fix(recovery): an agent step a restart interrupted continues where it stopped (#70) #72

Merged

cmoriarty merged 1 commit from worktree-restart-continues-in-place into main

2026-09-27 01:06:23 -04:00

cmoriarty commented

2026-09-27 01:06:19 -04:00

Owner

Closes #70.

Why. An osfd restart threw away every running agent step's work. On 2026-09-26 at 23:11 a deploy restarted osfd 15 minutes into openspec.apply[1] of run_01M3GB3KN38YT9TZJ7KA7RQ5M0. Recovery saw the session gone (opencode dies with the container), found resume_safe: false, and ruled rollback_retry, so attempt 2 started from attempt 1's entry tree (105ce4f8d9).

What.

"Died mid-turn" is now resume for every agent step. resume and reprompt put continue: true on the ready event, so the next attempt keeps the worktree and is seeded with the last handoff and git status/git diff. That is #60's run-resume machinery, reused unchanged.
A recovery gate answered Resume continues the same way.
These still roll back: a wedged session, a stall redo, Redo from entry, and scripted steps.
resume_safe is retired, since it no longer decides anything. The built-in pipelines and the generated apply chain drop it. The loader no longer requires it, and ignores it where an existing .osf/pipeline.yaml still writes it. rollback_retry stays readable for old logs.
tools/test.sh: the full lane with OSF_URL died on macOS bash 3.2 at the browser step, because of an empty array under set -u.

Tests.

A restart mid-edit continues in place with its handoff, and a stall redo still rolls back. With the continuation disabled, the first of these fails as (2, False, False, '').
Which verdicts continue, by step kind.
Resume and Redo from entry on a recovery gate.
The loader still accepts an old resume_safe.

Lanes.

Fast lane: green.
Full lane: backend 1,799, UI unit 416 and 81 browser specs pass. The two @live follow-the-tail specs failed against a stale local osfd serving its own UI, so they don't test this branch.

Spec: openspec/changes/restart-continues-in-place.

🤖 Generated with Claude Code

Closes #70. **Why.** An osfd restart threw away every running agent step's work. On 2026-09-26 at 23:11 a deploy restarted osfd 15 minutes into `openspec.apply[1]` of `run_01M3GB3KN38YT9TZJ7KA7RQ5M0`. Recovery saw the session gone (opencode dies with the container), found `resume_safe: false`, and ruled `rollback_retry`, so attempt 2 started from attempt 1's entry tree (`105ce4f8d9`). **What.** - "Died mid-turn" is now `resume` for every agent step. `resume` and `reprompt` put `continue: true` on the ready event, so the next attempt keeps the worktree and is seeded with the last handoff and `git status`/`git diff`. That is #60's run-resume machinery, reused unchanged. - A recovery gate answered **Resume** continues the same way. - These still roll back: a wedged session, a stall redo, **Redo from entry**, and scripted steps. - `resume_safe` is retired, since it no longer decides anything. The built-in pipelines and the generated apply chain drop it. The loader no longer requires it, and ignores it where an existing `.osf/pipeline.yaml` still writes it. `rollback_retry` stays readable for old logs. - `tools/test.sh`: the full lane with `OSF_URL` died on macOS bash 3.2 at the browser step, because of an empty array under `set -u`. **Tests.** - A restart mid-edit continues in place with its handoff, and a stall redo still rolls back. With the continuation disabled, the first of these fails as `(2, False, False, '')`. - Which verdicts continue, by step kind. - Resume and Redo from entry on a recovery gate. - The loader still accepts an old `resume_safe`. **Lanes.** - Fast lane: green. - Full lane: backend 1,799, UI unit 416 and 81 browser specs pass. The two `@live` follow-the-tail specs failed against a stale local osfd serving its own UI, so they don't test this branch. Spec: `openspec/changes/restart-continues-in-place`. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

cmoriarty added 1 commit

2026-09-27 01:06:19 -04:00

fix(recovery): an agent step a restart interrupted continues where it stopped (#70 )

deploy / deploy (push) Failing after 35m8s

Details

850dbab5f5

A deploy restarted osfd 15 minutes into openspec.apply[1] of
run_01M3GB3KN38YT9TZJ7KA7RQ5M0. Recovery saw the session gone (opencode dies
with the container), found resume_safe: false, ruled rollback_retry, and
attempt 2 started from attempt 1's entry tree.

- "Died mid-turn" is now resume for every agent step. resume and reprompt
  carry continue: true, so the next attempt keeps the worktree and is seeded
  with the last handoff and git status/diff, as a run-resume is (#60).
- A recovery gate answered Resume continues the same way. A wedged session,
  a stall redo, Redo from entry and scripted steps still roll back.
- resume_safe is retired. Nothing reads it now, so the built-in pipelines and
  the generated apply chain drop it, and the loader no longer requires it and
  ignores it where an older document still writes it. rollback_retry stays
  readable for old logs.
- tools/test.sh: the full lane with OSF_URL died on macOS bash 3.2 at the
  browser step (empty array under set -u).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

cmoriarty merged commit 850dbab5f5 into main

2026-09-27 01:06:23 -04:00

cmoriarty referenced this pull request

2026-09-27 09:27:08 -04:00

A restart of osfd throws away running steps' work: recovery should continue in place, as resume does #70