66 Commits

Author SHA1 Message Date
pliny
f6301765fb Merge pull request #72 from younger-plinius/docs/readme-privacy-accuracy
docs: make privacy and feature claims source-accurate
2026-07-15 10:48:02 -04:00
faber
4d8136bceb docs: make privacy and feature claims source-accurate 2026-07-15 10:07:38 -04:00
pliny
86e487829f Merge pull request #69 from younger-plinius/feat/local-models-and-actionable-errors
feat: local model mode and actionable provider errors
2026-07-14 11:47:42 -04:00
faber
43782a0a74 fix: harden local provider credentials and errors 2026-07-14 11:41:47 -04:00
faber
54d1204d3c docs: use volcano icon for ULTRAPLINIAN 2026-07-14 11:16:28 -04:00
faber
4f8efdf6e1 docs: add local model quick start 2026-07-14 11:14:18 -04:00
faber
45e40b57bb fix: block stale OAuth exchange in local-only mode 2026-07-14 10:59:04 -04:00
faber
f1851326b9 feat: add local model mode and actionable errors 2026-07-14 10:55:52 -04:00
pliny
4143f46dd2 Merge pull request #66 from younger-plinius/fix/telemetry-date-partition
fix: date-partition telemetry + 10k file compaction
2026-07-14 10:35:57 -04:00
faber
2e1cf58343 fix: handle missing archive/legacy dir on first compaction run
listAllFiles('archive/legacy') throws 404 EntryNotFound when the
directory doesn't exist yet on HF. Catch it so the first run
proceeds instead of dying at Step 2.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-14 10:16:56 -04:00
faber
9f714a09fc fix: use recursive listing when scanning archive/legacy for manifests
HF tree API defaults to non-recursive, so nested manifest.json files
under archive/legacy/<runId>/ were invisible. After a partial deletion,
the script would miss the original manifest and incorrectly rebuild
archives from the reduced file set.

Now listAllFiles accepts a recursive flag, and the archive scan uses it.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-14 10:16:56 -04:00
faber
9f4f898b8e fix: resumable compaction + restore memBufferFirstTs on flush failure
Compaction script (P0 restart-safe migration):
- Derives stable run ID from hash of sorted source file paths
- Archives stored under versioned archive/legacy/<runId>/ path
- Manifest includes full sourceFilePaths list for resume detection
- On rerun: detects existing manifest, verifies archives, resumes
  deletion from the original source list — never rebuilds from
  a partial file set
- Verifies ALL artifacts (manifest + archives + invalid-records)
- Post-deletion relist confirms every manifest source is gone
- Idempotent upload: skips if manifest checksum already matches

Telemetry writer (memBufferFirstTs bug):
- On flush failure, restore snapshotTs instead of Date.now() so
  cross-midnight retries don't change partition paths

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-14 10:16:56 -04:00
faber
2896b530ac fix: deterministic date partitions + compaction script rewrite
Telemetry writer: derive partition dates from batch timestamps (KV key
timestamp / memBufferFirstTs) instead of `new Date()` so retries always
produce the same path.

Compaction script: full rewrite addressing all audit P0/P1 findings:
- Paginated listing (follows rel="next" until all 10k+ files listed)
- Fail-closed: aborts entirely if ANY download fails
- Remote verification: re-downloads archives and checks SHA-256 before deleting sources
- Batched deletes: max 50 operations per HF commit
- Preserves FULL invalid lines in separate artifact (not truncated)
- Dry-run is default; --apply required for writes

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-14 10:16:56 -04:00
faber
70cb3bb188 fix: date-partition telemetry writes + legacy compaction script
The HF dataset's telemetry/ folder hit the 10k file limit.

Code fix: writes now target incoming_v2/YYYY/MM/DD/ instead of
telemetry/, distributing files across date-partitioned subdirectories.
Max ~200 files per day-folder means this can never saturate again.

Compaction script: scripts/compact-hf-telemetry.mjs reads the
existing 10k legacy files, groups events by UTC month, and writes
~12 archive files to archive/legacy/. Dry-run by default; --apply
commits archives and deletes legacy files. Verifies event counts
before any deletion.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-14 10:16:56 -04:00
pliny
1c18710862 Merge pull request #65 from younger-plinius/feat/venice-standalone
feat: Venice-only standalone mode
2026-07-14 09:58:36 -04:00
faber
15c8a755c5 fix: default telemetry on, honest UI copy, keep NoLog escape hatch
noLogMode now defaults to false so the HF dataset collects structural
metadata from 100k MAU. Settings copy clarifies: "helps improve model
routing and reliability. Does not send prompt or response text. Turn
off anytime." All NoLog protections remain intact.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-14 08:36:38 -04:00
faber
1ab3563133 fix: replace stale x-ai/grok-3-fast with x-ai/grok-4.5
grok-3-fast doesn't exist in OpenRouter's catalog. Swap to grok-4.5
in the combo registry and max tokens map.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-14 07:59:43 -04:00
faber
9434f0dd83 fix: actionable error diagnostics when all models fail
classifyModelError() categorizes errors (auth/credits/not_found/rate_limit/
timeout/server). diagnoseAllModelsFailed() picks the dominant failure type
and returns a specific message with provider links instead of the generic
"All models failed in ULTRAPLINIAN mode." Telemetry error_type now uses
the classifier too.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-13 21:51:02 -04:00
faber
72d9dfbbbb fix: abort race stragglers, harden backup import/export contract
- finish() now calls controller.abort() to cancel in-flight straggler requests
- Export bumped to v2: excludes plaintext API keys, fixes stale field names
- Import allowlist expanded to cover all settings, migrates v1 field names
- Credentials (apiKey, veniceApiKey) intentionally excluded from backup

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-13 21:01:12 -04:00
faber
15d3ae3c55 fix: restore harmPromise, cap grace threshold, refresh tier counts, replace grok-3
- Restore harmPromise as a defined variable (Promise.resolve(null) under No-Log)
  Fixes crash: harmPromise was awaited in 4 places but undefined when noLogMode=true
- Cap MIN_RESULTS_FOR_GRACE at 5 so large races don't require too many successes
- Call buildTierSelect() after credential decryption and on Settings open
- Replace stale x-ai/grok-3 with x-ai/grok-3-fast in combo registry + max tokens

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-13 15:39:10 -04:00
faber
e4fc62fb79 fix: skip harm classifier under No-Log, fix flush concurrency race
- classifyHarm() skipped when noLogMode is true (saves OpenRouter credits + up to 4s latency)
- flushTelemetry() no-log early return no longer resets _telemetry.flushing
  (prevents concurrent flushes when an in-flight request is still pending)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-13 13:53:27 -04:00
faber
9ff3082dc2 fix: scale grace threshold to race size, show accurate combined tier counts
- MIN_RESULTS_FOR_GRACE now scales to 50% of actual race size (min 2)
  Fixes Venice-only Fast (6 models) waiting full 45s when only 4 succeed
- Tier selector shows combined OpenRouter + Venice model counts via _tierCount()
- buildTierSelect() called on settings save so counts update when keys change

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-13 13:26:24 -04:00
faber
124eb8c51d fix: close No-Log telemetry race — guard flush, requeue, beacon, and import
- flushTelemetry() clears buffer and returns immediately when noLogMode is true
- Failed flush never requeues events if noLogMode became true during flight
- beforeunload beacon guarded by !state.noLogMode
- Import path clears telemetry buffer when imported backup has noLogMode enabled

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-13 12:01:26 -04:00
faber
466d489a2f fix: make No-Log Mode functional — gate telemetry collection, clear buffer on enable
- trackEvent() returns early when state.noLogMode is true (no events buffered or sent)
- Toggling No-Log on immediately clears pending buffer
- Remove dead no_log field from telemetry context
- Update UI description to accurately state it disables G0DM0D3 telemetry

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-13 08:59:31 -04:00
faber
ca70d4c9d4 fix: deep Magic schema sanitization, correct No-Log claim
- Extract _sanitizeMagic() shared helper — covers all nested structures
- Validate thinkingModels values (filter null entries, coerce status/score)
- Filter thinkingLogs, all_scores, scores arrays (remove non-objects, normalize fields)
- Coerce triggers_found to string array, validate tastemakerScore shape
- Normalize winnerModel, topModel, model, systemPrompt, enhancedUserPrompt as strings
- Guard AutoTune .toFixed() with Number.isFinite(), coerce top_k and all param grid values
- Coerce liquidIterations/liquidImprovement, Parseltongue summary counters
- Correct No-Log Mode description: local metadata only, not provider-side

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-13 08:36:38 -04:00
faber
e09dd6b222 fix: sanitize remaining Show Magic metadata — autoTune params, liquid counters, parseltongue summary
- Coerce variants_succeeded/variants_total in Parseltongue summaryText via parseInt
- Coerce msg.liquidIterations via parseInt and msg.liquidImprovement via Number
- Guard autoTune params with Number.isFinite before .toFixed() (prevents crash on crafted strings)
- Coerce at.confidence/at.conversationLength via Number, at.godmodeBoost via strict boolean
- Escape at.strategy/at.context fallback paths through escapeHtml(String(...))
- Add full magic metadata sanitization in both sanitizeLoadedState and import validator:
  mode, score, duration, variants_*, topScore, combos_*, autoTune.* all normalized on load

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-12 19:26:24 -04:00
faber
1f232cab9d fix: sanitize raceResponses + magic metadata, filter null array entries, purge credentials on Clear All
- Coerce current.score to Number in race navigator (prevents innerHTML XSS)
- Escape magic.mode in badge, coerce magic.topScore/magic.score/s.score/variants_* via Number/parseInt
- Escape magic.duration via escapeHtml(String(...)) in PLINY/CLASSIC mode
- Filter null/non-object entries from liquidVersions and strategyLogs.results (prevents crashes)
- Add raceResponses sanitization in both sanitizeLoadedState and import validator
- clearAllData now removes g0dm0d3-apikey and g0dm0d3-venice-apikey from localStorage

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-12 16:18:26 -04:00
faber
c13246ab13 fix: close all remaining XSS sinks, harden import + load schemas
Show Magic version tooltip: escapeAttr() on v.label and improvements
in title attribute — prevents quote-breakout to onmouseover.

Message role: hardcoded to 'user'|'assistant' in class attribute —
arbitrary role strings can no longer inject HTML via class="message X".

Strategy logs: all fields (model, strategy, error, score, duration)
now escaped/coerced before innerHTML insertion.

Image rendering: strict base64 data:image/ regex rejects embedded
quotes; both src and alt use escapeAttr(); onclick removed (was an
XSS amplifier for injected src values).

Import schema hardened:
- Non-object conversations/messages filtered out (null, string, array)
- msg.role restricted to user/assistant/system
- imageDataUrl strict base64 regex; imageName stripped of HTML chars
- liquidVersions: all fields typed/bounded
- strategyLogs: all fields typed/bounded
- Runs BEFORE size check (validate-then-commit preserved)

Load-time sanitization: sanitizeLoadedState() runs on every page load
with the same rules — existing stored XSS payloads are neutralized
even without a fresh import.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-12 11:09:58 -04:00
faber
118dde5ab4 fix: deep XSS paths, import validation, image URL sanitization
Liquid XSS (version history modal):
- All LLM-derived fields (label, grade, improvements) now escaped via
  escapeHtml() before innerHTML insertion
- score/delta coerced to finite numbers
- improvements filtered to string[], bounded to 10 items x 200 chars

Refiner JSON normalization:
- Parse result now strictly validated: numbers clamped [0,100],
  improvements filtered/bounded, refinedContent typed as string
- Malformed LLM output can no longer inject HTML via stored versions

Backup import hardening:
- Validate-then-commit: build candidate state, check size, sanitize,
  THEN apply to live state — rejected imports no longer corrupt state
- Sanitize imageDataUrl: only data:image/ prefixed URLs survive import
- Sanitize liquidVersions: improvements bounded, label/grade truncated,
  score/delta forced to finite numbers
- Imported API key increments _apiKeyGeneration

Image URL render-time guard:
- Inline image rendering now validates data:image/ prefix before
  inserting into src attribute — existing stored payloads blocked

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-11 19:19:02 -04:00
faber
53b00682ae fix: Liquid XSS, credential load race, Venice image badge
XSS (critical): updateLiquidIndicator now escapes all model-derived
strings (improvements, subtext) via escapeHtml() before innerHTML
insertion. LLM-produced JSON can no longer execute arbitrary JS.

Credential race: added _apiKeyGeneration counter incremented at every
state.apiKey assignment. The async decryptApiKey callback on page load
checks the generation matches before writing — if the user clears the
key before decryption finishes, the stale value is discarded.

Venice image badge: vision badge now only renders when
msg.visionAnalyzed is true (set on successful Gemini Flash call).
Venice-only users no longer see "analyzed via Gemini Flash" on
images that were never analyzed.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-11 18:07:58 -04:00
faber
eeff0ce2db fix: remove stale encrypted OpenRouter key when cleared
When the user clears their OpenRouter key (switching to Venice-only),
_flushSaveState now removes the g0dm0d3-apikey localStorage entry.
Previously the encrypted key survived and was restored on reload,
silently resuming OpenRouter requests.

Also guards against a race where an older async encryptApiKey resolves
after the key was already changed — only writes if state.apiKey still
matches the value that was encrypted.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-11 14:10:13 -04:00
faber
40c310f566 fix: guard all OpenRouter helpers for Venice-only + fix pong game leak
Venice-only guards added to:
- generateSmartPrefill: returns random prefill without LLM call
- llmRefusalCheck: falls back to regex-only detection
- llmAccuracyCheck: returns optimistic result without LLM verification
- liquidResponseLoop: early-returns (refinement requires OpenRouter)

Pong game leak: regenerateFromMessage now wraps both branches in
try/finally that always calls hidePongGame(), hideTyping(), and
resets streaming state — success, failure, and abort all clean up.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-11 13:00:31 -04:00
faber
cf6fddb00d fix: Venice-only regeneration + skip OpenRouter when no key
- regenerateFromMessage: routes through ULTRAPLINIAN pipeline when in
  ultraplinian mode (previously called OpenRouter directly with empty key)
- plinyImprovementLoop: early-return when !state.apiKey (no point coaching
  via OpenRouter that will 401)
- llmJudgeResponses: skip LLM judge, use tastemaker-only winner when no
  OpenRouter key
- classifyQueryWithLLM: early-return with default classification when no key

Venice-only users now get the full ULTRAPLINIAN race (via Venice models)
without any stalling on failed OpenRouter calls.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-11 11:02:03 -04:00
faber
7b1d534f59 feat: Venice-only standalone mode (no OpenRouter key required)
Users can now use G0DM0D3 with just a Venice API key:
- Send guard accepts either OpenRouter or Venice key
- Auto-switches to ULTRAPLINIAN when Venice-only (other modes need OpenRouter)
- Race entries conditionally include OpenRouter models only when key exists
- Warning banner updated to mention Venice as an option
- Regeneration guard accepts either key
- Vision processing degrades gracefully without OpenRouter

All three key combinations work:
- OpenRouter only: all modes, full feature set
- Both keys: ULTRAPLINIAN races both providers
- Venice only: ULTRAPLINIAN with Venice models, local scoring

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-11 09:19:06 -04:00
pliny
d64d3d4b3f Merge pull request #64 from younger-plinius/feat/model-refresh-venice-diagnostics-kv-telemetry
Model registry refresh, Venice API, diagnostic errors, KV telemetry
2026-07-11 08:56:21 -04:00
faber
6b2cd9da5d fix: bounded streaming body read + update stale fallback comments
Replace request.text() with a streaming reader that cancels at
512 KiB — prevents a headerless or chunked-encoding request from
buffering an unbounded payload into Worker memory. No WAF rule
dependency.

Also updates the file header and wrangler.toml comments to
describe the current in-memory fallback (not the old direct-commit
behavior).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-11 07:33:30 -04:00
faber
f0744d1f1b fix: in-memory isolate buffer replaces direct-commit fallback
When TELEMETRY_KV is not bound, telemetry events now accumulate in
the Worker isolate's module-level buffer (same persistence model as
the rate limiter) and flush to HF when the buffer reaches 50 events,
256 KiB, or 15 minutes — whichever comes first. This eliminates the
one-commit-per-request explosion without requiring any Cloudflare
Dashboard configuration.

KV is still preferred (durable across isolate evictions), but the
fallback now produces ~1 HF commit per 50 requests instead of 50.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 21:59:47 -04:00
faber
2658bf15e4 fix: remove incorrect alignment comment in utf8ToBase64
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 21:53:10 -04:00
faber
0fcc33f9eb fix: harden telemetry for 100k MAU — memory caps, IP rate limiting, size guards
P1 fixes:
- Per-request body capped at 512 KiB (was 24 MiB — would OOM Workers)
- Flush reads at most 50 keys capped at 5 MiB combined (was unbounded)
- Dual rate limiting: per-session + per-IP (blocks session_id rotation)
- Content-Length early reject before reading body into memory
- Chunked UTF-8 base64 encoding (avoids Array.from() on full payload)
- Max 100 events per batch (was 500)
- Rate limiter map eviction at 10k entries (prevents isolate memory leak)
- Direct-commit fallback warns loudly when TELEMETRY_KV not bound
- Deployment checklist in header comment (WAF, IP rate limit, KV plan)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 21:22:04 -04:00
faber
2be5eebb14 fix: replace retired gemini-2.0-flash-001 with gemini-2.5-flash
The old slug returns 404 from OpenRouter — caught during smoke testing.
Affects tastemaker FAST_MODELS and VISION_MODEL.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 20:10:06 -04:00
faber
25486c3155 fix: add api.venice.ai to CSP connect-src
Venice fetch requests were blocked by Content-Security-Policy in both
the HTML meta tag and the Cloudflare _headers file.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 19:40:52 -04:00
faber
affe3d80bc fix: handle Venice error format + expand error classifier
Venice returns {"error": "string"} while OpenRouter returns
{"error": {"message": "..."}}. All three error parse sites
(index.html queryModel, backend queryModel, frontend Venice helper)
now handle both shapes.

Error classifier now catches Venice-specific auth strings:
- "Authentication failed" (Venice bad key)
- "User not found." (OpenRouter bad key)

Diagnostic messages updated to reference both providers.

Verified 7/7 against real API error strings from smoke tests.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 19:00:32 -04:00
faber
1559e61c9f feat: port Venice support into live index.html, cap KV batch size
Live app (index.html):
- Venice API key field in Settings → API Keys tab
- 44 Venice models (VENICE_MODELS + VENICE_TIER_SIZES) race alongside
  OpenRouter in ULTRAPLINIAN when a Venice key is set
- queryModel() accepts provider param — routes to Venice or OpenRouter
  endpoint with correct auth headers
- veniceApiKey added to state, STATE_KEYS, save/load cycle

Telemetry:
- Reject batches exceeding 24 MiB with 413 before KV.put() (CF KV
  values cap at 25 MiB)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 18:51:20 -04:00
faber
f22aaf665d fix: port 60-model roster into live index.html, fix btoa Unicode crash
Live app (index.html):
- Remove 11 stale OpenRouter models (grok-code-fast-1, both MiMo,
  claude-3.5-sonnet, gpt-5.4-chat, gemini-3-pro-preview,
  llama-3.1-405b, grok-4, both grok-fast variants, qwq-32b)
- Add replacements: grok-4.5, claude-opus-4.8/fable-5/sonnet-5,
  gpt-5.3-5.6 series, qwen3.7-max, glm-5.2, gemini-3.5-flash, etc.
- Update TIER_SIZES to {fast:12, standard:27, smart:41, power:53, ultra:60}
- Update both model dropdowns (header + settings)
- Migrate HALL_OF_FAME combo from dead claude-3.5-sonnet to claude-sonnet-4.6

Telemetry:
- Replace btoa() with UTF-8-safe base64 encoding (btoa crashes on
  non-Latin-1 characters in telemetry JSONL)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 17:44:28 -04:00
faber
132ef2e2ac fix: derive flush filename from SHA-256 digest only (no timestamp)
Concurrent flushes reading identical KV keys now produce the exact same
filename (`flush_<sha256>.jsonl`), making HF overwrites truly idempotent.
The previous approach included a timestamp which broke idempotency — two
isolates hitting the threshold simultaneously would write different files
containing duplicate data.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 16:55:27 -04:00
faber
b5992aea15 fix: flush auth, dedup filenames, TS errors, stale UI counts
- Require Bearer TELEMETRY_FLUSH_SECRET on GET ?flush=force
- Derive flush filename from key fingerprint (concurrent flushes → same
  file → idempotent overwrite, no duplicate rows)
- Fix TS errors in Venice helper (explicit type assertions)
- Correct model counts in SettingsModal (smart=41, power=53, ultra=60)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 16:42:21 -04:00
faber
cebc27cf4c fix: remove fake KV lock, add GET ?flush=force, add .gitignore
KV lock removed — KV is eventually consistent and can't do atomic CAS.
Concurrent flushes may produce duplicate JSONL files on HF, which is
harmless (extra telemetry rows, not data loss or corruption).

Added GET /api/telemetry?flush=force so an external cron can trigger
guaranteed periodic flushes regardless of traffic volume.

Added .gitignore (repo had none).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 16:24:48 -04:00
faber
2a42b57cfb fix: address review blockers — streaming diagnostics, model counts, KV safety, telemetry schema
- Streaming path now returns diagnostic message (not blank) when all models fail
- All models_queried/models_total counts include Venice entries
- KV flush uses a 60s lock to prevent concurrent flush races
- TTL increased from 24h to 7 days (low-volume batches won't expire)
- Restored all previously-existing telemetry ALLOWED_FIELDS (winner_combo,
  stream, fast_stream, has_image, liquid_upgrades, encoding, etc.)
- Replaced AbortSignal.any with compatible signal forwarding pattern
- wrangler.toml no longer has fake namespace ID (commented out for local dev)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 16:11:42 -04:00
faber
64170bf9a2 fix: add per-model timeout (25s) and 429 retry with backoff
Addresses #50 #51 #52 — queryModel() now has a 25s per-model timeout
so a single hanging model can't stall the entire race, and retries
up to 2x on 429 with exponential backoff (respects retry-after header).
Same retry logic added to the standard chat streaming path.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 15:50:43 -04:00
faber
9420624c03 feat: expand Venice model list to maximize OpenRouter overlap
Cross-referenced all 100 Venice API models against the OpenRouter
ULTRAPLINIAN tiers. Venice list now has 44 models (up from 17) with
near-complete overlap where both providers carry the same model, plus
Venice-exclusive additions (deepseek-v4-flash, nemotron-3-ultra-550b,
grok-4-20, gemma-4-31b, sol-pro/terra-pro/luna-pro variants, kimi-k2-6).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-07-10 15:07:09 -04:00