine's semantic models cover PHP"
|
||||
rules = tmp_path / "ruleset" / "php"
|
||||
rules.mkdir(parents=True)
|
||||
rule = 'rule_id = "php-x"\ntitle = "x"\nlanguages = ["php"]\ncwe = "CWE-89"\ntags = []\npattern = "x"\n'
|
||||
(rules / "rules.toml").write_text(f"[[rule]]\n{rule}")
|
||||
assert alerts.quick_languages(tmp_path / "ruleset") == {"php"}, "a ruleset added for a language covers it"
|
||||
(rules / "rules.toml").write_text(f'[[rule]]\n{rule}status = "disabled"\n')
|
||||
assert alerts.quick_languages(tmp_path / "ruleset") == set(), "a disabled rule fires nowhere"
|
||||
|
||||
|
||||
def test_alerts_on_an_uncovered_language_record_coverage_none_not_a_clean_zero(tmp_path: Path) -> None:
|
||||
for side in ("v", "f"):
|
||||
(tmp_path / side).mkdir()
|
||||
(tmp_path / side / "app.rb").write_text('id = params["id"]\nDB.execute("SELECT " + id)\n')
|
||||
(tmp_path / side / "app.js").write_text("module.exports = 1;\n")
|
||||
out = tmp_path / "out"
|
||||
env = {
|
||||
"OUSAST_OUTPUT_DIR": str(out), "OUSAST_WORKSPACE_DIR": str(tmp_path / "v"), "OUSAST_FIXED_DIR": str(tmp_path / "f"),
|
||||
"OUSAST_VULNERABLE_PIN": VULNERABLE, "OUSAST_FIXED_PIN": FIXED,
|
||||
} # fmt: skip
|
||||
assert alerts.main(env) == 0
|
||||
summary = json.loads((out / "summary.json").read_text())
|
||||
assert (out / "alerts.jsonl").read_text() == "" and summary["alerts"] == {"vulnerable": 0, "fixed": 0}
|
||||
assert summary["coverage"] == {
|
||||
"javascript": {"coverage": "quick", "files": {"fixed": 1, "vulnerable": 1}},
|
||||
"ruby": {"coverage": "none", "files": {"fixed": 1, "vulnerable": 1}},
|
||||
}
|
||||
assert summary["uncovered"] == ["ruby"] and summary["pins"] == {"vulnerable": VULNERABLE, "fixed": FIXED}
|
||||
|
||||
|
||||
def test_php_is_quick_covered_and_alerts_on_a_php_sink(tmp_path: Path) -> None:
|
||||
for side in ("v", "f"):
|
||||
(tmp_path / side).mkdir()
|
||||
(tmp_path / side / "index.php").write_text('<?php\n$id = $_GET["id"];\nmysqli_query($c, "SELECT " . $id);\n')
|
||||
out = tmp_path / "out"
|
||||
env = {
|
||||
"OUSAST_OUTPUT_DIR": str(out), "OUSAST_WORKSPACE_DIR": str(tmp_path / "v"), "OUSAST_FIXED_DIR": str(tmp_path / "f"),
|
||||
"OUSAST_VULNERABLE_PIN": VULNERABLE, "OUSAST_FIXED_PIN": FIXED,
|
||||
} # fmt: skip
|
||||
assert alerts.main(env) == 0
|
||||
summary = json.loads((out / "summary.json").read_text())
|
||||
assert summary["coverage"] == {"php": {"coverage": "quick", "files": {"fixed": 1, "vulnerable": 1}}}
|
||||
assert summary["uncovered"] == [] and summary["alerts"] == {"vulnerable": 1, "fixed": 1}
|
||||
rows = [json.loads(line) for line in (out / "alerts.jsonl").read_text().splitlines()]
|
||||
assert {(r["rule_id"], r["line"]) for r in rows} == {("php-sql-call-composition", 3)}
|
||||
|
||||
|
||||
def test_fix_ranges_new_side_are_the_fixed_pins_lines(tmp_path: Path) -> None:
|
||||
repo = tmp_path / "r"
|
||||
repo.mkdir()
|
||||
git = ["git", "-C", str(repo), "-c", "user.email=t@t", "-c", "user.name=t"]
|
||||
subprocess.run([*git, "init", "-q"], check=True)
|
||||
(repo / "a.py").write_text("a\nb\nc\n")
|
||||
subprocess.run([*git, "add", "."], check=True)
|
||||
subprocess.run([*git, "commit", "-qm", "v"], check=True)
|
||||
old = subprocess.run([*git, "rev-parse", "HEAD"], capture_output=True, text=True, check=True).stdout.strip()
|
||||
(repo / "a.py").write_text("a\nB1\nB2\nB3\nc\n")
|
||||
subprocess.run([*git, "commit", "-qam", "f"], check=True)
|
||||
new = subprocess.run([*git, "rev-parse", "HEAD"], capture_output=True, text=True, check=True).stdout.strip()
|
||||
assert fix_ranges(repo, old, new) == {"a.py": [(2, 2)]}
|
||||
assert fix_ranges(repo, old, new, side="new") == {"a.py": [(2, 4)]}
|
||||
|
||||
|
||||
# --- the generator -------------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_every_case_ends_in_remember_and_the_loop_is_model_free(tmp_path: Path) -> None:
|
||||
plain = render([_case()], _templates(), "validation-46", "t", population="population-v2", split="validation")
|
||||
run, _ = reconciler.load_run(_write(tmp_path / "plain", plain))
|
||||
assert [e.name.removeprefix(f"{CASE}-") for e in run.tasks] == [*CHAIN[:4], "vc", "final", "features", "remember"]
|
||||
remember = run.task(f"{CASE}-remember")
|
||||
assert remember.inputs["agreed"] == f"{CASE}-final/agreed.json" and "alerts" not in remember.inputs
|
||||
assert set(remember.producers) == {f"{CASE}-{step}" for step in ("facts", "va", "vb", "vc", "final", "features")}
|
||||
assert set(run.task(f"{CASE}-features").producers) == {f"{CASE}-{step}" for step in ("facts", "va", "vb", "vc", "final")}
|
||||
|
||||
files = render([_case(fixed_ranges={"x.py": [(1, 2)]})], _templates(), "validation-46", "t", population="p", split="s", loop=LOOP)
|
||||
run, manifests = reconciler.load_run(_write(tmp_path / "loop", files))
|
||||
names = [e.name for e in run.tasks]
|
||||
assert names[:6] == [f"{CASE}-{step}" for step in CHAIN]
|
||||
assert names[6:] == [f"{CASE}-alerts", f"{CASE}-features", f"{CASE}-remember", *LOOP_TASKS]
|
||||
assert run.task(f"{CASE}-features").inputs["alerts"] == f"{CASE}-alerts/alerts.jsonl", "features read the alerts"
|
||||
assert run.task(f"{CASE}-remember").inputs["alerts"] == f"{CASE}-alerts/alerts.jsonl"
|
||||
assert run.task(f"{CASE}-remember").inputs["alerts_summary"] == f"{CASE}-alerts/summary.json", "the coverage travels too"
|
||||
measure_inputs = {f"remember_{CASE}": f"{CASE}-remember/memory.jsonl", "snapshot": "memory-snapshot/rows.jsonl"}
|
||||
assert run.task("loop-measure").inputs == measure_inputs
|
||||
assert run.task("loop-improve").producers == ("loop-propose",)
|
||||
assert run.task("loop-propose").producers == ("loop-measure", "memory-snapshot")
|
||||
assert run.task(f"{CASE}-alerts").serialize == run.task("loop-improve").serialize == "engine"
|
||||
fixed = manifests.workspaces[f"{CASE}-fixed"]
|
||||
assert fixed.pins == {"repo": FIXED} and manifests.workspaces["openultrasast-cccccccccccc"].pins == {"repo": "c" * 40}
|
||||
for entry in run.tasks[6:]:
|
||||
task = manifests.tasks[entry.task]
|
||||
assert entry.budget is not None and (entry.budget.usd, entry.budget.calls) == (0, 0), entry.name
|
||||
assert reconciler.MODEL_ANNOTATION not in task.metadata.annotations, f"{entry.name} binds no Model"
|
||||
docs = reconciler.render_task(run, entry, manifests, "http://127.0.0.1:1/")
|
||||
env = {e["name"]: e["value"] for e in docs[-1]["spec"]["env"]}
|
||||
assert "OUSAST_MODEL" not in env and float(env["OUSAST_BUDGET_USD"]) == 0 and env["OUSAST_BUDGET_CALLS"] == "0"
|
||||
guard = json.loads(manifests.tasks["memory-snapshot"].metadata.annotations[SNAPSHOT_ANNOTATION])
|
||||
assert guard == {"catalog": "catalog.toml", "manifest": "bench.toml", "populations": []}
|
||||
inputs = yaml.safe_load((tmp_path / "loop" / "workspaces" / f"{CASE}-inputs.yaml").read_text())
|
||||
case_json = json.loads(next(f["content"] for f in inputs["spec"]["files"] if f["path"] == "case.json"))
|
||||
assert case_json["fixed_ranges"] == {"x.py": [[1, 2]]}
|
||||
|
||||
|
||||
# --- the whole loop through the fake ax ----------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _project(root: Path) -> Path:
|
||||
"""This repository at the commit under test, cut down: a benchmark whose one expected finding the eval rule
|
||||
makes, and an empty pair catalog."""
|
||||
(root / "target").mkdir(parents=True)
|
||||
(root / "target" / "app.py").write_text("def f(x):\n return eval(x)\n")
|
||||
(root / "bench.toml").write_text(
|
||||
'name = "t"\nlanguage = "python_web"\n\n[source]\npath = "target"\n\n[[expected]]\ncwe = "CWE-95"\n'
|
||||
f'class = "code injection"\npath = "app.py"\nline = 2\nrule_id = "{EVAL}"\nsink = "eval"\nevidence = "eval"\n'
|
||||
)
|
||||
(root / "catalog.toml").write_text("")
|
||||
(root / "benchmarks" / "pairs").mkdir(parents=True) # the catalog loader's mechanism vocabulary, by relative path
|
||||
shutil.copy(PLANE.parent / "benchmarks" / "pairs" / "mechanisms.toml", root / "benchmarks" / "pairs")
|
||||
return root
|
||||
|
||||
|
||||
def _older_run(store: FileStore) -> None:
|
||||
"""Two false alerts of the eval rule at rejected candidates of another repository, from an earlier run."""
|
||||
rows = []
|
||||
for fn in ("f1", "f2"):
|
||||
base = {"repo": "example.com/other/app", "pin": OTHER, "run": "older", "task": "app-remember", "population": "p", "split": "s"}
|
||||
base["image"] = "sha256:" + "0" * 64
|
||||
cand = f"app.py::{fn}"
|
||||
rows.append({**base, "id": row_id("verdict", "older", "app-remember", cand), "kind": "verdict", "candidate": cand,
|
||||
"family": "code_injection", "final": "rejected", "site_match": False}) # fmt: skip
|
||||
rows.append({**base, "id": row_id("alert", "older", "app-remember", fn), "kind": "alert", "rule_id": EVAL,
|
||||
"rule_status": "enabled", "path": "app.py", "line": 3, "function": fn, "pin_role": "vulnerable"}) # fmt: skip
|
||||
store.ingest_rows("older", "app-remember", rows)
|
||||
|
||||
|
||||
def test_the_loop_runs_as_one_plane_run_and_the_gate_rejects_a_demotion_that_costs_recall(
|
||||
fake: Fake, # noqa: F811
|
||||
tmp_path: Path,
|
||||
monkeypatch: pytest.MonkeyPatch,
|
||||
) -> None:
|
||||
project = _project(tmp_path / "project")
|
||||
monkeypatch.chdir(project) # `ousast plane run` from the repository root: the snapshot's guard paths resolve here
|
||||
store = FileStore(tmp_path / "memory")
|
||||
_older_run(store)
|
||||
monkeypatch.setenv("OUSAST_MEMORY", f"file://{store.root}")
|
||||
vulnerable, fixed = tmp_path / "checkout-v", tmp_path / "checkout-f"
|
||||
for root, body in ((vulnerable, "def load(x):\n return x\n"), (fixed, "def load(x):\n return eval(x)\n")):
|
||||
(root / "lmdeploy" / "vl" / "media").mkdir(parents=True)
|
||||
(root / "lmdeploy" / "vl" / "media" / "other.py").write_text(body)
|
||||
case = _case(ranges={"lmdeploy/vl/media/other.py": [(1, 2)]}, fixed_ranges={"lmdeploy/vl/media/other.py": [(1, 2)]})
|
||||
files = render([case], _templates(), "validation-46", "t", population="population-v2", split="validation", loop=LOOP)
|
||||
manifest = _write(tmp_path / "plane", files)
|
||||
inputs = yaml.safe_load((tmp_path / "plane" / "workspaces" / f"{CASE}-inputs.yaml").read_text())
|
||||
(tmp_path / "case.json").write_text(next(f["content"] for f in inputs["spec"]["files"] if f["path"] == "case.json"))
|
||||
|
||||
base = reconciler.run_dir("validation-46") # the recorded chain, done: the loop reuses it and runs only itself
|
||||
shutil.copytree(FIXTURE, base)
|
||||
state = {"run": "validation-46", "tasks": {f"{CASE}-{s}": {"status": "done"} for s in CHAIN}}
|
||||
(base / "state.json").write_text(json.dumps(state))
|
||||
|
||||
project_env = {"PYTHONPATH": SRC, "OUSAST_PROJECT_DIR": str(project)}
|
||||
fake.script({
|
||||
f"{CASE}-alerts": {"exec": {"env": {"PYTHONPATH": SRC, "OUSAST_WORKSPACE_DIR": str(vulnerable), "OUSAST_FIXED_DIR": str(fixed),
|
||||
"OUSAST_INPUT_CASE": str(tmp_path / "case.json")}}},
|
||||
f"{CASE}-features": {"exec": {"env": {"PYTHONPATH": SRC, "OUSAST_WORKSPACE_DIR": str(vulnerable)}}},
|
||||
f"{CASE}-remember": {"exec": {"env": {"PYTHONPATH": SRC}}},
|
||||
"loop-measure": {"exec": {"env": {"PYTHONPATH": SRC}}},
|
||||
"loop-propose": {"exec": {"env": project_env}},
|
||||
"loop-improve": {"exec": {"env": project_env}},
|
||||
}) # fmt: skip
|
||||
|
||||
assert cli.main(["plane", "run", str(manifest), "--ax", str(fake.ax)]) == 0, (fake.root / "ax.log").read_text()[-3000:]
|
||||
|
||||
statuses = {name: info["status"] for name, info in json.loads((base / "state.json").read_text())["tasks"].items()}
|
||||
assert set(statuses.values()) == {"done"} and len(statuses) == 13
|
||||
assert json.loads((base / "memory-snapshot" / "index.json").read_text())["kept"] == 4, "seeded on the host, not run"
|
||||
delivered = [name for _, name in fake.events("delivered")]
|
||||
assert delivered == [f"{CASE}-alerts", f"{CASE}-features", f"{CASE}-remember", "loop-measure", "loop-propose", "loop-improve"]
|
||||
|
||||
measure = json.loads((base / "loop-measure" / "measure.json").read_text())
|
||||
assert measure["run_metrics"]["declared_sites_agreed"] == 1 and measure["run_metrics"]["candidates"] == 1
|
||||
assert measure["run_metrics"]["alerts"] == {EVAL: {"vulnerable": 0, "fixed": 1, "fixed_in_fix_range": 1}}
|
||||
assert measure["union_metrics"]["rejected"] == 2 and measure["rows"]["snapshot"] == 4
|
||||
|
||||
[proposal] = [json.loads(line) for line in (base / "loop-propose" / "proposals.jsonl").read_text().splitlines()]
|
||||
assert proposal["edit"]["rule_id"] == EVAL and (proposal["edit"]["from_status"], proposal["edit"]["to_status"]) == ("enabled", "shadow")
|
||||
assert proposal["provenance"]["counts"] == {"false_alerts": 3, "repos": 2, "agreed_hits": 0}
|
||||
cited = {(e["kind"], e["repo"], e["pin"]) for e in proposal["provenance"]["evidence"]}
|
||||
assert ("alert", "example.com/reserved/lmdeploy", FIXED) in cited, "the fixed-pin alert inside the fix is evidence"
|
||||
assert (base / "loop-propose" / "memory_proposals.jsonl").read_text().count("\n") == 1
|
||||
|
||||
gate = json.loads((base / "loop-improve" / "gate.json").read_text())
|
||||
assert gate["memory_edits"] == [f"rule:{EVAL}:enabled->shadow"] and gate["accepted"] is False and gate["reason"] == "reverted"
|
||||
assert gate["outcome"]["recall_after"] < gate["outcome"]["recall_before"], "demoting the eval rule loses the benchmark's finding"
|
||||
assert not (base / "loop-improve" / "rule_policy.json").exists(), "a ledger only when the gate accepts"
|
||||
assert json.loads((base / "loop-improve" / "journal.json").read_text())[0]["outcome"] == "reverted"
|
||||
assert not (project / "target" / ".openultrasast").exists(), "the repository's own ledger is never touched"
|
||||
|
||||
report = reconciler.attribution("validation-46")
|
||||
for row in report["tasks"]:
|
||||
if row["task"] in (f"{CASE}-alerts", f"{CASE}-features", f"{CASE}-remember", *LOOP_TASKS):
|
||||
assert row["model"] is None and row["usd"] in (0, None) and row["calls"] in (0, None), row
|
||||
ingested = FileStore(store.root)
|
||||
outcomes = ingested.rows(kind="proposal_outcome")
|
||||
assert {r.row["outcome"] for r in outcomes} == {"reverted"} and len(outcomes) == 2, "one per (repo, pin) of the evidence"
|
||||
[fixed_alert] = ingested.rows(repo="example.com/reserved/lmdeploy", pin=FIXED, kind="alert")
|
||||
features = {r.row["candidate"]: r.row for r in ingested.rows(kind="features")}
|
||||
assert set(features) == {"lmdeploy/vl/media/connection.py::_load_http_url"}, "no fixed-pin alert became a candidate"
|
||||
assert all(r["pin"] == VULNERABLE and "in_fix_range" not in json.dumps(r["x"]) for r in features.values())
|
||||
assert fixed_alert.row["in_fix_range"] is True and fixed_alert.row["function"] == "load"
|
||||
@@ -64,32 +64,6 @@ def test_the_removed_modules_are_gone_from_the_tree() -> None:
|
||||
assert not (SRC / relative).exists(), f"{relative} still exists"
|
||||
|
||||
|
||||
def test_the_mechanism_lever_is_gone_from_the_retained_improve_loop() -> None:
|
||||
"""`improve/` survives with its lever removed, not deleted.
|
||||
|
||||
The design's Removed list named `improve/evolve.py` whole, but the file is a mix: the mechanism lever
|
||||
(deleted here) and the deterministic rule-status loop that `ousast improve` still runs from benchmark
|
||||
evidence, which has no LLM in it and is not what this feature supersedes. Deleting the file would have
|
||||
taken a retained capability with it, so the lever is stripped and the loop stays — asserted by symbol so
|
||||
the lever cannot creep back.
|
||||
"""
|
||||
evolve = (SRC / "improve" / "evolve.py").read_text()
|
||||
validator = (SRC / "improve" / "validator.py").read_text()
|
||||
for symbol in (
|
||||
"MechanismEdit",
|
||||
"MechanismStore",
|
||||
"apply_mechanism_edits",
|
||||
"evaluate_mechanism_profiles",
|
||||
"propose_mechanism_edits",
|
||||
"validate_mechanism",
|
||||
"mechanism_store",
|
||||
"mechanism_candidates",
|
||||
):
|
||||
assert symbol not in evolve, f"{symbol} survives in improve/evolve.py"
|
||||
assert symbol not in validator, f"{symbol} survives in improve/validator.py"
|
||||
assert "def run_improvement(" in evolve and "def propose_status_edits(" in evolve, "the rule-status loop must survive"
|
||||
|
||||
|
||||
def test_the_re_homed_modules_are_where_the_model_layer_expects_them() -> None:
|
||||
for relative in ("model/taxonomy.py", "model/candidates.py", "model/endpoint.py", "model/specs.py"):
|
||||
assert (SRC / relative).is_file(), f"{relative} is missing"
|
||||
|
||||
Reference in New Issue
Block a user