Release 3.3.2 — perf + reliability mini release · nexiouscaliver/OmniForge

[3.3.2] - 2026-09-04

Note entries on omni_post_review.py — findings objects carrying ONLY {"body": ...} (no file_path/line_number) post as top-level MR notes (execution order: summary → threads → replies → notes), counted in the additive stdout key notes; mixed batches fine; a body-only entry that also carries anchor keys is an ambiguous-shape usage error (exit 2 — never silently skipped). Replaces the raw-glab general-notes path: glab api --input - silently drops note bodies — production !1388 lost 5/5 MR-meta notes to it. Live-proven on scratch MR !21: note-only batch (no --summary, no --skip-summary) exit 0 with notes: 1, the note re-fetched by id, then deleted (204)
--host on omni_post_review.py — same resolution order as the fetch script (flag > GITLAB_HOST > CI_API_V4_URL > https://gitlab.com), threaded through EVERY API call including the guard's notes listing and the note posts (parity with omni_fetch_mr.py)
Auto-summary-skip for no-new-thread batches — --summary is required ONLY when the batch has at least one new-thread entry; reply-only and note-only batches proceed without it (implied skip: no summary, guard not evaluated — same as reply-only today). Removes the production friction of reply-only batches dying on argparse exit 2 (seen twice in production, one wasted turn each); --skip-summary keeps its current meaning

Duplicate-summary guard paginates — the guard's notes listing now goes through omni_glab_api.get_all (per_page=100, page-following), so a busy MR (100+ notes) can no longer hide an OmniForge summary beyond page 1; a single page still costs exactly ONE notes GET (pinned by test)
omni_fetch_mr.py accepts bare full-path projects — non-numeric --project values are URL-encoded (urllib.parse.quote(safe="")) on every endpoint the script builds; numeric IDs and pre-encoded values pass through unchanged, and the CLI help says so. Production motive: the 3.3.1 A/B arm passed regenai-gitlab/regenai/regenai-base unencoded → 3× HTTP 404 ≈ 72 s before retreating to the numeric ID. Live-proven first-try on MR !21 with the full path (6 API calls, 4.1 s)
SKILL.md + posting-guide document note entries — Phase 6 posting instructions and the guide's findings-shape section name note entries as the poster's general-notes path (with the glab api --input - body-drop hazard as the reason); Phase 3 gains the stalled-subagent rule: attempt exactly ONE SendMessage resume before harvesting partials — never loop
Test suite 376 + 19 subtests (was 362 + 19)
Version bumped to 3.3.2

Headless MCP −32000 — the .mcp.json launch arg is pinned to mcp[cli]>=1.0.0,<2.0.0: the unpinned uv run --with mcp[cli] resolves mcp 2.x, where from mcp.server.fastmcp import FastMCP dies at import and every tool call returns −32000 (diagnosed end-to-end in 3.3.1 with this exact pin proven locally; now shipped under the operator's keep-MCP decision)

line_range positions (needs nested-param dev validation), adjudication-turns lever, dead-subagent watchdog automation