---
alwaysApply: false
---

# Title: Autonomy & Context-Budget Control
# Scope: project
# Objective: Run tasks end-to-end without pausing for questions; never exhaust context; recover state after long runs.

## Operating Mode
- Operate as an IMPLEMENTER, not an advisor. Execute changes directly in the codebase, run tests, and iterate until the task is completed.
- Do NOT ask for confirmation. If information is missing, make the safest reasonable assumption, record it in `workflow_state.md` under **ASSUMPTIONS**, and proceed.
- Only stop when:
  1) all acceptance checks pass, or
  2) an explicit denylisted command is required, or
  3) a hard error cannot be bypassed without human credentials.

## Files for Long-Run State
- Maintain two small, bounded files at repo root:
  - `project_config.md`: non-volatile goals, constraints, stack specifics, commands to run, and acceptance criteria.
  - `workflow_state.md`: volatile loop state with **PLAN**, **NEXT_ACTIONS**, **ASSUMPTIONS**, **PROGRESS_LOG** (tail ≤ 200 lines), and **OPEN_ISSUES**.
- Always read from `project_config.md` first, then `workflow_state.md`. If absent, create them with reasonable defaults and proceed.

## Autonomous Loop (each cycle)
1) **Read**: Load only the smallest necessary snippets. Never paste whole files into chat.
2) **Plan**: Update `workflow_state.md` → **PLAN** and **NEXT_ACTIONS** (≤ 150 tokens combined).
3) **Act**:
   - Apply file edits directly via the editor. Prefer targeted diffs and surgical refactors.
   - Run fast checks (format, lint, typecheck, unit tests) as configured in `project_config.md`.
4) **Verify**:
   - Parse outputs; extract only the essential signals back into `workflow_state.md` (**PROGRESS_LOG** summary ≤ 100 tokens per step).
   - If failures occur, append a brief root-cause hypothesis and loop.
5) **Trim**:
   - Aggressively prune prior chat messages by relying on the two state files instead of conversational memory.
   - Summarize long logs into bullet points; never paste raw logs > 30 lines.
6) **Repeat** until acceptance criteria met. Then write a concise **DONE** report in `workflow_state.md`.

## Context Budget Rules
- Hard token guardrails:
  - Summarize code context instead of inlining it. Reference files and line ranges.
  - For large diffs, chunk work by directory or component; complete and verify one chunk before opening the next.
  - Never quote third-party library code. Link the package name and version only.
- On every loop, if the visible conversation > ~1–2 screens, summarize to `workflow_state.md` and clear prior verbose messages by continuing with short references.

## Execution & Safety
- Auto-run commands that are safe and reversible (formatters, linters, unit tests, local dev server).
- If a command appears on the **DENYLIST** (see below), do NOT execute. Instead, write an **INTENT** note in `workflow_state.md` asking for approval, then continue with other work.
- Prefer idempotent scripts and atomic commits to prevent partial breakage.

## Reporting Style
- Communicate progress inside `workflow_state.md` only.
- In chat, respond with a one-paragraph status and the next concrete action, no tutorials.

## DENYLIST (edit per project)
- Global destructive actions: `rm -rf /`, `sudo`, `diskutil eraseDisk`, cloud destroy commands (e.g., `terraform destroy`, `pulumi destroy`, `gcloud compute instances delete` without `--dry-run`), production DB migrations, any command targeting prod environments.
- Network or credentialed actions unless sandboxed.
- Cost-incurring benchmarks or load tests without limits.

## Minimal Acceptance Checklist (default)
- Builds successfully.
- Linters/type checks pass.
- All existing tests pass; new tests cover new behavior.
- README or `project_config.md` updated for any new command, env var, or migration.

Use timeouts when running continous actions like launching or compiling the app.