research-workflow

name	research-workflow
description	Master meta-router for the 7-stage research pipeline. Use when user discusses research at the workflow level — 'how do I start research', 'what's the research pipeline', 'I have a research idea, what's next', 'guide me through a research sprint', 'what stage am I at', 'is my sprint working', '我有个研究想法', '下一步该做什么', '研究流程'. Routes to the right /research-* command based on sprint state + user intent.

research-workflow — meta-router for the 7-stage research pipeline

The user has 7 specialized research commands at ~/.claude/commands/research-*.md. This skill is the natural-language entry point when the user describes research-shaped work without specifying a stage. Map intent → the right command.

The 7-stage chain

0. /research-init <tag>          ← initialize sprint (creates branch + ledger + TODO chain)
1. /research-survey [topic]      ← Obsidian SOTA + Gemini Flash cross-domain mining + Codex novelty pre-screen
2. /research-axiom               ← Opus 4.7 first-principles distillation + hidden axiom enumeration
3. /research-propose             ← cross-product reframing + Codex novelty + Gemini challenger + 3-step MVE
4. /research-dispatch            ← parallel niche subagents (single-message peer Sonnet 4.6)
5. /research-loop "<idea>"       ← autoresearch experiment cycle (5min single-GPU + R3 seed + R7 failure + surface_implicit)
6. /research-review              ← triple-heterogeneous review (Codex correctness + Gemini challenger + Sonnet red-team)
                                 ↺ back to /research-dispatch for next cycle until SOTA / budget / BLOCK / human halt

Plus 2 admin commands:

• /research-status — read-only dashboard (ledger tail, niche coverage, entropy alert)
• /research-constitution — view/amend the 12-rule constitution (rare; Opus 4.7 for amend)

Intent → command routing table

User intent (paraphrased)	Sprint state required	Command to invoke
"Start research on X", "create new sprint"	NO active sprint	/research-init <tag>
"Survey literature on X", "what's been done in field Y"	sprint exists	/research-survey [topic]
"Apply first principles", "find hidden axioms"	survey done	/research-axiom
"Generate research proposal", "check novelty", "design MVE"	axioms done	/research-propose
"Dispatch parallel niches", "spawn 9 subagent branches"	proposals done	/research-dispatch
"Run experiment X", "try this idea on train.py"	dispatch done OR --standalone	/research-loop "<idea>"
"Audit kept result", "triple review", "red-team this"	results.tsv has keep row	/research-review
"Show me sprint status", "what stage are we at"	any	/research-status
"Edit the constitution", "view the 12 rules"	any	/research-constitution

Sprint-state detection (run BEFORE routing)

Before dispatching, check current state:

RGMARE_ROOT="${RGMARE_ROOT:-$HOME/research/rgmare-lite}"
CURRENT_BRANCH=$(git -C "$RGMARE_ROOT" branch --show-current 2>/dev/null)

# Has active sprint?
case "$CURRENT_BRANCH" in
  research/*)
    SPRINT_TAG="${CURRENT_BRANCH#research/}"
    SPRINT_DIR="$RGMARE_ROOT/ledgers/$SPRINT_TAG"
    test -f "$SPRINT_DIR/TODO.md" && grep -E '^\s*-\s*\[' "$SPRINT_DIR/TODO.md"
    ;;
  *)
    echo "no_active_sprint"
    ;;
esac

If no_active_sprint AND user intent is anything past stage 0 → route to /research-init first, ask for tag, THEN chain to user's intended stage.

Hard rules (DO NOT violate)

1. Don't skip stages. If user asks for /research-propose but axioms/axioms.md is missing, route to /research-axiom first; explain why. Each command's HARD GATE will refuse anyway.

2. Don't auto-chain past stage 6. After /research-review, the user picks: continue cycling (back to dispatch) or halt. Never auto-advance without user signal — this is autoresearch's NEVER STOP discipline applies inside a single /research-loop, NOT across review verdicts.

3. Honor halt conditions. If ~/.hermes/shared/.research-halt exists, OR last review verdict was BLOCK, OR results.tsv shows SOTA reached, refuse to advance. Surface to user.

4. Don't bypass externalization gate. Each command requires a disclosure: / surface_implicit block. If asked to skip it ("just give me the answer"), refuse — that's the load-bearing trust mechanism per user-directive 2026-04-27.

5. Cron orchestration is separate. The Hermes cron driver ~/.hermes/scripts/research-cycle.py runs unattended every 360m. This skill is for INTERACTIVE invocations. Never modify cron state here.

When to invoke vs decline

Invoke when:

• User describes research-shaped work (sprint, experiment, paper, proposal, axiom, novelty, review)
• User has a research workspace at ~/research/rgmare-lite/ or asks to create one
• User mentions any of the trigger keywords in this skill's description

Decline (route to other skills) when:

• User wants engineering code review → codex:codex-rescue or gemini:gemini-consult
• User wants to write a blog post about research → blog-* skills
• User wants to plan a non-research feature → planning-with-files:plan or superpowers:writing-plans
• User wants to debug a single training crash → gsd:gsd-debug or direct Claude work, NOT a full sprint

Cross-tool routing

Within the research pipeline, the academic-mode wrappers handle external models:

• Codex GPT-5.4 academic novelty / citation-verify / claim-vs-code-vs-numbers → ~/.hermes/skills/routing/codex-academic/
• Gemini 3.1 Pro / Flash challenger feasibility / cross-domain mining / arxiv-probe → ~/.hermes/skills/routing/gemini-academic/
• Hermes cron autonomous stage advance → ~/.hermes/skills/cron-jobs/research-cycle/

The 7 commands invoke these wrappers internally; this skill should not call them directly.

Anti-patterns

• ❌ DO NOT execute the entire 7-stage chain in one turn. Each stage has its own externalization + audit gate; skipping invocations defeats them.
• ❌ DO NOT invoke /research-loop without an active dispatch (or --standalone). It needs a niche-scoped proposal.
• ❌ DO NOT decide PASS / BLOCK on a kept experiment in this skill — that's /research-review's job; here we only route.
• ❌ DO NOT explain to the user "this is what stage X does" — invoke the command, let its body explain via the user-visible output.