Update the map engagement scoring formula to match plan §14.6:
- score = win_prob_crossings * 3.0 + critical_moments * 2.0 +
resource_contest_turns * 1.5 + survival_turns * 0.5
New metrics computed from replay data:
- resource_contest_turns: turns where energy is contested by multiple players
- survival_turns: turns where all players have at least one bot alive
The old formula used map_coverage_pct, closeness, and turn_pct which
did not match the specification.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The rating recovery CLI mode (-mode=recalc-ratings) was using
glicko2Tau (0.5) instead of glicko2DefaultSigma (0.06) for the
default sigma value when resetting ratings. This caused the reset
sigma to be ~8x higher than the schema-defined default.
Added glicko2DefaultSigma constant (0.06) and updated ResetAllRatings
and recalcRatings to use it correctly.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Implements the rating recovery procedure specified in plan §12.3.
Running 'go run ./cmd/acb-worker -mode=recalc-ratings' will:
1. Reset all bot ratings to Glicko-2 defaults (mu=1500, phi=350, sigma=0.06)
2. Fetch all completed matches from the database in chronological order
3. Replay each match to recompute Glicko-2 ratings from scratch
4. Update the bots table with the recalculated ratings
This is needed for disaster recovery when ratings are corrupted or lost.
Database functions added:
- ResetAllRatings: resets all bot ratings to defaults
- GetAllCompletedMatches: fetches completed matches chronologically with participants
- UpdateAllRatings: bulk updates all bot ratings in a single transaction
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Add engine.CalculateMapEngagement() to compute map engagement scores from replay data (win_prob_crossings, critical_moments, map_coverage_pct, closeness, turn_pct)
- Add DBClient.UpdateMapEngagement() to update map engagement using rolling average
- Worker now calculates and writes map engagement scores after each match
- Add test to verify win_prob array is non-empty in produced replays
This implements the win probability Monte Carlo array storage in replay JSON
feature. The engine already called ComputeWinProbability() in MatchRunner.Run(),
so this commit adds the missing map engagement tracking.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds combat_turns metric (distinct turns where ≥1 bot died from enemy
focus-fire, excluding self-collisions). Worker computes it after each
match; index builder sorts matches/index.json and the new most-combat
playlist descending by it, and bumps interest score for combat-heavy
matches so they surface in highlights.
Also switches homepage featured replay default view from influence to
standard so the actual bot-on-bot combat is visible.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two root causes prevented bots from making any moves:
1. SignRequest signing string included timestamp ({match_id}.{turn}.{timestamp}.{hash})
but all bots implement verifySignature without timestamp ({match_id}.{turn}.{hash}).
Fixed by dropping timestamp from the signing string; X-ACB-Timestamp header is still
sent for clock-skew checks but not in the HMAC.
2. The API stores bot secrets AES-GCM encrypted (184 hex chars) in the DB. The worker
was passing the ciphertext directly as the HMAC key, while bots use their plaintext
k8s secret (64 hex chars). Fixed by decrypting in the worker using ACB_ENCRYPTION_KEY.
Also tightens the home page winner filter to exclude winner_id="0" stalemates.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds R2 (Cloudflare) as a direct upload target alongside B2 (cold archive).
When ACB_R2_* credentials are configured, the worker uploads replays and
thumbnails to R2 immediately after each match, bypassing the index-builder's
B2→R2 promotion cycle.
This is necessary because ARMOR's B2 app key is write-only; reads via the
direct S3 path return 403. The Cloudflare CDN read path (armor-hub-b2.ardenone.com)
is dead post-hub-decommission. Direct R2 upload ensures replays are available
without waiting for a working B2 read path.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Raw game scores (capture points) are tied in most matches since the
winner is determined by an energy/bots-alive tiebreaker. This caused
Glicko-2 delta=0, leaving rating_mu frozen at 1500 for all bots.
Now winner gets 1.0, non-winners 0.0, draws 0.5 — correct pairwise
win/loss signal for Glicko-2 convergence.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace random 2-player pairing with the full §6.1 algorithm:
- Seed selection: bot with oldest last-match timestamp (tiebreak: lowest bot ID)
- Format selection: seed's least-played player count among {2, 3, 4, 6}
- Opponent selection: Pareto 80%/16-rank skill proximity + oldest last-pairing
with seed + fewest 24h games for game-count balance
- Map selection: least-recently-used active map for the chosen player count,
with map_scores.last_used_at updated after each match
- Random player slot assignment for all participant counts
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
PacifistBot never attacks; it survives by maximizing distance from enemies
and retreating toward own core when cornered. Pure evasion strategy that
wins via opponent elimination by third parties.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Worker now gzip-compresses replays before uploading to B2 with
key replays/{match_id}.json.gz and Content-Encoding: gzip.
Updated B2 client Upload to accept contentEncoding parameter.
Fixed downstream web consumers (matches, bot-profile, playlists)
to reference .json.gz URLs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Upload replays to B2 (Backblaze) instead of R2 for cold archive storage
- Write match results directly to PostgreSQL instead of HTTP API
- Perform Glicko-2 rating updates in worker after match completion
- Update config: ACB_R2_* env vars → ACB_B2_*
- Remove obsolete api_test.go (tested removed HTTP client)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds a metrics HTTP server to acb-worker exposing Prometheus text format
at /metrics, plus /health and /ready K8s probe endpoints. Tracks counters
(matches, errors, jobs, replays, polls, heartbeats) and histograms
(match duration, replay upload duration, replay size). Instruments the
full worker execution flow. Fixes .gitignore binary patterns to use
root-anchored paths so cmd/ subdirectories aren't incorrectly excluded.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Worker polls Cloudflare Worker API for pending match jobs
- Claims jobs and executes matches using the game engine
- Uploads replays to R2 via S3-compatible API
- Sends heartbeats during match execution
- Submits results back to Worker API
- Includes retry logic with exponential backoff
- API client tests for job coordination endpoints
Also fixes glicko2.ts: export g() and E() functions for testing
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
