refactor: pre-task-communication cleanup (C1–C14, one commit per item)

[juacker]

Pre-feature cleanup for the inter-agent task-communication work. Net about −1,690 lines (+4,646 / −6,335 across 92 files). One commit per item; every commit builds and passes its tests on its own, so any point of the branch can be checked out locally.

Plan and rationale: TODO.md → "Pre-feature cleanup" (workspace), data-model review plans/db-model-review-backend.md / -frontend.md.

Commits (in order)

#	Item	Commit	What
1	C1	drop assistant_session_links	Dead table (no producer since ce9751f). New DROP TABLE IF EXISTS migration; removes parent_session_id, include_ancestors, the ancestor-walking page loop (now build_message_page, unit-tested), FE parentSessionId/includeAncestors.
2	C4	drop write-only SessionContext fields	tool_scopes (3 writers, 0 readers), inter_agent_call + InterAgentCallContext (never constructed). No migration (serde-defaulted; test session_context_ignores_removed_keys). Also ends a session-row rewrite ping-pong between the two reconcilers that disagreed on tool_scopes.
3	C5	remove unused useAssistantSession hook	Dead FE hook (no importer; the live path is getOrCreateWorkspaceSession). −156 lines.
4	C2	task transcript panel follows live task status	Workspace.tsx passed a one-time copy (viewingTask) to the panel, so status/title/assignee/sessionId never refreshed after open. Now derives the live task from the polled details.tasks (fallback to the snapshot). Test fails without the fix.
5	C3	call hooks unconditionally in task panel and message block	Two rules-of-hooks violations (hook after if (!task) return null; useMemo inside a role branch). No behaviour change. Follow-up: enable react-hooks/rules-of-hooks in eslint.
6	C7	remove scheduled-run connection fallback	One enabled connection selected in configured order for each scheduled run; removes the retry loop and output-detection heuristic. Shared selector also serves task runs; config shape stays intact.
7	C6a	route session commands through explicit workspace DB	Session/run-scoped commands take workspace_id and open only that DB; deletes the session_pool/run_pool/pool_for_new_session probes and five FE-uncalled commands (assistant_create_session, _list_sessions, _load_session_messages, _list_tool_calls, _retry_run). One repository::find_manager_session (newest non-task session, NOT EXISTS, id tie-break) shared by runner and UI. Fix: the scheduled runner could run inside a newer self-assigned task's transcript.
8	C6b	drop workspace ids from SessionContext	workspace_id/agent_workspace_id removed from the persisted context and the tool context; the id now comes from AssistantDeps/ToolBinding/command args. Tool gate that read agent_workspace_id.is_some() now uses the run's resolved workspace root (test for both branches). emit_event takes the workspace explicitly; event workspaceId is always set. Old rows still load.
9	C13	type workspace agent wrappers as WorkspaceAgentDetail	FE wrappers were typed as the wrong response; get can return null. New TS type matches the Rust struct; removes an as unknown as cast; explicit "Agent not found." + test.
10	C6c	build run config at run start, drop session reconcile	SessionContext keeps only automation_id + CLI session id/provider. New non-persisted RunContext (MCP ids, execution, roster, name, description) built once per run from live config, for API and CLI paths (rides ToolBinding). Deletes the reconcile code; get-or-create never rewrites existing rows (fixes stale CLI session id write-back and spurious updated_at bumps). MCP toggle edits the main agent's config, errors visibly without one. Task runs load the assignee's live config. A run whose agent is gone fails at start.
11	sched	release scheduler slot when a scheduled run fails to start	Fix (pre-existing): seven early errors in run_next_agent after next_ready skipped complete_agent, so the agent never ran again until restart/config edit. On a start error the slot is now released, the run marked failed and retried in 5 min. Regression test fails without the fix.
12	C6d	task runs follow the assignee's live config	Deletes the throwaway AgentConfig in assign_task; shared first_enabled_connection. F5: AgentConfig loses selected_mcp_server_ids/execution; dead required_tools/agent_config_to_definition and unread WorkspaceAgentRow fields removed. Behaviour: a task whose assignee was removed fails instead of running on stale settings.
13	C8	delete dead provider/MCP/skill commands and UI, legacy run triggers, notices storage	Removes 10 FE-uncalled commands + commands/provider.rs and dead helpers; AppConfig.ai_provider dropped (old config still loads, tested). RunTrigger::Retry/InterAgentCall deleted with a migration mapping legacy rows (quoted and plain forms, tested); inter_agent_call_depth removed. notices_json no longer read/written; AssistantRun.notices gone; "completed with warnings" status kept. Column drop in C9. FE: legacy ai_provider wrappers, ProviderSettings, AgentCard, NoticesBanner, unused provider/MCP/skill wrappers removed; react-hooks/rules-of-hooks is now an error. FE and Rust halves are one commit so ipc_surface holds at every commit.
14	C9	drop write-only task, session and run columns	One migration: backfills user_response_at → attention_acknowledged_at (keeps old tasks out of the attention count), then drops workspace_tasks.user_response/user_response_at/result_json, assistant_sessions.kind/title, assistant_runs.notices_json. In-memory notices become one had_warnings flag (RunNotice/RunNoticeKind deleted; same status). SessionKind gone. Migration tests seed old schemas with raw SQL. Agent-visible: task results no longer include resultJson (it duplicated resultSummary).
15	C10	one task mapper and one agent-row loader	New assistant/workspace_tasks_repo.rs holds the only task row/columns/load/list/create/update and WorkspaceTaskResponse; the tool-side copies are gone. One WorkspaceAgentRow loader with display_name(). Visible: agent task-tool JSON now equals the UI response (adds display names, omits empty optionals); workspace_listAgents orders Main first then join time; blank agent names show the definition id.
16	C12	return only latest and active runs; drop unused config/summary fields	assistant_list_runs/workspace details return ≤2 runs (newest + newest active) instead of the whole history every 5 s poll. Drops WorkspaceConfig.preferred_provider_connection_id (old config loads, tested), WorkspaceAgentSummary.agent_definition_id/role, WorkspaceAgentResponse.role; manager check uses agent.id (tested). Context bar: the provider picker and MCP toggle are disabled while a save runs, the values are reloaded from the backend afterwards, and the bar remounts per workspace (key), so overlapping saves can't happen.
17	C14	store enum columns as plain text	Migration unquotes assistant_messages.role, assistant_runs.status/trigger, assistant_tool_calls.status (guarded by json_valid/json_type), rebuilds assistant_compactions with plain CHECKs. enum_text/parse_enum_text bind/read every enum; one ACTIVE_RUN_STATUSES; no hand-written SQL special-cases quoted values any more. Migration test seeds quoted + plain rows. One-time rewrite of all message rows on upgrade.
18	—	break started_at ties by id in get_active_run	Matches current_runs ordering.
19	—	treat completed-with-warnings runs as terminal	Fix (pre-existing): assistant_cancel_run did not treat CompletedWithWarnings as finished, so cancelling such a run overwrote it as Cancelled and emitted RunCancelled. One RunStatus::is_active()/is_terminal() rule now.

| 20 | — | move workspace agent roster out of the command layer | Roster code moves to assistant/workspace_agents.rs; the assistant no longer imports from commands/. |
| 21 | — | tool context holds the run context | ToolExecutionContext holds Arc<RunContext> instead of field copies. |
| 22 | — | drop scheduler entries by workspace id on delete | Scheduler instances carry workspace_id; deleting a workspace removes them (otherwise the C6c start-failure path retried them every 5 min). |
| 23 | — | fail a task whose worker run is missing or still active | Fix: such a task was reported as completed. The BLOCKED: rule applies only to terminal runs. |
| 24 | — | AssistantDeps::emit for session events | One helper replaces repeated emit boilerplate (−58 lines). |
| 25 | — | generate workspace agent detail types from Rust | ts-rs bindings replace hand-written TS copies of WorkspaceAgentDetail and related types. |
| 26 | — | drop unused tabId / getActiveSessionForTab from the session store | Leftover of C5. |
| 27 | — | resolve a scheduled agent in its own workspace only | Looks up the agent in its own workspace instead of scanning all of them, so another workspace's unreadable config can't block it. |
| 28 | — | close the workspace database before deleting its directory | Fix: on Windows, deleting a workspace failed with os error 32 because the DB pool was still open. |

Review

Every commit had a read-only Rust or frontend review (production_quality) before the first squash. Four whole-PR review rounds followed. On owner direction (simplicity first), the fixes from those rounds were cut back to the items above (rows 20–28): regressions and changes that remove code. Hardening of older behaviour was dropped from this PR. That covers the unreadable-config sweep on MCP/skill delete, ConfigManager::try_update, MCP secret-clear ordering, the manual-run retry cap and the optimistic-save hook. Each of the 28 commits passes cargo check --all-targets and tsc --noEmit on its own.

Not verified

• The DROP TABLE migration against a real user DB that still has the table (the migration chain creates then drops it in the test DB).

Upgrade note: migrations 20261013000000–20261015000000 rewrite stored enum/trigger values and drop columns; a binary built from main cannot read a DB migrated by this branch.