Changelog.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog. Versioning: YYYY.M.D (date-based, e.g. 2026.2.22). Multiple releases per day use -N suffix (e.g. 2026.2.22-2).

2026.8.7

Fixed

• fix/957-store-lifecycle — hoist the LangGraph store to a per-process singleton, and remove the async with <singleton-store> re-entry that hoisting would otherwise multiply. get_store_db() returned an async context manager that built a fresh AsyncPostgresStore on every call, each with its own psycopg pool of DB_POOL_MIN_SIZE (5, max 20) connections that psycopg opens eagerly. Six non-test sites called it: the FastAPI lifespan, WorkerState, both APScheduler jobs, and two paths in workers/tasks.py. services/db.py now owns one store per OS process behind get_shared_store()/close_shared_store() — an async def returning the store rather than a context manager, precisely so no call site can be written as async with again — with double-checked locking, no caching of a failed construction, and an explicit close wired into the lifespan and WorkerState. Per-process, not per-app: correcting the issue's suggested app.state sharing, app.state is unreachable from the taskiq worker processes and from the module-level pickled scheduler job functions, which have no Request in scope. Also correcting the issue's stated acceptance check — measured against the pre-fix code, the connection count does not step up across sequential task runs, because each async with get_store_db() closes its pool on exit. The real cost is two-shaped: five connections established and torn down on the critical path of every task and job, and 5 x C simultaneous backends at concurrency C. tests/integration/test_store_pool.py measures the second directly against a live Postgres — three concurrent callers hold 15 backends on the old pattern and 5 on the new one. Folded in, and the reason this is one PR rather than two: services/assistant.py:227, services/assistant.py:294 and services/prompt/__init__.py:168 already did async with self.store on the app-wide singleton handed out by get_store — the same defect PR #964 fixed in the health probe, and the three sites that entry explicitly deferred. Hoisting the four factory sites without this cleanup would have converted them into re-entry too, taking 3 latent sites to 7 and spreading them across the worker and scheduler paths. All three now await asearch on the injected store directly; search_public's isinstance(..., InMemoryStore) guard collapses because both arms became identical, while the two guards whose arms genuinely differ — the 3-attempt "connection"/"closed" retry, and the public-namespace derivation in PromptService — are kept verbatim, along with both sort keys. Still latent, not an active outage, on the lock-pinned langgraph-checkpoint-postgres 3.1.0: AsyncPostgresStore.__aenter__ is a bare return self and __aexit__ only stops a TTL sweeper that is never started, because ttl= is never configured. One langgraph bump or one ttl= makes all of it live at once, which is why get_store_db's docstring now says not to add one. Two pool-kwargs additions come along: TCP keepalives copied from get_checkpoint_connection_kwargs (a pool that lived for one call did not need them; a process-lifetime pool behind a pooler does, or hoisting the lifetime just trades a connection leak for a half-open-connection staleness bug), and an application_name of orchestra-store-<pid> so the pg_stat_activity count is attributable per process — consumed by the new tests rather than merely offered. _RunScopedStore moves from workers/tasks.py to services/db.py as RunScopedStore and now wraps the shared store at all four migrated sites: six sites assign store.fields, and although nothing reads that attribute today (it is not a LangGraph API), sharing one store object across concurrent runs is what would make it matter. Its fields are now seeded explicitly instead of from the shared object, which would have let one run inherit another's value. The lifespan additionally shuts the scheduler down before closing the store — its jobs consume the same singleton and the distillation job carries misfire_grace_time=3600, so one can fire during teardown — and builds the store before scheduler.start(), so a restored misfiring job cannot be the first caller. 33 tests across four new files: singleton identity by construction count rather than object identity, failed-construction recovery, close/rebuild/idempotence/mock-tolerance, RunScopedStore isolation, the lifespan wiring (which nothing previously executed, since ASGITransport emits no lifespan events), the live-Postgres pool counts, and the re-entry pins. The re-entry doubles deliberately do not subclass InMemoryStore — copying test_health_store.py's RecordingStore would route into the in-memory arm, never reach the async with, and pass against the pre-fix code — and the route test installs its get_store override on both app and api_app, because main.py splices *api_app.routes by reference so an override on app alone is inert. All four re-entry tests fail against the pre-fix code on the entry counter itself. An autouse fixture in tests/conftest.py resets the singleton around every test; it is synchronous because pytest does not apply async autouse fixtures to the unittest.IsolatedAsyncioTestCase classes in test_distill.py, which build a fresh event loop per method — without it the first test to construct the store pins its double for the whole session and silently voids four existing patch("src.services.db.get_store_db", ...) sites. Left for a follow-up: ScheduleService.__init__ defaults to get_store_in_memory(), so the module-level schedule_service holds an InMemoryStore — but self.store is never read anywhere in the class, so it is a dead field rather than a wrong-backend data path; and routes/v0/schedule.py mutates schedule_service.user_id per request on that same module-level instance.
• fix/968-banner-contrast — make destructive colour readable at the palette level. components/lists/ChatMessages.tsx:293 rendered the run-error banner as border-destructive/40 bg-destructive/10 text-destructive, measuring 3.30 / 1.93 / 1.08 (light/dark/gray) against WCAG 1.4.3's 4.5:1 — at text-sm, so not large-text exempt, and effectively invisible in gray. Two findings widened the issue as filed. First, the tint is barely a factor: text-destructive measures 3.76 / 1.99 / 1.05 on plain --background too, so there was no "compliant on opaque surfaces" bucket — --destructive is a fill token, designed to sit behind --destructive-foreground on a button, and is not legible as ink anywhere. Second, the border failed as well: border-destructive/40 measures 1.72 / 1.19 / 1.08 against WCAG 1.4.11's 3:1, so in dark and gray the banner had no perceptible red at all. Rather than redefine --destructive (which would move every variant="destructive" button and badge), this adds a second token --destructive-accent — light 0 74% 42%, dark 0 91% 71%, gray 0 100% 92% — contracted as ink on a surface: 6.41 / 7.24 / 5.68 on the background and 5.63 / 7.04 / 5.86 on the bg-destructive/10 tint, with border-destructive-accent/70 at 3.82 / 3.98 / 3.69 (/70 is the first opacity clearing 3:1 against both the page and the tint in all three themes). Inside tinted error containers the container carries the signal, so prose moves to text-foreground (17.47 / 18.53 / 6.14) and the accent goes on the icon and border, matching the chip shipped in #965; standalone inline field errors, where the red is the message, take text-destructive-accent. Every opacity modifier was stripped from destructive text — alpha composites toward the surface and always lowers the ratio (text-destructive/70 measured 2.39 / 1.50 / 1.01). Applied across all 39 non--foreground call sites, led by the two shared primitives ui/alert.tsx and ui/form.tsx, which fix every <Alert variant="destructive"> and <FormMessage/> consumer at once; ui/alert.tsx's dark:border-destructive had to be deleted rather than swapped, since Tailwind emits it at specificity (0,2,0) and it would have silently kept the 1.99:1 border in dark. The replay button keeps bg-destructive/text-destructive-foreground — the one legitimate fill use in the banner — and gains only the focus ring it never had (5.29 / 6.77 / 4.39 against its own fill). Two checked-in vitest gates replace the eyeballing that let this survive a year: destructive-contrast.test.ts parses the real token values out of globals.css and asserts every threshold per theme, and destructive-usage.test.ts fails CI on any new text-/border-destructive that is not -foreground or -accent — the arithmetic test pins the palette, but #968 was a usage bug, so the usage gate is the one that prevents the recurrence. Deliberately left, now tracked with measured numbers: destructive fills are untouched, so --destructive-foreground on --destructive still measures 3.60:1 in light (#969), and gray's --muted/--secondary/--accent (220 15% 45%) cannot host any label at 4.5:1 — its own --accent-foreground reaches only 4.12 — which is also why gray's accent is a pale pink rather than a red (#970).
• fix/966-toast-audit — audit the ~101 toast call sites that #965 made live in one change, and stop the four that storm. Highest severity was hooks/useChat.ts: toast.error("MCP sandbox unreachable", { duration: Infinity }) fired per SSE payload with no id and no guard, so three of them permanently occupied every slot under visibleToasts={3} and starved the notification surface app-wide — a real outage could silence the very surface you need during the debugging session. It is no longer a toast at all. An unreachable sandbox is a terminal failure of a run the transcript is visibly waiting on, which is exactly what the existing RunErrorBanner was built for (the streamMode === "error" branch 15 lines above already routes there), and a setRunError state write is idempotent by construction — N events collapse to one banner with no id bookkeeping to get wrong later. Found while implementing it: the branch was near-unreachable in the first place. The backend emits mcp_sandbox_unreachable (agents/__init__.py:353) but convertEventToLegacy had no case for it and both parseEvent switches in lib/utils/fetchStreamReader.ts fell through to default: return null, so on the unified stream path the event was silently dropped and the run just stopped with no explanation — a worse bug than the one reported. The event is now typed in lib/entities/stream.ts, wired through both readers and the legacy converter, and added to both terminal-event sets (streamSource.ts's hasTerminalEvent and clearDistributedRecovery) — it had been surviving only because the distributed emitter happens to append a trailing done that the sync emitter does not, and without that accident the reader would have retried five times and then fallen into the non-recovery onError branch, firing a blocking alert() and wiping the user's message behind the new banner. ChatMessages now imports RunError instead of redeclaring its own copy, and the DLQ Replay button — which rendered unconditionally while recoverable was declared and never read — is now gated on recoverable !== false && runId, because replaying a run that never reached the DLQ dead-ends. Related: metadata is captured by closure at startManagedStream render time while run_id arrives in a later event, so both error branches now read a metadataRef and the banner's correlation id is actually populated.
• fix/966-toast-audit — stop the remaining three storms. hooks/useChat.ts:319: a recoveryMode error event toasted per event and returned without closing the source; it now aborts the controller, closes the stream, and notifies once per stream behind a closure latch and a stable id — per-stream rather than global, so a genuinely new stream still gets its own notification. The stream.onError sibling shares that latch and id; the 404/409 informational toast deliberately does not, since sonner merges by id and would have overwritten the error text with a reassurance. hooks/useMessageQueue.ts: the drop-after-max-retries toast reschedules itself every 100ms, so one outage emitted ~30 toasts; drops inside a 5s window now accumulate into one toast under a fixed id ("3 messages dropped after max retries"), and the per-retry warning keeps its copy under a second, deliberately distinct id — "still trying" and "gave up" are different states and must not clobber each other. The per-message query preview is dropped from the plural case, where it can only describe one of N. pages/memories/edit.tsx — not in the issue, same class, cheap — was a mount effect that toasted and navigated with no id and no guard, giving two of each under <StrictMode>; it now has both a ref guard and a stable id.
• fix/966-toast-audit — fix the retry loop underneath the context/ChatContext.tsx save toast. lastSavedPersistentSignatureRef was only advanced inside the try, so after a failed PATCH the signature stayed dirty and every subsequent keystroke re-fired a failing request and an identical toast, roughly one per 500ms idle window with the backend down. Advancing the ref in the catch is the tempting one-liner and is a data-loss bug — that ref is also the dirty indicator and the autosave-skip guard, so advancing it marks unsaved text as saved. Instead, autosave failures now back off exponentially (2s doubling to a 30s cap) and re-arm a real retry at the deadline, because the only other autosave trigger is a signature change: a user who types, sees the failure, and stops typing would otherwise be stranded with unsaved content and no pending timer, and hasUnsavedPersistentChanges has no UI consumer to tell them. Manual saves bypass the backoff entirely, so the Save button is always a live escape hatch, and their toast carries no dedupe id — deduping a click response would violate the surrounding policy. The issue's alternative (follow useScheduleExecutions.ts:40 and decline to toast at all) was rejected: unsaved user text is at stake, so this earns exactly one toast.
• fix/966-toast-audit — apply one policy, "the hook owns the toast", across the schedules stack, replacing eight double-toast pairs where the hook toasted and rethrew and the component caught and toasted again ("Schedule created successfully" and "Schedule created successfully!" both fired). pages/schedules/index.tsx's delete handler was already correct and is the pattern the other twelve call sites now match. The try/catch blocks and their console.error stay — deleteSchedule rethrows, so removing them would trade a duplicate toast for an unhandled rejection — and the page's bare await deleteSchedule(...) gains the catch it was missing. Two related defects surfaced while applying it. useAgentSchedules.createSchedule/updateSchedule early-returned on a falsy agentId before any toast and before any API call, so the panel was reporting success for a write that never happened; they now report the error and throw, which also keeps the dialog open. And a create whose refresh failed emitted 2 success + 1 error simultaneously, leaving the user staring at "created successfully" beside "failed to load" and a list without their new schedule — the rational read is that it failed, and the retry creates a duplicate; post-write refreshes are now silent, so the write reports its own outcome once. Load-failure toasts get stable ids (they fire from mount effects, and there are two live useSchedules consumers — the page and the sidebar panel — so an outage stacked one toast each, doubled again under StrictMode); create/update/delete deliberately get none, since they answer a click and must fire every time.
• fix/958-store-health-reentry — stop GET /api/info/health/store from entering the application-wide store as an async context manager. get_store returns req.app.state.store, the single AsyncPostgresStore created once in the lifespan, so the handler's async with store as s: ran __aenter__/__aexit__ on a singleton whose lifecycle it does not own — a diagnostic endpoint manipulating the resource it measures. The handler now awaits store.asearch(...) on the injected instance directly, inside the same 5s asyncio.timeout; the response body, the timeout→503 branch, and the "connection"/"closed"→503 mapping are all unchanged. Correcting the issue as filed: this was latent, not an active outage. On the pinned langgraph-checkpoint-postgres 3.1.0, AsyncPostgresStore.__aenter__ is a bare return self and __aexit__ only stops a TTL sweeper task that is never started (the store-level ttl= is never configured, so _ttl_sweeper_task is always None); the pool is closed solely by from_conn_string's own async with, which unwinds at lifespan exit. The reported pool exhaustion therefore does not reproduce today — but a future ttl= config, or a langgraph version whose __aexit__ releases resources, would have degraded the process-wide store silently. The endpoint had zero tests, which is how the pattern survived; backend/tests/integration/test_health_store.py adds five, covering the happy path, both 503 branches, the 500 fallback, and — the one that pins the fix — a store double that is a perfectly valid context manager and simply counts entries, so the pre-fix handler fails on aenter_calls == 1 rather than on an incidental error. All five fail against the pre-fix handler. Scoped deliberately to the health probe: services/assistant.py:227,294 and services/prompt/__init__.py:168 re-enter the same singleton and are left for a follow-up.
• fix/960-toast-and-model-picker — stop a backend outage from rendering as a silently vanishing model picker. Two independent defects combined to produce it. First, the app could not display an error at all: sonner was a dependency and components/ui/sonner.tsx exported a Toaster, but that component was imported nowhere, so all 101 non-test toast.* calls across 25 files were no-ops (the issue says 66; the real count is 101). <Toaster /> is now mounted once in main.tsx under ThemeProvider, at top-center — every bottom position collides with the bottom-anchored composer, the split-view panel, AgentMenu's mobile sheet, or the query devtools, and below ~600px sonner goes full-width and blankets the composer entirely. Second, useModel destructured only data and defaulted it to an empty payload, so ChatInput's {displayModel && ...} gate collapsed error, empty, and still-loading into one output: nothing. The picker is extracted into components/inputs/ModelPicker.tsx and now renders five distinct states — nothing when unauthenticated, a skeleton while loading, an enabled retry chip on error (a disabled control would say "not for you" rather than "not right now"), a static No models available chip on a zero-model success, and the normal picker otherwise — all at a uniform 24px height so the action row does not reflow. The composer stays fully usable during an outage: displayModel is display-only and the server resolves the default, so the textarea and submit button are never disabled.
• fix/960-toast-and-model-picker — make ui/sonner.tsx read the app's own theme. It imported useTheme from next-themes, but no NextThemesProvider is mounted anywhere, so theme was always undefined and fell back to "system" — a user on the default dark UI with a light OS got white toasts. It now reads ThemeProviderContext via @/hooks/useTheme and maps the app's four-value Theme union onto sonner's three, with "gray" mapped to "dark" (verified against .gray's --background: 220 15% 35%, not inferred from the name). Measured in-browser: the toast background now matches the app background exactly in light, gray, and dark. next-themes is left in package.json for a follow-up.
• fix/960-toast-and-model-picker — guarantee one toast per outage instead of a stack. New lib/utils/connectionToast.ts exposes notifyConnectionLost/notifyConnectionRestored behind a single fixed sonner id, so N concurrent failures replace in place rather than stacking, and recovery replaces the error rather than sitting contradictorily beside it. It fires from one effect in useModel, ref-guarded on the false → true transition of isError so <StrictMode>'s double-invoked effects stay idempotent, and never when unauthenticated. The six mount-time and edit-load failure paths in useSchedules, useAgentSchedules, MemorySettings, memories/edit, schedules/index, and AgentSchedulesPanel route through the same helper only when the error carries no HTTP response; anything the server actually answered keeps its specific message, and every user-initiated toast is untouched. Verified in-browser: /schedules and /memories each show exactly one toast with both their own endpoint and /llm/models failing.
• fix/960-toast-and-model-picker — set retry: 0 on the models query. apiClient already retries once and the global queryClient retried again on top, compounding to roughly 40s before isError ever flipped — far past the point a user has given up. The global default is unchanged; retry resolves per query rather than per observer in TanStack Query v5, so useModelsList, which shares the ["models"] key, carries the same override or it would restore retry: 1 whenever it initiated the fetch. refetchOnWindowFocus and refetchOnReconnect are enabled on this query so recovery is hands-free.
• fix/960-toast-and-model-picker — fix two adjacent states that also rendered as nothing. Hiding every model in Settings now still renders the normal picker (user-caused and user-fixable, so it must not look like an outage) with CommandEmpty reading All models are hidden. Turn some on in Settings → Model Visibility.; a successful load with no default renders the picker labelled Select model rather than disappearing. Accessibility: the error chip carries a WifiOff icon and distinct text so it never relies on color alone, the picker trigger gains an aria-label carrying the current model (its only accessible name was title="Change default model"), and a polite live region announces the outage while rendering empty — rather than unmounting — when healthy. The chip's label uses text-foreground rather than the run-error text-destructive: measured on the composer, the inherited triple scores 1.93:1 (dark), 3.30:1 (light) and 1.08:1 (gray) at 12px, all failing WCAG 4.5:1, while text-foreground measures 18.5 / 17.5 / 6.1.
• fix/960-toast-and-model-picker — add the missing coverage. ChatInput.tsx had no tests at all (ChatComposer.test.tsx stubbed it out entirely); it now has 13 covering every picker state, keyboard reachability of the retry control, and submission during an outage. useModel.test.ts gains its first failure cases, pinning that the toast fires exactly once across re-renders and not at all when unauthenticated, and ui/sonner.tsx gains a suite asserting the theme mapping for all four values.

2026.8.6

Changed

• task/948-consolidate-docs — retire the wiki git submodule and consolidate documentation into a plain docs/ folder. The submodule was never initialized by CI, so every clone left an empty wiki/ directory that tooling flagged as missing. Vendors the documentation markdown (27 pages) and the 49 screenshots it references into docs/, rewriting root-absolute /img/... links to page-relative paths so they render on GitHub. Drops the docs/ entry from .gitignore (added by de842557 in March, when docs/ was a stale duplicate of decks/) so the path can serve as the documentation home. Updates the dangling README.md link, the AGENTS.md component list, and both issue templates, and adds docs/** to test.yml's paths-ignore so documentation edits no longer trigger the backend and frontend suites, plus a workflow_dispatch trigger so the suite can be re-run manually. Also gates test-e2e behind that manual trigger (if: github.event_name == 'workflow_dispatch'): at ~8m50s it was the workflow's entire critical path against ~2m for test-backend and test-frontend combined, while already carrying continue-on-error: true — so it cost wall-clock on every push without gating anything. Run it from the Actions tab when a change warrants full-stack coverage. The Docusaurus app stays in mifunedev/wiki, so the docs.ruska.ai deploy is unchanged.
• task/950-config-env-path — move the environment file location from ~/.env/orchestra/ to ~/.config/orchestra/. The primary backend env is now ~/.config/orchestra/.env (was .env.backend); the test and frontend envs move to the same directory but keep distinct names (.env.test, .env.frontend) so make test and the pre-commit hook keep running against the test database rather than the development one. Updates backend/Makefile, frontend/package.json, infra/docker-compose.yml, .pre-commit-config.yaml, backend/.vscode/launch.json, the four evals/probes/resiliency-*.sh, the example notebooks, and the agent docs. Existing checkouts need mkdir -p ~/.config/orchestra and to move their env files across.

Fixed

• task/955-db-pool-exhaustion — stop the SQLAlchemy connection pool from exhausting and returning 500 for every authenticated request. verify_credentials and the two get_optional_user* wrappers took the session as a FastAPI dependency, and dependency teardown runs after the path operation function returns — so one pooled connection stayed checked out for the whole request, which for /llm/invoke is an entire agent turn and for the SSE path the whole stream. Auth now opens its own short-lived session around the user lookup only. The engine was also created with no pool arguments at all, silently running on SQLAlchemy's defaults (size 5, overflow 10, timeout 30s); it is now configured explicitly via DB_SQLA_POOL_SIZE/DB_SQLA_POOL_MAX_OVERFLOW/DB_SQLA_POOL_TIMEOUT/DB_SQLA_POOL_RECYCLE with pool_pre_ping. pool_timeout drops to 5s deliberately: queueing silently for 30s is what turned one burst into a cascade of 500s. Symptom was an empty model picker — GET /llm/models 500s for a signed-in user while the same endpoint returns 200 unauthenticated, because the anonymous path never touches the pool.

• task/953-consolidate-api-deps — fold the optional api extra into the backend's default dependencies so a plain uv sync produces a runnable API. fastapi, uvicorn, slowapi, fastapi-cache2, fastmcp, and python-multipart move from [project.optional-dependencies] into [project] dependencies, and --extra api is dropped from all six call sites that passed it (test.yml x2, build.yml, backend.Dockerfile, and both docker-compose.yml services). This repairs .github/actions/backend-install.yml and .github/workflows/deploy-vm.yml, which ran uv sync --frozen --no-cache --no-dev without the extra and so started python main.py against an install with no web framework in it. No documentation changed — plain uv sync is what README.md, AGENTS.md, backend/README.md, and backend/CLAUDE.md already documented. Trade-off: infra/backend.Dockerfile's worker-builder stage deliberately installed without the extra and ran fine that way, so the worker image grows by the 35 packages the extra pulls in (+28 MB, 771 MB → 799 MB measured --no-dev).

• task/961-reasoning-effort — stop OpenAI reasoning models 500ing on every chat. openai:gpt-5.6-luna, the default model, returned 400 Function tools with reasoning_effort are not supported for gpt-5.6-luna in /v1/chat/completions for every request: init_graph passed no reasoning configuration, so ChatOpenAI used Chat Completions, and every Orchestra agent binds function tools (deepagents ships todo/filesystem tools). OpenAI reasoning models now route through the Responses API, the only transport that accepts function tools alongside reasoning — forcing reasoning_effort="none" instead would have kept the old transport by disabling reasoning on the very models chosen for it. Applied in init_graph, so every construction path (sync invoke, SSE stream, distributed worker, scheduled runs) is covered. Non-reasoning models, Anthropic, Gemini, and Bedrock are untouched; ChatBedrockConverse sets extra="forbid", so it is excluded by construction.

Added

• task/955-db-pool-exhaustion — GET /api/info/health/db reports SQLAlchemy pool size, checked-in/checked-out counts, overflow, and configured limits. Pure introspection of the pool object, so it still answers while the pool is exhausted.
• task/961-reasoning-effort — first-class reasoning-effort support. POST /llm/invoke and POST /llm/stream accept reasoning_effort; PATCH /settings/default accepts it as a saved default (request value wins); GET /llm/models gains a reasoning map of the effort values each model accepts plus default_reasoning_effort. Accepted values are read from the models.dev catalogue that already backs the model list rather than hard-coded, because they differ per model (o3 stops at high, gpt-5.6-luna reaches xhigh/max, gpt-5.2-chat-latest takes only medium). An unsupported effort named explicitly in a request is a 422; a saved default a later model does not accept is dropped rather than rejected, so switching models never breaks existing chats. Efforts are sent only to providers whose LangChain class accepts one (OpenAI, xAI, Groq) — Anthropic and Gemini expose thinking budgets instead and are tracked separately. The chat input gains an effort picker beside the model badge, rendered only for models that publish effort values.

2026.6.15

Changed

• feat/mermaid-zoom-support — add zoom in/out/reset controls to rendered Mermaid diagrams in markdown and Mermaid file previews.

2026.6.14

Fixed

• fix/940-mermaid-navigation — preserve Mermaid diagram source before rendering so slide diagrams stay visible after navigating between nested slides.
• task/938-mifune-slide-brand — align the GitHub Pages slide deck and one-page handout with Mifune branding, green-only accents, mifune.dev links, and support@mifune.dev contact details.
• task/936-pages-workflow-source — deprecate the legacy GitHub Pages branch source in favor of the slides.yml GitHub Actions deployment and correct stale deck deployment documentation.

2026.6.13

Changed

• feat/infra-consolidation — consolidate all Docker/build artifacts into a single infra/ directory (one docker-compose.yml + thin docker-compose.test.yml); move backend/Dockerfile to infra/backend.Dockerfile; drop the dev/storage/services/debug overlays and the dozzle service; update Makefile, CI (build.yml/test.yml), and docs.

Fixed

• fix/thread-delete-silent-failure — thread deletion no longer silently fails while returning 204. ServiceContext.delete_thread now deletes the thread record first (the operation the thread list reads from) and treats checkpoint cleanup as best-effort, instead of deleting checkpoints first, aborting the whole delete on any checkpoint failure, and swallowing the exception (return e) so the route reported success regardless. Genuine failures now propagate (404/500) instead of a false 204. Also implements adelete_thread on ResilientAsyncPostgresSaver, which previously inherited BaseCheckpointSaver's NotImplementedError and made every thread delete fail when CHECKPOINT_USE_RESILIENT=true.
• fix/onboarding-tour-persist-on-close — persist onboarding completion when the welcome tour is dismissed via the X button or overlay click (ACTIONS.CLOSE), so it no longer re-fires on every login. Previously only Skip/Done persisted; closing hid the tour for the session but left onboarding_completed unset.
• fix/walkthrough-visibility-zindex — make the onboarding walkthrough's highlighted element clearly visible (darker overlay dim + a vivid spotlight ring) in both light and dark themes, and stop the Help/walkthrough button from overlaying the Files drawer by only elevating it above the tour overlay while the tour is running.

2026.3.6-2

Changed

• feat/840-disable-mcp-server-toggle

Fixed

• fix/855-files-map-not-persisted-across-threads (2026-03-11)

2026.3.4-2

Changed

• feat/838-backend-benchmarks

2026.2.23-5

Changed

• feat/830-model-visibility-backend
• feat/821-docs-screenshots

2026.2.23-4

Changed

• bug/823-schedule-calendar-mobile-overflow

2026.2.23-3

Changed

• feat/826-dedup-ci-jobs

2026.2.23-2

Changed

• feat/826-dedup-ci-jobs (2026-02-23)
• feat-822

2026.2.22

Changed

• feat/820-exec-server-template-edit-mcp
• Adopt YYYY.MM.DD-RR versioning scheme

0.1.0-rc20

Changed

• feat/803-persist-assistant-subagent-selection (2026-02-20)
• feat/807-docs-update-memories-onboarding-mcp (2026-02-20)
• feat/787-migrate-memories-seeder (2026-02-18)
• feat/801-search-threads-tool (2026-02-18)
• feat/736-rlm-skill (2026-02-05)
• feat/722-frontend-schedule-refactor (2026-02-01)
• feat/724-add-watcher-to-worker-reload (2026-02-01)
• feat/715-ubuntu-execution-env (2026-01-31)
• feat/442-able-to-edit-memorys (2026-01-31)
• feat/555-add-compacting-middleware (2026-01-31)
• feat/665-allow-user-configure-default-account-settings (2026-01-30)
• feat/707-ticket-md-ralph-loop (2026-01-30)
• feat/694-show-subagent-tool-calls (2026-01-24)
• feat/504-frontend-queue (2026-01-20)
• feat/680-update-readme-out-of-sync (2026-01-16)
• feat/664-edit-project-information-settings (2026-01-16)
• feat/675-cleanup-presidio-service (2026-01-16)
• feat/673-web-scrape-dump-tool-results (2026-01-15)
• feat/663-auth-user-share-private-thread-link-anon (2026-01-14)
• feat/666-suport-aws-models (2026-01-14)
• feat/658-site-title-status (2026-01-13)
• feat/656-distro-workers-taskiq (2026-01-11)
• feat/634-web-search-tavily-fallback (2026-01-13)
• feat/654-inference-dication (2026-01-10)
• feat/637-user-can-persist-files-to-agent (2026-01-10)

v0.0.2-rc141

Changed

• feat/650-file-treeview-sidebar (2026-01-10)
• feat/633-public-agents-display (2026-01-08)
• feat/471-public-agents (2026-01-08)
• FEAT: Render the todos state from useThread in Session
• FEAT: Not Required to pass assistant to ruska chat
• FEAT: CLI Should have Command for getting current version info and heath info
• FEAT: Truncate Tool Calls. Allow extension (cli) #54

Fixed

• bug/612-web-scrape-cannot-fetch-plain-text-pages (2026-01-31)
• bug/705-fallback-searxng-from-exa (2026-01-29)
• bug/662-fix-db-timeout (2026-01-19)
• bug/677-nonetype-items-attributeerror (2026-01-16)

v0.0.2-rc138

Changed

• feat/560-ctrl-click-thread-sidebar (2025-12-30)
• feat/514-add-context-to-input (2025-12-30)

v0.0.2-rc135

Changed

• feat/622-docs-to-netlify (2025-12-28)

Fixed

v0.0.2-rc134

Changed

Fixed

• bug/619-fix-editor-mode-home-pag (2025-12-27)

v0.0.2-rc133

Changed

• Made updates to synchronize the external branding

v0.0.2-rc132

Changed

• feat/614-native-mcp-support-fast-mcp (2025-12-25)

v0.0.2-rc131

Changed

• feat/602-can-add-files (2025-12-22)

v0.0.2-rc130

Changed

• feat/602-can-add-files (2025-12-22)
• feat/595-tokens-and-secrets (2025-12-21)
• feat/596-deselect-tools (2025-12-21)
• feat/589-correlation-matraix (2025-12-20)
• feat/588-home-page-facelift (2025-12-20)

Fixed

• bug/592-chart-render-issue (2025-12-21)

v0.0.2-rc129

Changed

• feat/595-tokens-and-secrets (2025-12-21)
• feat/596-deselect-tools (2025-12-21)
• feat/589-correlation-matraix (2025-12-20)
• feat/588-home-page-facelift (2025-12-20)

v0.0.2-rc125

Changed

• feat/585-user-defaults (2025-12-17)
• feat/578-api-token-support (2025-12-14)

Fixed

• bug/541-gemini-stream-stop-fix (2025-12-19)
• bug/575-api-docs-not-showing-in-docker (2025-12-14)

v0.0.2-rc120

Changed

• feat/572-obfuscate-build-image (2025-12-13)

Fixed

v0.0.2-rc118

Changed

Fixed

• bug/570-undefined-agent (2025-12-13)

v0.0.2-rc117

Changed

• feat/446-configure-schedule-via-assistant-id (2025-12-10)
• feat/513-create-thread-api (2025-12-10)

Fixed

v0.0.2-rc116

Changed

• feat/513-create-thread-api (2025-12-10)
• feat/566-langchain-sandbox-dx (2025-12-09)

Fixed

• bug/419-schedule-modal (2025-10-12)

v0.0.2-rc115

Changed

• feat/564-create-tool-as-tool (2025-12-09)

Fixed

v0.0.2-rc114

Fixed

• bug/562-select-model-glitchy (2025-12-08)

v0.0.2-rc113

Changed

• feat/558-dedicated-tools-pages (2025-12-07)

Fixed

v0.0.2-rc107

Changed

• feat/553-select-models-modal (2025-12-06)
• feat/550-create-tools-ui (2025-12-06)
• feat/479-api-as-a-tool (2025-12-05)
• feat/547-more-stock-tools (2025-12-04)

Fixed

v0.0.2-rc104

Fixed

• bug/538-fetch-threads-directly (2025-12-02)
• bug/536-error-search-threads-files (2025-12-01)
• bug/518-on-new-token-is-broken (2025-11-29)
• bug/516-fix-google-stop-reason-stream (2025-11-21)
• bug/505-claude-stream-response (2025-11-19)
• bug/456-select-tools (2025-11-02)
• bug/444-anon-achat (2025-10-23)
• bug/423-fix-checkpointer-conn-closed (2025-10-10)
• bug/406-day-mode-theme-fix (2025-10-05)
• bug/400-make-login-suck-less (2025-10-01)
• bug/379-fail-to-edit-mcp-settings (2025-10-01)
• bug/389-cannot-unselect-from-mobile-agentmenu (2025-09-30)
• bug/392-new-chat-buttons-no-auth (2025-09-30)
• bug/370-anthropic-streaming (2025-09-16)

Changed

• feat/545-agent-threads-to-sidebar (2025-12-03)
• feat/543-relocate-agent-pages (2025-12-03)
• feat/525-pass-instructions (2025-12-03)
• feat/530-semantic-search-over-threads (2025-11-30)
• feat/534-mermaid-diagram (2025-11-30)
• feat/528-pagination-for-threads (2025-11-30)
• feat/519-react-window (2025-11-23)
• feat/507-add-projects-ui (2025-11-20)
• feat/503-add-bettter-tool-ui (2025-11-20)
• feat/494-can-render-html-files (2025-11-20)
• feat/509-search-model (2025-11-20)
• feat/498-data-model-for-prorject-sources (2025-11-14)
• feat/501-update-stale-docs-with-placeholder (2025-11-16)
• feat/420-project-endpoints (2025-10-28)
• feat/464-update-input-support-multiple-keys (2025-11-11)
• feat/490-can-read-deepagents-agent-files (2025-11-10)
• feat/485-user-env (2025-11-08)
• feat/482-teams-webhook (2025-11-05)
• feat/476-html-chart-construct-tool (2025-11-02)
• feat/468-support-ms-presidio-analyzer (2025-11-02)
• feat/465-update-sidebar (2025-10-31)
• feat/457-checkpoint-stream (2025-10-26)
• feat/448-tokens-per-sec (2025-10-24)
• feat/442-edit-memories (2025-10-22)
• feat/431-assistant-id-optimize (2025-10-15)
• feat/429-prompt-endpoints (2025-10-12)
• feat/425-editor-for-system-message (2025-10-10)
• feat/415-schedules-page (2025-10-08)
• feat/411-tools-api (2025-10-05)
• feat/408-edit-existing-schedules (2025-10-05)
• feat/216-shot-at-scheduled-tasks (2025-10-02)
• feat/382-move-threadid-gen-to-backend (2025-10-01)
• feat/380-subagent-configure (2025-09-29)
• feat/387-ip-rate-limit (2025-09-28)
• feat/381-delete-agent (2025-09-28)
• feat/375-interface-for-agent-config (2025-09-24)
• feat/375-readd-mcp-a2a (2025-09-20)
• feat/372-query-params (2025-09-19)
• feat/361-separate-user-threads (2025-09-16)
• feat/367-checkpointing (2025-09-15)
• feat/359-pre-commit-hooks (2025-09-07)
• feat/357-graph-selection (2025-09-02)

v0.0.1

Fixed

• bug/314-async-session-no-attr (2025-05-17)
• bug/311-multiserver-mcp-client-context (2025-05-17)
• bug/307-db-call-cause-cpu-leak (2025-05-16)
• bug/294-mem-lead-from-mcp (2025-05-13)
• bug/280-session-tool-card-decouple (2025-05-04)
• bug/276-resolve-json-parsing-on-tools-message (2025-05-01)
• bug/254-subsequent-messages-work (2025-04-21)
• bug/227-streaming-output-for-anthropic-fix (2025-04-02)
• bug/215-async-db-timeout (2025-03-25)
• bug/204-async-db-timeout (2025-03-20)
• bugfix/168-remove-system-message-metadata-redundancy (2025-03-04)
• bugfix/106-resolve-break-on-mobile (2025-02-25)
• bugfix/111-resolves-break-on-mobile (2025-02-25)
• bugfix/137-migrations-auto-up (2025-02-07)
• bugfix/131-action-log-overflow (2025-02-04)
• bugfix/128-fix-remote-container-exec (2025-02-03)
• bugfix/109-open-sidebar-mobile (2025-01-17)
• bugfix/105-fix-chat-selection (2025-01-12)
• bugfix/104-model-select-stopped-working (2025-01-12)
• bugfix/85-static-users-failing-to-parse-json-env (2025-01-10)
• bugfix/72-default-vite-api-url (2024-12-25)
• bugfix/68-close-stream-client-properly (2024-12-24)
• bugfix/65-can-refresh-from-sub-route (2024-12-22)
• bugfix/28-cannot-auth-tools-list (2024-11-30)

Changed

• feat/332-add-memories (2025-06-15)
• feat/290-api-as-a-tool-v2 (2025-06-08)
• feat/327-agent-select-from-chatinput (2025-06-06)
• feat/322-add-react-voice-viz (2025-05-29)
• feat/321-knowledge-source-page (2025-05-25)
• feat/319-integrate-langconnect (2025-05-24)
• feat/316-select-arcade-tools-from-ui (2025-05-20)
• feat/303-audio-recorder-on-chat-page (2025-05-14)
• feat/308-arcade-dev-tools (2025-05-14)
• feat/295-speech-to-text (2025-05-12)
• feat/296-update-available-models (2025-05-11)
• feat/288-add-contact-create-to-oauth (2025-05-05)
• feat/287-limit-input-buttons (2025-05-04)
• feat/281-public-server-display-on-landing (2025-05-03)
• feat/270-create-server-from-dash (2025-04-26)
• feat/268-react-flow-test-run (2025-04-26)
• feat/264-auth-user-home-page (2025-04-26)
• feat/265-mcp-server-docs-page (2025-04-25)
• feat/247-public-ensos-shareable (2025-04-25)
• feat/258-user-can-create-server-config (2025-04-22)
• feat/256-servers-table (2025-04-21)
• feat/250-update-docs (2025-04-20)
• feat/248-update-readme-dev-steps (2025-04-19)
• feat/245-add-a2a-logic (2025-04-14)
• feat/241-better-ui-for-tool-select (2025-04-13)
• feat/238-public-use-page (2025-04-11)
• feat/231-collect-register-event-to-slack (2025-04-04)
• feat/225-simplify-chat-state-to-reducer (2025-03-31)
• feat/219-persist-mcp-to-settings (2025-03-27)
• feat/213-updates-to-create-agent-page (2025-03-23)
• feat/210-create-agent-page (2025-03-21)
• feat/171-able-to-update-system-message (2025-03-20)
• feat/182-mcp-tool-adapter (2025-03-13)
• feat/196-add-init-chat-model (2025-03-15)
• feat/189-fix-settings-db-table (2025-03-14)
• feat/173-test-a-tool-call (2025-03-04)
• feat/165-oauth-init (2025-03-01)
• feat/153-clear-chat-inputs (2025-02-27)
• feature/61-make-pwa (2025-02-25)
• feature/151-add-settings-table (2025-02-23)
• feature/148-add-storage-endpoints (2025-02-18)
• feature/145-update-settings-to-remote-dropdown (2025-02-17)
• feature/76-user-can-configure-their-own-keys (2025-02-13)
• feature/132-autosize-textarea (2025-02-11)
• feature/139-individual-user-accounts (2025-02-09)
• feature/135-google-llm-provider (2025-02-07)
• feature/130-edit-system-message (2025-02-04)
• feature/125-add-03-mini (2025-02-02)
• feature/120-popout-drawer-action-log (2025-01-29)
• feature/114-add-chat-ollama (2025-01-22)
• feature/115-remove-shell-local-from-demo (2025-01-21)
• feature/112-can-select-tools-from-app (2025-01-20)
• feature/102-frontend-build-cd (2025-01-11)
• feature/99-fix-vite-api-url-append (2025-01-05)
• feature/93-open-drawer-on-mobile (2024-12-30)
• feature/95-fix-docker-build (2025-01-01)
• feature/86-model-select-in-react-app (2024-12-28)
• feature/83-add-endpoint-to-display-models (2024-12-27)
• feature/52-ability-to-switch-llms (2024-12-27)
• feature/70-basic-auth-json-defined-users (2024-12-26)
• feature/70-add-shad-cn (2024-12-25)
• feature/62-clean-up-message-formatting (2024-12-21)
• feature/50-add-mkdocs-site (2024-12-09)
• feature/48-sql-agent (2024-12-07)
• feature/46-create-docs-from-data-sources (2024-12-07)
• feature/44-indexing-vectors (2024-12-06)
• feature/23-multimodal-query (2024-12-05)
• feature/35-switch-pip-for-uv (2024-12-01)
• feature/33-deploy-app-using-terraform (2024-12-01)
• feature/30-tf-do-deployment (2024-11-30)
• feature/25-add-basic-auth (2024-11-29)
• feature/19-deploy-to-dev (2024-11-28)
• feature/17-ci-push-to-registry (2024-11-28)
• development