Architect

Single skill, two model tiers:

• Planning tier (Phases 1–3): clarify, research, create the plan and task file
• Worker tier (Phases 4–5): execute the task file, run lint checks, fix feedback

Follow AGENTS.md communication + reference-doc rules.

Phase 1 — Clarify

If requirements are unclear, ask one grouped message. Skip if already clear.

Phase 2 — Research

Build context from graft first, then read only what you need:

1. bin/harness graft ask "<research question>" --source to locate/understand · bin/harness graft grep "<symbol>" to find every occurrence · bin/harness graft callers <sym> (or --depth all for multi-file changes) to trace edges · bin/harness graft skeleton <file> for an API skim. Add --in <dir>/ to narrow a subtree. Always through bin/harness graft — a bare graft answers an empty graph as "no matching nodes" and exits 0.
2. Open source files only at the exact file:line graft names — never whole files to rebuild understanding graft gives.
3. .agents/notes/ for the same slug
4. AGENTS.md for the project's layout and its commands — it is on every project; workflow.md beside it when the project has one Follow AGENTS.md reference-doc rules (never load a whole file under .agents/docs/; grep one section). Live by the graft skill: "use graft before raw grep/code-read" is authoritative.

Phase 3 — Plan (planning tier)

Create .agents/tasks/YYYY-MM-DD-<type>-<slug>.md. Generate uid from the current time (YYYYMMDD-HHMMSS) if known; otherwise use the date plus a random suffix (YYYYMMDD-<random>).

---
type: <type>
slug: <kebab-case>
date: YYYY-MM-DD
uid: YYYYMMDD-<random>
status: planning
review_required: yes | no
risk: []
---

## Goal
## Context
## Tasks
- [ ] 1. ...

## Edge cases
- [ ] 1. ...

## Session checkpoint
## Log
## Feedback

Enumerate 3–5 edge cases relevant to the change; ensure the plan covers each. REST/SQL/auth/user-input surfaces must cover at minimum: missing/invalid input, capability/nonce failure, and the empty/oversize boundary.

Set review_required: yes and risk: per AGENTS.md when REST, SQL, auth, export, multisite, user input, or new admin UI is involved.

Show a ~10-line summary of the plan. End with: "Ready to start, or would you like to adjust anything?" Ambiguous SQL/auth/REST design: Tell the user to re-run planning on audit tier before any coding.

Phase 4 — Execute (worker tier)

On approval, run all tasks in order.

Lint tiers

Task touches	Lint
REST, SQL, auth, sanitization, PHP logic, JS/React	Immediate after task: run composer run verify:php and npm run lint + npm run lint:css
Comments, docs, pure CSS, no new surface	Batch every 5 tasks
All tasks done	Run npm run test:all
When lint is skipped, say: "Lint skipped — no PHP/JS surface changed."

Re-evaluate review_required after implementation

After all tasks, compare the actual code against the task file's review_required/risk. If it touches a surface (REST, SQL, auth, user input, export, multisite, new admin UI) the plan did not flag, update the frontmatter so the review gate matches what shipped.

Thread rotation

After 3 tasks or ~20 turns, update ## Session checkpoint and tell the user: start a new worker tier thread with /resume and this task file.

Retry ceiling

After 3 failed attempts on the same task (any failure type), STOP and re-plan — update ## Tasks/## Log, tell the user what's wrong. Do not grind past three iterations; different error types accumulate toward the same ceiling.

Self-review before "done"

Before reporting completion, silently argue against your own solution (redundancy, unused code, simpler alternative, missed edge case). Fix or note anything surfaced. Then add a one-line self-review note: checked [redundancy/unused code/edge cases/alternatives] — nothing surfaced, or noted [issue] — see Log.

Stop conditions

Stop only for uncovered design decisions, frozen documentation, version bumps, or major scope creep.

Completion message

On completion: status: done, full test gate, one-line summary of what was done and what files changed. Ask user to test.

Phase 5 — Feedback (worker tier)

Fix reported issues; append to ## Feedback. When stable: review_required: yes → "Open /reviewer on audit tier (planning tier OK only for low-risk optional review)."; else "Review optional; ship when satisfied."

Implementation reflexes

ABSPATH guard · the project's own prefix on every hook — zenpress_ is derived from the text domain, which is the same string for the projects this was written for and a different one for any project that was renamed, forked or shortened its prefix, so confirm the prefix with the owner before relying on it, and ask rather than assume when the code disagrees with it · REST current_user_can() · $wpdb->prepare() · sanitize in / escape out · docblocks on public surfaces · never reword existing i18n strings.

Project specifics

The reflexes above are the baseline. What this project adds — its own capability names, REST namespace, tables, prefix exceptions, block bindings — is in AGENTS.md, and it wins where the two disagree. Read it before your first edit, not after.