Skip to content

Latest commit

 

History

History
103 lines (73 loc) · 8.72 KB

cursor.md

File metadata and controls

103 lines (73 loc) · 8.72 KB

Harness: Cursor (Grok parent + Composer implementer)

Status: live. Roles: ../. Executable defs: .cursor/agents/. Isolation is opt-in. Local subagents share the parent checkout unless you ask for a worktree or /in-cloud.

When to use

  • Human is in Cursor Agent (this repo’s default while Cline free quota is exhausted).
  • Parent model: Grok 4.6 (orchestrator + leader; planner inherits unless you skip the child).
  • Implementer model: Composer 2.5.

Do not recreate three Herdr panes. Cursor has no wait/prompt protocol between named long-lived workers.

Topology mapping

Role This harness Notes
orchestrator parent Agent chat (Grok 4.6) only human-facing session
leader folded into the parent same duties + LOG as ../leader.md
planner subagent .cursor/agents/planner.md model: inherit (Grok). PLAN.md/LOG.md only — not readonly because Part A writes PLAN.md
developer subagent .cursor/agents/implementer.md model: composer-2.5 · own worktree
verifier subagent .cursor/agents/verifier.md readonly: true, model: gpt-5.6-sol-medium — run after PR-DRAFT. Covers planner Part B. Do not also spawn planner review unless verifier asks.

Small phases may skip spawning planner and refine PLAN.md in the parent. Still write PLAN-READY into LOG.md.

Alternate (two pinned chats)

Agents Window: Chat A = Grok (plan/ask), Chat B = Composer + /worktree. The human is the bus. Use this when you want to watch implementation without Task-tool folding. Same role contracts.

Isolation

Mode Use
Shared parent checkout verifier (readonly); planner may share because it only writes PLAN.md/LOG.md
Git worktree implementer on product code (/worktree, Agents Window, or agent --worktree)
/in-cloud long implementer runs; own VM + branch. Cloud MCP ≠ local .cursor/mcp.json
/best-of-n competing implementations, not a role team
/multitask only when file sets do not overlap

Never two write-capable agents on the same checkout.

Model pinning

Role Pin Rationale
orchestrator / leader Grok 4.6 in the parent picker long tool loops, instruction following
planner inherit same judgment as parent
developer composer-2.5 (composer-2.5-fast if the Fast variant is the picker name) edits + terminal
verifier gpt-5.6-sol-medium draft review (SPEC + gates). OpenAI models in Cursor: proposed shutoff 2026-11-12 (OpenAI). After that, fallback inherit (Grok) and LOG it.

On legacy request-based plans without Max Mode, Cursor may ignore model: and run subagents as Composer. If that happens, run planner/verifier in the parent Grok chat instead of Task, and LOG the fallback.

If Grok quota is exhausted: keep this harness, switch the parent picker to whatever reasoning model is available, and LOG it. Do not silently start Cline panes from a Cursor session.

OpenAI / GPT window and usage

  • Official proposed shutoff of OpenAI models in Cursor: 2026-11-12. Do not start a Sol review on or after that date; switch verifier to Grok (inherit) first.
  • After every Sol review, append to .tasks/phase-N-*/LOG.md:
    • model: gpt-5.6-sol-medium
    • start/end timestamps
    • Cursor usage if the UI shows it (request cost on the review turn, or Settings → Usage delta). There is no billing API in this harness — if the number is not visible, write usage: not visible and still record the model + time.
    • verdict (REVIEW APPROVE|CHANGES)
  • If one review looks expensive relative to a Grok/Composer pass, stop and ask the human before the next Sol review. Candidate fallbacks: Grok (inherit), then Composer.

Spawn sequence (per phase)

  1. Parent (Grok) reads SPEC + PLAN + this file. Creates or updates .tasks/phase-N-*/LOG.md (leader duties).
  2. Use the planner subagent (or /planner) with the Part A brief. Wait for PLAN-READY. Human may still be asked to approve PLAN deltas. Planner is not readonly (it writes PLAN.md); treat product-code edits as a bug.
  3. After PLAN-READY: Run the implementer subagent on Composer in its own worktree with the developer brief (../developer.md). Isolation phrase is mandatory. Implementer must open --draft and stop at PR-DRAFT.
  4. On PR-DRAFT: /verifier on GPT 5.6 Sol medium (covers planner Part B). Do not spawn a second reviewer. Record usage in LOG. Do not mark the PR ready yet.
  5. On REVIEW APPROVE: implementer (or parent) runs gh pr ready, then PR-READY.
  6. Parent records ready + asks the human to merge. Do not gh pr merge unless the human explicitly asked for that phase (Phase 0 was that exception).

Parent prompt (copy):

Harness: cursor. Fold leader into this chat.
Use the planner subagent first (PLAN.md only). After PLAN-READY and human ack,
run the implementer subagent on Composer in its own worktree.
Implementer opens a draft PR (PR-DRAFT). Then run verifier on GPT 5.6 Sol medium.
After REVIEW APPROVE, gh pr ready. Do not merge. Do not implement product code in this chat.

Invoke explicitly with /planner, /implementer, /verifier when automatic delegation picks the wrong child.

MCP / skills

Project MCP: .cursor/mcp.json — Astro Docs at https://mcp.docs.astro.build/mcp.

Parent and local subagents inherit it. Cloud children do not; register the same URL as a Team MCP if implementer is /in-cloud.

Failure handling

  • Approval / destructive: parent asks the human (Cursor Run Mode). Never tell implementer to --force git or skip hooks.
  • Stall: Stop, or resume the child by agent ID (Resume agent <id> …).
  • Implementer dirty-tree conflict: stop, move to a worktree, do not rebase the human’s checkout.
  • Nested spawn: main + direct children only (no grandchild Task). If implementer needs search, it uses built-in Explore — it must not spawn another implementer.

Do not

  • Pin muse-spark / glm-5.3-flash in this file.
  • Treat /best-of-n as leader/planner/developer.
  • Assume cloud children see local MCP or Herdr.