From 084ad924f02fc3e0cf9a69a65e370a06db84d71a Mon Sep 17 00:00:00 2001 From: Vuong Hoang Date: Tue, 18 Aug 2026 16:36:13 -0700 Subject: [PATCH] =?UTF-8?q?memory:=20snapshot=20=E2=80=94=20ESH=20fiber=20?= =?UTF-8?q?live,=20esh-pve-nas=20migrated+patched,=20v6=20mapped?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Session captured: ESH cut over to Cityside 2Gb symmetric fiber and was fully provisioned on it; esh-pve-nas completed its ZFS-root migration off the USB DOM and took its 225-package security backlog with the reboot deferred; ESH<->colo IPsec was rebuilt as a dialup tunnel with NAT-T after CGNAT broke the statically-pinned one; IPv6 was mapped across all three sites. Auto-archival fired at the soft cap: 7 entries moved to archival-memory.md (Recent decisions 3, Tried and abandoned 4), all verified-complete arcs, with two detail files moved and removed. The remaining pre-Aug-05 entries were held back by the open-deferred-work guard, so the index stays slightly over cap at 313 lines rather than losing a live pointer. Also records the one self-inflicted outage of the session (missing --make-rslave on a chroot rbind) and that three long-dead things surfaced incidentally: pvestatd down 82 days, a vzdump hung 126 days, and a VM sitting in prelaunch for four months. --- archival-memory.md | 73 +++++++++++++++++++ .../2026-08-02-mimir-inbox-arc.md | 32 -------- .../2026-08-03-worldtree-b168-384-385-arc.md | 17 ----- persistent-memory.md | 17 ++--- 4 files changed, 78 insertions(+), 61 deletions(-) delete mode 100644 persistent-memory.d/2026-08-02-mimir-inbox-arc.md delete mode 100644 persistent-memory.d/2026-08-03-worldtree-b168-384-385-arc.md diff --git a/archival-memory.md b/archival-memory.md index 60b8cfc..26b2ba4 100644 --- a/archival-memory.md +++ b/archival-memory.md @@ -2392,3 +2392,76 @@ See [[2026-07-25-wt-376-per-instance-config-arc]] for the incident that produced - `[2026-07-25]` **Peer green-light ≠ operator consent for a managed-box mutation.** Auto-mode guard blocked a config-replace+restart on the Worldtree-team demo box that was authorized only by worldtree-dev's althing message — correctly: a persistent change to shared infra needs the *operator's* yes for that specific change, not a peer's. Surface it; don't route around the guard. (The operator then stood the whole change down — the guard's hold was the right call.) _Archived 2026-08-16._ + +## Recent decisions (archived) + +**Worldtree b168/#384/#385 arc — COMPLETE 2026-08-03.** A long peer-driven arc across worldtree-dev / muninn-dev / mimir-dev / ratatoskr-dev, all on corviduo-dev's demo+personal instances. Sequence: providers.yaml boot-gate pre-sync → b168 deploy → DCC #384 reindex → round-2 full re-ingest → #381 restart → operator-approved production dedup sweep. Landed clean; three of MY foot-guns along the way, each caught + hardened into a fleet runbook rule (see Tried-and-abandoned: `mv -t`, `docker exec -u 1000`, shared-containerd race). + +## providers.yaml pre-sync (boot-gating config) +b168 (commit `293f8f3`) added a `summarization` capability block that in-image `agents/muninn/config.yaml` references → boot-blocking if the host bind-mounted providers.yaml lacks it. Synced both hunks (summarization block + deep-reasoning desc) into demo+personal via `deploy-wt-config`; instance-configs commit `53349f8`. +- **deploy-wt-config runbook:** `~/development/worldtree-instance-configs/scripts/deploy-wt-config {diff|deploy|capture} --file providers.yaml` (per-instance dirs demo/personal/pinned; `deploy` = host write + api/matrix restart + 90s health-gate + auto-rollback; `diff`/`capture` safe). demo+personal providers.yaml are byte-identical. +- **GOTCHAS:** (1) an UNPUSHED source commit → `git show ` 404s and a gitea `raw?ref=` silently falls back to the default branch; verify the commit exists (`/git/commits/`) before trusting a fetch, else ask the peer to paste hunks. (2) a peer's hunk paste may be mis-indented (8-space vs the block's 4-space) → invalid YAML; always YAML-validate after a paste-sourced edit. +- **Config-delta pre-sync rule (verified via `docker inspect`):** worldtree containers bind-mount ONLY `config/` host-side (`/opt/worldtree-*/config/` → providers/model_roles/matrix/policies/defaults/env.public = the pre-syncable set); `agents/` (schemas.yaml, prompts) + all code ship IN-IMAGE. So only a `config/*.yaml` change is boot-blocking-pre-syncable; an `agents/`-or-code delta needs NO host pre-sync (CI carries it). b169's schemas.yaml (#387) was correctly no-pre-sync. + +## #384 reindex + #381 restart + verify +DCC job `mimir-6351554e8e8f`. Reindex: `sudo docker exec -u 1000 worldtree-personal-worldtree-muninn-1 python -m core.muninn --reindex ` (⚠️ MUST `-u 1000` — default-root writes contaminate the uid-1000 KB tree; see Tried-and-abandoned). Then **#381 restart** (stale-Chroma-client fix): `sudo docker restart worldtree-personal-worldtree-api-1` (plain bounce, NO compose up / no image repoint) → healthz/readyz 200 ~25s. +- **Chroma-verify runbook:** `sudo docker exec -i python -` (MUST pass `-i` or stdin never reaches `python -`) → `chromadb.PersistentClient('/data/kb/.chroma').get_collection('fiction').get(where={'job_id':}, include=['metadatas'])`. Chroma persists at container `/data/kb/.chroma` = host volume `worldtree-personal_worldtree-kb`. +- **Retrieval-visibility check (NOT grounding — that's ratatoskr's):** a Mimir session — admin token `~/.config/worldtree/personal-admin-token` (wildcard scope) → POST `/sessions` (agent_id=`mimir`, `record_tool_intermediates=true`) → POST `/sessions/{id}/messages` (STREAMS SSE, not JSON) → parse SSE `tool_result` for `search_library` wing hits → DELETE session. + +## Production dedup sweep (operator-approved) +Deleted the 785 April-era DCC orphan rows (`job_id=b59c147c5ce0`, no wing/source_identity metadata → predate identity tracking) from the `main` collection. Supervised protocol: read-only verify count == 785, back up all rows (ids+docs+embeddings) to `corviduo-dev:/tmp/main-sweep-backup-b59c147c5ce0.json` (reversible), `main.delete(where={job_id})` (assert target==785 first), verify `main` 4009→3224, then **bounce the api** (a separate-process delete leaves the api's in-memory HNSW index holding the vectors until reload — the #381 pattern generalizes to deletes), confirm search now fiction-only. Backup left for /tmp natural cleanup (fiction wing is canonical; `~/archives` has the historical record). + +Result: fiction wing 166 → 1,372 concepts; three consumer verify rounds 0/5 → 5/5 → saturated; #385 budget fix validated (705 vs April's 785 control, extraction AND indexing, zero truncations). worldtree-dev filed #388 for a deploy concurrency-lock (the shared-containerd race fix). See [[2026-08-02-mimir-inbox-arc]]. +_Archived 2026-08-18._ + +`[2026-08-02]` **The mimir-inbox / #377-read-path arc — deploy, four bugs found+fixed+verified, a cloned voice, all in one long session (2026-08-01→02).** + +The browser-facing half of the #377 Muninn ingestion arc, end to end: mimir-inbox stood up, the write path proven, the read path chased through four defects to a verified-working state, and a character voice cloned into the TTS zoo. Peers: mimir-dev (the app), muninn-dev (gate/watcher spec), worldtree-dev (Worldtree app layer + the #380/#381/#382/#383 fixes), ratatoskr-dev (a consumer + the rigorous verifier). + +## mimir-inbox deployed (#377) +- **New infra-ops stack, canonical eshpfi `stacks/mimir-inbox/`; live corviduo-dev `10.250.50.152:8091`** (co-located w/ muninn-gate :8090 + the worldtree-personal muninn watcher). Full deploy detail + procedures → auto-memory `reference_mimir_inbox_deploy`. +- **Placement decision (operator, reversed):** 7-31 he ruled mimir-inbox stays OFF corviduo-dev (shared/NFS mount); 8-01 he REVERSED to CO-LOCATE. Trigger: muninn-dev's code-check showed staging is NOT same-fs-constrained (gate reads staging metadata + passes path strings; `os.replace` is inside `ingestion_root`) — staging's real constraint is **path-identity across writer/gate/watcher**, which co-location buys outright while dodging NFS failure modes. I HELD the reversal for the operator's direct word (data/hosting on a team-managed box, reversing his own ruling) even against 3 peer relays — vindicated as the right instinct; muninn-dev agreed. +- Build: **`uv sync --no-dev --frozen`, SINGLE-STAGE** (project installs editable-linked to `src/`, so src/ MUST stay beside .venv — a multi-stage "copy only .venv" dies at import/404s assets). uid 1000, host-net bind 10.250.50.152:8091, TCP-liveness healthcheck (deliberately NOT gate-coupled). Redeploy = refresh build context (**preserve the on-server `.env`!**) → `docker build -t mimir-inbox:0.0.1 -t mimir-inbox: .` → `compose up -d`. Version stays 0.0.1 across dev commits → tag the image w/ the source SHA too. Live commit progression `0478452`→`c8ab38f`→`2dcc77e`→**`8ece117`** (3 redeploys). +- mimir-inbox key on the gate bumped [read,submit]→**[read,submit,control]** (cancel/retry); brokered via a 0600 drop on nh3-dev (never on the althing bus). + +## The read-path bug chain (worldtree-dev's, all found via this arc) +- **#380 wing-blind indexing:** the book-ingest path upserted concepts into a hardcoded `main` Chroma collection while wing search reads the `fiction` collection → P&P written to disk but `search_library` returned total 0. A silent-success defect ("complete/69 indexed" was right about the WRITE, wrong drawer). Root-caused off MY physical evidence (files on disk + search empty). Fixed b164 + a one-shot `--reindex ` (re-upsert into the right wing collection + delete stray `main` rows). +- **#381 stale Chroma client:** the personal api opens its Chroma client before the watcher's cross-process writes → **a freshly-ingested/re-indexed book is NOT queryable until the api is restarted.** Proven by my restart-diagnostic (pre-restart total 0 → post-restart hits, same index). Workaround until fixed: `docker restart worldtree-personal-worldtree-api-1` after any ingest/re-index. Filed as #381. +- **#382 unreliable Mimir grounding (the subtle one):** post-#380-fix the index was correct, but Mimir's grounding was INTERMITTENT — some sessions navigated the opaque job-hash dir (`mimir-f3887c9b97b7`) to the content, others distrusted the correct vector hits and **silently answered from training knowledge** (worst of the looks-fine-isn't family). ratatoskr-dev caught it; I'd been over-confident ("Mimir read Austen back to you") having verified the INDEX, not the GROUNDING. Fixed b166 with BOTH shapes: a self-describing `_index.md` per wing job-dir (resolves the hash dir to its title) + a Mimir prompt rule (wing-scoped hits ARE library content, never discard on a name mismatch, never substitute training). **Verified: ratatoskr-dev re-ran 3× fresh sessions → 3/3 grounded**, citations in note-extracted language not raw Austen. #382 CLOSED. +- **DCC (Dungeon Crawler Carl, job `b59c147c5ce0`) backfill:** `--reindex` FAILED ("job not found in any state dir" — predates state-tracking). SETTLED = **no re-file** (the b166 prompt rule already grounds it even without an `_index.md`; ratatoskr confirmed incidentally); an `_index.md` rides whenever DCC is next re-ingested. +- **#377 mimir-inbox banner bug (mimir-dev's, `8ece117`):** `/health-banner` misattributed an unwritable `ingestion_root` to the WORKER, rendering "The worker is not running." for a running worker — a false lead pointed at infra-ops's half of #377. Fixed (guard split into two banners); I confirmed from the DEPLOYED handler (not just the test) that `ingestion_root_writable:False` now renders "The ingestion root is not writable." + +## muninn-gate → muninn-dispatch 0.1.5 +Rebuilt `muninn-gate` off `vh/muninn-gate` main `bc04c4c` (dispatch 0.1.4→0.1.5) so the gate serves the new `concept_schema`/`concept_schema_source` row fields (computed gate-side). Gate version unchanged 0.0.14 (dual-tag the SHA). Build needs the vh gitea token as a BuildKit secret (`--secret id=gitea_pw`, UV_INDEX_GITEA_USERNAME=vh, drop+shred). Recreate with `compose up -d` (NOT bare restart — needs the new image). Verified: P&P job serves `concept_schema='fiction'`, `concept_schema_source=null` (null correct — pre-b164 job). Registry tags by commit SHA — `v1.0.0bNNN` docker tags don't resolve; use the deployed SHA (confirm `--reindex` present before using an image for a data-op). + +## donut voice (65-frost → Zonos gateway) +Operator: "pick up 65-frost, use that bundle as a voice for a character named donut." 65-frost = a **Booth id** (`~/booth-data/65-frost/`) holding a curated yt-voice-clipper dataset (`dataset-…-curated.zip`: 4 clips + manifest, all SPEAKER_02 = Princess Donut). **Zonos gateway voice registry = a filesystem drop:** `.wav` in the voices dir (44.1kHz mono s16 PCM) auto-registers as `voice:""` on **startup** (needs a restart). The LIVE dir is the bind mount `/opt/docker/compose/zonos-gateway/voices/` (lkraven-writable), NOT the working tree. Built `Donut.wav` from seg000 (best clip), dropped it, restarted → `voice:"donut"` live in the gateway AND the Asset Engine's make form. Also copied to the build-source tree `~/zonos-gateway/voices/` for rebuild-durability (true canonical = the gitea repo, not yet CI-wired). Auditioned in booth `donut-voice`. **Expanded 2026-08-02 (onyx-58 bundle):** operator curated a 2nd Booth bundle `onyx-58` (`dataset-467d2cf8…curated.zip`, 3 Donut clips) as additions. Rebuilt the reference = **seg000 (65-frost) + seg101/seg110/seg148 (onyx-58)** ffmpeg-concat + resampled 24k→44.1k mono s16 = **52.0s**. `seg148` was diarized SPEAKER_03 but is Donut (operator-confirmed misdiarize → included). Assembly is NOT `assemble_voice.py` (that `-c copy` can't resample + caps ~15s); used a manual `aresample=44100,aformat=…,concat=n=4` filter. Backed up old ref → `irv-ml1:~/Donut.wav.pre-onyx58`; dropped to live bind-mount + build-source tree; `docker compose restart` (healthy 2s, `voice:"donut"` still 1 of 9). A/B booth `donut-onyx58` (A=old 16.3s ref, B=new 52s ref, same line). Longer ref is fine mechanically: gateway passes it as `speaker_audio_base64` → speaker *embedding*, not an audio prefix. **BUT auditioned → REVERTED same day:** pinned-seed neutral A/B (5 pairs, booth `donut-onyx58`) showed the single-clip seg000 (16.3s) beats the 52s 4-take concat on timbre — concatenating disparate takes muddied the embedding more than the range helped. Reverted both live + build-source to seg000-alone. Lessons (→ Tried-and-abandoned): more reference ≠ better when takes vary; and **emotion steering pulls output away from the clone fast** (operator craft rule) — keep clones emotion-neutral; bare `{input,voice}` calls send NO emotion (gateway only enables it on an explicit `emotion_*`/`preset` dial). + +## Zonos streaming (no gateway change needed) +ratatoskr wanted play-as-it-arrives. `/v1/audio/speech` ALREADY streams — chunked `StreamingResponse`, opens native `/tts/generate` with `stream=True`, wraps as a streaming int16 WAV with `0xFFFFFFFF` placeholder sizes (meant for progressive `