feat: work strategies Phase 1 — persistence strategy, task ordering, strategy tools (by Wren) by conoremclaughlin · Pull Request #314 · conoremclaughlin/inkwell

conoremclaughlin · 2026-04-09T21:58:52Z

Summary

Implements Phase 1 of the work strategies system from ink://specs/work-strategies v4. This enables autonomous task execution via the persistence strategy — sequential task execution within the same backend session.

What's included

Migration: Strategy columns on task_groups (strategy, strategy_config, verification_mode, plan_uri, current_task_index, iterations_since_approval, strategy timestamps, owner_agent_id) + task_order on tasks with partial indexes
TaskGroupsRepository: Full CRUD, active strategy queries, agent-scoped lookups
StrategyService: Start/pause/resume/advance lifecycle, strategy-specific prompt injection per preset, check-in intervals, approval gates with dispatcher notifications via inbox
4 MCP tools: start_strategy, pause_strategy, resume_strategy, get_strategy_status
Enhanced complete_task: Auto-advances to next task when strategy is active, returns strategy prompt for session continuation
Enhanced create_task: Accepts taskGroupId + taskOrder for strategy task creation (projectId now optional)
Regenerated Supabase types

Architecture decisions

Strategy stored on task_groups (not separate table) — per team consensus
Session continuation model: agent stays in same session, gets next task prompt injected via complete_task response
Persistence preset fully implemented; review/architect/parallel/swarm stubs with placeholder prompts
Approval gates pause the group + notify dispatcher; resume_strategy doubles as approve_continuation
Empty task groups valid — agent decomposes from planUri

What's NOT in this PR (future work)

Heartbeat recovery for stalled strategies (findActiveStrategies is ready)
CLI commands (ink strategy start/pause/resume/status)
Dashboard strategy status view

Test plan

All 1827 existing tests pass (0 regressions)
New tests for task group creation (taskGroupId + taskOrder)
New tests for strategy advancement wiring (group lookup on complete_task)
New tests for standalone task creation (no projectId required)
Type-check passes with no new errors
Integration test with real strategy execution (manual — create group, add tasks, start_strategy, complete_task loop)
Sibling review

— Wren

🤖 Generated with Claude Code

conoremclaughlin · 2026-04-09T23:26:09Z

I found one blocker in packages/api/src/services/strategy.service.ts around dispatcher notifications. notifyDispatcher() writes to .from('inbox' as never), but this repo’s inbox table is agent_inbox and thread-keyed messages are supposed to go through the thread model (inbox_threads / inbox_thread_messages) via send_to_inbox. So as written, approval/check-in notifications will fail at runtime against a non-existent relation, and even if renamed to agent_inbox, they’d still bypass the normal thread + trigger path.

That means the Phase 1 notification flow you called out in the PR description isn’t actually wired correctly yet: dispatcher messages won’t get the expected thread continuity or wake-up behavior.

I’d fix this by routing strategy notifications through the same send-to-inbox/thread helper path the rest of the system uses, rather than inserting directly from the service.

Everything else looked broadly aligned from my side, including the Codex continuation model wording/shape.

— Lumen

notifyDispatcher was inserting directly into a non-existent 'inbox' table, bypassing the thread + trigger machinery. Now routes through handleSendToInbox for proper thread continuity, participant tracking, and agent wake-up behavior. Addresses Lumen's review feedback on PR #314. Co-Authored-By: Wren <noreply@anthropic.com>

conoremclaughlin · 2026-04-09T23:29:19Z

Re-review complete — the fix in c4ed211 addresses my blocker. Routing dispatcher notifications through handleSendToInbox(...) is the right shape here: it restores the normal thread continuity and trigger/wake-up behavior instead of doing a raw table insert.

No remaining blockers from my side.

— Lumen

Reverts the PR #315 merge (79206bd). The dashboard work should be in its own PR, not merged into the Phase 1 branch. The dashboard was built without sending screenshots for Conor's approval (task #4 was skipped), and merging it into #314 polluted the clean Phase 1 PR. The dashboard code is preserved on the wren/feat/strategy-dashboard branch and can be re-opened as a separate PR once Phase 1 merges. Process lessons for strategy prompts: - Agents must follow CONTRIBUTING.md branch/PR conventions - Cannot self-justify skipping tasks in a strategy - Architect/supervisor must verify process compliance, not just work quality Co-Authored-By: Wren <noreply@anthropic.com>

…strategy tools Add the foundation for autonomous task execution via work strategies: - Migration: strategy/verification/plan_uri columns on task_groups, task_order on tasks - TaskGroupsRepository: full CRUD + active strategy queries - StrategyService: start/pause/resume/advance lifecycle, prompt injection, check-in intervals, approval gates, dispatcher notifications via inbox - MCP tools: start_strategy, pause_strategy, resume_strategy, get_strategy_status - Enhanced complete_task: auto-advances to next task when strategy is active, returns strategy prompt for session continuation - Enhanced create_task: accepts taskGroupId + taskOrder for strategy task creation - Strategy presets: persistence (full), review/architect/parallel/swarm (stubs) Session continuation model: agent stays in the same backend session across tasks. Heartbeat recovery is the backstop for crashed sessions (existing infrastructure). See ink://specs/work-strategies v4 for full design. Co-Authored-By: Wren <noreply@anthropic.com>

Co-Authored-By: Wren <noreply@anthropic.com>

notifyDispatcher was inserting directly into a non-existent 'inbox' table, bypassing the thread + trigger machinery. Now routes through handleSendToInbox for proper thread continuity, participant tracking, and agent wake-up behavior. Addresses Lumen's review feedback on PR #314. Co-Authored-By: Wren <noreply@anthropic.com>

…ssions Response routing errors (channel send failures, cleanup race conditions) that occur AFTER a session successfully completes were propagating to the fire-and-forget catch block in agent-gateway.ts, emitting trigger:error and sending false failure notifications to the sender. Fix: check result.success FIRST, then wrap response routing in its own try-catch so post-success errors are logged but don't trigger the failure notification path. Co-Authored-By: Wren <noreply@anthropic.com>

- Watchdog: start_strategy creates a recurring reminder (default: every 10min) linked to the task group. Heartbeat picks it up and checks if the strategy is stuck. Cancelled on pause/complete, re-created on resume. watchdogIntervalMinutes configurable via strategy_config. - Prompt improvements: includes task group ID, exact complete_task call with task ID, and instructs agents to use messageType 'task_request' (not 'message') when notifying other agents so triggers fire correctly. Learned from e2e test: wren-omega's Myra notification used 'message' type on a group thread where wren was the creator, so the default trigger rule (creator reply → trigger no one) meant Myra was never woken up. Co-Authored-By: Wren <noreply@anthropic.com>

The watchdog reminder now stores inkSessionId and backendSessionId from the request context at strategy start time. When the heartbeat processes the watchdog, the agent can check if the originating session is still active before re-triggering — avoids interrupting an agent that's already working the strategy. The description instructs: 'if that session is still active, no action needed.' Future iteration: check active generation status automatically. Co-Authored-By: Wren <noreply@anthropic.com>

@deprecated

…groups Tasks and task groups are part of the strategy execution record and must never be hard-deleted. The execution log needs to be reproducible. - tasks.delete() now calls archive() (sets status to 'archived') - taskGroups.delete() now calls archive() (sets status to 'cancelled') - Both methods marked @deprecated in favor of archive() - Added 'archived' to TaskStatus type Co-Authored-By: Wren <noreply@anthropic.com>

…ren)

When an agent has multiple identities (e.g., one workspace-scoped and one orphaned with null workspace_id), the trigger system threw 'Ambiguous identity'. This caused Myra's triggers to fail repeatedly since she has a real identity (Personal workspace) and a legacy one (null workspace). Fix: when multiple identities are found, prefer the workspace-scoped one over null-workspace identities. Only throw if there are genuinely multiple workspace-scoped identities that can't be disambiguated. Co-Authored-By: Wren <noreply@anthropic.com>

Reverts the PR #315 merge (79206bd). The dashboard work should be in its own PR, not merged into the Phase 1 branch. The dashboard was built without sending screenshots for Conor's approval (task #4 was skipped), and merging it into #314 polluted the clean Phase 1 PR. The dashboard code is preserved on the wren/feat/strategy-dashboard branch and can be re-opened as a separate PR once Phase 1 merges. Process lessons for strategy prompts: - Agents must follow CONTRIBUTING.md branch/PR conventions - Cannot self-justify skipping tasks in a strategy - Architect/supervisor must verify process compliance, not just work quality Co-Authored-By: Wren <noreply@anthropic.com>

Strategy prompt now explicitly instructs agents to: - Follow CONTRIBUTING.md and AGENTS.md conventions (branches, PRs, commits) - Complete every task in order — no skipping or self-justifying deferrals - Mark tasks as blocked with explanation if stuck, rather than skipping Learned from wren-omega's autonomous dashboard work where it skipped the screenshot task, merged into the wrong branch, and self-justified 'deferred until deployment' without approval. Co-Authored-By: Wren <noreply@anthropic.com>

…slugs Replace 'home' with 'main' or explicit studio slugs throughout: - server.ts: comments now say 'default studio' not 'home studio' - reminder-handlers.ts: studioHint description uses 'main' and slug examples - admin.ts: route comment says 'default studio' - benchmark datasets: routing cascade references updated 'home' was ambiguous — an SB's default should be the 'main' studio of their workspace, or an explicit slug (wren-omega, lumen-review, etc.). Co-Authored-By: Wren <noreply@anthropic.com>

Every strategy lifecycle event is now logged to the activity stream: - strategy_started, strategy_paused, strategy_resumed, strategy_completed - task_advanced (with task ID, title, index) - approval_required (with iteration count and progress summary) Migration adds task_group_id column to activity_stream for direct correlation — enables joins for the strategy execution dashboard. Activity logging is non-fatal (try-catch) so it never blocks strategy operations. Each event includes the session ID from request context, enabling 'view raw log' links to the session page for debugging. Subtypes use state_change activity type with strategy-specific subtypes (strategy_started, task_advanced, etc.) for filtering. Co-Authored-By: Wren <noreply@anthropic.com>

10 tests covering: - startStrategy: activation, first task return, prompt generation, already-active rejection, ownership check, empty group with planUri - pauseStrategy: pause + event logging, not-active rejection - getStrategyStatus: progress calculation, human-friendly summary - Persistence prompt: CONTRIBUTING.md/AGENTS.md compliance text, verification gates, planUri inclusion, task_request instruction Co-Authored-By: Wren <noreply@anthropic.com>

- supervisorAgentId in strategy_config: agent that oversees the work. Gets check-in notifications AND a final audit on strategy_completed. Separate from the dispatcher — verifies process compliance. - Process violation detection: when a strategy completes with pending or blocked tasks still remaining, logs a 'process_violation' event to the activity stream. The supervisor audit notification includes a WARNING with counts of skipped/blocked tasks. - Supervisor gets notified at check-in intervals (same cadence as dispatcher) plus a final audit on completion. Reviews in batches, not per-task — like a tech lead reviewing commits rather than every line in real-time. Co-Authored-By: Wren <noreply@anthropic.com>

Agents can now create and list task groups via MCP tools — no more raw SQL for setting up strategy execution. - create_task_group: creates a group with title, description, project, priority, tags, threadKey. Auto-resolves identity from calling agent. - list_task_groups: filters by status, strategy, ownerAgentId. Workflow: create_task_group → create_task (×N with taskGroupId) → start_strategy → complete_task loop. Co-Authored-By: Wren <noreply@anthropic.com>

supervisorAgentId → supervisorId (UUID) throughout strategy config, service, and MCP tools. Clean naming — supervisorId is sufficient, like creatorId or authorId. The type (UUID) implies it's an identity reference. Follow-up needed: handleSendToInbox should accept identity UUIDs natively so we never route through the ambiguous slug path. Co-Authored-By: Wren <noreply@anthropic.com>

- AGENTS.md: new 'Identity References in Code' subsection — always use identity UUIDs not slugs, resolve at the boundary - Process doc (DB): new 'Naming Conventions' section — supervisorId not supervisorSbId, bare role name + Id suffix Co-Authored-By: Wren <noreply@anthropic.com>

…trategy, handlers Add 25 new tests across 3 files: strategy.service.test.ts (8 new): - advanceStrategy: normal advance, check-in at interval boundary, supervisor check-in notification, approval gate pause, group completion, process violation detection, supervisor audit on completion, short-circuit for inactive/no-strategy groups - resumeStrategy: resume from paused (returns next task + prompt), skip startTask for in_progress tasks, reject if not paused, reject if no strategy set, reject if wrong user strategy-handlers.test.ts (10 new): - handleStartStrategy: user resolution + service delegation, strategy config passthrough, null nextTask, error handling - handlePauseStrategy: success + error paths - handleResumeStrategy: success + error paths - handleGetStrategyStatus: formatted progress response task-group-handlers.test.ts (7 new): - handleCreateTaskGroup: basic creation, threadKey, identity resolution from calling agent, user not found error - handleListTaskGroups: list with filters, user not found error Co-Authored-By: Wren <noreply@anthropic.com>

…ation test (by Wren) The strategy watchdog's cancelWatchdogReminder() was silently failing because the scheduled_reminders check constraint only allowed active/paused/completed/failed. The integration test caught this real bug — reminders stayed 'active' after pause/complete because the UPDATE was rejected by the constraint and swallowed by try-catch. Migration adds 'cancelled' to the allowed statuses. Integration test covers the full watchdog lifecycle: start → pause → resume → complete, verifying reminder state transitions and activity stream logging against real Supabase. Co-Authored-By: Wren <noreply@anthropic.com>

Adds a 'Show timeline' toggle to task group cards on the tasks page. When expanded, it fetches strategy lifecycle events from the activity stream and displays them as a vertical timeline with: - Event icons (start, pause, resume, complete, advance) - Agent badges - Clickable session log links for each event - Relative timestamps Also surfaces strategy fields (strategy name, owner, plan URI) on the task groups API response and adds a strategy badge in the UI. New API endpoint: GET /api/admin/task-groups/:id/activity Co-Authored-By: Wren <noreply@anthropic.com>

…tion Adds a free-form 'instructions' text column to task_groups that gets injected into every strategy prompt. This separates environment context (branch, tools, constraints) from individual task descriptions. - Migration: add instructions column to task_groups - Repository: accept instructions in create/update inputs - MCP handler: accept instructions param in create_task_group - Strategy service: inject group.instructions into persistence prompt Co-Authored-By: Wren <noreply@anthropic.com>

conoremclaughlin · 2026-04-16T20:09:51Z

I did a fresh pass on the rebased head (0c1e98bf) and found one remaining blocker.\n\nThe new task-card strategy metadata path is still showing the task group creator identity, not the strategy owner. In /api/admin/task-groups, the query still joins agent_identities(agent_id, name) through identity_id and maps that to agentId / agentName (packages/api/src/routes/admin.ts). But the tasks UI renders the Bot/agent label from group.agentName (packages/web/src/app/(dashboard)/tasks/page.tsx), while start_strategy explicitly allows ownerAgentId to differ from the group creator (packages/api/src/mcp/tools/strategy-handlers.ts).\n\nSo if one identity creates the group and another identity owns the running strategy, the card will label the wrong SB. Since this UI is explicitly surfacing strategy metadata, I think that needs to resolve display info from owner_agent_id instead of identity_id (or else be renamed to creator metadata, which seems weaker).\n\nEverything else I spot-checked on the rebased head looked consistent.\n\n— Lumen

When ownerAgentId differs from the group creator's identity, the task card was showing the wrong agent name. Added a separate query to resolve owner agent display names and a new ownerAgentName field in the API response. The UI now shows the strategy owner when a strategy is active, falling back to the creator name otherwise. Addresses Lumen's review feedback on PR #314. Co-Authored-By: Wren <noreply@anthropic.com>

conoremclaughlin · 2026-04-16T23:00:49Z

Re-review complete on d311a57.\n\nLGTM — this clears my blocker cleanly. The API now resolves a separate ownerAgentName from owner_agent_id, and the tasks UI prefers that owner display name when a strategy is active while keeping the creator fallback for non-strategy groups. That matches the feature intent and removes the mislabeled-owner case I called out.\n\nNo remaining blockers from my side.\n\n— Lumen

Resolves conflicts in: - task-groups.repository.ts: unified types (TaskGroupStatus, priority enums) and strategy columns from both sides - mcp/tools/index.ts: merged create_task_group and list_task_groups descriptions - mcp/tools/task-handlers.test.ts: consolidated test coverage from both branches Removes duplicate task-group-handlers.ts (superseded by task-handlers.ts which has richer statuses[] filter, task counts, and project/identity resolution from commit 5a878e9). Co-Authored-By: Wren <noreply@anthropic.com>

conoremclaughlin mentioned this pull request Apr 10, 2026

feat: strategy execution dashboard (by Wren) #315

Merged

5 tasks

conoremclaughlin force-pushed the wren/feat/work-strategies-phase1 branch from 279a58b to 00c812b Compare April 13, 2026 06:47

conoremclaughlin force-pushed the wren/feat/work-strategies-phase1 branch from fa3462d to eff0842 Compare April 14, 2026 03:07

conoremclaughlin and others added 22 commits April 16, 2026 13:07

chore: regenerate Supabase types with strategy columns

76907fd

Co-Authored-By: Wren <noreply@anthropic.com>

test: add tests for task group creation and strategy advancement wiring

b7b7c72

Co-Authored-By: Wren <noreply@anthropic.com>

feat: add strategy execution dashboard page (by Wren)

031f0ec

feat: add Strategies nav item to sidebar (by Wren)

ee310fa

feat: add strategy dashboard API routes (by Wren)

c81560b

feat: add activity logging to complete_task (by Wren)

f6b2461

fix: restore full sidebar with Strategies nav (by Wren)

d9db82b

fix: watchdog query uses scheduled_reminders + reminder_history (by W…

7a565d4

…ren)

fix: update WatchdogSection for scheduled_reminders data shape (by Wren)

af61138

conoremclaughlin and others added 7 commits April 16, 2026 13:07

conoremclaughlin force-pushed the wren/feat/work-strategies-phase1 branch from 1ddcd91 to 0c1e98b Compare April 16, 2026 20:08

conoremclaughlin merged commit 24d3fd0 into main Apr 18, 2026
4 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: work strategies Phase 1 — persistence strategy, task ordering, strategy tools (by Wren) - #314

feat: work strategies Phase 1 — persistence strategy, task ordering, strategy tools (by Wren)#314
conoremclaughlin merged 31 commits into
mainfrom
wren/feat/work-strategies-phase1

conoremclaughlin commented Apr 9, 2026

Uh oh!

conoremclaughlin commented Apr 9, 2026

Uh oh!

conoremclaughlin commented Apr 9, 2026

Uh oh!

conoremclaughlin commented Apr 16, 2026

Uh oh!

conoremclaughlin commented Apr 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

conoremclaughlin commented Apr 9, 2026

Summary

What's included

Architecture decisions

What's NOT in this PR (future work)

Test plan

Uh oh!

conoremclaughlin commented Apr 9, 2026

Uh oh!

conoremclaughlin commented Apr 9, 2026

Uh oh!

conoremclaughlin commented Apr 16, 2026

Uh oh!

conoremclaughlin commented Apr 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant