fix(runner): steered sessions keep their workspace mount (pre-existing)

[mmabrouk]

Context

A steered session died on its next turn. Start a long turn, send a second message mid-flight (the steer path), and the third turn on that session failed with Internal error: Path /tmp/agenta/mounts/<id> does not exist. The session stayed broken until its pool entry aged out.

This bug predates the agent-config-editing stack: the identical repro fails against a git archive extraction of release/v0.109.0, and a variant where the displaced turn aborts promptly (as it would after the separate API-side heartbeat fix) still fails. The E2E campaign for the stack found it; the stack did not cause it.

What was broken

A cold turn was invisible to the keep-alive pool while it ran; nothing was inserted until park, after the turn returned. A steer therefore found no pool entry and cold-acquired a second environment for the same session. Both derive the same durable working directory, and mounting is idempotent, so the second adopted the first's live mount. Whichever environment tore down first unmounted and deleted the shared directory under the other. The next turn took the warm path, which never re-mounts, and crashed.

Changes

Two halves, independently tested:

• A running cold turn now occupies its pool slot: SessionPool.reserve() seats the environment as busy before the turn runs, and the park converts the reservation (or an identity-checked evict removes it). A concurrent same-session request then takes the existing supersede branch, which completes the old teardown before the new acquire mounts. The reservation is refusable so the pool's capacity bound holds; on refusal the turn runs unreserved, exactly as before.
• Warm reuse now probes the parked environment's mount for liveness (on both the idle and the approval-resume branches; an approval park holds its environment far longer). A dead mount becomes a mismatch, evicts, and rebuilds cold, which re-mounts. The probe fails open: a probe error reuses rather than sacrificing warm reuse to a flaky check.

Log-visible changes: a steer now logs evict (supersede-busy) instead of miss, and tears the displaced turn's sandbox down mid-flight rather than letting it finish beside the new one. New lines: reserve, evict reason=pre-acquire, mismatch (mount-lost), and mount-probe threw, reusing.

Tests / notes

• 14 new tests in session-steer-mount-loss.test.ts, including the exact production sequence and the approval-park variant. Each half bite-verified independently (reverting the reservation fails 4, the idle probe 1, the approval probe 1, the pre-acquire claim 1).
• Full runner suite on this tree: 116 files, 1959 tests, 0 failures; typecheck clean.
• The API-side half of the steer story (a displaced turn's heartbeat keeps is_current_turn) is a separate bug with its own repro and is not addressed here.

Part of the agent-config-editing stack; targets its stack parent. Reading order bottom up.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview	Aug 7, 2026 10:18am

[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: ddb60e23-36d8-446b-b03f-16072e1b4aa4

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

---

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]

@coderabbitai review

[coderabbitai]

⚠️ Action not completed

Review rate limited.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[github-actions]

Railway Preview Environment

Preview URL	https://gateway-pr-5764.up.railway.app/w
Project	agenta-oss-clone-spike
Image tag	pr-5764-d4099ec
Status	Deployed
Railway logs	Open logs
Workflow logs	View workflow run
Updated at 2026-08-07T10:29:03.426Z

[mmabrouk]

lgtm