docs(sessions): collect overnight reviews, spikes, and test evidence - #6505
docs(sessions): collect overnight reviews, spikes, and test evidence#6505mmabrouk wants to merge 32 commits into
Conversation
Two reviews of the session-control RFC from 2026-09-02, a merged review with the decisions Mahmoud must make, and the protocol log of the overnight work. Claude-Session: https://claude.ai/code/session_01GAqSs7fw6QRi2n1ZJ2tmAV
…ation plan Copies of the four spike documents and five slice documents from their branches, the implementation plan for version one, and the updated reading order and status. Claude-Session: https://claude.ai/code/session_01GAqSs7fw6QRi2n1ZJ2tmAV
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. 🗂️ Base branches to auto review (1)
Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Repository YAML (base), Organization UI (inherited) Review profile: CHILL Plan: Team Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
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. Comment |
Takes the RFC branch's rewritten README, status, decisions, plan, requirements, and rfc as they are, and keeps only a pointer to the evidence files on this branch. Claude-Session: https://claude.ai/code/session_01GAqSs7fw6QRi2n1ZJ2tmAV
…p after a durable Stop Claude-Session: https://claude.ai/code/session_01GAqSs7fw6QRi2n1ZJ2tmAV
Agent-generated, low weight. Draft. Docs and evidence only.
Two rounds of evidence on the session-control RFC (PR #6495), on one branch.
Round two, 2026-09-03 daytime: the complete report
Start with
docs/design/session-control-and-live-events/evidence-2026-09-03/report.md. It answers the seven scenario groups Mahmoud asked for, with a table of every scenario by provider, harness, commit, result, timing, and the lane report that holds the raw evidence.In one paragraph: warm Stop works on Daytona and on the local provider, with Pi and with Claude Code, during model output, during a long tool, and while an approval waits; the next message continues in the same sandbox and the same native harness session every time (Daytona: Stop answered in 70 to 104 ms, harness cancel 19 to 151 ms; local Claude Code: cancel settled in 111 to 165 ms). The round found and fixed seven defects the overnight run could not see, each on the branch that owns the code: a named Stop refused during an approval, an outcome report losing a race, a Stop after finish evicting the sandbox, the session row keeping
is_running: true(all PR #6503, head38cbc92201); a stopped turn never writing its continuity record and the runner never releasing its owner claim at shutdown, so a restart lost the native session and refused messages for two minutes, plus the Codex shell child left running after a Stop (all PR #6496, head5a10e6b100); late runner output after a watchdog ending, now quarantined at records ingest (PR #6501, head3f25f06d64). A runner killed with SIGKILL still waits for the 120 s owner lease, by design.Lane reports in
evidence-2026-09-03/:daytona-pi-claude.md,local-claude-and-restart.md,child-process-cleanup.md,watchdog-stale-tail.md,post-stop-liveness.md,post-stop-mirror.md,cancel-continuity.md,integration-refresh.md(thirteen of thirteen cells pass on the re-merged branch, head7f2ef31307), andREADME.md(the timeline). How credentials were obtained, without values, is a section of the report: the Daytona key in the session env files cannot manage Secrets, the one in the mainagentacheckout can; the OpenAI and Anthropic keys were loaded from the QA secret files into each test project's vault through the vault endpoint.Decisions the round leaves to Mahmoud are the last section of the report: quarantine versus reject, the Codex reap versus a pin bump,
not_runningversuslostpast teardown, the runner stop grace period, the ungated Claude Code shell tool, and the post-Stop liveness rule to paste into the RFC.Round one, 2026-09-02 overnight
overnight-2026-09-02.mdis the log of the unattended night: every branch, stack, decision made on Mahmoud's behalf, and the revert path for each.review-2026-09-02.mdmerges the two RFC reviews;implementation-plan.mdis the build plan and merge order. The four spike reports and six slice reports are copies from their branches.Code lives on the sibling draft PRs #6496 to #6504 and the integration PR #6506. Nothing targets
mainor a release branch.https://claude.ai/code/session_01GAqSs7fw6QRi2n1ZJ2tmAV