Questlock 0.3.0 — measured delegation and reuse

Questlock 0.3.0 — measurable bounded tasks and local delegation

Repeated tasks need observable resource use and an acceptance gate before a
worker returns generated text. Version 0.3.0 adds current-run measurements,
optional content-free Langfuse export, deterministic evaluation, and a Go/MCP
worker with an explicit local-first policy. Published 0.2.0 and older releases
remain available.

Release and downloads ·
Installation · Container ·
Validation record

New in 0.3.0

• Per-invocation Runner phase metrics, separate from durable result identity and
  historical generation timing. Optional local Ollama streaming observes first
  and last content arrival; reuse has no fresh generation or TTFT.
• Opt-in Langfuse model metadata export: one root plus one generation child for a
  fresh attempt, one root with zero new usage for reuse, exclusive cache/reasoning
  token buckets, and optional declared-check scores. No prompt or answer text is
  exported. Export failures never retry a model call. MCP uses a silent bounded
  best-effort queue that can drop telemetry.
• model evaluate with three simple and three complex public cases, explicit
  configurations, private reports, and separate fixture/live modes. A report's
  selection changes no model or future routing policy.
• worker run and stdio worker serve; the sole MCP tool accepts an approved
  request path inside a declared workspace. No shell, file editing, file discovery,
  or full-session interception is exposed. Concurrent calls return worker_busy.
• At most one explicit escalation after an error-free complete, durable,
  valid-known-usage answer fails declared checks. Unknown, incomplete, budget,
  provider, publication, and settlement outcomes stop. Hosted escalation requires
  process --allow-paid, explicit prices, and per-call/daily budgets.
• Local Ollama work neither consumes nor clears the API dollar ledger; it can
  continue when paid admission is frozen. Hardware and runtime costs remain.
• An isolated disposable Langfuse v4.59.0 stack with six content-pinned service
  images and a reusable required CI contract. Actual API readback verified three
  roots, two generation children, exclusive usage/cost fields, streaming timing,
  content-free metadata, and root quality scores 1/1/0.

Try it

After extracting your platform's archive:

./questlock model quest
./questlock model evaluate
./questlock model install --agents codex,claude

Both demo commands use simulated public fixtures and need no model account.
The worker needs an existing authenticated broker, an approved request, and
installed local weights. See measurements and Langfuse setup,
bounded delegation and native MCP registration, and the
installation guide.

To build the code you checked out instead, run:

go build -trimpath -o bin/questlock ./cmd/questlock
./bin/questlock worker run --request "$HOME/questlock-workspace/approved-task.json"

Observed result and release limits

A bounded MCP ticket extraction passed strict JSON and exact-value checks with
58 input / 21 output tokens. One exact MCP repeat returned the saved passing
result with zero new usage and no generation. The SDK calls took 5,524.343 ms and
27.584 ms respectively, including transport/provider setup; these are two
observations, not a general speedup or invoice saving.

The broader live local suite completed twelve fresh durable generations and
twelve exact reuses with zero new generation usage. compact-64 passed 4/6 task
checks; balanced-256 passed 3/6. No configuration qualified, so the live
evaluation exited 1. A separate feature extraction withheld its failed answer,
including on exact reuse. These observations demonstrate measured reuse and the
quality guard, not general local-model quality or a reliable latency ranking.
See the content-free measurement record.

The usage counters cover delegated provider/local generation and exclude the
surrounding hosted assistant's prompt, tool, and context tokens. Request
preparation and result review can still consume its normal allowance. Langfuse
export needs an explicit project, endpoint, and keys; the CI stack uses a separate
disposable project and cleans it up.

Direct OpenAI, Anthropic, and local Ollama remain supported. A capability-aware
OpenRouter companion adapter is future work. Questlock does not train models,
choose them automatically, replace an entire hosted chat, or bypass paid limits.
Native archive and container publication evidence is recorded in the
validation record.

---