[matthewtsmith]

Problem

The planner had to predict the issue numbers its new tasks would get, because .issues.json had no way for one new task to name another. Those predictions went into every Depends on line and e2e file name. Any issue or PR opened between drafting and approval shifted the real numbers.

In litart phase 9 the plan predicted that task N would be #N+279. Two builder PRs (#280, #281) opened while the plan waited for approval, so the tasks landed on #N+281. Every new-to-new dependency then pointed two numbers too low, mostly at closed issues. #288 and #289 were dispatched before their real dependencies existed, and both builders blocked.

Fix

• Task bodies in .issues.json write {{task:K}} for the plan's K-th task (1-based, so a dependency reads #{{task:3}}) and {{issue}} for the task's own number (for example e2e/{{issue}}-<slug>.spec.ts).
• applyPlan creates or reuses every issue first, then fills the placeholders with the real numbers. Each task gets agent-ready only after its own body is filled.
• A {{task:K}} outside the plan refuses the whole plan before any write. The error names the file and the task, and the tick fails the same way a schema error does today.
• If an apply dies between creating and filling, the retry reuses the issues by title and fills them.
• A plan with no placeholders makes the same gh calls as before.
• roles/planner.md explains the placeholders and tells the planner never to predict numbers. Its JSON example uses the placeholders. The README documents them and adds a troubleshooting entry for the new error.

Tests

• src/phase.test.ts: filling, detection, and out-of-range errors.
• src/loop.test.ts: apply_plan fills created and existing tasks with the numbers they actually got, and does not re-edit a created task that has nothing to fill. Each task's body is filled before it is armed. A bad placeholder throws with no gh writes.
• src/roles.test.ts: {{issue}} is allowed in planner.md as a placeholder that apply_plan fills, not the prompt composer.

pnpm lint, pnpm typecheck, pnpm test (714 passed) and pnpm build are clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JrFSqr3eJsNihHcgYph1hB