14 integration tests covering full fight lifecycle with real in-memory
DB: SSE event ordering, polling bot challenge/response flow, human
player response submission, and 3 concurrent fights without interference.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
8 tests covering: complete fight lifecycle, HP progression, ELO
updates, concurrent fight prevention, round data validity, ELO
conservation, status transitions, and win streak tracking.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Critical finding: when all bots answer correctly, even a 50ms speed
advantage wins 95.4% of fights. At 100ms+ gap it's 100% deterministic.
ELO separation reaches 450+ after just 50 fights. Speed completely
dominates the "all correct" meta.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Simulated 1000 fights with 80% accuracy and equal speed. First-to-lead
wins 60.4% of decided fights, confirming combo system is balanced.
No decay or comeback mechanics needed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When both bots answer wrong, the one with higher checkAnswer confidence
(closer to correct) gets +1 point advantage. Rewards trying over
timing out. Equally wrong remains a pure draw.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When both bots are correct, the one with higher checkAnswer confidence
(exact match 1.0 vs fuzzy match 0.8) gets up to +1.0 bonus points.
This rewards precise answers over approximate ones, adding another
competitive dimension beyond pure speed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Both-correct fights could almost never produce critical hits because
max margin at typical speed differences was ~2.2, far below threshold
of 4. Lowering to 3 yields ~15% critical hit rate (target 10-20%),
making speed differences produce more exciting fight dynamics.
Research findings: simulated 1000 fights with both bots answering
correctly. Critical rate went from 0% to 15.3% with new threshold.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Multi-turn simulation, authority impersonation, encoding tricks
(Base64, ROT13), fake JSON system prompts, red team framing,
reward manipulation, inverted instruction logic. All difficulty: hard.
Zero audit failures after addition.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Test roundToDifficulty: rounds 1-2 easy, 3-4 medium, 5+ hard
- Test pickChallenge difficulty filtering works with round numbers
- Audit prompt difficulty tags across all 16 challenge types
- 8 types lack hard prompts, 3 lack medium prompts
- 82.7% of prompts are untagged (no difficulty attribute)
- Report written to loop/difficulty-distribution.md
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- checkAnswer now returns highest score across all accepted answers
instead of first match, fixing 95 false-low-confidence results
- Skip string containment for purely numeric strings to prevent
false positives like "1000" matching inside "10000"
- Preserve decimal points in normalize() (42.0 no longer becomes 420)
- Use word-boundary regex for number matching in responses
- Fix 47 wrong choices scoring too high (comma-formatted numbers,
verbose choices matching terse answers)
- Fix 17 prompts where no choice matched any accepted answer
- Challenge audit now reports zero failures across all 1472 prompts
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
useNostr (9), useFightCache (5), useOnlineStatus (4) composable tests.
Added fake-indexeddb dev dependency for IDB tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Only invalidates __alltime__ and current season cache keys on fight completion,
preserving historical season caches. Test verifies selective behavior.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Global app.onError handler already in app.ts catches all unhandled route
exceptions. Production mode returns "Internal server error" only.
Tests verify no stack traces or file paths leak in responses.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Both bots timing out every round: verified draw with zero damage
- Perfect victory (10-0): verified correct ELO calculation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
blacklistJwt() adds token to in-memory blacklist until its natural expiry.
verifyJwt() checks blacklist before signature verification.
Cleanup interval removes expired entries every 10 minutes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Documents finding: no replay protection in NIP-98 verification.
Token replay within 120s window succeeds (mitigated by JWT issuance being idempotent).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Tests the actual rateLimit middleware with production mode via dynamic import.
Covers: under-limit allows, over-limit returns 429, window reset, per-IP
isolation, and poll endpoint config (30 req/1s window).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Validates challenge type against the full CHALLENGE_TYPES enum before
processing. Invalid types now return 400 instead of silently falling
back to speed_blitz.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Production now requires JWT_SECRET env var. Added comprehensive JWT
tests: creation, verification, expiry, tampered payload, tampered
signature, and malformed token rejection.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Added getDecodedToken validation that rejects malformed tokens with
400 before any DB lookups. Tests cover empty, non-base64, truncated,
and random base64 tokens.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add rateLimit(1_000, 30) middleware to GET /poll endpoint
- Fix Challenge type errors in human-responses test files (missing baseDamage)
- Add rate-limit unit test verifying 429 after exceeding limit
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
better-sqlite3 driver is synchronous — .get() does NOT need await.
Added tests for unknown pubkey (404) and missing pubkey (400) on join.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Frontend: vitest.config.ts with vue plugin + jsdom, dummy component test
- Server: in-memory SQLite test DB factory + Hono testClient helper + smoke test
- CI: add pnpm audit and server coverage threshold steps
- Root: vitest workspace config for multi-project test discovery
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Scoring: both-wrong is now a DRAW — equal scores (3/3), no winner,
symmetric minimal damage. Garbage answers no longer beat reasonable
ones just by being faster. Both-wrong narrations reflect the draw.
Entrance: removed duplicate announceDeepIntro() call from FightViewer
(was already called inside playEntrance). Removed _resetPositions()
after entrance (entrance already places fighters at home positions,
the extra reset caused a visible snap/reset).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Polling bots have no external script running during practice mode,
so the poll would time out giving empty answers. Now overrides the
webhook URL to human.local so the browser UI handles challenges.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Allow huggingface.co in CSP connect-src (fixes Kokoro TTS model download)
- Add registerSW.js route (fixes PWA service worker 404)
- Add _resetPositions() safety after entrance (fixes invisible fighters)
- Fight end sequence works without canvas scene (KO/overlays/log always play)
- Pre-fight instructions in battle log for human players
- NIP-55 visibility sync and cleanup handlers
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
