vh
7fdaf3bd23
The long-turn "swaps to German" garble was NOT a language leak (infra-ops's
initial framing) and NOT the sampling entropy my interim curb targeted. The real
cause, signal-measured by infra-ops: the Chatterbox Turbo model over-runs its
generation TAIL — a long single generation degrades into garble/dead-air in its
final ~2-3s (voiced-tail zero-crossing rate 1.58x the middle). The gateway's
unbounded chunk-size ratchet built 300-600 char mega-chunks that landed in that
zone, and streaming concatenated each bad tail.
My interim curb (top_k 1000->80, top_p 0.95->0.85, temp 0.8->0.5) made it WORSE:
tight sampling pulls the degradation onset to a shorter length (~200 chars vs
~300 at defaults), so it fights the server-side fix rather than helping.
Fixed server-side (infra-ops, chatterbox-fast image :v2): a max_chunk_chars=250
cap bounds each generation below the ~300-char onset -> clean prosodic chunks
(verified ZCR 1.58x -> 0.64x; operator ear-confirmed clean audio + clean joins).
Consumer side, this commit:
- Revert the sampling knobs: gateway_body back to {text, voice, format, stream},
send full text with the gateway's default sampling. The server chunks at 250.
- Keep the /api/tts empty-200 -> 503 guard as hygiene (DEC-9a; the shared-3090
OOM that produced empty 200s is also resolved — Zonos moved off the card).
Contract DEC-9 rewritten with the resolved root cause + the two wrong hypotheses;
DEC-9a marked kept-as-hygiene. 520 green.
Ratatoskr
A Worldtree Conversation API debug console. Runs up and down Worldtree's API surface — sessions, turns, persona, tools, admin events, Bifrost state — carrying messages between layers. Like the squirrel.
The product is the observability surface; chat is the input mechanism.
Devs run Ratatoskr against a local Worldtree to watch a turn flow through
every layer of the system, side-by-side. The interactive surface is the
web console (ratatoskr-web); a headless --send CLI drives scripted smokes.
Status
v0 scaffold. Design locked; implementation hasn't started. The dev team owns the implementation pass.
Read in this order
docs/design-brief.md— the locked design. Read this first. Every architectural decision is recorded with its rationale, the alternatives considered, and (where relevant) the operator's lock-in moment.docs/SPEC-PIN.md— what Worldtree spec version Ratatoskr is built against, where the vendored snapshot lives, and how to bump the pin.docs/conversation-api-spec.md— the vendored Worldtree spec snapshot. Read this to understand the API surface Ratatoskr consumes. Do not import anything from a Worldtree checkout — the boundary is the spec, not the code. Seedocs/design-brief.md§2.CLAUDE.md— Claude Code conventions for this repo (mostly inherited from the Corviduo template).persistent-memory.md— durable intent across context resets. Update as decisions and state evolve.
Quickstart
# 1. Environment
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"
# 2. Verify the spec pin
cat docs/SPEC-PIN.md # documented Worldtree SHA + bump procedure
# 3. Tests (none yet; scaffold only)
uv run pytest
# 4. Run against a local Worldtree (Worldtree must be running)
# Interactive web console:
ratatoskr-web --host 0.0.0.0 --port 8765
# Headless CLI (scripted smoke):
ratatoskr --send "hello" --new --agent mimir --api-key "$WORLDTREE_API_KEY"
What this repo is NOT
- NOT a polished consumer for Worldtree's end-users — that's the web app.
- NOT a Worldtree admin tool — admin CLI is separate (
sessions_cli.pylives in Worldtree). - NOT a featuretracking shadow of the web app — when the Conversation API surface grows, Ratatoskr does not necessarily grow with it.
- NOT a remote-Worldtree debug client — file-tail surfaces (persona log, server log) assume local-dev posture.
The full negative-clause list lives in docs/design-brief.md §6.
Boundary rule
Ratatoskr depends on a small, fixed surface:
httpx+httpx-sse(network layer)starlette+uvicorn(the web console; thewebextra)- Worldtree's published Conversation API spec at the pinned SHA
Hard rule: no imports from a Worldtree checkout. No core.* imports,
no from worldtree.*, no shared models, no submodule of Worldtree. The
spec is the entire surface. Boundary smoke test at tests/test_no_worldtree_imports.py
enforces this in CI.
Version-skew strategy
The dev team is decoupled from Worldtree's dev team. To detect drift when Worldtree changes the API:
- Spec-version pin in
pyproject.toml(worldtree-spec-rev). Bump explicitly; bumps are a tracked action. - Recorded-SSE snapshot tests at
tests/snapshots/. Captured against a live Worldtree; replayed in CI. Re-record after every pin bump. - Conformance smoke test — boots Worldtree via Docker compose in CI, runs a one-turn happy path. Catches integration-level drift.
See docs/SPEC-PIN.md for the bump procedure.
Consumer-side discoveries
If you find a spec gap, ambiguity, or missing-but-needed endpoint while working on Ratatoskr: route the discovery back to worldtree-dev via althing rather than via PR on Worldtree directly. The separate-team boundary is intentional and helps catch spec gaps that an in-tree consumer would paper over.
Related repos
- Worldtree — the API Ratatoskr consumes
- brokkr-smithy — authored this design brief
- Skaldsong — another Worldtree API consumer (Python; render-and-aggregate pattern)
- mead-hall — another Worldtree API consumer (TypeScript; server-side broker)