fix(sweep): replace the reply-reaction shortcut with the four-condition trigger gate

[carlos-alm]

The bug

Step 2g decided "is Greptile satisfied?" by checking for a positive reaction on the most recent non-Greptile comment — that is, on one of your own replies. A 👍 on a reply is not a re-review of your code. Under that rule the re-trigger is skipped while the fix sits un-reviewed — which is how a sweep reports a PR as converged when Greptile has never seen the change.

The fix

Transplant the four-condition gate the fleet standardised on. Satisfied now requires all of:

1. an @greptileai trigger comment exists — and really mentions Greptile. A literal inside a code span, fenced block, indented block or HTML comment notifies nobody and is not a trigger.
2. Greptile reacted positively to that trigger (never to a reply).
3. Greptile posted no issue-style or inline comment since.
4. Greptile has reviewed the current head, established from its own Last reviewed commit marker — not a commit timestamp. Greptile re-reviews by editing its summary in place, so timestamps cannot observe a completed re-review; a timestamp proxy re-triggers commits Greptile has already approved. Timestamps survive only as the fallback when no marker parses.

Every fetch failure fails safe by posting rather than assuming satisfaction.

Fitting it to this repo

• Step 2i (final mandatory re-trigger) added; Return result renumbered to 2j. The gate is designed to run twice, and its own header names Step 2i. That header doubles as the sentinel the shared test suite locates the gate by, so aligning this repo's numbering keeps one identical suite proving every fleet copy. I first tried the reverse — rewording the sentinel to say "Step 2h.1" to match the existing layout — and it immediately broke the suite's locator, which is the right outcome: that line is a fleet contract, not prose.
• The 50-trigger hard cap is preserved, and Step 2g's prose now says explicitly that the gate is not exempt from it — if the cap is spent, follow the cap's instructions instead of running the gate.
• The Step 2g prose was corrected. It still told the reader to "skip the actual trigger only if Greptile already reacted to your most recent reply" — the precise rule the gate exists to remove, and a direct contradiction of the code beneath it.

Verification

test_sweep_greptile_gate.py ported alongside. It extracts the real fenced block from SKILL.md — no copy — and runs it against fixture PR state behind a stub gh, under both bash and zsh (zsh because it is the maintainers' interactive shell, and a /sweep guard in a sibling repo once silently passed under zsh while failing its job, since zsh does not word-split an unquoted parameter).

Three of its six cases are mutation tests: they revert the gate to the pre-fix behaviour — remove the marker lookup, break the SHA regex, ignore a marker mismatch — and require the regression assertion to fail. Without them that assertion could pass with and against the bug it names.

6/6 on this branch. Wired into CI as sweep-gate.yml, path-filtered to the skill + suite — without it nothing ever executes this gate: it lives in Markdown, so it is shipped, load-bearing and otherwise entirely unproven.

Part of the fleet-wide sweep tracked by optave/data-retrieval-storage-svc#1021.

[greptile-apps]

Greptile Summary

The PR replaces reply-reaction satisfaction with a four-condition Greptile trigger gate and adds a cross-shell fixture suite plus path-filtered CI. Two fetch-error paths still fail to uphold the gate’s stated fail-safe behavior.

• Classifies genuine trigger mentions while excluding code and comment literals.
• Uses Greptile’s reviewed-commit marker before falling back to timestamps.
• Tests the embedded shell under bash and zsh, including mutation cases.
• Adds a dedicated GitHub Actions workflow.

Confidence Score: 3/5

The PR should not merge until failed comment fetches consistently attempt the mandatory trigger rather than aborting or permitting timestamp-based satisfaction.

The new gate can terminate without posting on its first fetch failure and can silently treat a later marker-fetch failure as an absent marker, allowing the documented inexact timestamp proxy to skip review of the current head.

Files Needing Attention: .claude/skills/sweep/SKILL.md

Important Files Changed

Filename	Overview
.claude/skills/sweep/SKILL.md	Adds the four-condition gate and mandatory final invocation, but two comment-fetch failures can abort or incorrectly fall back instead of safely posting.
.github/scripts/test_sweep_greptile_gate.py	Adds substantive cross-shell and mutation coverage, though it does not isolate the marker-fetch failure that permits a false skip.
.github/workflows/sweep-gate.yml	Adds an appropriately scoped workflow that installs zsh and executes the standalone gate suite.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Run Step 2g or Step 2i gate] --> B[Fetch trigger comments]
  B -->|Failure| X[Exit without posting]
  B -->|Success| C[Check trigger reaction and later comments]
  C --> D[Fetch current head]
  D --> E[Fetch reviewed-commit markers]
  E -->|Marker matches head| F[Skip trigger if all conditions hold]
  E -->|No marker or fetch failure| G[Timestamp fallback]
  G -->|Timestamp not newer| F
  G -->|Timestamp newer or unavailable| H[Post @greptileai]

Loading

_{Reviews (1): Last reviewed commit: "fix(sweep): replace the reply-reaction s..." | Re-trigger Greptile}

[greptile-apps]

Trigger fetch aborts fail-safe gate

When the issue-comments request fails during Step 2g or mandatory Step 2i, this branch exits before attempting to post @greptileai, causing the sweep to stop without completing the required final review trigger.

[greptile-apps]

Marker fetch failure permits false skip

When this issue-comments request fails after the initial fetch succeeded, the pipeline silently produces an empty reviewed_shas and enters the timestamp fallback. If the commit timestamp is no later than the trigger, the gate treats the head as reviewed and skips the mandatory trigger even though Greptile's current-head marker was never retrieved.