fix(#1184): strip Signed-off-by trailers instead of rejecting

[fullsend-ai-coder]

Closes #1184

post-code.sh §3b, post-fix.sh §1b and the precommit-gate re-check treated a Signed-off-by: trailer on an agent commit as fatal: the run was discarded, nothing pushed. The trailer was the only thing wrong with the commits, and one such run cost $14.29 (fullsend-ai/fullsend#7031). This PR strips the trailer and continues, failing only if the rewrite itself cannot run or a trailer survives it.

What changed

Scoped to agent-authored commits. On the post-fix rebase path SCAN_RANGE widens to merge-base and can cover a human's commits on the PR branch. A human's sign-off is a DCO attestation the DCO app requires (it waives bot authors only), so those commits pass through untouched — matched on author, because a rebase re-stamps the committer onto commits the human wrote.

The rewrite preserves everything but the trailer. Author and committer identity and dates, commit count and order, the range base, and the rest of the message. Multi-commit ranges go through git filter-branch --msg-filter; a single commit through git commit --amend --only, with the original committer carried across so the two paths agree.

Detection on %B. A trailer on line 2 with no blank line is folded into the subject and invisible to %b; the old check would have shipped it.

Fails closed, under an honest category. Anything that stops the rewrite from running — unstaged changes, a dirty index (filter-branch refuses both), a missing agent identity — reports as Signed-off-by strip failed (signoff-rewrite-failed). Signed-off-by rejected (signed-off-by) now means only that a trailer survived a rewrite that reported success.

The validation loop no longer burns an iteration on it. validate-code-output.sh is the validation_loop.script for both harnesses and still hard-failed on a trailer; that was the half of #1184's cost the strip alone did not address (the issue's evidence has the trailer surfacing on iteration 2). It soft-passes now, in the same bucket as secret-scan — a rewrite there is impossible anyway, TARGET_REPO_DIR is an extracted copy. This is a fourth site beyond the three the issue names; included because the issue's purpose is "stop discarding runs for the trailer", and this site discards iterations for it.

Recorded where a human will see it. PR body on creation; a comment on the existing-PR path (which exits before the body is built); the post-fix summary comment.

The logic lives in three helpers in scripts/lib/precommit-gate.lib.sh, which all four sites already source, so the copies cannot drift.

The SKILL.md prohibition is unchanged — agents still must not sign off. The guard is the safety net, not the enforcement.

Testing

All under podman (debian:bookworm-slim + git, jq, python3, python3-jsonschema) from the committed HEAD. macOS cannot run these suites — both die on a pre-existing BSD-sed incompatibility, identical on main.

$ make check-bundle
check-bundle OK

$ bash scripts/signoff-strip-test.sh          # new, real git
signoff-strip-test        EXIT=0 PASS=54  FAIL=0
$ bash scripts/post-code-test.sh
post-code-test            EXIT=0 PASS=202 FAIL=0    (main: 208/0 — six stub cases removed)
$ bash scripts/post-fix-test.sh
post-fix-test             EXIT=0 PASS=56  FAIL=0    (main: 56/0)
$ bash scripts/validate-code-output-test.sh
validate-code-output-test EXIT=0 PASS=24  FAIL=0    (main: 24/0; signoff case flipped to soft-pass)
$ bash scripts/post-failure-report-test.sh
post-failure-report-test  EXIT=0 PASS=37  FAIL=0
$ python3 scripts/process-fix-result-test.py
Ran 40 tests — OK  (EXIT=0)
$ make lint
skillsaw 0.18.0 --strict: Errors 0, Warnings 0, Grade A
$ pre-commit run --all-files
all hooks Passed (shellcheck, end-of-file-fixer, gitleaks, actionlint, pinact, …)

The original PR's run_signoff_test cases were string-matching stubs — detect_signed_off_by was reimplemented in the test file and never touched git, so signoff-multi-commit-stripped passed a single string. They are replaced by scripts/signoff-strip-test.sh, which drives the real helpers against real repositories and asserts each of the following:

Path	Asserts
Single bot commit	trailer gone; subject and body intact; author+committer identity and dates unchanged; count unchanged
Multi-commit (3, trailers in 2)	both stripped, third untouched; order, subjects, identities, dates unchanged; range base SHA and tree untouched; mid-body prose around the trailer preserved
Idempotence	second pass exit 0, SHAs unchanged
Human commit in range	sign-off survives, SHA unchanged; bot trailer still removed
Human commit rebased by the agent (committer = bot, author = human)	sign-off survives; bot trailer removed
GPG-signed human commit below the agent tip	gpgsig header and SHA unchanged; bot trailers above it removed; count unchanged
Agent trailer below a human commit	exit 1, diagnostic; nothing rewritten
Trailer glued to the subject (line 2, no blank line)	invisible to %b, detected, stripped, subject preserved
Staged-but-uncommitted file, single commit	not swept into the commit, still staged, trailer removed
Unstaged tracked change	exit 1, diagnostic names "unstaged changes"; an untracked file does not block
Staged change, multi-commit	exit 1, diagnostic names "staged changes"
No agent identity	exit 1, diagnostic names GIT_BOT_EMAIL; human sign-off untouched
Clean range	count 0, present false

End-to-end shape of the rebase case, from a fixture through the real helpers (123+fullsend-ai-coder[bot]@… is the agent):

--- before ---
6f56c9e 123+fullsend-ai-coder[bot]@… | fix: agent change | signoff=fullsend-ai-coder <123+…>
c07aea4 human@example.com            | feat: human work  | signoff=Real Human <human@example.com>
--- after (exit 0) ---
d72d44c 123+fullsend-ai-coder[bot]@… | fix: agent change | signoff=
c07aea4 human@example.com            | feat: human work  | signoff=Real Human <human@example.com>

Before this scoping, the same fixture stripped the human's trailer and the push fell back to --force-with-lease.

Known limits

• Scoping is by author identity, not provenance. An agent that commits with --author=<a human> and appends that human's sign-off is out of scope and the trailer ships; before this PR any trailer killed the run, which incidentally blocked that. Authorship forgery itself is pre-existing and the SKILL.md prohibition plus human merge review remain the controls. Provenance scoping ("commits this run created") is not available on the rebase path, which is exactly where the widening happens.
• The PR-body checkbox line in post-code is wired but not integration-tested: post-code-test.sh's build_pr_body is a local reimplementation, and a case there would test the copy rather than the script.
• Lowercase signed-off-by: is intentionally not detected (pre-existing, case-sensitive contract). Under the new semantics that means it ships and fails gitlint rather than being rejected.
• filter-branch leaves a trailing blank line that --amend -F (--cleanup=whitespace) does not; harmless for gitlint.
• refs/original/ is left behind by filter-branch -f. Nothing in either script traverses all refs and both pushes are single-refspec, so it stays local to the ephemeral runner.
• The precommit-gate re-check amends HEAD only. §3b has already cleaned the rest of the range by the time it runs, and it fails closed if that ever stops being true.

[fullsend-ai-review]

🤖 Review · ⚠️ Cancelled · Started 2:21 AM UTC · Ended 2:52 AM UTC

Commit: f359597 · View workflow run →

[waynesun09]

/fs-fix-stop

[github-actions]

Fix agent disabled for this PR. Remove the fullsend-no-fix label or use /fs-fix to re-engage.

[fullsend-ai-review]

🤖 Review · ⚠️ Cancelled · Started 2:53 AM UTC · Ended 3:16 AM UTC

Commit: b9cced1 · View workflow run →

[fullsend-ai-review]

🤖 Review · ⚠️ Cancelled · Started 3:18 AM UTC · Ended 3:26 AM UTC

Commit: cbd8c10 · View workflow run →

[fullsend-ai-review]

🤖 Finished Review · ❌ Failure (validation failed after 2 iteration(s)) · Started 3:27 AM UTC · Completed 4:09 AM UTC

Commit: 9dd355b · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high

[waynesun09]

Approved after takeover: rewrite scoped to the contiguous agent-authored tip, real-repository tests incl. GPG-signed and below-human cases, validator soft-pass, two squad rounds, CI green on 9dd355b.

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 1:11 PM UTC · Completed 1:22 PM UTC

Commit: 9dd355b · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $4.07

[fullsend-ai-retro]

Retro: PR #1185 — Strip Signed-off-by trailers instead of rejecting

Timeline

1. 01:49 UTC — Issue #1184 opened: agent commits with Signed-off-by trailers were rejected entirely, wasting full runs (one cost $14.29).
2. 01:50–01:55 — Triage agent completed ($0.81 opus). Severity: high. Identified 3 code paths and related issues.
3. 01:57–02:14 — Code agent completed ($5.30 opus). Created PR #1185 with 6 commits, 17 files changed, +1403/−178 lines. Comprehensive implementation with 54 new test assertions, GPG-signed commit handling, and human-vs-bot commit scoping.
4. 02:14–03:16 — Four review dispatch runs triggered and cancelled as new commits were pushed to the branch.
5. 02:48 — Human (waynesun09) stopped the fix agent with /fs-fix-stop and took over the PR.
6. 03:25–04:09 — Final review dispatch ran. The review agent completed all substantive work — risk assessment, 5 dimension reviews, synthesis, and challenger adjudication producing 5 findings (all low severity) — but timed out at the 20-minute budget before writing agent-result.json. Validation failed: output file not found.
7. 12:51 — Human approved after manual review and "two squad rounds." CI green on 9dd355b.
8. 13:10 — PR merged.

Key Findings

Review timeout wasted completed work. The review agent finished all its analytical work — the challenger returned adjudicated findings at ~17 minutes — but had insufficient time remaining to write the output file. The 20-minute budget killed a review that was substantively complete. This is additional evidence for #1177 (per-iteration 20-minute budget kills large-PR reviews and retries from scratch). This PR's 17 changed files and +1403 lines represent the kind of large, complex change that #1177 predicts will fail.

Sonnet model alias resolution failed for review sub-agents. Five sub-agents (risk-assessment, intent-coherence, style-conventions, docs-currency, cross-repo-contracts) use model: sonnet in their frontmatter. The alias resolved to the CLI default claude-sonnet-4-5@20250929 instead of the fleet-pinned claude-sonnet-4-6 (set by ANTHROPIC_DEFAULT_SONNET_MODEL in env/gcp-vertex.env). The agent recovered by retrying with fallback models, but this added approximately 3–5 minutes of delay that directly contributed to the timeout. See proposal below — distinct from #1100 which covers the retro harness model config specifically.

Code agent quality was high. The code agent produced a thorough implementation with real-repository tests (54 assertions across 7 test groups), proper conventional commits, and an exceptionally detailed PR body documenting rationale, test evidence, and known limitations. The human took over for additional refinements and squad review.

No agent-vs-human review delta possible. Because the review timed out before writing output, no autonomy-readiness comparison between agent and human review is available for this PR.

Proposals filed

• Review sub-agent sonnet alias resolves to unavailable model despite fleet pin (in fullsend-ai/agents)