- Persist structured verdict (flags/severity/confidence/evidence) on
moderation_actions so the public web can show WHY a message was moderated.
- Add a persistent Qdrant archive collection (gmw_message_archive); embed
every captured message at capture time (fire-and-forget, best-effort).
- Public semantic search over the archive (backend oRPC + FE toggle on the
messages view). Both features are read-only/public and fully automatic.
Migration: 0015_add_moderation_explainability.sql
- Memoize buildSystemPrompt by (mode|channelCulture); identical signatures
now reuse the ~5k-token core instead of rebuilding per sub-batch call
(textBatchProcessor rebuilt it inside the loop; a 200-msg batch re-sent
the full system prompt ~4x). Correction tail stays per-attempt (uncached).
- Hoist URL-image -> vision evidence out of the per-sub-batch loop in
textBatchProcessor: it depends only on fetched images + full target set,
so compute once per whole batch, not per sub-batch.
- Compact system instructions: collapse 3x-duplicated 'evaluate by content
alone' statements into one standalone rule; trim output.ts channel-culture
+ context framing already covered by rules.ts/system.ts; drop duplicate
programming-error-log few-shot (id 17, covered by rules AMAN list).
- Fix misleading config default: AI_LLM_BASE_URL default -> omniroute
(gateway already runs omniroute via BWS; 9router was dead/misleading).
typecheck + lint + build green.
- resetOffensiveNickname: skip when target role sits above bot
(member.manageable) instead of hammering a doomed setNickname PATCH
that Discord rejects with 50013 'Missing Permissions'. Log the
Discord error code on failure for clear diagnosis.
- llmCaller: include contentPreview (first 200 chars) in the parse-
failure warning so non-JSON LLM responses are debuggable.
- Add wikipediaClient.ts: native fetch to Wikipedia REST/Action APIs
(search + summary), no extra npm dependency.
- Extract shared Redis cache into cacheStore.ts (decoupled from search).
- Term glossary now uses wikipediaSummary for direct article lookup.
- Remove searxngSearch.ts entirely; drop SEARXNG_BASE_URL config,
add WIKIPEDIA_LANG / WIKIPEDIA_TIMEOUT_MS.
- Rename backend searxngCalls metric to webSearchCalls.
User: 'jangan ada reputasi juga' — no profile, no reputation in the prompt,
raw messages only.
- textBatchProcessor: drop initializeUserReputation fetch + <user_reputation>
tag injection (kept the minimal <message> tag + reply/reference context).
- visionAnalyzer (prepareMediaMessage): same removal.
- prompts/system.ts + prompts/output.ts: replace <user_reputation>/<user_history>
instructions with an explicit 'no per-user profile/reputation context'
note so the LLM judges purely on message content + conversation/web/location.
- mediaBatchProcessor: fix stale comment.
Trust/infraction state is STILL written to the DB (userReputationsTable) for
enforcement — only the LLM context injection is removed, so moderation
actions (mute/ban via infraction thresholds) keep working.
Net: even smaller prompts (no per-user context at all) → more messages fit
per request, and one fewer DB round-trip per unique user per sub-batch.
tsc, biome, vitest (129) all clean.
User insight: personal profile summaries bloat the prompt (less room per
request) and add a per-user DB/Redis round-trip for little moderation signal.
Only the behavioural <user_reputation> history is kept.
- textBatchProcessor: stop fetching getUserProfile; remove <user_profiles>
block + <user_profile_ref> from message tags. Keep <user_reputation>.
- mediaBatchProcessor + visionAnalyzer: same removal (profile fetch + ref).
- prompts/system.ts + prompts/output.ts: drop stale <user_profiles>/
<user_profile_ref> instructions; point LLM at <user_reputation> instead.
- aiAnalyzer: gate userProfileLearner behind AI_USER_PROFILE_LEARNING_ENABLED
(default false) — generates profiles nobody reads, pure LLM/DB waste.
- Add AI_USER_PROFILE_LEARNING_ENABLED config knob.
Net: smaller prompts (more messages fit per request), fewer DB round-trips
per sub-batch, and no background LLM calls learning unused profiles.
tsc, biome, vitest (129) all clean.
- Parallelize per-user reputation/profile fetches in textBatchProcessor
(was a serial ~2N DB/Redis round-trip loop per sub-batch; now Promise.all
over unique users). Cuts per-batch latency, biggest win on small/quiet
batches.
- Make the LLM concurrency semaphore dynamic (cached per config value) instead
of frozen at import time, so AI_LLM_MAX_CONCURRENT is tunable without code
change and reflects current config.
- Bump AI_LLM_MAX_CONCURRENT default 5 -> 8 (gemini-flash-lite is cheap; helps
throughput when busy).
- Lower AI_ANALYSIS_DEBOUNCE_MS 500 -> 250 (snappier first-message analysis
when quiet).
- Lower AI_ANALYSIS_RECOVERY_INTERVAL_MS 15000 -> 10000 (stuck/errored
messages re-analyze sooner).
tsc, biome, vitest (129) all clean.
The standalone image analysis path (analyzeSingleMediaImage → llmVision →
llmChat) previously had no request-level timeout of its own — it silently
inherited the shared OpenAI client default (60s), and AI_LLM_MEDIA_ANALYSIS_
TIMEOUT_MS only governed the text+media *batch*, not a single vision call.
- Add AI_LLM_VISION_ANALYSIS_TIMEOUT_MS (default 60000) to config.
- llmChat now accepts an optional per-request `timeout` in LlmCallOpts,
forwarded to the OpenAI request options (falls back to the 60s client
default when omitted).
- llmVision passes config.AI_LLM_VISION_ANALYSIS_TIMEOUT_MS, so a single
image/sticker/emoji analysis gets a guaranteed 1-minute budget and is
independently tunable from the text path.
Verified: tsc + biome green, 129 gateway tests pass.
Co-Authored-By: Claude Opus 5 (Nous Research)
buildCorrectedFewShotExamples() (a getRecentCorrectedModerations(5)
DB hit) was called inside the per-sub-batch buildContent closure in
textBatchProcessor.ts — re-queried for every sub-batch (≈10× for a
200-msg burst) AND re-fired on each parse-error retry. mediaBatchProcessor
already hoisted it once. Mirror that: fetch once per runTextOnlyBatch,
reuse the cached string inside the closure.
No behavior change — identical content, fewer identical DB reads.
tsc + 129 tests + biome green.
Co-Authored-By: Claude Opus 5 (Nous Research)
The 32 few-shot examples each re-echoed score/confidence/
recommended_action/categories/policy_version inline (~150 chars ×
32). Those fields carry zero moderation-decision signal — the schema
and their ??-default coercion already live in OUTPUT_INSTRUCTIONS +
moderationResponseParser.ts. Removed 96 redundant key/value pairs.
Kept per-example: message_id, status, flags, severity, evidence,
analysis — the fields that actually teach decisions. Parser derives
the rest via ?? fallback, so real output shape is unchanged.
examples.ts: 21.7K→18.5K chars; FEW_SHOT(mixed) 15.3K→13.4K.
Total mixed system prompt now 33.9K (was 39.3K at audit start,
~14% leaner). tsc + 129 tests + biome green.
Co-Authored-By: Claude Opus 5 (Nous Research)
- prompts/system.ts: merge 3 overlapping framing blocks (Blok Data /
Konteks Pengguna / Framing Konteks vs Target) into 1 tight block —
same coverage, no duplicated "standalone judgment / profile-is-
reference-not-evidence" prose.
- prompts/output.ts: trim duplicated user_history/standalone paragraph
in PERSONALITY & MEMORI (keep concrete per-case lessons).
- prompts/examples.ts: drop 2 exact-duplicate-lesson few-shots (LGBT id=19
dup of id=30; weapons-tech id=33 dup of id=32). All teaching signals
retained via the surviving example of each lesson.
Static system prompt: text 32.7K→29.2K, mixed 39.3K→35.8K chars
(~10% smaller). No moderation rule, zero-tolerance category, or decision
tree altered — accuracy-controlling content untouched. tsc + 129 tests +
biome green.
Co-Authored-By: Claude Opus 5 (Nous Research)
- gateway-metrics: collectors now run per scrape so Prometheus sees real
data (process memory/uptime + live AI-analysis pipeline gauges) instead
of an always-empty stub. bootstrap registers the pipeline collectors.
- systemd: MemoryMax 512M -> 1G (live RSS ~500MiB, peak 508MiB; 512M left
~2% headroom and risked an OOM-kill restart; host has 8GB free).
- config: POSTGRES_POOL_MIN 2 -> 0 so main + 4 Piscina worker threads don't
hold ~10 permanently-open idle pg connections against PgBouncer.
- docs: rewrite stale ARCHITECTURE.md / MODULE_STRUCTURE.md (winston ->
pino, removed mock-crc/indonesianTextNormalizer, renamed
aiAnalysisWorker/llmModerationClient).
Verified: tsc clean, 129 vitest pass, biome clean on changed files.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds per-message AI moderation analysis time (ai_analysis_duration_ms)
so operators can see how long the LLM took to moderate each message.
Gateway:
- messagesTable: new ai_analysis_duration_ms (bigint) column.
- AIAnalysisUpdate + buildAIAnalysisSet: carry analysisDurationMs through
both single and bulk update paths.
- ai-analysis-worker: measure wall-clock time around runModerationAnalysis
and attach it to every result in the batch.
Backend:
- Mirror schema column; messageMapper maps ai_analysis_duration_ms;
moderation-types + MappedMessage expose it.
Frontend:
- message.ts type gains ai_analysis_duration_ms.
- AiBadge (messages view) shows 'status · 1.2s' when duration is present;
analysis view badge mirrors the same formatting.
DB:
- scripts/add-ai-analysis-duration.sql (idempotent ADD COLUMN IF NOT EXISTS).
No behavior change for moderation logic; null until new gateway build
records values.
Qdrant upserts were failing with 'This operation was aborted' ~32x/2h,
so semantic moderation cache entries were silently dropped. Root cause:
upsertQdrantPoint ran ensureQdrantCollection() on EVERY call — a GET
(and sometimes DELETE+PUT) round-trip — while the request AbortController
had only a 10s timeout. Under moderation load Qdrant is busy (the
gmw_text_moderation collection is not yet HNSW-indexed, so searches are
full-scans), the extra round-trips pushed the upsert past 10s, and the
client aborted it.
- Memoise ensureQdrantCollection() at module scope so the collection is
verified exactly once per process (resetQdrantCollectionCache() for
tests / config reload).
- Bump the upsert request timeout 10s -> 30s so a transiently busy
Qdrant no longer aborts the write.
Qdrant server itself is healthy (<100ms for direct upsert; collection is
green), so no server-side change is needed. Semantic cache should now
populate reliably.
Address every remaining biome lint/format warning across both services
so the codebase ships warning-free:
- textCacheStore: drop unused deleteExpiredQdrantPoints import; hash
image cache key (sha256[:32]) so long/base64 URLs no longer blow the
text_analysis_cache PK B-tree 8191-byte index (was aborting the media
analysis lock INSERT).
- bootstrap: drop unused unhandledRejection promise param.
- moderationOrchestrator: drop unused destructure at L197.
- mediaDownloader / textBatchProcessor / transmitter: replace non-null
assertions with proper null guards (stickerName ?? '', urlImages.get
guard, backpressureQueue.shift guard).
- backend utils: throw lastError ?? fallback instead of lastError!.
- message-capture: remove unused (retentionDb), (moderationActionsDb,
reviewsDb); simplify renderDiscordMentions guard to optional chain.
- transmitter: remove dead write-only field + its assignments.
No behavior change beyond the cache-key hashing (now deterministic
fixed-length) and the intentional null-safety guards.
imageResizer.ts had a line exceeding the print width that biome flagged
as a formatter error, failing the Build & Deploy biome check. Re-format
the file. No logic change.
Two root causes behind 'all image analysis failing':
1. imageResizer still emitted lossless PNG for vision input. A 1024px
Facebook photo balloons to multi-MB PNG base64 that the vision model
silently rejects ('Vision API null response'). Switch to JPEG q85
(no upscaling) — same photo drops to ~100-400KB, model processes fine.
Re-encodes even already-small images so raw originals never bloat the
data URL. Added tests/imageResizer.test.ts covering both cases.
2. acquireMediaAnalysisLock INSERT aborted with 'index row requires N
bytes, maximum size is 8191'. text_analysis_cache.text is the PK in a
B-tree index (8191-byte/row cap); callers pass the raw image URL as the
key, and base64 data URLs / very long URLs blow past the limit, so the
lock INSERT fails and every media analysis is skipped. Hash the URL in
makeImageCacheKey (image:<sha256[:32]>) — fixed-length, deterministic,
well under the limit. All store/get/lock/delete callers already route
through this function so lookup stays consistent.
- config: add AI_LLM_VISION_BASE_URL + AI_LLM_VISION_API_KEY (separate from text router)
- llmClient: llmVision() now calls dedicated vision endpoint when configured
(axios POST to integrate.api.nvidia.com, model nvidia/nemotron-3-nano-omni-30b-a3b-reasoning,
reasoning_budget 16384, non-stream), falls back to router combo otherwise
- keeps text/moderation on omniroute, vision on NVIDIA direct
yt-dlp 2026.07.04 rewrites the --cookies file on close. Handing it the
root-owned /etc/.../ytcookies.txt (not writable by the gmw service user)
caused PermissionError -> exit 1 on every screen-share download attempt.
- buildCookieArgs on-disk branch now copies the system cookie file into a
per-run temp file (like the env branch) so write-back lands somewhere we
own; unreadable -> anonymous.
- resolveInputWithRetry Invidious fallback regex now also matches
permission|EACCES|cookie, so a cookie failure triggers the link-alternative
(no-auth Invidious mirror) path instead of failing all retries.
- adds regression test asserting the original cookie path is never passed to yt-dlp
The live pipe (yt-dlp -o - -> ffmpeg) delivers data at network speed with
unreliable PTS, which defeats ffmpeg -re and made x264 -r 30 force-duplicate
held frames -> ~1fps video (the patah-patah symptom). Per user suggestion,
download the FULL clip to a temp file first (downloadScreenInput), then feed
that FILE PATH to prepareStream. String inputs already get -re, so the
encoder now paces cleanly at 1x against a monotonic-PTS file — proven
reliable in local tests (vs the live pipe which always bursted). Temp file
is removed on stream end / stop.
- getDirectScreenInput -> downloadScreenInput (returns file path)
- resolveInputWithRetry now awaits a completed file + retries on failure
- screenShareController.stops/cleanup removes the per-run tmpdir
- screenShareInput.test.ts updated to the file-download contract
- Add examples for technical discussions (kinetic energy, drone weapon
engineering, physics simulations) that should be marked clean
- System rule: physics/engineering topics (kinetik, gravitasi, energi,
drone, senjata, drone warfare, CAD, CNC, 3D printing, robotics, aerospace)
are safe when in technical context — flag only if explicit threat
- Riwayat pengguna dengan pelanggaran sebelumnya tidak memengaruhi
penilaian pesan bersih yang terpisah dan tidak mengandung pelanggaran
- Removed getUserRecentInfractions usage in textBatchProcessor.ts and visionAnalyzer.ts
- Removed buildUserHistoryXml import and calls
- Messages are now evaluated standalone, not influenced by past violations in other channels
- Updated moderation prompts with clearer instructions about user_history usage
- Fixes issue where benign messages like 'tubuh manusia vs gravitasi' were incorrectly flagged due to carryover from previous drone weapons discussion
The user history context was causing the LLM to interpret unrelated current messages
as threats because it conflated them with past violations. Now each message is judged
on its own merit with only channel-specific context.
Add debug logging to trace cacheKey + messageId + content length on
every vision cache HIT and MISS, so we can detect if the vision model
returns duplicate analysis for different images (provider issue vs
cache collision). Includes the phash on cache miss (new analysis cached).
Follow-up to 9f7ce7d which fixed makeImageCacheKey to hash full data
URL instead of just first 128 chars (root cause of all images sharing
the same cached 'konten judi' verdict due to hash collision).
Root cause: makeImageCacheKey() only hashed the first 128 chars of the
data URL. Since all resized images use the same MIME prefix
('data:image/png;base64,') + identical base64 header bytes, nearly every
image got the same 16-char hash → 'image:<same-hash>' → all images reused
the first cached vision analysis (often a gambling-detection verdict).
Fix: hash the entire data URL instead of just the prefix. Verified
114 stale 'image:' entries + 745 stale 'phash:' entries purged from prod
DB. tsc --noEmit clean, 133 tests pass.
bws-exec exposes the BWS secret as env GMW_YT_DOWNLOADER_COOKIES.
Materialize to temp Netscape file (yt-dlp --cookies needs a path).
Falls back to /etc/gmw-discord-gateway/ytcookies.txt written by deploy.
YouTube now blocks anonymous embeds (403 'Sign in to confirm you're not a
bot'). Resolve with account cookies via --cookies.
- mediaSource: buildCookieArgs() reads GMW_YT_COOKIES_PATH (default
/etc/gmw-discord-gateway/ytcookies.txt) and injects --cookies into
resolveMediaUrl + getDirectScreenInput + extractMediaInfo. Falls back
to anon if file missing (graceful 403, not crash).
- bws-exec now writes cookies file from BWS secret gmw_yt_downloader_cookies
on service start (systemd ConfigFile).
Root cause of "langsung left": YouTube bot-block/403 on u_c1tRmj7E4 (live
stream, LOGIN_REQUIRED) made yt-dlp timeout in resolveInputWithRetry (12s).
The timeout handler did cleanup() (removing once() listeners) THEN
tee.destroy(new Error(...)) — the PassThrough emitted 'error' with NO
listener left → unhandled stream 'error' event → uncaughtException →
gracefulShutdown → bot left voice.
Fix:
- resolveInputWithRetry: tee.destroy() silently after cleanup (error carried
in the rejection only); add permanent no-op tee.on('error') safety.
- prepareStream: output.on('error') no-op so ffmpeg spawn failure before
playStream attaches a demux listener never crashes the gateway.
- bootstrap: serialize uncaughtException/ClientError/DB errors with
{err, errorMsg, stack} (pino only serializes the 'err' magic key — the old
{error: err} key printed {} so crashes were invisible).
- Add term_glossary_cache table + migration 0014: resolved definitions are
stored permanently (definitions rarely change); misses stay ephemeral in
Redis/LRU with 1h TTL so transient failures get retried
- Lookup flow: LRU -> Redis -> Postgres (permanent) -> live SearXNG; DB hits
re-warm the fast caches; stale Redis miss sentinels no longer shadow DB
- Rate-limit-aware live lookups: concurrency 2 + stagger, retry once on empty
results, strict definition filter (Wikipedia preferred, rejects
disambiguation/ads/translate-homepages)
- Make SEARXNG_BASE_URL configurable via env (default unchanged)
Screen share showed a single frozen frame: the GoLive pipeline sent video
only (-"-an", h264 muxer cannot carry audio) so the audio SSRC never
transmitted and Discord kept the stream in thumbnail state.
- prepareStream: mux NUT when includeAudio (h264 muxer drops audio) and
return the actual container format
- Demuxer: support NUT input with a second output pipe (fd3) carrying
Ogg Opus; parse OGG pages into opus frames (20ms, 48kHz) emitted as
GoLiveFrames; fix metadata parsing that dropped the audio stream line
when it arrived in a later stderr chunk (early parsedMeta return)
- playStream: pipe audio.stream into AudioStream → RTP on the audio SSRC
- Encoders: -x264-params repeat-headers=1 → SPS/PPS inline before EVERY
IDR (NUT remux drops container extradata; also enables PLI recovery)
- screenShareController: includeAudio true
- tests: demuxerNut.test.ts — OGG parser unit test + real ffmpeg NUT
integration (video access units + parsed opus frames)
Second root cause (2026-08-12): even with yt-dlp http_headers forwarded,
YouTube still returns 403 when a signed DASH URL from --dump-single-json is
fetched raw by ffmpeg/curl on some videos (verified on fONoh7Pc6VU: curl
with the EXACT headers got 403; yt-dlp's own downloader succeeded). The
signature is tied to the extracting client context (po_token/visitor), not
just UA/IP.
Fix: getDirectScreenInput now spawns 'yt-dlp -o -' and returns its stdout
as a Readable — the same mechanism resolveMediaUrl already uses for music.
yt-dlp handles auth, cookies and transient retries internally. Merge
fragments go to /tmp/gmw-ytdlp-tmp (Nix store CWD is read-only → EACCES).
Removed resolveScreenInput + mergeScreenStreams (dead code).
Controller resolveInputWithRetry unchanged: tees the stream, waits for the
first byte (12s), retries with a fresh yt-dlp run up to 3x on error/EOF/
timeout, and destroys stuck inputs (EPIPE) so no process leaks.
Tests: rewritten for streaming (yt-dlp emits bytes; fail mode = exit 8
without stdout → stream must terminate with zero bytes).
Root cause (2026-08-12 11:50 test): merge ffmpeg hit a transient YouTube
403 and exited code 8 BEFORE prepareStream attached its input listeners
(voice release+join takes ~10s). The input's end/error events fired into
the void, the encoder stdin never received EOF, demux resolved with
fallback 0x0 metadata, setSpeaking fired anyway → stream 'started' with
zero frames for 8+ minutes (black tile, both ffmpeg processes hung).
Fixes:
- mediaSource: pass yt-dlp http_headers (UA/referer) to the merge ffmpeg
via -headers to suppress transient 403s; destroy the returned stream
with an error when the merge exits non-zero before producing bytes.
- screenShareController: resolveInputWithRetry — tee the merge stream and
wait for the first readable byte (12s timeout) before proceeding; on
error/EOF/timeout retry the whole resolution with a FRESH yt-dlp run
(signed DASH URLs expire fast) up to 3 attempts. Stuck merges get
EPIPE via input.destroy() so no process leaks per attempt.
- prepareStream: race guard — if the input already ended/destroyed before
listeners attach, EOF the encoder stdin immediately; first-frame
watchdog in playStream rejects 'started but nothing flowing' after 10s
instead of resolving with a silent black stream.
Tests: +2 (merge-fail zero-byte terminal state, -headers forwarding).
ROOT CAUSE of empty GoLive tile (finally): prepareStream ran with
includeAudio: true + output -f h264. The h264 muxer cannot mux audio
('h264 muxer does not support any stream of type audio') → header write
fails -22 → stdout empty → Demuxer ffmpeg 'Invalid data found when
processing input' → 0 frames → black tile. Reproduced locally end-to-end
(13s backpressure delay + prepareStream + demux).
Fixes:
- screenShareController: includeAudio: false (video-only GoLive; demux
path never delivers audio anyway)
- Demuxer: pin input format -f h264 for stream inputs (raw AnnexB H264
has no magic header → auto-detect unreliable on delayed pipes)
- Streamer.signalStream: self_video: false — stop flipping on the bot's
camera in Discord (user request; screen share ≠ camera)
Verified: local repro now emits 644 frames 1280x720 (was 0); tsc/biome/
vitest all green.
Root cause of 'tile appears but content empty': demux() spooled the live
NUT/H264 input to a temp file and awaited stream 'finish' — but the merge
ffmpeg output never ends during playback, so demux deadlocked, no probe,
no transcode, 0 frames sent.
- Demuxer: pipe input straight into ffmpeg stdin (-i pipe:0), parse NAL
frames live from stdout; parse video metadata from ffmpeg stderr with a
1.5s race (fall back to H264 defaults). No spool, no await-end.
- screenShareController: pass width/height/frameRate (1280x720@30) to
playStream — matches the prepareStream encode settings, so setVideoAttributes
gets real dimensions even when ffmpeg can't report metadata on an open pipe.
- Add tests/golive-demux-live-e2e.ts: proves frames flow while input is
still open (regression test for the deadlock).
prepareStream (from @dank074/discord-video-stream) unconditionally appends
audio filters 'volume@internal_lib' + 'azmq' that exist ONLY in its custom
node-av jellyfin-ffmpeg build. The Nix deployment runs plain ffmpeg-headless
on PATH, so fluent-ffmpeg died instantly with 'Filter not found' (exit 8),
the NUT output stream stayed empty, and playStream's node-av demux failed
with 'Failed to open input from Readable stream: Invalid data found when
processing input' — every screen share failed ~100ms after start.
Fix: pass customFfmpegFlags ['-filter:a','anull'] — ffmpeg applies the LAST
-filter:a for a stream, so the trailing no-op filter overrides the custom
chain (verified: command ends with '-filter:a anull', transcode runs, node-av
demux finds video+audio). Realtime volume control was already removed from
GMW (a690e5b), so dropping the filters is lossless.
Verified end-to-end with the failing URL (youtu.be/fONoh7Pc6VU, AV1+Opus
DASH): getDirectScreenInput → NUT merge → patched prepareStream → node-av
demux finds H264 video + Opus audio streams.
Root cause (3rd layer after 50371bd + 4f4c435): a vision model run
(2026-08-10) returned 'Maaf, saya tidak melihat gambar apapun yang terlampir...'
and that text was cached as a VALID vision_llm result (image + phash keys,
24h/7d TTL). Every subsequent analysis of the same image (same hash/phash)
hit the poisoned cache, so image analysis looked broken forever even though
9router responded fine — the moderation LLM wrote 'lampiran yang gagal
terbaca' from a cache hit.
Also: mimo via 9router streams reasoning in delta.reasoning +
delta.reasoning_details[].text (content:"") — extractChunkText only read
delta.reasoning_content, so those runs aggregated empty → 'Vision API null
response' (observed 08:54/09:07/09:38).
Fixes:
- llmClient.extractChunkText: fall back to delta.reasoning and
reasoning_details[].text (mimo), on top of reasoning_content (gemma).
- visionAnalyzer: isNoImageSeenText() detects 'no image' style outputs;
such results are NEVER cached, and poisoned entries are purged when hit
(LRU/DB/phash) so re-analysis actually re-runs vision.
- Tests: reasoning/reasoning_details extraction + isNoImageSeenText
(Indonesian + English, no false positives on real descriptions).
Root cause (2nd layer after 50371bd): the analysis worker could pick up an
image message while its attachment upload was still in flight
(upload_status='pending'). downloadAndExtractFrame then fell back to the
Discord CDN URL (cdn.discordapp.com), which often 404s for old/purged links,
and 'if (!res.ok) return' silently dropped the image — no log, no vision
call, empty image map, and the LLM produced a text-only verdict like
'lampiran yang gagal terbaca oleh sistem'.
Fixes:
- ai-analysis-worker: skip targets whose attachment upload is still pending
(both batch + individual paths) — they stay ai_status='pending' and the
next 15s cycle analyzes them after the upload lands.
- mediaDownloader.downloadAndExtractFrame: try uploaded_url first, then
discord_url as fallback; log non-OK responses (status + host) instead of
silently returning; log when all candidate URLs fail.
Root cause: 9router combo 'multimodal' routes to cloudflare-ai/@cf/google/
gemma-4-26b-a4b-it which streams ALL output in delta.reasoning_content
(content:"") and finishes with 'length' at max_tokens. llmClient only read
delta.content, so llmVision returned empty → every image moderation fell back
to text-only analysis ('Meskipun analisis gambar gagal' in every ai_analysis).
Fix: extractChunkText() prefers delta.content then falls back to
delta.reasoning_content (also handles message/text/response fields), with
unit tests for the exact 9router chunk shape. Verified live against a real
DB image: oc/mimo-v2.5-free (new first model in the multimodal combo) returns
a proper description in delta.content.
- <message> targets now carry time (ISO), repetitions (N identical short texts = spam signal), bot and edited flags; escape id/user XML
- rich <user_reputation>: total_infractions, clean_streak, last_offense_days_ago, repeat_offender (7-day window)
- <user_history> with last flagged messages for repeat offenders (wires dead getUserRecentInfractions)
- <user_profile as_of> staleness signal; <location_context topic> from captured channel topic
- prompt framing + output instructions teach the LLM to use the new signals without treating history as proof
- tests: contextEnrichment.test.ts (13) + topic cases in conversationContext.test.ts
asepharyana