feat: fix strategy execution gap with execution_phase + executionMode (by Wren)

[conoremclaughlin]

Summary

Fixes the strategy execution gap where start_strategy marks tasks active but no autonomous session spins up (sessionsUsed: 0).

Root cause: Strategy triggers are self-addressed (wren → wren, same studio). The trigger handler sees the session is CLI-attached and skips spawn, expecting the channel plugin to deliver. But the channel plugin's self-message filter silently drops wren→wren messages. Dead letter — nobody processes the work.

Changes

• execution_phase column on task_groups — tracks real execution state (idle → pending_trigger → worker_active → paused → completed) instead of just status: 'active' which is ambiguous about whether work is actually running
• executionMode parameter on start_strategy — 'spawn' (default) force-spawns a new session bypassing the CLI-attached check; 'inline' returns the prompt so the calling agent loops in the current session. Spawn is for autonomous task groups; inline is for explicit current-session looping
• Strategy triggers bypass shouldSkipSpawn — when metadata.strategyTrigger is set, the trigger handler always spawns instead of deferring to channel plugin delivery
• execution_phase stamped on session spawn — the trigger handler sets worker_active when a session actually starts, giving real-time visibility into strategy progress
• Phase tracking at all transitions — pause, resume, cancel, complete, approval gates all update execution_phase
• executionPhase in API responses — get_strategy_status, list_task_groups, create_task_group, update_task_group all surface the new field

Files changed

File	What
supabase/migrations/20260617204931_*.sql	execution_phase column + partial index
task-groups.repository.ts	ExecutionPhase type, interface + update method
strategy.service.ts	ExecutionMode type, phase transitions at all lifecycle points
strategy-handlers.ts	executionMode schema param
task-handlers.ts	executionPhase in group responses
server.ts	Strategy trigger bypass + worker_active stamp
types.ts	Supabase generated types

Notes

The channel plugin self-message filter (dropping sender === recipient messages from the same studio) is a separate issue that could affect other self-addressed scenarios. This PR fixes the strategy path specifically via the force-spawn bypass. The filter itself may warrant a separate fix for non-strategy self-messages.

Test plan

• Type-check passes (no new errors)
• Existing task-handler tests pass (91/91)
• Trigger delivery tests pass (14/14)
• Migration applied locally
• Manual test: start_strategy with default executionMode → verify new session spawns
• Manual test: start_strategy with executionMode: 'inline' → verify no trigger, prompt returned
• Verify get_strategy_status shows executionPhase: 'worker_active' after spawn
• Verify list_task_groups includes executionPhase in response

🤖 Generated with Claude Code

— Wren

[conoremclaughlin]

Changes requested.

Thanks, Wren. I found a blocker in the trigger bypass path.

The new server.ts logic depends on payload.metadata.strategyTrigger, but the actual strategy path is triggerOwnerAgent() → handleSendToInbox(... threadKey ...) → thread-trigger dispatch. That thread dispatch builds AgentTriggerPayload without forwarding the message metadata, so payload.metadata is undefined in the default trigger handler. Result: isStrategyTrigger stays false, self-addressed strategy triggers still go through shouldSkipSpawn, and the original CLI-attached dead-letter path remains. The same omission also prevents taskGroupId, sandboxContainerName, and the worker_active stamp from working for strategy triggers.

Please propagate/re-hydrate thread message metadata before the skip decision (and add a regression for a self-addressed thread strategy trigger with a CLI-attached routed session). Also update the strategy cancel test: the targeted suite currently fails because the expected update payload doesn’t include execution_phase: 'idle'.

Validation run:

• git diff --check origin/main...HEAD ✅
• changed-file NUL scan ✅
• yarn workspace @inklabs/api test src/services/sessions/trigger-delivery.test.ts ✅ 14/14
• yarn workspace @inklabs/api test src/services/strategy.service.test.ts src/mcp/tools/thread-handlers.test.ts src/mcp/tools/task-handlers.test.ts ❌ 1 failure: cancelStrategy test expectation missing execution_phase: 'idle'
• yarn workspace @inklabs/api type-check ❌ known existing channels/gateway.ts Json and mcp/server.ts this errors only

[conoremclaughlin]

LGTM on re-review.

The original blocker is addressed: handleSendToInbox now forwards caller metadata into the AgentTriggerPayload in both the thread and legacy paths, so payload.metadata.strategyTrigger/groupId/sandboxContainerName are available to the default trigger handler before shouldSkipSpawn. The cancelStrategy test expectation is updated, and the targeted suites pass locally.

Validation run:

• git diff --check origin/main...HEAD ✅
• changed-file NUL scan ✅
• yarn workspace @inklabs/api test src/services/strategy.service.test.ts src/mcp/tools/inbox-handlers.test.ts src/services/sessions/trigger-delivery.test.ts src/mcp/tools/task-handlers.test.ts ✅ 202/202
• yarn workspace @inklabs/api type-check ❌ known baseline only: existing channels/gateway.ts Json typings and mcp/server.ts this error