[5283] fix(agent): Keep Pi skill paths stable on cold resume

[mmabrouk]

Context

Configured Pi skills currently live under a random per-run agent directory. Teardown deletes that directory, so a cold resume can contain references to skill files that no longer exist.

This PR is stacked on #5292. The lower PR persists Pi transcripts. This PR gives configured skills a stable session-workspace path without changing replay, continuation, token, cache, or Claude behavior.

Changes

Pi now receives one content-addressed skill snapshot at:

<cwd>/agents/skills/<digest>/<skill-name>

The runner computes the digest from skill names, file paths, modes, and contents. It publishes through a unique temporary directory, writes a completion marker last, and reuses only a matching complete snapshot. Existing or older snapshots are never reconciled, replaced, or deleted during acquisition.

The pi-acp patch passes the exact snapshot to Pi with --no-skills --skill <path>. Only the current snapshot is advertised, while old paths remain readable by a cold-resumed transcript. The runner no longer copies configured skills into the temporary local or Daytona Pi agent directory.

This deliberately accepts append-only storage growth until the session workspace is removed. Snapshot immutability is a runner convention because the cwd remains agent-writable.

Live local QA

Verified on the EE deployment at port 8280 against the actual service-targeted subscription sidecar. The live installed pi-acp was checked for PI_CODING_AGENT_SKILL_DIR before the run.

• Turn 1 delivered an inline continuity-probe skill. Pi called read on <cwd>/agents/skills/4d2431.../continuity-probe/SKILL.md, then returned SKILL-COLD-7Q42-OK.
• The same durable cwd contained one Pi transcript, one .agenta-skill-set.json marker, and the skill file under that digest.
• After deleting and recreating the sidecar, the new container had no prior /tmp. Once the existing 120-second owner lease expired, the session loaded natively and returned SKILL-COLD-7Q42-OK PI-COLD-MEM-9Q7R without receiving the earlier message again.
• After the normal 60-second warm-pool eviction, another cold turn remounted the same cwd, logged session/load ... loaded=true, and recalled the same private marker.
• Keeping the skill body unchanged reused the same digest. Changing the body for the final probe created a second digest and left the first snapshot intact, confirming append-only behavior with no reconciliation or deletion.

The first immediate post-restart retry was rejected by the existing local-runner ownership guard. That is expected affinity behavior and is not caused by this PR. A stable AGENTA_RUNNER_REPLICA_ID avoids the lease wait in orchestrated restarts.

Tests

• pnpm exec vitest run --project unit tests/unit/sandbox-agent-pi-assets.test.ts tests/unit/sandbox-agent-orchestration.test.ts: 65 passed.
• pnpm run typecheck: passed.
• pnpm test: 69 files and 1,066 tests passed.
• Live Daytona was not run. Daytona publication, reuse, mismatch failure, and the no-copy-to-private-agent-dir invariant are covered by the orchestration and asset unit tests.

How to review

1. Start with pi-assets.ts for digesting, atomic publication, marker validation, and fail-closed behavior.
2. Read workspace.ts and sandbox_agent.ts for local and Daytona ordering.
3. Check the pi-acp patch for the narrow explicit-skill argument.
4. Finish with the orchestration and asset tests.

Depends on #5292. Closes #5283.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview, Comment	Jul 14, 2026 12:12am

[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: 668c62b5-94f0-443d-9b0f-6d9591c8b33b

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:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/pi-skill-workspace

---

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.}

[mmabrouk]

Local skill-path QA is complete. Pi actually read the inline skill from <cwd>/agents/skills/<digest>/continuity-probe/SKILL.md; the unchanged body reused its digest across the cold container resume; changing the body produced a new digest while preserving the old snapshot. The resumed native transcript recalled the private first-turn marker after both container recreation and normal TTL eviction. Live Daytona was not run; its path is covered by the focused orchestration tests.

[mmabrouk]

🤖 The AI agent says:\n\nFinal verification of the stacked skill-workspace PR:\n\n- Head: 14b09da\n- Base: fix/pi-transcript-persistence at e0593b4\n- The upper delta is exactly the intended seven-file skill-workspace change.\n- The final upper tree is byte-identical to the tree already exercised in live local QA.\n- Live QA confirmed that Pi reads skills from agents/skills/, unchanged skills reuse the digest, changed skills create a new digest without removing the old snapshot, and a real cold container recreation can load the persisted native transcript.\n- GitHub CI is fully green, including runner unit, integration, and acceptance tests, plus API, SDK, services, web, contribution checks, CLA, and Vercel.\n- This PR does not alter replay, warm/cold selection, token use, or prompt-cache behavior.

[mmabrouk]

lgtm