Retry after rollback-auto never names the attempt-preserve-dirty snapshot, so attempt 2 rebuilds attempt 1 from scratch

[wsquared]

Summary

After a failed dev attempt under worktree isolation, rollback-auto resets the worktree to baseline and attempt-worktree-preserved snapshots the discarded tree to refs/attempt-preserve-dirty/<slug>-<baseline>-<attempt>. The retry launches with a prompt identical to attempt 1 and no reference to that snapshot, so attempt 2 rebuilds everything.

Observed (0.11.1)

Setup: an unattended nightly sweep run with [limits] session_timeout_min = 180, max_dev_attempts = 2 and [scm] isolation = "worktree"; the dev adapter is Claude Code.

dev-decision            attempt 1  action: retry   reason: dev session timeout
rollback-auto           baseline <baseline>  note: reverting tracked changes + run-created untracked files
attempt-worktree-preserved  ref: refs/attempt-preserve-dirty/<run>-<baseline>-1
session-start           <story>-dev-2  prompt: <byte-identical to dev-1's prompt>

tasks/<story>-dev-1/prompt.txt and tasks/<story>-dev-2/prompt.txt are byte-identical (330 bytes). Attempt 1's snapshot held a complete 486-line implementation whose changed test files passed. Attempt 2 spent 20 minutes re-authoring the spec and 52 minutes re-implementing, reached the same verification phase, and timed out there too. Both attempts consumed about 114M raw tokens (from the session transcripts) for a product attempt 1 had already produced.

Proposal

When a retry follows a preserved attempt, name the snapshot in the retry prompt, for example:

A previous attempt of this story was rolled back after <reason>; its worktree is preserved at <ref> (git diff <baseline> <ref> shows what it had). Judge that work against the spec before writing anything new: reuse what passes, discard what does not, and say in your report which you did.

Optionally a [scm] retry_from_preserve = true that checks the snapshot out onto the fresh worktree, leaving the prompt unchanged.

The snapshot already exists, is journalled and is pruned by preserve_keep; only the hand-off to the next attempt is missing. Related: #701 (reverts with no preserve artifact), and #776 (dev.on_timeout = salvage-if-done), which would make the retry unnecessary in the timeout case.

[wsquared]

We have contained this downstream without forking, in case the shape is useful.

Since nothing can rewrite a retry prompt from a plugin hook, the hand-off went into a persistent fact the dev workflow loads (a persistent_facts entry in our bmad-build-auto override) rather than into the harness. At step-01, before writing anything, the session:

1. derives the run id from the .bmad-loop/runs/<run>/worktrees/ segment of its own working directory, and the baseline from git rev-parse --short=8 HEAD, skipping silently when either cannot be derived;
2. lists git for-each-ref refs/attempt-preserve-dirty/ filtered to <run>-<baseline8>-*, taking the newest by committer date when several match;
3. reads git diff <baseline> <ref> --stat, judges the snapshot against the spec, and reuses what passes with git checkout <ref> -- <paths>;
4. names the ref, the paths reused and the paths rejected in its final result.

Two things needed pinning explicitly, both raised in review:

• Reuse is additive. git checkout <ref> -- <paths> restores and overwrites but never replays a deletion, so a deletion the snapshot made has to be re-derived. Left unsaid, an agent assumes the snapshot applies wholesale.
• Reused work is not done work. The deterministic gates still have to pass afterwards, and a failure means discarding the reuse rather than debugging it.

The self-validating part is the ref name. Because it embeds the baseline, matching on <run>-<baseline8>-* stops matching once the session has committed anything, so the lookup is only ever answered against the tree the snapshot was taken from.

This works, but it is a prompt asking an agent to do the right thing, which is weaker than the harness naming the ref in the retry prompt as proposed here. The [scm] retry_from_preserve variant would be stronger again for our use, since it removes the judgement step.

[nidelson]

+1