Added
- UberLLM is available as its own OpenAI-compatible provider seat. Select a
model explicitly through its environment variables orproviders.json.
Readiness, fleet selection and MCP honor that model and its allowlist.
Gateway seats remain outside Council voting. See #1288.
Fixed
- Named OpenAI-compatible presets retain their own endpoint, credential and
model settings when the generic slot is configured. Direct helper calls no
longer send a preset's credential to the generic slot's host or inherit its
model. Atlas Cloud admission and dispatch share configured defaults and exact
model pins. Explicit flags keep precedence. Set the preset-specific model
variable or provider default if you previously usedOPENAI_COMPAT_MODEL.
See #1300 and
#1302, addressing the
isolation reports in #1299.
Custom-provider registration remains a separate open enhancement. - Managed Tangle validates declared output artifacts separately from read-only
context references. Missing outputs and unavailable coverage evidence remain
blocking. See #1289. - Council reviews no longer reject reviewer-created test inputs as fabricated
source quotations. Explicit source claims remain checked. Optional partial
quote refinement has bounded work budgets and cannot prevent required
evidence checks. See #1260. - Research summaries track each seat separately, so one seat's timeout cannot
hide another seat's successful output. Interrupted workers and reclaimed
output claims keep accurate status and preserve saved results. See
#1243. - Quota and authentication failures retain their cause and retry behavior.
Cleanup stops owned processes, preserves external termination status and
leaves other seats' status intact. Circuit breakers recover without turning
an expired cooldown into an indefinite block. See #1266. - The token-extraction CLI starts through its TypeScript runner, uses checked
options and handles CSS names and TypeScript output generation correctly.
Its Node.js requirement matches its dependencies. See #1298.
Documentation and maintenance
- Dispatch documentation explains fallback selection and fail-loud behavior
consistently. See #1227. - Full unit coverage retains every suite and its timeout while using measured
timings across six macOS shards and four Linux shards. - The Tangle fix is listed here because it landed after the published
11.22.2tag, despite its earlier changelog placement.