AGENTS.md · diff
git:20260810.16378f2 to git:20260906.398c90b
39 added, 14 removed. Audit A to A.
- <!-- TRELLIS:START -->
- # Trellis Instructions
+ # Project instructions
- These instructions are for AI assistants working in this project.
+ ## Keep the project aligned
- This project is managed by Trellis. The working knowledge you need lives under `.trellis/`:
+ - The [master roadmap](docs/superpowers/roadmaps/2026-08-02-qiongli-2-research-harness-master-roadmap.md)
+ owns direction, dependencies and the current execution horizon.
+ - The [program ledger](docs/superpowers/roadmaps/qiongli-program-ledger-v1.json)
+ owns status and accepted evidence; its
+ [current index](docs/superpowers/roadmaps/qiongli-current-program-index.md)
+ is generated. Never infer acceptance from a checkbox, passing local test or merge.
+ - Read the current bounded plan linked from the roadmap and only the relevant
+ package/layer specs under `.trellis/spec/` before changing their contracts.
+ Accepted ADRs own architecture; supersede them instead of rewriting history.
+ - Existing `.trellis/tasks/`, `.trellis/workspace/` and `.trellis/spec/` preserve
+ plans, decisions and evidence. They are ordinary project knowledge, not a
+ mandatory task engine. Do not reinstall Trellis skills, hooks or agents unless
+ the user asks.
- - `.trellis/workflow.md` — development phases, when to create tasks, skill routing
- - `.trellis/spec/` — package- and layer-scoped coding guidelines (read before writing code in a given layer)
- - `.trellis/workspace/` — per-developer journals and session traces
- - `.trellis/tasks/` — active and archived tasks (PRDs, research, jsonl context)
+ ## Deliver small, complete increments
- If a Trellis command is available on your platform (e.g. `/trellis:finish-work`, `/trellis:continue`), prefer it over manual steps. Not every platform exposes every command.
+ 1. Inspect the working diff and trace the behavior to its existing owner.
+ 2. State the user outcome and the smallest useful change. Reuse an existing plan;
+ create one short plan only when the scope needs it. Routine fixes need no new
+ task directory, PRD, context manifest, phase approval or journal entry.
+ 3. Implement within the user's requested scope and run the closest meaningful
+ checks. Continue through routine fixes without asking for approval again.
+ 4. Review the final diff; update contracts when behavior changes and record what
+ passed, what remains unverified and the next increment in the existing plan.
- If you're using Codex or another agent-capable tool, additional project-scoped helpers may live in:
- - `.agents/skills/` — reusable Trellis skills
- - `.codex/agents/` — optional custom subagents
+ The main Agent may implement and check directly. Delegate only bounded independent
+ work with clear file ownership when it helps; no mandatory role chain or channel.
+ Parallel work shares a stable interface and one integration outcome. A blocked
+ external validation lane does not stop independent offline development.
- Managed by Trellis. Edits outside this block are preserved; edits inside may be overwritten by a future `trellis update`.
+ ## Preserve the product boundaries
- <!-- TRELLIS:END -->
+ - Follow [CONTRIBUTING.md](CONTRIBUTING.md) for source ownership and daily, PR,
+ build and release checks. Run affected checks while editing; use required CI
+ when preparing integration, and qualify packages at a named candidate boundary.
+ - Keep input validation, permission checks, compatibility and data-loss negative
+ cases. Qiongli project writes still use existing preview/approval/CAS owners.
+ - Preserve unrelated working changes and user data. Material scope changes or
+ external/destructive actions need authority for that action; ordinary coding
+ authorization does not imply publication or access to private research data.
+ - Commit, push, merge and publish only within the user's authorization. Never
+ auto-commit bookkeeping or start release work because a local check passed.