Open_Ultra_SAST/ops/ax/README.md

ax on this host

The service plane (.kiro/specs/ai-service-plane/) runs every task on google/ax over Agent Substrate in a single-node kind cluster. ax is the only executor; ousast plane run submits, resumes, starts and collects.

Bring-up

ops/ax/up.sh          # idempotent: kind + registry, Substrate, ax, gVisor worker pool, egress gateway,
                      # receiver Service, runner image (digest pinned), ax CLI, smoke Task
ousast plane doctor   # the profile's cluster (ops/k8s/profiles/kind.toml), Substrate, ax, runner image, memory store
ops/ax/smoke-run.sh   # end to end: repo-facts on ax, artifacts back on the host, attribution table
ops/ax/down.sh        # deletes the cluster and its registry

Tools live in ~/go/bin (kind, ax, ko, kubectl-ate) and ~/.local/bin (kubectl); sources and rendered files in ~/.cache/ousast/ax-src/. The cluster's context, registry and pool sizes come from the kind profile (ops/k8s/profiles/kind.toml); up.sh writes the runner's digest pin to that profile's images file (ops/k8s/profiles/kind-images.json), which the templates are re-pinned from and which is committed with them. up.sh and smoke-run.sh run the checkout's code with OUSAST_PYTHON (default .venv/bin/python).

The kind profile runs on a 7.7 GB server VM. Observability is off by default (kind_observability = false in ops/k8s/profiles/kind.toml). This saves the measured 930 MiB used by otel-system, about 40% of the cluster's 2.3 GB. The smoke Run still completes with those deployments stopped. To enable the stack, run OUSAST_KIND_OBSERVABILITY=1 ops/ax/up.sh or set kind_observability = true in the profile.

up.sh stages copied Substrate manifests in a temporary checkout view and installs through hack/install-ate-kind.sh. The local ops/ax/substrate-kind-lean/ overlay removes the collector, Prometheus, Jaeger and their supporting resources before installation. This preserves the installer's secret, CRD and image setup. The upstream checkout is unchanged. The shared ate-otel-config disables trace, metric and log export and the OTEL SDK; it removes the old collector endpoint and metric timing keys. On an existing cluster, up.sh scales the three observability deployments to zero, replaces the ConfigMap data and rolls its consumers when the configuration changes. Opting in restores the upstream ConfigMap and observability manifests.

How a Run executes

A Run (plane/runs/*.yaml, kind openultrasast.io/v1alpha1 Run) is a DAG of steps; each step names an ax Task (plane/tasks/), its inputs and outputs, and its own budget: {usd, calls}. Every Task binds its Workspaces (plane/workspaces/) and one ax Model (plane/models/): deepseek-flash for chat, openrouter-embedding (OpenAI text-embedding-3-small through OpenRouter) for embeddings. ousast plane run (plane/reconciler.py) applies Workspaces before the Tasks that bind them, resumes each Task, sends its start signal, collects its artifacts and records state under $OUSAST_RESULTS/plane/<run>/ (default ~/ousast-results/plane/); a rerun skips tasks already done, and --workers bounds the tasks in flight.

Runner contract (plane/runner.py, PID 1 of the runner image). ax hands the Task in AX_TASK_YAML and the Workspaces in AX_WORKSPACES_YAML; the runner serves /healthz and /readyz (503 until every workspace is materialised). Agent Substrate boots each template once as a golden actor to snapshot it, so the command never starts by itself: it waits for POST /ousast/v1/start with {run, task, credentials}, which the reconciler sends through Substrate's atenet-router to that actor only (plane/router.py). The provider key exists only in that request: the runner keeps it in memory, passes it to the command's environment, redacts it from echoed stderr and never writes it. The reconciler reads the variable the Model's secretKey.key names from the operator's environment or .env (which never overrides an exported variable). After the command exits, the runner posts OUSAST_OUTPUT_DIR as a tar to the receiver; that delivery is the task's completion, and a completion marker stops a resumed actor from repeating billed work.

Budgets. The metered client refuses the next call once a task's usd or calls ceiling is reached; the task ends unfinished and resumes on a rerun with a larger budget. An account or authentication error ends it failed. Model-free tasks run with {usd: 0, calls: 0}, so a stray call fails loudly.

Egress (plane/egress.py). One EgressPolicy per actor through the atenet-egress gateway, deny by default, hostnames only: plain HTTP to the artifact receiver, TLS passthrough to the source hosts of the bound Workspaces and to the hosts the bound Model declares (openultrasast.io/egress-hosts). The policy is deleted with the actor.

Token attribution. ousast plane status <run> prints per task its status and calls, prompt, cache_hit and output tokens and usd from the tasks' summary.json, and writes <run dir>/attribution.json; --units adds per-unit rows.

Tasks

Task	Model	Does
repo-facts	none	functions per product file, and each candidate's call sites in other files
verify (passes a, b, c)	deepseek-flash	batched tool hunt per file over the candidates, with their known callers
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
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); 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.

Memory store

plane/memory.py. One store keyed by repository and pin holds what runs learned: index.jsonl, facts/ by content hash, repos/<host>__<owner>__<name>/<pin>.jsonl rows, and content-addressed blobs (excerpts, embeddings, cached model responses, compiled programs). ousast plane remember <run> ingests a run's rows; a repeated ingest is skipped by the index. A facts entry is reused for the same repository, pin, candidates and runner image.

OUSAST_MEMORY selects the backend:

• file:///path (FileStore), default $OUSAST_RESULTS/plane/memory; it refuses to write below 1 GiB free;
• s3://<bucket>[/<prefix>] (S3Store, the s3 extra: boto3 against any S3-compatible server; RustFS is the 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.

Measured footprint (2026-09-29, idle, two workers)

Item	Value
kind node container memory	1.7 GiB (27% of the 7.7 GiB host)
registry container memory	34 MiB
node volume (images, containerd)	4.9 GB on disk
worker pool	2 gVisor workers, 1 CPU / 1.5 GiB limit each

Disk is the binding constraint: the Substrate install and ko builds need several GB of Go caches. Clear them afterwards (go clean -cache -modcache); rerunning the full Substrate install on a near-full disk fails in ko resolve.

What this host taught (each one cost a failed live run)

• Substrate installs no WorkerPool ("no free workers") -> workerpool.yaml.tmpl.
• Actor images must be pinned by digest; ax's default runner image on gcr.io needs credentials.
• A release ax CLI skews from the server; build it from the deployed checkout.
• ax's API rejects unknown fields; metadata is name + atespace; Git entries have no commit.
• ax apply leaves a Task Suspended until ax resume; the first resume may time out while Substrate builds the template's golden snapshot. ax has no Completed phase: completion is our artifact delivery.
• Substrate boots each template once as a golden actor with the task's env, so the runner waits for a start request sent through atenet-router; the provider key travels only in that request.
• Actor egress goes through the atenet-egress gateway, deny by default, one EgressPolicy per actor, hostnames only; the gateway dials the address the actor connected to, so the receiver is a Service dialled by ClusterIP with its name as Host. The Envoy gateway needs a Rust build; the agentgateway variant needs none.
• Apply a Task after the Workspaces it binds; ax copies them at create time.
• Every Task is its own actor template with a golden actor (~24 MB under /var/lib/ate/actors on the node). Deleting them failed while the snapshot bucket ax's templates name (dberkov-gke-dev3, from AX_SNAPSHOTS_BUCKET) did not exist in rustfs: golden actors piled up in DELETING and filled the disk during a 600-task harvest (2026-10-01). Create that bucket in rustfs (an aws-cli pod with the rustfs-bucket-init env); even then a deleted actor's directory stays on the node, so long Runs need a janitor that removes directories no live actor owns.
• A Workspace's files reach the actor inline in one environment variable: above ~20 KB of content the template fails with "actor template not found" (an 86 KB and a 31 KB excerpt did; 7 KB ran).

Memory store on S3 (RustFS is the tested server)

OUSAST_MEMORY=s3://<bucket>[/<prefix>] (or s3:// for the bucket in S3_BUCKET) puts the plane's memory store (plane/memory.py) in an S3-compatible bucket through boto3 (the s3 extra). Any store with versioning, lifecycle rules, object tags and S3 Select works; RustFS is what the contract tests run against. Endpoint and credentials come from .env (S3_ENDPOINT as a full URL, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, optional AWS_SESSION_TOKEN and S3_REGION, which avoids a GetBucketLocation call) and are never printed.

The store never configures its bucket. An admin sets it up once. Each time an S3Store is opened it checks the setup (S3Store.verify_bucket), using reads plus one small probe object at <prefix>/_probe/select.jsonl. If anything is missing, it refuses to start with a MemoryStoreError that lists each missing piece and the admin command that fixes it. Nothing is skipped quietly. The store needs:

• versioning Enabled, because provenance cites object version ids;
• an enabled lifecycle rule expiring <prefix>/runs/ (raw run outputs) after a number of days, 30 by default. The rule must be a plain prefix filter, and no rule may expire the whole store (repos/, facts/, index.jsonl and the blobs are kept);
• S3 Select (SelectObjectContent over JSON Lines). Every filtered read is pushed down to the server, and there is no fetch-and-filter fallback. A server without Select is refused;
• object tags readable, because rows are filtered by their kind tag.

One-time admin setup (admin credentials, bucket sast-memory, store at the bucket root):

aws s3api put-bucket-versioning --endpoint-url "$S3_ENDPOINT" --bucket sast-memory \
    --versioning-configuration Status=Enabled
aws s3api put-bucket-lifecycle-configuration --endpoint-url "$S3_ENDPOINT" --bucket sast-memory \
    --lifecycle-configuration '{"Rules":[{"ID":"ousast-runs-expiry","Status":"Enabled",
      "Filter":{"Prefix":"runs/"},"Expiration":{"Days":30}}]}'

put-bucket-lifecycle-configuration replaces every rule on the bucket, so merge this rule with any rules already there. The agent account's policy can read the bucket's configuration but not change it. It reads and writes objects and their tags, and it gets no Put* on versioning or lifecycle:

{"Version": "2012-10-17", "Statement": [
  {"Effect": "Allow", "Action": ["s3:ListBucket", "s3:GetBucketLocation", "s3:GetBucketVersioning",
                                 "s3:GetLifecycleConfiguration"],
   "Resource": ["arn:aws:s3:::sast-memory"]},
  {"Effect": "Allow", "Action": ["s3:GetObject", "s3:GetObjectVersion", "s3:PutObject", "s3:DeleteObject",
                                 "s3:GetObjectTagging", "s3:PutObjectTagging", "s3:GetObjectVersionTagging"],
   "Resource": ["arn:aws:s3:::sast-memory/*"]}]}

Measured state on 2026-10-02 (agent sast-memory-agent): versioning was Enabled and a 30-day rule on runs/ was in place. verify_bucket first refused because GetObjectTagging returned AccessDenied for the agent; with the three tagging permissions above (s3:GetObjectTagging, s3:PutObjectTagging, s3:GetObjectVersionTagging) the real-server contract tests pass.

Fixed schema: every row carries every queryable field. RustFS's Select infers an object's JSON schema from its first 1000 rows (measured 2026-10-02). Two failures follow:

• a where field missing from those rows, even if row 5000 has it, fails with EvaluatorBindingDoesNotExist;
• a column that is null in all of those rows and set in a later one fails the whole object with JSONParsingError, whatever the where names (999 leading nulls answer, 1000 fail). A column holding two JSON types (a string, then a number) fails the same way at any size.

So plane/memory.py declares per kind the fields a where may name (QUERY_FIELDS, besides the fields every row has). They are strings, and the store writes all of them on every row at write time, "" where one does not apply (never null). rows() refuses an unknown kind, a where on a field no kind declares, and where = null, before any read: a new query declares its field first. Rows written before the rule are rewritten once with ousast plane memory-normalise [--store URL] [--dry-run], which is idempotent and prints its counts; on a bucket it writes new object versions, and versioning keeps the old ones. Run it when a store moves to RustFS. The store never answers "no rows" for a failed Select; it raises a MemoryStoreError naming the object and the cause.

The real-server contract tests (OUSAST_MEMORY_TEST_S3=1, bucket OUSAST_MEMORY_TEST_BUCKET, else S3_BUCKET) verify the bucket at its root and write only under a fresh contract-<id>/ prefix. They never configure the bucket.

Moving to a separate Kubernetes cluster (planned, not done)

No deployment outside this host's kind cluster has been made. The step-by-step guidance, with what is implemented and what is planned, is docs/deployment.md (the site page "Deployment"). The Model and Workspace manifests under plane/ move as they are and the Task templates need their image reference re-pinned to your registry; these parts assume this server VM and must change first (checked 2026-10-02):

Assumption	Where	Needed in a real cluster
Artifact receiver runs on the developer host, reached through the ousast-receiver Service with an EndpointSlice to the kind network's gateway address on OUSAST_ARTIFACT_PORT	egress.py, receiver-service.yaml.tmpl	the receiver as an in-cluster Deployment (or object storage, e.g. Substrate's S3-compatible store), with the reconciler reading from it
The kind-local registry (registry in ops/k8s/profiles/kind.toml), rewritten by Substrate for kind	up.sh (KO_DOCKER_REPO), the profile's images file, the image of every template under plane/tasks/	the k3s profile's registry; keep digest pins (Substrate rejects tags); --runner-image FILE re-pins generated Tasks
kubectl context	kube_context in the profile, read by doctor.py, egress.py, router.py (plane-on-kubernetes 1.2); up.sh reads the same profile	the k3s profile's context; nothing else changes
ax's snapshot bucket: AX_SNAPSHOTS_BUCKET in ax's deploy/ax-server.yaml points at the ax authors' GCS bucket	ax deploy manifest	your own bucket, set before deploying ax
Egress gateway applied by hand (agentgateway variant, no Rust build)	this README	the Substrate-installed gateway; per-task EgressPolicies work unchanged
Worker pool of 2 x 1 CPU / 1.5 GiB for the 7 GB host	workerpool.yaml.tmpl	sized to the cluster; --workers to match
Plane memory store and results under ~/ousast-results/ on the host	reconciler.py (OUSAST_RESULTS), memory.py (OUSAST_MEMORY)	a persistent volume or bucket shared by the reconciler; the s3:// store already works against any reachable S3-compatible server

The provider key already travels only in the start request through atenet-router, which works the same through the kube-context tunnel to any cluster.

What a larger ax deployment needs

A Kubernetes cluster with Agent Substrate (and its egress gateway), the ax control plane with AX_SNAPSHOTS_BUCKET pointing at your own bucket, a registry the workers can pull from, the runner image pinned by digest, the receiver reachable as a Service, and the provider host allowed per task (which each task's EgressPolicy already does). The Model and Workspace manifests under plane/ move unchanged; the Task templates get a new image reference; the kube context and doctor's registry check are not configurable yet (the table above). Step by step: docs/deployment.md.

Search executor and verifier task contract (operator verified 2026-10-05)

There are two workers and no queue. no free workers available is capacity back-pressure: the dispatcher retries the same apply with 15–30 seconds of jitter, up to the workload deadline. It does not retry a sandbox execution or route to the VM because of this response. A shared reservation limits all AX workloads in one host process to two tasks; separate host processes coordinate through AX's rejection.

Search executors build on the disk-backed /workspace, export a single archive, and are deleted before verifier repetitions start. The archive contains checkout/, products/, and spec.json with a none build recipe. A verifier receives only VERIFY_INPUT_URL and RESULT_URL; it never installs dependencies or fetches source. The executor's host-scoped PREPARED_PUT_URL carries the built output. The host checks its SHA-256 before passing it to fresh verifier tasks. Each side has three fresh runs.

Verifier egress to files.because-security.com:443 is requested but pending. Until the operator enables it, live verifier input/output is blocked. The first input download tolerates up to 60 seconds of delayed policy activation.

/ and /tmp consume the worker's 3 GiB RAM; /workspace has about 2 GiB usable disk. Entrypoints set TMPDIR=/workspace/tmp, npm_config_cache=/workspace/.npm, MAVEN_OPTS=-Dmaven.repo.local=/workspace/.m2, COMPOSER_HOME=/workspace/.composer, PIP_CACHE_DIR=/workspace/.pip, and GRADLE_USER_HOME=/workspace/.gradle. OUSAST_SCRATCH_BYTES overrides the default 2147483648-byte workspace guard. The guard includes caches and open output logs, checks running children and completed writes, and reports could_not_build / scratch limit. It is a monitored guard, not a hard filesystem quota; programs that hardcode /tmp still consume the worker's RAM.

Repository clones must use HTTPS git clone https://github.com/...git or codeload archives, never api.github.com: tasks share one public IP. Search task inputs are presigned checkout archives, so these entrypoints do not clone repositories themselves.

Image layers, publication and node cache budget

The task family has one shared runtime base, ousast-task-base, built from the public AX runner v0.3.1. base-browser adds Chromium to that exact published base (by digest); engine and replay do not carry Chromium. The VM/kind openultrasast image uses the same base and Joern recipe. Replay inherits the published engine digest in CI, so even a cold build cannot accidentally duplicate its Joern layers.

Java is Temurin 21 JDK, copied from eclipse-temurin:21-jdk-jammy. ops/frontend-retention/install.py explicitly requires Java 21 and invokes java -cp ... dotty.tools.dotc.Main using Joern's bundled Scala compiler. The shared JDK also provides javac for Maven/Gradle project builds. Curl/unzip and retention build scratch stay in a discarded Joern build stage. Joern stays above the base in the engine image (also used by replay and the VM lane); its compiler jars and retention provenance must not be blindly pruned. Runtime dependencies from pyproject.toml (core plus semantic/s3 for the VM lane) precede project code. Package installation uses --no-deps --no-cache-dir; its isolated packaging backend is temporary.

The base retains upstream Python 3.12 with venv/pip. Its explicit apt package list is nodejs npm node-typescript php-cli composer maven sqlite3 bubblewrap util-linux git ca-certificates libstdc++6 zlib1g. Composer is the trixie apt package. Chromium is installed only in base-browser. Apt caches/lists and docs/man/info are cleaned; dpkg excludes new docs/man/info. Files already in inherited AX layers still occupy space even if deleted.

Executor preparation supports pip, npm, Composer, Maven and Gradle builds. There is no installed Gradle binary: the executor invokes the selected project's gradlew through /bin/sh, including for wrappers without executable permissions. Gradle-wrapper distribution downloads and dependency downloads go through public HTTPS from the executor, subject to its egress policy and deadline. The verifier receives prepared products and performs no dependency downloads.

Rebuild policy:

• Dependency/toolchain changes: run Task base image (also triggered by changes to plane/Dockerfile.base, pyproject.toml, or uv.lock). It publishes base-<short sha> and base-content-<input hash> tags, plus their -browser variants, and prints immutable digest references. Content tags identify build inputs; they do not claim reproducible upstream apt/PyPI resolution. The lockfile triggers refreshes but pip installs the declared runtime constraints, not a uv frozen environment. Deliberate upstream refreshes use manual dispatch.
• After both targets publish, ensure ousast-task-base is public in GHCR's package settings. Copy the workflow summary's JSON into plane/task-base.digest and submit that pin update for normal review. CI does not commit on your behalf. The initial null pins mean unpublished, not a usable bootstrap image.
• A pin update or plane/Dockerfile.*/root Dockerfile change on main runs Engine image. For source, entrypoint or retention-installer changes, dispatch it once for the desired revision, or push a v* release tag. Source-only merges no longer publish images automatically. Never rebuild per task/run. All builds use scoped GHA caches; pin checks precede builds and require anonymous access plus identical base-layer ancestry for the browser variant.
• Final package names must also be public in GHCR. Publication checks anonymous access after logout and fails if the package is private. Set visibility on first publication and rerun as needed. Deploy only the reported image@sha256:... references in the cluster profile/Task manifests, never main or a release tag. Publication does not automatically change deployed Task pins.

For a local task build, pass TASK_BASE_TAG and TASK_BASE_DIGEST from the base pin; search instead takes TASK_BROWSER_TAG and TASK_BROWSER_DIGEST. Replay's ENGINE_REF defaults to its local engine stage; set it to the published engine reference to reuse exactly those layers. When building base-browser locally, pass TASK_BASE_REF=<full base tag@digest> for the same ancestry as CI.

The operator's capacity rule is all images in rotation, unpacked, together at most approximately 9 GiB, including rollback revisions. Working builds take priority over minimizing an individual executor image; its previous 2.13 GB unpacked (~1.98 GiB) was acceptable. Retire unused Task templates and image revisions before rotating a large base/Joern generation. Shared layers occupy storage once where the runtime deduplicates them, but do not assume that turns a sum above 9 GiB into an acceptable rotation. Record both the sum of unpacked image sizes and actual unique-layer storage. Never prune images used by live tasks.

Expected footprint (planning estimates only; no images built for this change):

Image	Expected unpacked size	Basis / uncertainty
Shared base	about 1.3–1.9 GiB	~454 MB inherited AX layers + ~281 MB JDK baseline + npm/Composer/Maven, language runtimes and venv/extras; apt dependency closure unmeasured
Browser base	about 1.65–2.45 GiB	Shared base + roughly 350–550 MB Chromium/dependencies
Search executor/verifier	about 1.7–2.5 GiB	Browser base + application; previous 2.13 GB unpacked is a baseline, not a ceiling
Engine	roughly 3.5–6 GiB	Restored shared base + Joern; the supplied 2.16 GB compressed engine cannot establish its unpacked size
Replay	engine + a few KiB	Same engine digest plus replay entrypoint
VM/kind openultrasast	roughly 3.5–6 GiB plus benchmark assets	Same base/Joern, runtime extras shared, benchmark payload retained

An engine + search rotation therefore sums to about 5.2–8.5 GiB before rollback images. Adding replay as another unpacked image brings the sum to about 8.7–14.5 GiB; this rotation is not demonstrated to fit 9 GiB. Engine/replay share almost all layers, but the operator must confirm the accounting and measured footprint before selecting a rotation. Extra revisions and the VM image add to the budget if retained there.

The previous search history included a 1.37 GB apt layer and 281 MB JDK; no per-package apt breakdown is available. Removing system Gradle and docs/caches may save space, while retained build tools and shared Python extras cost space. These ranges are not measured savings or a capacity guarantee. After publication, measure unpacked images and unique layers for the entire intended rotation, and verify build recipes, PHP/JavaScript retention, runner startup and Chromium before rollout. There is no separate <1.5 GiB executor gate.