3.4.0 — Round-2 Latency: Deterministic Pre-Dispatch, Judgment-Only Adjudication, Fix Brief, Cost-Weighted Partition

What's New

Round-2 latency release, validated GO by the R2-E A/B run (round2-validation.md): the review lead-in, adjudication, and reviewer balance are now deterministic scripts, and every posting run ends with a paste-ready fix brief for the developer's coding agent.

Added

• omni_prepare.py deterministic pre-dispatch (R2-A, PR #34) — one stdlib-only invocation runs the whole Phase-1 lead-in (gather file, partition, digest, per-agent dispatch briefs) and emits prepare.json + one-line machine receipts with phases.jsonl prepare rows. Pre-dispatch: 827 s fleet median → 265 s raw / ~193 s production-equivalent (prepare itself is O(seconds) at any MR size — 3.73 s / 4.05 s measured). Exit codes are a soft-fail contract: a nonzero exit falls back to the improvised Phase-1 path immediately — no retry, no STOP — proven end-to-end by the R2-E force-exit-1 fallback probe.
• omni_adjudicate.py mechanical pre-pass + judgment-only Phase 5 (R2-B, PR #33) — the worklist's mechanical rows (single-agent autos, near-dup merges, prior matches) are decided script-final without the LLM; only cross-finding and validity rows reach judgment. Adjudication: 461 s / 40 turns fleet median → 239 s / 22 turns — best ever measured on a real MR (17 of 26 rows auto-decided, 9-row judgment core).
• Fix-brief note (R2-P, PR #32) — the poster's last artifact is a self-contained "OmniForge fix brief — paste this into your coding agent" general MR note (template v1): frozen guard clause first, MR-intent line, then one FIXED in <short-sha> — / SKIPPED — / QUESTION — line per finding linking its exact #note_<id> thread. R2-E verified 20/20 links map 1:1 to posted artifacts (0 dead links, set-equality with the new-adjudicated set); poster stdout gains fix_brief: true.

Changed

• Cost-weighted 3-agent partition (R2-D, PR #35) — omni_partition.py balances reviewer assignments by a per-file cost model (1.25× cap, security spill) instead of file count: real-MR per-agent balance 1288/1296/417‰ → 1000/1000/1000‰ (76/75/76 cost units) — deterministic, ending the 2× reviewer wall skew; dispatch briefs trimmed −51.4 % with tripwire tests against over-trimming.
• SKILL.md Phases 1–3 rewritten script-first — prepare is the primary lead-in; the fetcher executes only inside prepare and at the two sanctioned --verify-head checkpoints (R2-E gate 2: zero double-fetch on the knob-on path).

Tests

• Full suite at the release commit: 560 passed + 128 subtests (was 383 + 19 at 3.3.3) — python3 -m pytest plugins/omniforge/tests.

Validation

• R2-E A/B verdict: GO (frozen head of MR !21; A = 3.3.3 dry vs B = Round-2 real-posting). Every functional, safety, and quality gate passed: prepare-first structure with machine receipts, zero double-fetch, script-final auto rows, 8/8 planned threads anchored first-try, poster failures: 0, zero sleeps, zero MCP on the data path (MCP remains only for the Phase-2/7 worktree lifecycle), waiter unchanged; findings 21 vs 21 (≥70 % of baseline), every verdict flip attributed to reviewer nondeterminism — 0 attributable to the Round-2 levers.
• Fallback probe PASS (gate 6) — omni_prepare.py stubbed to exit 1 → improvised-path fallback invoked 9 s later (no retry, no STOP); the run completed engine exit 0 with a full 23,806-byte report, harness SUCCESS.
• Fix-brief checks PASS (gate 5) — guard clause first, MR-intent line, 20/20 exact #note_<id> links, listed set == new-adjudicated set, FIXED/SKIPPED/QUESTION vocabulary verbatim, posted last (note 3790163250).
• Two turn-count calibration misses, disclosed — pre-dispatch 45 turns vs ≤10 and adjudication 22 turns vs ≤12: decomposition shows harness-only token friction (≈12 turns), retrospective re-render absent on first reviews (≈8), and agent inspection verbosity (≈25) — not lever regressions (the 3.3.3 baseline measured 42 pre-dispatch turns under the same harness). R2-F fleet telemetry is the production confirmation — watch phases.jsonl prepare rows and Phase-5 walls.

Known limitations (stated honestly)

• The ≤10/≤12 turn targets are unreachable under the local A/B harness for ANY arm including the 3.3.3 baseline; they are re-anchored to production telemetry for R2-F.
• 3.4.x backlog (disclosed): ADJ_ARGS snippet ↔ argparse first-call hiccup (2-turn self-correcting tax per run), retrospective prepare double-gather (+4 s), fix-brief >25-findings cap path awaits its first real >25-findings MR (unit-covered only).

Contributors

• @nexiouscaliver (Shahil Kadia) — Round-2 levers, tests, and R2-E validation.

Full changelog: see CHANGELOG.md § 3.4.0.