The fake Idunn signer can lose the lease once a job's meta.json exists, which
pins the guards in startWorker (no logs or embeddings directory, no worker).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hand-run lock is a bind in the abstract socket namespace named by the state
root: atomic, exclusive, gone with its owner, so the probe, the unlink and the
takeover race are deleted. Every write to the state root and every spawn calls
guard(), which asks held() at that moment; Idunn's held() reads the lease each
time and reports a loss through the handler the daemon registered, so a request
admitted before the loss stops at its next write and a respawn cannot outlive it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Red at 4fd3f60 for stale-lock-takeover-race, fence-admits-in-flight-request and
bind-port-65536-accepted; green guards for reclaim-root-unpinned and the SIGTERM
property. The daemon's stop() no longer leaves a daemon holding the run open.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
serve(runtime) listens first, writes the state root only once the runtime
holds it, answers Idunn's route challenge ahead of the cross-site refusal,
and fences on lease loss. --health is deleted; the pid file and /proc
reclaim move behind the hand-run runtime. Adds scripts/idunn-entry.mjs,
the kimodo and SOMA-X gitlinks, encode_texts for the child, and tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WSL, Docker, GPU and Raven launchers leave main. Not deployed: admission on Yggdrasil is the next cut. s2 residue in motion-pipeline:follow_up:vili-cpu-body-s2-residue.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Closes the survivors StrykerJS reported over the new supervisor code (61 mutants, 41 killed, 20 survived): the pid
file is removed when its worker exits and not when it spawned nothing, the restart kill is SIGKILL (a stubborn stray
and a SIGTERM-proof encoder child), the group match is the pid file's, not any marked process. killGroup loses its
unused return value.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Appends a dated correction to both probe notes, citing motion-pipeline:finding:cut-vili-cpu-body.s1.ab-premise-inverted.
The measurements stand as measured.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The front door drops its claim about the Raven instance and describes the encoder's explicit chat header and the
daemon's ownership of the worker's process group. docs/verse-service-contract.md, which the front door names as an
entry point, is rewritten for the CPU tree: witnesses, command boundary and worker boundary, with the request fields
left to the README.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The header-on vector now matches Kimodo's encoder loaded from the repo ids (cosine 1.0; the bare header gave
0.9974) because the wrapped text carries LLM2Vec's instruction/query separator, which keeps the header out of
the pooling. live_identity reads the installed versions itself, so the worker passes no version it could fake,
and the tests pin installed_versions against substituted modules, the prompt form and device in the identity,
and a text that differs past any prefix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The worker's process group is killed when the worker exits for any reason, so no encoder child survives it.
A pid file in the state root names the group; the next daemon start kills a group left by a crashed
predecessor if one of its members carries this state root's embeddings directory, before serving. The
retired operator route's 404 test returns (built from parts, so the grep stays empty) and the test script
gets a per-test timeout.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The child wraps every text with text_encoding.chat_header and tokenizes it itself, so the prompt form no
longer rides on the model's _name_or_path. The identity names the assembly truthfully (header form
included, as a hash of the template) and re-keys the cache. The worker builds its identity through
live_identity and installed_versions, which the tests reach; the identity test now uses live-length
revisions, so a truncated identity or key fails. An env-gated test compares a fresh child vector with
Kimodo's own encoder loaded from the repo ids.
Supersedes encoder-assembly-b (motion-pipeline:ruling:encoder-chat-header-on).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Mutation triage (StrykerJS, 81 survivors on the first full run) showed worker.model was
written and never read. The new tests pin the survivors that were behaviour: the loopback
default, --host, stray flags, the models-root resolution, the SIGTERM grace, a worker that
dies while the daemon waits for its health, the POST route, and lastError after a
requested stop.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured on Yggdrasil: loading by repo id leaves the model's _name_or_path
equal to the Llama repo id, so prepare_for_tokenization wraps the text in
the chat header and the vector matches assembly A (cosine 1.0, 0.909 to the
reference B). Loading from the snapshot paths reproduces the reference.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
vili-daemon.mjs spawns the Python worker itself (no shell, port 0 from the
ready line, nice 10, one child at a time, respawn after exit, group kill on
SIGTERM) with a closed environment. WSL/Docker, the Eve surfaces and deck
WebSocket, the provider advertisement, /smoke, CORS, JSON sidecars and the
pid file are deleted. POSTs with an Origin or a non-JSON body are refused
403 cross-site. cultcache-ts comes from the vendor/CultLib submodule at
b549d731; ws is dropped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The encoder child is CPU only (GPU placement, GpuBusy and plan_placement
are deleted) and assembles LLM2Vec B: Llama base, MNTP merged, then the
supervised adapter. The cache identity names assembly, device, dtype,
torch/transformers/peft/accelerate versions, Kimodo commit and the child's
own sha256 beside the three revisions. VILI_THREADS (1..16) caps torch in
the worker and the child. The worker binds port 0 and reports it in its
ready line; errors name a class and place, never the request.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Raven scheduled-task installer, launchers, restart shims and the smoke client
(GET /smoke on 10.77.0.4) leave main under the Loud Rebuild Contract
(motion-pipeline:ruling:vili-moves-to-yggdrasil-now).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured on Yggdrasil CPU under other agents' load; Raven untouched.
A/B eye-check BVHs and a Blender side-by-side viewer under docs/research/assets/.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gpu-busy-blind (High) is deferred to the GPU tenancy gate cut, which owns 'is the GPU free'; fist-clip bound recorded as bf16 numerics. Raven runs 02ca1cd.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replaces the 2026-06-12 timing paragraph with the cache, the child, and the numbers
measured on Raven on 2026-10-22. Adds the tests hand mutation showed missing: the
exact 2 GiB boundary, a child answer with too few vectors, and the 4096 dimension.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The worker builds CachedTextEncoder and passes it to load_model, so Kimodo's own
encoder selection (and the CPU encoder behind TEXT_ENCODER_DEVICE) is never built.
generate() returns the per-request encoder record (identity, hits, misses, seconds,
vector hashes) and the worker identity; /health adds the identity and stored count.
startKimodoWorker drops TEXT_ENCODER_DEVICE and mounts the scripts directory, the
cultcache-py source and .vili/embeddings, with the Kimodo commit and image id read
once per worker start. The encode clock is perf_counter: time.time went backwards
under WSL during a probe.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
text_encoding.py holds the CultCache embedding store (gamecult.vili.text_embedding.v0
through cultcache-py), the encoder identity and key, plan_placement with its typed
gpu-busy refusal, and CachedTextEncoder, which sends a request's misses to one child
call. kimodo-text-encode.py is that child: it assembles Kimodo's LLM2Vec encoder in
bf16 with its own loader (operator ruling encoder-own-loader; peft's offload path
raises TypeError in Kimodo's loader), encodes each text alone, writes a MessagePack
answer and exits. GPU reserve 3 GiB: four cold runs on Raven's 16 GB card produced
identical vectors with no CUDA error. The worker and daemon are not wired yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Serves motion-pipeline fork: which motion model and text encoder Vili runs
(follow_up kimodo-v1-2).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It published idunn.daemon_health to 10.77.0.2:17870 (the retired Windows Idunn)
every 15 s and failed each time; Idunn on Yggdrasil binds no RUDP health ingress.
Deletes idunn-rudp.cjs, the publisher and its status, the --idunn-* args in the
daemon and the three launcher scripts, operator_state slot 4, and the
cultnet-ts and direct @msgpack/msgpack dependencies. transportProfile now states
local-cultcache-only. All commits build (node --check passes); not yet run on Raven.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Campaign motion-pipeline, opened 2026-10-20 at the operator's word; first
consumer is the StreamPixels RPS duel.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
