3.3.1 — W4 GLAB I/O

[3.3.1] - 2026-09-03

Added

• omni_fetch_mr.py one-shot MR gather (skills/omnireview-gitlab/scripts/) — ONE stdlib-only invocation gathers everything review Phase 1 needs into ONE JSON file (MR metadata, paginated diffs, unified-diff headers synthesized where the API omits them, diff_line_map, commits, discussions envelope, versions, notes) with bounded retry/backoff (2 s / 4 s on 5xx/429/network, fail-fast on 400/401/403/404), atomic write, and a one-line stdout JSON receipt (elapsed_ms, api_calls); replaces the improvised dozens-of-calls Phase-1 gather
• --verify-head mid-run STOP guard on omni_fetch_mr.py + the deterministic STOP protocol pinned in SKILL.md (Phase 3 after the waiter, and again immediately before Phase 6 posting): a moved diff_refs.head_sha → exit 4, no re-partition/re-dispatch/re-gather, one flat top-level addendum note, no inline threads on stale anchors — fixes the !1403 class (3 passes / 9 dispatches, watchdog kill, nothing posted)
• Gather-file shims — omni_partition.py and omni_digest.py detect the gather file's shape (data + discussions at top level) and operate on the embedded envelope; legacy single-/two-file invocations byte-identical
• --skip-summary on omni_post_review.py — threads/replies-only resume path for mid-batch failures (--skip-summary --force), closing the guide's old unanchored raw-command escape hatch
• omni_glab_api.py shared direct REST transport (skills/omnireview-gitlab/scripts/) — one tested host/token/auth/retry/pagination contract (stdlib urllib, Bearer auth, token redaction) used by both scripts

Changed

• omni_post_review.py posts via the direct GitLab Discussions API — every glab subprocess call replaced with omni_glab_api.request; inline threads carry the full documented position[...] payload, so every thread posts ANCHORED first-try (live-proven on scratch MR !21: the probe thread's first note carries position.new_path; see w4-glab-io-validation.md); all 3.3.0 poster contracts preserved (exit codes 0/1/2/3, stdout JSON + elapsed_ms, duplicate-summary guard, reply routing, --dry-run executes nothing and works without a token, refs fetched once per invocation)
• SKILL.md is script-first — Phase 1 names omni_fetch_mr.py as the primary gather (MCP fetch demoted to optional-interactive); posting-guide names the shipped poster script as the Implementation path (MCP posting tools optional in interactive installs)
• Test suite 362 (was 325)
• Version bumped to 3.3.1

Fixed

• Unanchored inline threads — root cause: glab api --raw-field "position[base_sha]=..." drops nested position[…] keys, so threads landed as unanchored general notes (measured ~209 s / 27 turns of delete+repost anchor repair per review); the direct-API transport sends literal bracket keys form-encoded and the guide's unanchored raw fallback command is removed
• Headless MCP −32000 diagnosed (record-only) — the .mcp.json launch line uv run --with mcp[cli] resolves the latest mcp (2.x), where from mcp.server.fastmcp import FastMCP dies at import; proven end-to-end, and a one-line pin (mcp[cli]>=1.0.0,<2.0.0) fixes it locally — NOT shipped in 3.3.1 (operator keep/drop decision pending; the standalone script path is primary either way, so headless runs are unaffected). Full evidence: w4-mcp-diagnosis.md
• posting-guide's trailing "MCP post_full_review remains the recommended primary" claim removed — contradicted the script-first Implementation line

A/B

• A/B e2e (same MR !21, same head, both arms real-posting, inter-arm reset, w4 arm first): total wall 1335 s vs 1380 s (−45 s, within single-pair noise); gather = 6 API calls / 4.3 s once parameterized vs 5 improvised calls + 7 inline heredocs (~104 s); posting = ~48 s (dry-run validation + one 20.8 s real post) vs ~86 s + an ~80 s token detour + 1 failed post; threads anchored first-try 24/24 (100%) vs 0/30 (3.3.0 posted every finding as an unanchored top-level note and never repaired — silent anchor loss, worse than the known delete+repost loop); improvised /tmp executables 0 in both arms; sleeps 0 (w4, main + all subs) vs 2 in one 3.3.0 subagent; STOP-guard --verify-head ran at both checkpoints, exit 0 both. Honest misses: 3 failed fetch invocations (~72 s — unencoded full-path 404s + a token detour) before the run settled on the numeric project ID (3.3.2 candidate: accept/URL-encode full project paths); adjudication turns 31 vs 24 (the ≤12 target remains unmet — unchanged lever, not a W4 regression). Full tables: w4-glab-io-validation.md

---