[chore] Clean up big-agents platform Part I

[junaway]

Context

This is Part I of the platform cleanup for big-agents: the breaking renames and the non-dev wiring that bring the merged agent-runner and durable mounts up across every deployment surface. It also fixes a root-cause defect that made durable mounts fail to sign credentials on any real S3-compatible store.

The store signed scoped runner credentials with S3 STS GetFederationToken. SeaweedFS (the bundled dev store) does not model GetFederationToken: it returns actionless tokens that the data plane then denies, so a git clone into a durable working directory dropped its FUSE channel and surfaced as ENOTCONN. The wiring also predated a second problem. Nothing about the object store was mounts-specific, yet its env namespace, adapter location, and key layout all read as "mounts' private storage."

Changes

STS now uses AssumeRoleWithWebIdentity, the only path SeaweedFS authorizes. The API acts as its own OIDC issuer. It holds an RSA private key, mints a short-lived RS256 web-identity token, and serves the matching public JWKS at /.well-known/jwks.json. SeaweedFS fetches that JWKS to verify the token on each sign. A bucket-wide role grants the store; a per-request inline session policy narrows each credential to one mount prefix, so effective permission is role intersect session. The API fails closed: it refuses to sign when it cannot build a prefix-scoped policy. The signing key is shared across API replicas (env var in prod, a documented local-dev fallback otherwise) so the store's cached JWKS never desyncs.

Before: one GetFederationToken call, no roles, denied on SeaweedFS.

After: AssumeRoleWithWebIdentity against the API's JWKS, scoped to agenta-store/mounts/<project_id>/<mount_id>/*.

geesefs now runs in the foreground as a runner-tracked child (-f). The detached daemon was reparented to init and died under write-heavy load, which was the proximate cause of the ENOTCONN. The mount path also force-unmounts a stale node before remount and verifies the mount is alive with a real I/O probe.

Store generalization (W6). AGENTA_MOUNTS_STORAGE_* becomes AGENTA_STORE_*. The adapter and STS signing move from core/mounts/ to a shared core/store/, and MountStorage becomes ObjectStore. Keys gain a mounts/ top-level prefix so a second store consumer can coexist. No backward-compat shim: nothing was deployed, so every old name is deleted in the same change.

Runner rename (W5). services/agent becomes services/runner and the sandbox-agent service becomes runner across compose, Helm, and Railway. Runner-infra env vars lose the AGENT infix (AGENTA_AGENT_RUNNER_* to AGENTA_RUNNER_*); agent-behaviour vars keep AGENTA_AGENT_*. The MCP flag aligns with the renamed template field (MCP_SERVERS to MCPS). The CI job becomes run-runner-tests.

Non-dev wiring (W4) across every surface. Dev compose bundles SeaweedFS configured for the OIDC IAM engine (pinned image, -volume.max=64, generated iam.json). gh/prod compose and the private platform repo use external S3 (STS native, no bundled store). Helm gets the full OIDC statefulset, JWT issuer and shared key injected into API pods only when the bundled store is enabled, schema entries, and example values. Railway gets a SeaweedFS wrapper image that generates iam.json and the JWT wiring it needs.

Docs (W7). The self-host store reference documents the web-identity mechanism, the complete AGENTA_STORE_* table, and the shared-key-across-replicas requirement. Architecture and networking pages cover the runner sidecar on :8765 and the bundle-or-external store toggle.

Tests / notes

• Verified end to end against a freshly nuked-and-rebuilt EE dev stack: the JWKS endpoint serves the key, signing returns scoped credentials, a write to the mount's own prefix succeeds, and a write to another prefix is denied.
• Runner suite green: 377 unit tests pass, tsc typecheck clean. API ruff format and ruff check clean.
• Helm renders and lints with the bundled store enabled; the new jwtIssuer and jwtPrivateKey values pass schema validation.
• The Railway store wiring is syntax and JSON validated but deploy-untested; there is no Railway environment in this setup. It mirrors the proven compose and Helm config.
• The committed local-dev RSA key and the dummy SeaweedFS signing key are allowlisted in .gitleaks.toml by path and by value, not by exempting whole files.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview, Comment	Jun 30, 2026 6:11pm

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 2f557a28-444d-413f-9e31-fd365f879a05

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• ✅ Review completed - (🔄 Check again to review again)

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch chore/big-agents-platform-integration

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[junaway]

What lands on big-agents with this merge

This is the Part I platform cleanup. It brings the merged agent runner and durable mounts up across every deployment surface and fixes the one defect that kept mounts from working on a real store. The store now signs scoped runner credentials with AssumeRoleWithWebIdentity instead of GetFederationToken, which SeaweedFS does not model. The API became its own OIDC issuer (it mints an RS256 web-identity token and serves the JWKS the store verifies it against), a bucket-wide role is narrowed per request by an inline session policy, and signing fails closed when no prefix-scoped policy can be built. geesefs also runs foreground as a runner-tracked child, which removes the ENOTCONN that a git clone into a durable cwd used to trigger.

Two breaking renames ride along, with no backward-compat because nothing is deployed. The mounts store generalizes into a shared platform store: AGENTA_MOUNTS_STORAGE_* becomes AGENTA_STORE_*, the adapter moves to core/store/ as ObjectStore, and keys gain a mounts/ prefix so a second consumer can coexist. The runner rename moves services/agent to services/runner, renames the sandbox-agent service to runner everywhere, drops the AGENT infix from runner-infra env vars, and aligns the MCP flag to MCPS. Non-dev wiring follows the settled per-tier choice: bundled SeaweedFS on compose-dev / Railway / k8s-when-enabled, external S3 on gh, preview, and live.

Verification ran end to end against a freshly nuked-and-rebuilt EE dev stack: JWKS served, signing returns scoped credentials, a write to the mount's own prefix succeeds, and a write to another prefix is denied. The runner suite is green (377 unit tests, typecheck clean), the API passes ruff, and Helm renders and lints with the bundled store on. One caveat for reviewers: the Railway store wiring is syntax and JSON validated but deploy-untested here, since there is no Railway environment in this setup; it mirrors the proven compose and Helm config. Platform-repo items (records worker, the streams:records exporter entry, the env templates) already landed in the private repo.