feat(#108): short-circuit review on rebase-only pushes

[fullsend-ai-coder]

Add step 2a-1 to the pr-review skill to detect rebase-only pushes early and skip sub-agent dispatch. When PRIOR_REVIEW_SHA is set, provenance is app-verified, the compare API returns 0 changed files, and the base branch diff is unchanged, the orchestrator reuses the prior review's findings directly — skipping steps 3-6 and proceeding to step 7 to write the result.

This avoids the token cost of dispatching opus-tier sub-agents for a result identical to the prior review, as observed on PR #90 where review run 4 (triggered by a rebase) produced the same output as review run 3.

Guards fall through to full review when the compare API fails, provenance is untrusted, files changed between the prior SHA and HEAD, or the base branch has diverged.

Note: pre-commit could not run in sandbox (network restriction blocked shellcheck-py install). One pre-existing test failure in post-triage-test.sh is unrelated to this change.

---

Closes #108

Post-script verification

• Branch is not main/master (agent/108-rebase-short-circuit)
• Secret scan passed (gitleaks — 8fb64a25b70e4d364f6d535bb252c1bad521629d..HEAD)
• Pre-commit hooks passed (authoritative run on runner)
• Tests ran inside sandbox

[ralphbean]

This sounds dangerous to me. I don't want to introduce it without pre-existing evals.

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 7:07 PM UTC · Completed 7:17 PM UTC
Commit: dbe1379 · View workflow run →

[fullsend-ai-retro]

Retro: PR #111 — short-circuit review on rebase-only pushes (closed without merge)

Timeline

1. Issue #108 created by the retro agent (from PR fix(ci): persist fullsend version and validate before use #90 retro) proposing to skip review agent sub-agent dispatch on rebase-only pushes. A valid optimization idea backed by real evidence.
2. Triage (run 29104087079) labeled it ready-to-code within 5 minutes — sufficient verdict, correct understanding of the proposal.
3. Code agent (run 29104398019) implemented in 12 minutes — single commit, 75 additions to skills/pr-review/SKILL.md adding step 2a-1 with well-structured guard conditions.
4. PR feat(#108): short-circuit review on rebase-only pushes #111 created at 15:50 UTC. All CI checks passed.
5. Human reviewer ralphbean closed the PR at 19:04 UTC: "This sounds dangerous to me. I don't want to introduce it without pre-existing evals."
6. Issue Short-circuit review on rebase-only pushes to avoid redundant sub-agent dispatch #108 remains open with ready-to-code label, making it eligible for re-dispatch.

What went well

• The code agent produced a clean, well-structured implementation in one pass — correct scope, good guard conditions, clear fallback logic.
• The PR description was thorough and honest about limitations (pre-commit sandbox issue, pre-existing test failure).
• The human reviewer correctly identified the fundamental blocker: no review evals exist to validate behavioral changes to the review skill.

What went wrong

• Wasted cycle: The full triage→code→PR pipeline ran (~17 minutes agent time) for a change that was rejected on architectural grounds (missing eval infrastructure). This was predictable — the repo has triage evals (eval/triage/) but zero review evals.
• Issue Short-circuit review on rebase-only pushes to avoid redundant sub-agent dispatch #108 still ready-to-code: Known gap tracked in fullsend-ai/fullsend#3818 — rejected PRs don't update originating issue labels. If triggered again, the code agent would produce the same rejected implementation.
• No review agent ran on PR feat(#108): short-circuit review on rebase-only pushes #111: The PR received no automated review — only the human caught the eval prerequisite gap.

Existing issues with new evidence

• Remove ready-to-code label from issues when implementing PR is rejected by maintainer fullsend#3818 (remove ready-to-code on PR rejection): PR feat(#108): short-circuit review on rebase-only pushes #111 is a clear new instance — issue Short-circuit review on rebase-only pushes to avoid redundant sub-agent dispatch #108 remains ready-to-code after the implementing PR was rejected, creating re-dispatch risk.
• Issue Review agent should detect when previously-fixed findings regress after a rebase #112 (detect regressions after rebase) is architecturally in tension with Short-circuit review on rebase-only pushes to avoid redundant sub-agent dispatch #108 (skip review on rebase). A naive implementation of Short-circuit review on rebase-only pushes to avoid redundant sub-agent dispatch #108 would prevent Review agent should detect when previously-fixed findings regress after a rebase #112's regression detection from running on rebase pushes. Both issues need coordinated design.

Proposals filed

• Retro agent should assess eval coverage prerequisites when proposing behavioral changes to agent skills (in fullsend-ai/agents)
• Triage agent should treat absent eval coverage as a prerequisite for issues that modify agent skill behavior (in fullsend-ai/agents)

[ben-alkov]

This sounds dangerous to me. I don't want to introduce it without pre-existing evals.

Agreed, though it is a feature which we need to have eventually. Time for 🤔.

[ben-alkov]

What about something deterministic on the invocation side? If the "changes" to the PR are ONLY commit SHA changes (no diff changes), fail early and don't launch the review agent...