Skip to content

AICOM V2 Cutover Status

rayswaynl edited this page Jul 6, 2026 · 2 revisions

AICOM V2 Cutover Status

Tracks the five-step one-shot cutover from the V1 commander to the unified V2 implementation. Owner ruling: Ray, 2026-07-05. Binding source: docs/design/v2/AICOM-V2-CUTOVER-AND-RECONCILIATION.md on branch claude/fable-completion-push.

As of 2026-07-06 — step 1 (mapping) in progress; steps 2–5 not started.


The ruling (summary)

AICOM V2 deploys in full, at once — not as a permanent parallel system. After cutover the old V1 commander implementation and its telemetry are mapped, then shelved, then removed. A single mission-global flag (WFBE_C_AICOM_V2_ENABLE) governs the transition window only; it retires at step 5. For the authoritative text see the cutover brief cited above.


Five-step sequence

Step Name Gate condition Current state
1 Map AICOM-V2-MIGRATION-MAP.md accounts for every V1 worker, constant, and telemetry emitter In progress — telemetry census complete (2026-07-05); record/function mapping table outstanding
2 Cut over One-shot build ships; V2 is the only running logic; V1 inert behind flag Not started
3 Parity soak N clean T3 nights; capture-rate >= baseline; FPS parity; zero filtered RPT errors; telemetry grammar conformance Not started — soak-gate tools not yet created (see blockers)
4 Shelve V1 commander files + V1-only AICOMSTAT emitters removed onto tagged branch; Shelved-AICOM-V1 wiki record created Not started
5 Remove Rollback flag retires; docs re-anchor to V2-only; V1-citing wiki pages marked historical Not started

Current milestone: Step 1 — Mapping.


The fork — resolved by owner ruling

Two V2 implementation lines existed before 2026-07-05:

Spec line (Codex pack, PRs #700–705) Live line (PR #713, branch fable/aicom-v2-l1)
Records AICOMV2_* (PullWorldState, THIRD_VIEW, …) AICOM2_Snapshot / AICOM2_Allocate (wfbe_aicom2_snap, WFBE_SNAP_*)
Telemetry `AICOMSTAT v3` grammar (PLAN/WHY/INTEL families)
Method Clean-room, harness-first, lanes 415–426 Incremental milestones M0–M5 on live code
Validation Acceptance harness + fixtures Shadow-mode + live soak evidence

Resolution (owner, 2026-07-05): The live AICOM2 machinery is the base. It is shadow-validated against live soak data (M0 snapshot, M1 allocator, M5 latch) and already coexists with the mission's locality and HC constraints. The spec-pack behaviours, doctrine rules, schema discipline, WHY-row explainability, volume protection, and acceptance harness all still bind — ported onto AICOM2 vocabulary.

One telemetry grammar ships. The AICOM2|v1| family grows the V3 features (WHY correlation, INTEL families, harness-parseable enums) rather than a parallel AICOMSTAT|v3 stream.


RPT evidence confirming V2 live readiness (2026-07-05)

From the 2026-07-05 live Chernarus session (live server + HC RPT, direct reads):

  • AICOM2 telemetry is flowing correctly: SNAP/FISTPOOL/ALLOC/ORDER lines healthy from round start (46 neutral towns) through tick 820+. ASSAULT_DISPATCH active on both sides.
  • AICOMSTAT|v2 EVENT stream healthy: ECONOMY/COMBATSTAT/CAPTURE_TRACE families present.
  • Server perf healthy: FPS 25–48 under 339–403 AI. antistack_main is the highest-cost op (500–600 ms/call) — known design cost, not a regression.
  • GRPBUDGET well under cap (peak: west 67/144, east 47/144, GUER 44/144).
  • HC delegation degraded in this session: remotePct fell to 21% (vs 92–95% in 2026-07-04 sessions) with partial recovery to 37%. Hypothesis: GUER + server-local groups crowding the AI budget at low TOWNS_ACTIVE=4 / AI=403. Research packet open — no HC architecture changes without owner approval.

Pre-soak blockers (must close before step 3)

Blocker Required action Owner
`AICOM2 ` has zero tool consumers Extend Tools/Soak/analyze_soak.py to parse the unified grammar
Score-AicomRounds.ps1 does not exist Create in Tools/PrTestHarness/Aicom/; wire to unified grammar Agent B / Codex
aicom-watch.ps1 does not exist Create in Tools/PrTestHarness/Ops/; wire to unified grammar Agent B / Codex
GRPBUDGET / SRVPERF emitters live in AI_Commander.sqf Relocate to server_groupsGC.sqf (byte-identical format) before shelve — TP-4 is an approved ungated packet Agent B
AICOM-V2-MIGRATION-MAP.md record/function table incomplete Finish mapping every V1 worker to V2 home or deliberate drop Cutover build lead

What happens at step 4 (shelve)

  • V1 commander files move to a tagged branch (shelved/aicom-v1-<date>).
  • V1-only AICOMSTAT emitters (those whose content was not ported to the unified grammar) are removed from the mission tree. Any V1-only emitter still in the tree after step 4 is a review FAIL.
  • A Shelved-AICOM-V1 wiki record is created in the established Shelved-* pattern (what / where / how-to-revive).
  • GRPBUDGET, DELEGSTAT, TOWNSTAT, and the groupsGC audit families are not removed — they are not commander-owned and survive unchanged (GRPBUDGET emission moves host but keeps its format).

What is sequenced after cutover

  • GUER Director (PR #715) — approved including relief waves; sequenced after cutover stabilises unless the owner pulls it earlier.
  • MATCH|v1| match-facts family — new telemetry for Stats V2; emission begins after the unified grammar is confirmed and the ingest pipeline is wired.
  • V1-citing wiki pages marked historical at step 5; pages affected include AI Commander Logging And AICOMSTAT Telemetry and AICOMSTAT V2 Event Vocabulary Census.


HQ-strike finisher status (from PR #418, 2026-07-06)

Lane 82 queried whether B69 Patch A's HQ-strike finisher was present. Confirmed resolved: B69 Patch A (35547c471) is already an ancestor of claude/build84-cmdcon36. Current AI_Commander_Strategy.sqf sends HQ-strike teams to getPos _enemyHQ, stages/masses strikers, weights vehicle punch, and can close via base overrun. Chernarus, Takistan, and Zargabad files hash-match. No further action needed on this item before V2 cutover. Full audit note: Audit Archive 2026-07.


Related pages

Sidebar

Clone this wiki locally