Every internal failure was mapped to a fixed sanitized code before anything
recorded the cause, so an intermittent execution_failed was undebuggable: an
exhausted model-call budget, a model turn with no text, and a genuine crash all
looked identical.
The bridge now writes one bounded line to container stderr on failure with the
code, the internal reason, the state and both call counters, and the failure
sites pass a reason (budget exhausted, empty final answer with its content block
types, token counts against the limits). The wire error is unchanged.
The backend logs the agent's terminal code together with its own call counters,
and the sanitized tail of container stderr rather than only its byte count.
deploy/logging.json gives every logger a timestamp (uvicorn's default config
leaves non-uvicorn loggers on logging's fallback handler); override with
CONFLUENCE_WEB_LOG_CONFIG.
The bridge counted tool and model calls against fixed 100 and 50 caps, so
CONFLUENCE_WEB_MAX_CONFLUENCE_CALLS / CONFLUENCE_WEB_MAX_MODEL_CALLS above the
defaults were cut short inside the container, and the model cap failed the run
outright. The backend now passes both values into the container at start, the
supervisor forwards them to the bridge, and LIMITS reads them with the same
1-1000 validation as the backend, falling back to 100/50 when absent or
invalid. Add a test that loads LIMITS in fresh processes with good and bad
values for both budgets and the deadline.
The 100 Confluence / 50 model call caps per query were fixed Settings
defaults. CONFLUENCE_WEB_MAX_CONFLUENCE_CALLS and CONFLUENCE_WEB_MAX_MODEL_CALLS
now set them (1-1000 each, validated at startup). Cache hits stay free. The
request history still keeps at most 100 tool entries, so a Confluence budget
above 100 drops later entries with the existing history_overflow warning;
documented next to the setting.
In-memory AdmissionController (idle/reserved/running, FIFO tickets keyed by
the session cookie, reservation and heartbeat expiry, promotion after
cleanup, EMA wait estimate) behind POST /api/v1/queue/join,
GET /api/v1/queue/status and DELETE /api/v1/queue/ticket. POST /api/v1/query
claims the session's reservation first and joins implicitly when idle.
Settings: CONFLUENCE_WEB_QUEUE_{RESERVATION_SECONDS,HEARTBEAT_SECONDS,MAX_LENGTH}.
The 180 s query cap was enforced independently by the backend clamp, the
agent limits, and the container supervisor. All three now follow
CONFLUENCE_WEB_MAX_DEADLINE_SECONDS (default 900, allowed 60-3600): the
backend passes it into the container at launch, the supervisor reads it and
forwards it to the bridge, and both fall back to 900 s on invalid input. The
query timeout must not exceed it (startup fails otherwise). The supervisor
keeps a separate 180 s guard for a container that never receives a start
frame.
Add assets/book.svg as the tab icon: the backend serves assets/ and the CSP
allows same-origin images (the sanitizer still never emits <img>).
scripts/make-tls.sh creates a CA whose critical nameConstraints permit only the
listed IPv4 addresses (and the .invalid DNS subtree), plus an end-entity server
certificate; run-backend.sh passes CONFLUENCE_WEB_TLS_CERT/KEY to uvicorn.
- tests/integration: frontend served by backend, runtime error mapping,
backend + real pi image with scripted model/Confluence (shared example,
variants, failure paths, cancellation/busy gate, isolation canaries,
HTTP download ownership, 16 MiB prompt round trip), real OpenAI-compatible
adapter + real image over a scripted transport, real Chrome against the
real backend with a scripted runtime peer, backend crash/restart
reconciliation, and an opt-in live model check (marker: live).
- backend: map runtime terminal codes (model_output_limit,
model_context_exceeded, query_timeout, connectivity_failed) to the
contract's HTTP statuses; make the artifact 404 body identical for
no-session, wrong-session, unknown and expired IDs.
- Makefile, scripts/run-backend.sh, deploy/confluence-web.env.example,
root README for the integrated application; integration pytest marker.
