create_pull_request: git am fallback also fails and cancels unrelated safe outputs — needs stronger retry

[IEvangelist]

Problem

A create_pull_request safe-output failed because git am could not apply the generated patch, and the built-in fallback (re-checking out the original base commit and re-running git am) also failed. The job then exited with a hard error and the second safe-output message (notify_source_pr) was cancelled — even though it was a completely independent operation.

The agent's tokens/context spend were already paid by the time this failure happened in the post-processing step, so a transient failure here is especially expensive.

Failing run

https://github.com/microsoft/aspire/actions/runs/26078776570/job/76676900386

The agent ran in microsoft/aspire and targeted a PR in microsoft/aspire.dev (cross-repo create-pull-request against release/13.4).

Observed behavior

The fallback path that's supposed to "save" a failed git am by recreating the branch at the patch's original base commit also failed:

Attempting fallback: create PR branch at original base commit...
Original base commit from patch generation: 1cb508fd49bedc4afeb0b8fc008d51689756d853
...
Switched to a new branch 'docs/pr-17234-eager-config-migration-33637d63a4a37bea'
Created branch ... at original base commit 1cb508fd49bedc4afeb0b8fc008d51689756d853
/usr/bin/git am /tmp/gh-aw/aw-microsoft-aspire.dev-docs-pr-17234-eager-config-migration.patch
error: patch failed: src/frontend/config/sidebar/docs.topics.ts:867
error: src/frontend/config/sidebar/docs.topics.ts: patch does not apply
error: patch failed: src/frontend/scripts/check-data-files.mjs:21
error: src/frontend/scripts/check-data-files.mjs: patch does not apply
error: src/frontend/src/content/docs/app-host/hot-reload-and-watch.mdx: already exists in index
error: patch failed: src/frontend/src/content/docs/dashboard/index.mdx:17
error: src/frontend/src/content/docs/dashboard/index.mdx: patch does not apply
... (many more files) ...
error: patch failed: src/frontend/src/data/aspire-integrations.json:2
error: src/frontend/src/data/aspire-integrations.json: patch does not apply
error: patch failed: src/frontend/src/data/github-stats.json:1
error: src/frontend/src/data/github-stats.json: patch does not apply
error: patch failed: src/frontend/tests/e2e/ui-regressions.spec.ts:266
error: src/frontend/tests/e2e/ui-regressions.spec.ts: patch does not apply
Applying: Merge main into release/13.4 (#893)
Patch failed at 0001 Merge main into release/13.4 (#893)
...
Warning: Fallback to original base commit failed: The process '/usr/bin/git' failed with exit code 128
Error: ✗ Message 1 (create_pull_request) failed: Failed to apply patch
Warning: ⚠️ Code push operation 'create_pull_request' failed — remaining safe outputs will be cancelled
⏭ Message 2 (notify_source_pr) cancelled — Cancelled: code push operation failed (create_pull_request: Failed to apply patch)

Two distinct problems are visible:

1. Patch contains a mix of "modify existing file" hunks and one "create new file" hunk (hot-reload-and-watch.mdx: already exists in index) — even at the original base commit 1cb508fd…. This smells like the cross-repo / different-tree patch-generation issue from Cross-repo safe-output PRs fail: patch generated as "create new file" when target file already exists #17969, but it surfaces here as the fallback also failing rather than the initial git am.
2. Unrelated safe outputs are cancelled because one code-push message failed. notify_source_pr has no dependency on create_pull_request succeeding, but it never runs.

Expected behavior

When a git am patch application fails after the agent has already finished its work, the safe-outputs runner should be more resilient:

1. Stronger retry / repair before giving up. Cheap recovery attempts that don't require re-prompting the model should be exhausted first, e.g.:
   • git am --3way (in addition to plain git am).
   • git apply --3way --reject followed by git am --continue after staging the clean hunks.
   • For files where the only failure is already exists in index, fall back to applying the hunks as a modify against the file already present in the tree (re-derive a "modify" diff from the patch's + body, using the existing file as base).
   • For cross-repo targets, regenerate the patch against the target repo's tree (the long-standing root cause from Cross-repo safe-output PRs fail: patch generated as "create new file" when target file already exists #17969) — even as a one-shot fix-up after git am fails.
2. Don't cancel independent safe outputs. A failure in create_pull_request should not implicitly cancel notify_source_pr (or any other message that doesn't depend on the PR being created). Either declare dependencies explicitly, or have non-code-push messages keep running.
3. Optionally, push the partial result as a draft PR / artifact so a human can finish the apply with a clear diff in front of them, instead of throwing the whole agent run away.

Why this matters

The agent already spent its context window and tokens producing this output. A git am mismatch in the post-processing step is a very cheap problem (no model calls needed to retry with --3way, no model calls needed to regenerate the patch against the target tree). It's the worst possible place to bail out with no retry, because re-running the workflow means re-spending all those tokens.

Tool-like / mechanical steps that run after the model has finished should have aggressive retry logic precisely because the expensive part of the run has already been paid for.

Possibly related

• Cross-repo safe-output PRs fail: patch generated as "create new file" when target file already exists #17969 — Cross-repo safe-output PRs fail: patch generated as "create new file" when target file already exists. The already exists in index line above looks like the same underlying patch-generation bug, but here it manifests in the fallback path too.

Environment

• Repo: microsoft/aspire (source) → microsoft/aspire.dev (target)
• Base branch: release/13.4
• Failing job: https://github.com/microsoft/aspire/actions/runs/26078776570/job/76676900386