Deliverable Planner: Fails with “max turns” or 600s Timeout

Status: ✅ Fixed (2026-05-05) — requires strategy re-approval to test
Severity: Was Critical — deliverable plan never generated; all retries exhausted
File: packages/agents/src/workers/strategy.worker.ts

Symptom

The deliverable planner job fails in multiple ways across successive retries:


ERROR [strategy-worker] Deliverable planner job failed
  message: "Reached maximum turns per run"     ← ~6-8 min runs (with tools or with max_turns:1)

ERROR [strategy-worker] Deliverable planner job failed
  message: "Process timed out after 600s"      ← ~10 min runs (generation too slow)

The plan is never created. The onboarding flow stalls after strategy approval.

Root Cause (Confirmed — Three Combined Issues)

Issue 1: Missing `allowedTools: []`

The config was missing allowedTools: []. With --dangerously-skip-permissions and no --allowed-tools restriction, Claude had full access to all tools. Claude used tools during generation (browsing CWD), then needed turn 2 to continue after the tool result, which was blocked by maxTurnsPerRun: 1.

Fix: Added allowedTools: [] to disable all tool use. This resolves the spurious error_max_turns from tool use.

Issue 2: `maxTurnsPerRun: 1` blocks legitimate multi-turn output

The deliverable planner generates large JSON (goals + deliverables + action items + planning notes). The JSON is too long to fit in a single Claude response. The Claude Code CLI tries to do a continuation (turn 2), which is blocked by --max-turns 1.

Fix: Raised maxTurnsPerRun: 1 → maxTurnsPerRun: 3.

Issue 3: 600s timeout insufficient for large prompts

The prompt includes the full strategy document (~20K chars) + full client context (~30-50K chars). With this large input, Claude Sonnet 4.6 takes 3-4 min per turn × 3 turns = 9-12 min. The 600s (10 min) timeout fires before the generation completes.

Fix: Raised timeoutSec: 600 → timeoutSec: 900 (15 min) and lockDuration: 720_000 → lockDuration: 1_080_000 (18 min).

Fix Applied


// strategy.worker.ts — execute config
config: {
  cwd,
  model: "claude-sonnet-4-6",
  dangerouslySkipPermissions: true,
  allowedTools:   [],   // pure generation — disable all tool use
  timeoutSec: 900,      // 15 min — large input requires more time
  maxTurnsPerRun: 3,    // 3 turns: allows long JSON output to span continuations
}
 
// strategy.worker.ts — Worker config
lockDuration: 1_080_000,  // 18 min (15 min timeout + 3 min buffer)

How to Test

After an agents server restart:

Go to the dashboard → Strategy page for the test tenant
Admin approves the strategy document
Watch logs/agents-restart.log for Deliverable planner job completed
Check the dashboard for the generated goals and deliverable plan

The job that exhausted retries on 2026-05-05 CANNOT be retried automatically. It requires re-approval of the strategy document to trigger a new enqueue.

insight-worker-base.ts was fixed at the same time (maxTurnsPerRun: 1 → 3) for the same continuation issue.
Both require an agents server restart to take effect.