fix(sweep): mine the Greptile summary body for findings, not just inline comments

[carlos-alm]

Problem

The /sweep skill's addressing logic is inline-comment-centric. Greptile often raises actionable findings in its top-level summary — the ### Greptile Summary issue comment carrying a Confidence Score: N/5 and prose review — that are not mirrored as inline review comments. A sweep that only iterates the inline comments gathered in step 2d therefore silently misses these summary-only findings. This is the single most common way a Greptile concern survives a sweep unaddressed.

Fix

• New step 2d.1 ("Mine the Greptile summary body"), inserted immediately after the gather-comments step (2d) and before the address-comments step (2e). It treats the summary as a source of findings, not a single comment to reply to: pulls the full summary body (from both the issue-comments and reviews endpoints, since Greptile sometimes posts it as a review body), then extracts every distinct finding — the Confidence Score justification, "First… Second…" gaps in a "Safe to merge after…" paragraph, per-row concerns in the Important Files Changed table, and "Note on…" / "Caveat…" / "mismatch" lines. Each finding is reconciled against the inline comments; any with no matching inline comment is still fixed (or tracked as a follow-up issue).
• New Rules bullet reinforcing that you must mine the summary body, not just inline comments, and reconcile fixes against the summary before declaring a PR ready.

Docs-only change to .claude/skills/sweep/SKILL.md. Part of a cross-repo sweep-skill fix applying the same "mine the summary" step across repos that use the /sweep skill.

[greptile-apps]

Greptile Summary

This PR adds a new "step 2d.1" to the /sweep skill that explicitly mines the Greptile top-level summary body for actionable findings, addressing a gap where summary-only findings (not mirrored as inline comments) were silently skipped during sweeps.

• New step 2d.1 inserts a targeted gh api filter step between the gather-comments and address-comments phases, running two queries (issue-comments + review bodies) filtered to Greptile's user to extract the full summary body, then enumerates every distinct finding type (confidence-score justification, numbered gaps, table rows, "Note on…" caveats) and reconciles each against the inline-comment list.
• New Rules bullet reinforces the mining requirement and makes it a first-class rule alongside the existing "address ALL comments" rule.
• The previously flagged bug (wrong repo ops-codegraph-tool in the gh api calls) is fixed; both calls now correctly target repos/optave/codegraph.

Confidence Score: 5/5

Safe to merge — docs-only change to a skill file with no production code impact.

The change is confined to a single markdown skill file. The new step slots cleanly into the existing numbered sequence, both gh api calls target the correct repo (optave/codegraph), and the reconciliation logic is consistent with the pattern used elsewhere in the skill. No behavioral regressions are possible from a docs-only update.

No files require special attention.

Important Files Changed

Filename	Overview
.claude/skills/sweep/SKILL.md	Adds step 2d.1 to mine the Greptile summary body for findings and a corresponding Rules bullet; both gh api calls correctly target repos/optave/codegraph, and the logic fits cleanly into the existing step ordering.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Step 2d: Gather all review comments\nfetch /pulls/comments, /reviews, /issues/comments] --> B[Step 2d.1: Mine Greptile summary body]
    B --> C{Filter Greptile comments\nfrom both endpoints}
    C --> D[issues/comments\ngreptile summary as issue comment]
    C --> E[pulls/reviews\ngreptile summary as review body]
    D --> F[Extract distinct findings\nConfidence score gaps\nNumbered paragraphs\nImportant Files table rows\nNote on / Caveat lines]
    E --> F
    F --> G{Matching inline\ncomment exists?}
    G -- Yes --> H[Handle in address-comments step 2e]
    G -- No --> I[Add to actionable list\nfix it or file follow-up issue]
    H --> J[Step 2e: Address ALL comments]
    I --> J

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[Step 2d: Gather all review comments\nfetch /pulls/comments, /reviews, /issues/comments] --> B[Step 2d.1: Mine Greptile summary body]
    B --> C{Filter Greptile comments\nfrom both endpoints}
    C --> D[issues/comments\ngreptile summary as issue comment]
    C --> E[pulls/reviews\ngreptile summary as review body]
    D --> F[Extract distinct findings\nConfidence score gaps\nNumbered paragraphs\nImportant Files table rows\nNote on / Caveat lines]
    E --> F
    F --> G{Matching inline\ncomment exists?}
    G -- Yes --> H[Handle in address-comments step 2e]
    G -- No --> I[Add to actionable list\nfix it or file follow-up issue]
    H --> J[Step 2e: Address ALL comments]
    I --> J

Loading

_{Reviews (4): Last reviewed commit: "Merge branch 'main' into fix/sweep-grept..." | Re-trigger Greptile}

[greptile-apps]

Wrong repository in step 2d.1 gh api commands

The new bash block references repos/optave/ops-codegraph-tool (this file's home repo), but every other gh api call in the skill targets repos/optave/codegraph — the repo whose PRs are actually swept (Step 1 uses --repo optave/codegraph). When a subagent executes step 2d.1, it queries the wrong repo and silently returns no Greptile comments, so the entire new mining logic becomes a no-op. Both occurrences of ops-codegraph-tool should be replaced with codegraph.

[carlos-alm]

Fixed in commit ad63cd2 — both gh api calls in step 2d.1 now reference repos/optave/codegraph, matching the file's convention for all other API calls in the skill. Verified there are no remaining ops-codegraph-tool references in the step 2d.1 bash block; the one remaining occurrence (in the follow-up-issue permalink template) is pre-existing and outside this PR's diff. Your own summary review of ad63cd2 confirms the calls now correctly target codegraph.

[carlos-alm]

@greptileai