…tions (#15445)

Long-running queries (e.g. an insert stuck for 700s on lock contention)
could block indefinitely because Postgres' statement_timeout defaults to
0 (no limit) and neither the node nor neon pool configured one.

Add an optional DATABASE_STATEMENT_TIMEOUT env (milliseconds, no default)
applied to both NodePool and NeonPool as statement_timeout and
idle_in_transaction_session_timeout, so Postgres aborts a stuck statement
or idle transaction on the server side. Unset keeps the previous behavior.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

… DeepSeek compatibility (#15413)

…builtin agent (#15443)

Move the self-iteration skill-management action off the inline policy
implementation onto an execAgent-dispatched builtin agent (slug
`skill-management`), mirroring the S3/S4 memoryWriter + self-iteration
migration. Adds the `agentSignalSkillManagement` serverRuntime, the
builtin-tool-agent-signal skill-management manifest/systemRole, and the
builtin-agents skill-management agent; strips the ~3.5k-line inline
skillManagement policy down to the dispatch shim.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* 🗃️ feat(database): add workspace_id columns to existing tables

Add a nullable `workspace_id text` column to user-owned business tables
(agents, sessions, topics, messages, files, tasks, RAG/eval, RBAC, devices,
connectors, etc.) so records can later be scoped to a workspace. Workspace
tables themselves already landed on canary via 0105_add_usage_agent_share_workspace.

Also folds in the additive device schema from #15356: the structured
`working_dirs` jsonb column + `WorkingDirEntry` type (recent_cwds kept,
now @deprecated).

Scope is deliberately column-only — the lowest-risk slice:
- migration 0106 is pure `ADD COLUMN IF NOT EXISTS` (metadata-only, ~ms locks
  per table, online-safe, no app code change since columns are all NULL).
- FKs, btree indexes, and the per-user→workspace-scoped unique-constraint
  conversions are intentionally deferred to follow-up PRs so each can use the
  production-safe execution path Drizzle can't express (NOT VALID + VALIDATE,
  CREATE INDEX CONCURRENTLY, atomic unique swap).

Scoping notes:
- devices / user_connectors / user_connector_tools: scoped (user-owned resources).
- push_tokens: left user/device-level — an Expo token is one per app install and
  receives a person's notifications across all their workspaces.
- agent_shares: no workspace_id — scoped transitively via agent_id → agents.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* 🐛 fix(database): satisfy inferred row types after adding workspace_id

Adding workspace_id made it a required key in the Drizzle-inferred row types
($inferSelect), breaking call sites that build those shapes by hand:
- rbac.getUserRoles: include workspace_id in the explicit select projection
- session action: add workspaceId to the constructed chat-group literal
- test mocks (apiKey / generation / generationBatch / generationTopic): add
  workspaceId: null

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* ✅ test(database): use toMatchObject for topic.create row assertions

The two `expect(createdTopic).toEqual({ ...full literal })` snapshots broke
on every new column (here: workspace_id). Switch them to toMatchObject so the
returned row may carry extra columns without churning the expected literal.
The dbTopic↔createdTopic strict comparisons are left as toEqual.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ge (#15457)

* ♻️ refactor(message): prefer dedicated usage column over metadata.usage

Token usage was promoted out of metadata.usage into a dedicated messages.usage
column, but nothing populated it and all reads still went through metadata.usage.

- Centralize write-side promotion in the DB model (update / updateMetadata /
  create), so all executor callers populate the usage column from a top-level
  usage payload, falling back to metadata.usage. metadata.usage stays dual-written
  for backward-compatible reads.
- Reads prefer the usage column and fall back to metadata.usage: message queries,
  getTokenHeatmaps, recomputeTopicUsage, the usage record service, and context
  token accounting.
- Add top-level usage to UpdateMessageParams + DBMessageItem types.
- Mark metadata.usage and the legacy flat token fields as @deprecated, pointing
  to the top-level usage field.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🐛 fix(message): dual-write metadata.usage for top-level usage updates

When a caller passed the new top-level `usage` param without also sending
`metadata.usage`, the update wrote only `messages.usage` and left
`metadata.usage` stale/absent — legacy readers and rollback paths still consume
it during the dual-write transition. Fold the resolved usage into the metadata
patch so `metadata.usage` stays in sync regardless of how usage was passed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* 🗃️ db(database): add workspace_id FK constraints (migration 0107)

Phase 2 of workspace_id rollout: add the FK constraint on the 70 tables
that gained a bare `workspace_id` column in Phase 1 (0106), referencing
workspaces(id) ON DELETE CASCADE.

- schema: add `.references(() => workspaces.id, { onDelete: 'cascade' })`
  to all 70 nullable workspace_id columns
- 0107_add_workspace_id_fk.sql: idempotent drizzle migration
  (DROP CONSTRAINT IF EXISTS + ADD), runs in CI / dev / self-host
- 0107_concurrent.sql: production-safe out-of-band runbook
  (NOT VALID + VALIDATE) to avoid write-blocking locks on large tables;
  NOT run by drizzle

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🔥 db(database): remove stray 0107_concurrent migration file

* 🐛 fix(database): break user/workspace schema circular dependency

Move userInstalledPlugins from user.ts into connector.ts to break the
user.ts <-> workspace.ts import cycle flagged by dpdm. connector.ts
already imports both users and workspaces, and consumers import the
table from the schemas barrel, so no call sites change.

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Stdio MCP servers live on the user's machine, but in gateway (cloud) mode
the agent runs server-side and `executeMCPTool` tried to spawn the stdio
binary on the cloud server — which has neither the binary nor access to the
user's machine, so local MCP tools (e.g. tasks calling a local kimi-datasource
MCP) always failed.

Add a dedicated `executeMcpCall` path that forwards the stdio connection
params (command/args/env) to a connected device, which spawns the MCP server
and runs the call locally. It rides the existing `/api/device/tool-call`
relay — the gateway forwards `toolCall` opaquely — so the device-gateway
worker needs no changes; the device routes on the presence of
`toolCall.mcpParams`.

Server-side only: when no device is connected, behavior is unchanged
(standalone Electron still spawns in-process). The desktop-side receiver that
runs the forwarded call lands in a follow-up.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* 🗃️ db(database): add workspace_id indexes (migration 0108)

Phase 3 of the workspace DB migration (LOBE-9961). Adds a btree index on
workspace_id to 70 tenant tables, plus 7 workspace-scoped partial unique
indexes (WHERE workspace_id IS NOT NULL) that pre-build the "new" side of the
Phase 4 (0109) unique-constraint cutover.

A separate production-safe runbook (0108_concurrent.sql, CREATE INDEX
CONCURRENTLY, ordered smallest->largest) is intentionally NOT committed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🗃️ db(database): make 0108 index migration idempotent

Add IF NOT EXISTS to all 70 CREATE INDEX + 7 CREATE UNIQUE INDEX statements,
per the db-migrations standard flow (defensive/idempotent SQL), matching how
0107 used DROP CONSTRAINT IF EXISTS. Safe to re-run and safe if the concurrent
runbook already built the indexes before the auto-migrator reaches 0108.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

🐛 fix: prefer internal app url for comfyui calls

…ls (#15473)

* ✨ feat(gateway): add explicit type discriminator to tunneled tool calls

The device-gateway relays builtin local-system calls and tunneled stdio MCP
calls over one `tool-call` channel. The device was meant to tell them apart by
sniffing whether `toolCall.params` exists — fragile: any future builtin tool
that grows a `params` field would be misrouted to the MCP client.

Add an explicit `toolCall.type` discriminator (`'builtin' | 'mcp'`). The HTTP
client stamps it: `executeToolCall` → `'builtin'`, `executeMcpCall` → `'mcp'`.
The device routes on `type`, never on payload shape. Optional + back-compatible:
an older server that omits it is treated as `'builtin'`.

The desktop receiver switches to this discriminator in a follow-up.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* ✨ feat(desktop): execute tunneled stdio MCP calls from the gateway (#15470)

Receiving half of the gateway stdio-MCP work. When the cloud server tunnels a
stdio MCP tool call to this device (a `tool_call_request` carrying
`mcpParams`), run it locally instead of falling through to the builtin
local-system tool switch (which keys on apiName and has no MCP context, so it
rejected these as "not available on this device").

- `gatewayConnectionSrv`: add a dedicated `mcpCallHandler` + `setMcpCallHandler`;
  `handleToolCallRequest` routes on the presence of `toolCall.mcpParams`,
  sharing the existing response-envelope path.
- `GatewayConnectionCtr`: wire `setMcpCallHandler` → `executeMcpCall`, which
  maps the wire payload to `McpCtr.runStdioMcpTool`.
- `McpCtr`: extract `runStdioMcpTool` core from the `callTool` IPC method so
  both the renderer and the gateway tunnel share one stdio execution path
  (no SuperJSON round-trip for the in-process caller).

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…5472)

* 🗃️ db(database): migrate unique constraints to workspace scope (migration 0109)

Replace the legacy user-scoped UNIQUE constraints with workspace-scoped
partial unique indexes across agents, agent evals, agent skills,
documents, sessions, tasks, and rbac roles/user-roles. Adds migration
0109_migrate_unique_constraints and updates the affected schemas.

* 🐛 fix(database): match partial unique index in getBuiltinAgent upsert

Migration 0109 turned `agents_slug_user_id_unique` into a partial index
(WHERE workspace_id IS NULL). A plain `ON CONFLICT (slug, user_id)` no longer
matches it (Postgres 42P10), breaking getBuiltinAgent. Add the same predicate
via onConflictDoNothing's `where` option; builtin agents are always
workspace-less so the predicate always holds.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🔨 chore(database): use bare onConflictDoNothing in getBuiltinAgent for 0109 transition

Index-shape-agnostic upsert so the builtin-agent path works whether
agents_slug_user_id_unique is the legacy full unique or the 0109 partial,
removing the deploy-ordering coupling. Re-tighten to { target, where } in a
follow-up once 0109 has flipped the index everywhere.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…d submenus (#15433)

* 🐛 fix: bypass audits for headless tool calls

* 🐛 fix: block high-risk headless tools at execution

* Revert "🐛 fix: block high-risk headless tools at execution"

This reverts commit 1d4b534.

* 🐛 fix: restore headless audit bypass

* 🐛 fix: resolve headless blocked tools

* 🐛 fix: simplify blocked tool results

* 🧹 chore: remove unrelated prompt diff

* 🐛 fix: narrow blocked tool instruction type

* 🐛 fix: split security blacklist policies

* 🐛 fix: simplify security blacklist policy rules

* 💄 style: tighten security blacklist diff

* 💄 style: reduce agent config doc diff

* 💄 style: tighten headless audit diff

* 💄 style: minimize audit policy diff

* 💄 style: clarify global audit match naming

* 🐛 fix: auto-run required global audits in headless

* 💄 style: clarify headless intervention comments

* 💄 style: clarify headless global audit comment

* 💄 style: use blocked tool instruction type

* 💄 style: clarify headless audit tests

* 💄 style: annotate headless blocked tool tests

* 🐛 fix: type security blacklist policy filter

* 💄 style: clarify local system 403 guidance

* 🐛 fix: use current persist error helper

…15475)

🔨 chore(database): re-tighten getBuiltinAgent onConflict to the 0109 partial index

Now that migration 0109 has flipped agents_slug_user_id_unique to a partial
index (WHERE workspace_id IS NULL) in all environments, restore the precise
conflict arbiter { target: [slug, userId], where: isNull(workspaceId) } so
unexpected unique violations surface instead of being silently swallowed by the
bare onConflictDoNothing() transition form.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…odel fetch logic (#15376)

* 🗑️ chore(opencode-go): remove MiMo V2 Omni and MiMo V2 Pro models

* ✨ feat(opencode-go): fetch model list from API with models.dev enrichment

- Try API /models first for real-time available models
- Enrich with models.dev data (pricing, abilities, SDK routing)
- Fallback to models.dev + model-bank if API fails
- Dynamic Anthropic SDK routing via provider.npm field

* 💰 fix(opencode-go): update MiMo pricing to match models.dev

- mimo-v2.5: input $0.14, output $0.28, cache_read $0.0028
- mimo-v2.5-pro: input $1.74, output $3.48, cache_read $0.0145

* ✨ feat(opencode-go): add MiniMax M3 and remove deprecated Qwen3.5 Plus

- Add minimax-m3: 512K context, vision support (image+video), 131K output,
  pricing 0.6/2.4/0.12 USD per M tokens, released 2026-05-31
- Remove qwen3.5-plus: marked deprecated in models.dev

* 🐛 fix(opencode-go): restore Anthropic routing fallback when models.dev is unreachable

Codex P2 review on #15376:
- `routers` is called with `ClientOptions` (no `client` field), so
  `options.client?.models.list?.()` silently returned `undefined` via
  optional chaining; the `catch` never ran and `modelIds` stayed `[]`.
- In API + models.dev double-failure scenarios, `getAnthropicModels([])`
  returned an empty list, regressing Anthropic SDK routing for MiniMax /
  Qwen models.

Fix:
- Make `getAnthropicModels` self-contained: takes no parameters.
- Fallback chain: models.dev → static model-bank prefix match → `[]`.
- `routers` no longer touches `options.client`.

* ✨ feat(opencode-go): enrich model list with models.dev metadata

The model list pipeline previously forwarded only `{ id }` from the API
and models.dev, so displayName / pricing / context / modalities all came
from the static model-bank. When models.dev disagrees with model-bank
(e.g. a price update or new model), the runtime would show stale data.

Map models.dev fields into the flat shape that `processModelCard`
understands, so each card is enriched with:
  - displayName (dev.name)
  - contextWindowTokens / maxOutput (dev.limit)
  - releasedAt (dev.release_date)
  - functionCall / reasoning / vision / structuredOutput (dev.flags +
    dev.modalities.input)
  - pricing (dev.cost → flat input/output/cachedInput/writeCacheInput;
    processModelCard's formatPricing converts it to units)

Fields models.dev doesn't have (description, organization, settings
.extendParams, etc.) still fall back to the model-bank entry via
processModelCard's knownModel lookup, keeping the static config as the
source of truth for UX-only fields.

* ✨ feat(opencode-go): drive reasoning_content handling from models.dev

The `reasoningInterleavedModels` list was hardcoded and drifted from
models.dev:
  - Missing: kimi-k2.5, kimi-k2.6, mimo-v2-omni, mimo-v2-pro
  - Stale: qwen3.7-max (no longer has `interleaved` in models.dev)

Move the source of truth into the models.dev cache. `fetchModelsDevData`
now also builds an `interleavedIds: Set<string>` from `m.interleaved.field`
alongside `anthropicModels`, so every derived field stays in sync with
a single fetch.

The new `getInterleavedModelIds` sync accessor lets `buildOpenAIPayload`
keep its sync signature; it returns the cached set when populated and
falls back to a hardcoded snapshot of the last-known models.dev state on
the very first chat request before any fetch has run.

…15353)

Consume the `working_dirs` column: model `updateDevice`, tRPC `updateDevice`
input + `listDevices` output, and the client cwd pickers now operate on
`WorkingDirEntry[]` instead of the flat `recentCwds: string[]`.

- model / tRPC: `workingDirs` (input capped at 20, validated `{ path, repoType? }`)
- client `deviceCwd`: `nextRecentCwds` → `nextWorkingDirs`
- UI: DeviceWorkingDirectory / WorkingDirectory / DeviceDetailPanel / DeviceItem
  render the detected repo type via the shared `renderDirIcon`

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* ✨ feat(task): auto-ensure qstash schedule

chore: cleanup code

chore: cleanup code

chore: cleanup code

* chore: migrate qstash init workflow to startServer

chore: migrate qstash init workflow to startServer

* fix: set default QSTASH_URL to eu region, same as SDK

fix: set default QSTASH_URL to eu region, same as SDK

* 🐛 fix(agent-documents): render system docs in editor

* ✨ feat(agent-documents): autosave highlight editor with safe unmount flush

Add debounced autosave to the non-markdown highlight editor and a StrictMode-safe
unmount flush via queueMicrotask, plus a beforeunload guard against dirty buffers.

* ✅ test: fix agent document PR type checks

… triggers a new run (#14873)

* ✨ feat(task-detail): split task panel comment from topic-thread reply

CommentInput in TaskActivities stays as-is on canary — avatar + EditorCanvas
+ attachment + send button, posting a plain task-level comment.

TopicChatDrawer footer becomes a FeedbackInput that calls the in-scope
ConversationProvider's sendMessage, continuing the existing topic
conversation instead of attaching a comment + restarting the run.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* ✨ feat(task-detail): keep FeedbackInput visible while topic is running

Drop the canLeaveFeedback gate so the in-thread reply box renders even
when the topic is pending/running. ConversationStore.sendMessage already
queues messages during an in-flight stream, so this just exposes the
queue affordance to the user — letting them steer the next step
without waiting for the current run to terminate.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 💄 style(task-detail): collapse FeedbackInput behind a follow-up button + add attach action

FeedbackInput now starts collapsed as a full-width "Send follow up message"
button. Click expands a ChatInput shell with EditorCanvas inside and a footer
that carries an AttachmentUploadButton on the left (+ icon) and the send
button on the right. Files are inserted inline into the editor (same
pattern as CommentInput) so they ride along on sendMessage's editorData.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 💄 style(task-detail): tighten CommentInput card & switch follow-up button to filled

- CommentInput card: padding-block 8px → 4px, editor placeholder fontSize 14px
- FeedbackInput collapsed button: default size + variant="filled" for a less
  obtrusive look that sits flush in the chat footer

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 💄 style(task-detail): drop top padding above FeedbackInput in topic drawer

Use paddingBlock="0 12px" so the follow-up button hugs the last message
instead of floating with a 12px gap above.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 🐛 fix(task-detail): clear FeedbackInput editor before awaiting sendMessage

Previously the editor cleanup ran after the awaited sendMessage call, so
the box kept the just-sent text on screen until the entire send + stream
lifecycle resolved. Move clearContent / collapse before the await so the
input feels responsive (sendMessage already snapshots markdown and
editorData for its optimistic update).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 🐛 fix(task-detail): keep FeedbackInput expanded after sending

Drop the setExpanded(false) call in handleSubmit so the ChatInput
remains open once the user has opened it. Collapsing it back to the
"Send follow up message" button right after every reply was disruptive
mid-conversation; the button only makes sense as the initial resting
state of the drawer.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* ✨ feat(chat): add forceRuntime override to SendMessageParams

Plumb a new optional forceRuntime field through SendMessageParams →
ConversationLifecycle.sendMessage → selectRuntimeType(parentRuntime).
parentRuntime already wins over every other signal in the dispatcher,
so callers can pin a send to 'gateway' / 'client' / 'hetero' regardless
of the agent's local/cloud config.

Also propagate forceRuntime through the message queue (QueuedMessage +
MergedQueuedMessage + mergeQueuedMessages + both drain sites in the
client and hetero executors) so a follow-up queued during an in-flight
run keeps its runtime pin when it eventually fires.

FeedbackInput in TopicChatDrawer passes forceRuntime: 'gateway' so
task-topic follow-ups stay on the server-side path that runTask
originally used, even if the user's global runtime preference is local.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

* ✨ feat(topic): add one-click collapse/expand all groups in topic sidebar

Add a toggle button in the topic sidebar header (next to Filter and the
more-actions menu) that collapses or expands all topic groups at once.
It reuses the existing `expandTopicGroupKeys` global status, so it stays
in sync with manual per-group toggling, and hides itself when there are
fewer than two groups (e.g. flat mode).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🐛 fix(topic): hide group toggle in flat mode

In flat mode, groupedTopicsForSidebar falls through to time grouping so
the computed group count can exceed one, but List renders FlatMode with
no accordion for the toggle to affect. Hide the control explicitly when
topicGroupMode === 'flat' instead of relying on the group count.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 style(topic): use 2-corner minimize/maximize icons for group toggle

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…ch doesn't time out (#15634)

* 🐛 fix(cli): handle agent_run_request in `lh connect` so device dispatch doesn't time out

`lh connect` auto-registers the CLI as a device, so the gateway can pick it
as the dispatch target for a heterogeneous agent run (`agent_run_request`).
But the connect daemon only listened for `system_info_request` and
`tool_call_request` — it never handled `agent_run_request`, so it never sent
`agent_run_ack`. The gateway waited out its ack window and returned
`{error:'TIMEOUT',success:false}`, surfaced server-side as "Hetero agent
device dispatch failed".

Add an `agent_run_request` handler mirroring the desktop app: spawn
`lh hetero exec` fire-and-forget and ack `accepted` immediately. The spawned
process owns the full execution + server-ingest pipeline. It re-invokes the
current CLI entry (process.execPath + argv[1]) rather than relying on `lh`
being on PATH, so it works inside the detached daemon.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix: bump the cli version

* chore: bump the cli manifest

* 🐛 fix(cli): ack agent run only after spawn succeeds, reject on spawn error

`child_process.spawn` reports a missing/inaccessible cwd asynchronously via
the child's `error` event, after the handler had already sent an `accepted`
ack. The gateway/server then recorded dispatch success while no `lh hetero
exec` process existed to emit `heteroFinish`, leaving the assistant message
stuck instead of surfacing a failure.

`spawnHeteroAgentRun` now resolves on the child's outcome: `accepted` on the
`spawn` event (stdin is written only then), `rejected` on an early `error`. A
rejected ack returns the gateway 422 → execAgent writes a ServerAgentRuntimeError
onto the assistant message, so a failed dispatch is visible. Still resolves in
milliseconds, well within the gateway's 10s ack window.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… aborted (#13677)

* 🐛 fix(model-runtime): emit stop:abort instead of error when stream request is aborted

When user cancels a streaming request, the provider SDK throws abort errors
(e.g. "Request was aborted"). Previously these were propagated as error chunks,
causing the client to display a provider error message. Now abort errors emit
a stop:abort event through the SSE pipeline, allowing the client to handle
cancellation gracefully.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* 🐛 fix(model-runtime): fix type error in abort pipeline test

Use `as const` for type literal to satisfy StreamProtocolChunk union type.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* ✅ test(fetch-sse): add planUpgradeAfterFinish to onFinish expectations

#15616 added planUpgradeAfterFinish to the onFinish context but missed
updating fetchSSE.test.ts, breaking 13 tests on canary.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🐛 fix(model-runtime): harden abort detection against non-Error throws

isAbortError assumed error.message is always a string, but catch
clauses receive unknown — a non-Error throw (string, object without
message) would make the abort check itself throw inside the stream
error handler, swallowing both ABORT_CHUNK and the first-chunk error.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ueue mode (#15620)

* 🐛 fix(agent): deliver sub-agent resume bridge via QStash webhook in queue mode

The callSubAgent completion bridge was a handler-only hook, which lives in
process memory: in queue mode (AGENT_RUNTIME_MODE=queue) HookDispatcher only
delivers webhook-configured hooks, so the bridge never fired — the parent op
stayed parked in waiting_for_async_tool forever after all sub-agents finished.

- Give the bridge hook a webhook config (delivery: qstash) targeting the new
  /api/agent/webhooks/subagent-callback endpoint; local mode keeps the
  in-process handler. Both paths converge on
  AgentRuntimeService.completeSubAgentBridge (backfill + barrier/CAS resume).
- Park-time self-check: after the parked state and operation row are
  persisted, re-run the resume barrier once to recover children that
  completed before the parent finished parking.
- One-shot verify watchdog: when a completion finds the parent not yet
  resumable, schedule a delayed verifyAsyncToolBarrier re-check (no step
  lock, CAS-idempotent, never re-arms).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* 📝 docs(agent): correct verify-watchdog rationale comment

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* 📝 docs(agent): clarify eventFields trimming rationale

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* ♻️ refactor(agent): align subagent-callback with workspace-scoped step worker

Post-rebase adaptation to canary's runtime restructure (#15609):

- Route the webhook bridge through AiAgentService (like the /run step
  worker) so the runtime's models stay workspace-scoped — a bare
  AgentRuntimeService would be personal-scoped and the tool-message
  backfill / resume barrier could miss workspace-scoped rows.
- Extract SubAgentBridgeParams into agentRuntime/types and add the
  completeSubAgentBridge passthrough next to executeStep.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* 🐛 fix(agent): fail sub-agent callback loudly on backfill or delivery failure

Address two review findings on the resume bridge:

- completeSubAgentBridge now checks updateToolMessage's { success } result
  (it swallows transaction errors instead of throwing) and propagates all
  infrastructure failures. The webhook endpoint then returns non-2xx so
  QStash redelivers the whole bridge — previously a failed backfill was
  acked with 200 and the parent stayed parked forever, since the verify
  recheck only re-reads the barrier and cannot retry the backfill.
- New AgentHookWebhook.fallback: 'none' opts a qstash-delivered hook out of
  the unsigned plain-fetch fallback, which can never authenticate against a
  QStash-signed endpoint and only masked publish failures as silently
  dropped 401s. The bridge hook uses it; dispatch escalates such delivery
  failures to console.error instead of the debug namespace.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

…15638)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

* ✨ feat(model-bank): add claude-fable-5 to Anthropic models

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* 🐛 fix(agent): allow adding directory topics on web when agent targets a bound device

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

feat: support workspace (full) — store→business-hook + workspace router

#15643)