research-plan

name	research-plan
description	AnchoredAutoResearch planning wizard for Codex. Builds the PLAN section with scope, metrics, verify/guard commands, predictions, and human approval gates. Supports --effort low|middle|high and --search N for literature-pool size.
version	1.0.0

AnchoredAutoResearch — PLAN Entrypoint

Use this skill when the user invokes $research-plan in Codex.

Optional flags:

• --effort low|middle|high controls ideation breadth. Default is low (current behavior). See ### Effort levels in ../research/references/phase-protocol.md for the per-level protocol. Reject any other value with: "unknown effort level — use low, middle, or high".
• --search N sets the literature discover-pool size when this invocation triggers /research:context search <topic>. Integer in [10, 500]; clamped. Default by effort: low ignores it, middle honors if passed, high defaults to N=200. Deep-read count is min(30, N // 5). See knowledge-sources.md "Search mode" for the two-tier protocol. Reject non-integer values with: "--search must be a positive integer in 10..500".

MANDATORY: Read Protocol Before Any Action

1. Read AGENTS.md for project rules
2. Read RSD.md to determine current state (create if missing only via INIT in ../research/references/phase-protocol.md)
3. Read context/SOURCES.md if it exists
4. Follow the PLAN section in ../research/references/phase-protocol.md
5. Follow ../research/references/knowledge-sources.md for source-reading requirements

Execution

1. Treat any remaining user text as inline goal or planning context. Parse --effort low|middle|high (default low) and --search N (integer in [10, 500], clamped). Resolve the effective --search value per effort level: LOW ignores, MIDDLE honors if passed, HIGH defaults to N=200.
2. Verify the project is in PLAN or can safely enter PLAN from INIT
3. Read unread sources in context/
4. Branch on effort level per ### Effort levels in ../research/references/phase-protocol.md:
   • low → one proposal tied to the user's goal (current behavior)
   • middle → 2-3 candidates, silent auto-pick, log alternatives
   • high → cost gate first; 5-8 candidates spanning sub-topics with at least one cross-field candidate; present via interactive selection
5. Build the PLAN entry with:
   • goal and hypothesis linkage
   • grounded-in source citation (every candidate at every effort level)
   • prediction before execution
   • verify and guard commands
   • execution mode
   • Effort level: and Alternatives considered: fields (use N/A (low effort) for the default level)
6. Set Status: WAITING_HUMAN
7. Run bash scripts/compile_rsd.sh
8. Commit the phase boundary
9. Stop and tell the human to approve or revise in chat or RSD.md using any clear wording