You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Scope: first-class Docker deployment support, stale working-agent recovery, official direct-API preset onboarding, Kimi Code provider support, and development-default API assignment across hydrated office packs.
Highlights
1. Docker Deployment Is Now a First-Class Path
Added a production-oriented Dockerfile, .dockerignore, and docker-compose.yml.
Container runtime now uses a non-root user and standardized writable data locations.
Runtime path resolution was updated so database/log paths behave correctly inside Docker deployments.
README deployment guidance was added across English, Korean, Japanese, and Chinese.
2. Stale working Agents Now Recover Automatically
Startup and lifecycle sweeps now detect agents that are still marked working even though their current task is missing or no longer in_progress.
Those agents are reset back to idle with current_task_id = NULL.
Result: server restarts and interrupted runs no longer leave misleading “still working” agents behind in the UI.
3. Official Direct-API Presets Reduce Manual Setup Friction
Settings > API now includes official presets for:
OpenCode Go (OpenAI)
OpenCode Go (Anthropic)
Bailian Coding Plan (OpenAI)
Bailian Coding Plan (Anthropic)
Presets now:
lock and manage the expected Base URL
seed fallback model lists immediately
keep refresh/retry explicit instead of silently retry-looping after failures
show preset identity directly on provider cards
Existing provider-card model search remains available, so large cached model lists are still easy to filter before assignment.
4. Kimi Code Is Now Treated as a First-Class Provider
Kimi Code is now supported through the CLI execution path alongside the existing local coding providers.
The integration was completed across:
provider selection and display labels
task execution/provider validation
prompt-skill rendering
skill learn/unlearn provider routing
video-preprod remotion bootstrap
provider checks in existing database schema/runtime migration paths
Result: Kimi is no longer just a partial label-level addition; it is wired through the execution and skills flow end to end.
5. API Model Assignment Now Starts With Development and Includes Initialized Packs
The assign flow now always loads the development office as the default baseline.
If a user has already completed initial setup for other office packs and those packs have been hydrated into runtime state, their agents and departments also appear in the assignment modal.
Pack sections are now grouped safely, so packs that reuse the same department IDs no longer collapse into each other inside the modal.
Result: default development assignment remains predictable, while already-initialized office packs can also receive direct API model mappings from the same settings flow.
6. Local E2E Validation Is Safer and Leaves Less Residue
pnpm run test:e2e now resets the isolated .tmp/e2e-runtime database both before and after the Playwright run.
Playwright no longer reuses an already-running local 8810 server by default. Reuse is now opt-in via PW_REUSE_EXISTING_SERVER=1.
CRUD-heavy E2E scenarios were hardened so local validation does not keep leaving stale departments, agents, and similar CI residue behind during normal development verification.
Result: local E2E runs are more predictable, are less likely to touch the wrong runtime, and clean up their temporary state more aggressively.
PR Attribution
Includes PR #56: Docker deployment support (Dockerfile, compose, runtime path handling, multilingual docs)
Includes PR #60: stale/orphan working-agent recovery during lifecycle startup and interval sweeps
Includes intent-preserving integration of current open PR work:
PR #62: development-pack-scoped API assignment departments
PR #63: Kimi Code provider support and migration/runtime completion
PR #64: official API presets, preset refresh hardening, and provider-card UX updates
post-integration follow-up: API assign modal now includes hydrated office packs while keeping development as the default baseline