菜单
Use when the user invokes /research-op or needs guarded changes to an existing research package or Scope SSOT: row/card/section edits, inventory field updates, checks/lints, artifact event fanout, registry additions, scope transitions, or natural-language status fixes.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when the user invokes /research-auto or asks to run an autonomous research campaign over one Direction toward a measurable gate.
The Step-3 direction-formation on-ramp. Use when the user has only a vague or partial research idea and cannot yet state a clear Direction, or types /research-brainstorm, or asks to brainstorm / shape / explore a research direction before committing. Captures cheap pre-package, pre-SSOT ideas onto the dashboard brainstorm lane, automatically generates an English brainstorm HTML detail page for each idea, and converts one or more ideas into a single Direction proposal submitted through Triage. The agent only PROPOSES the Direction — the PM ratifies. Project-agnostic. Requires a committed Project node (run /research-onboard or /research-scope first).
Set up, repair, or validate the global research_html/ HTML dashboard scaffold for any research project. Use this skill whenever the user types /research-dashboard, asks to create or initialize a research dashboard, repair a broken research_html/ scaffold, install the four lane pages (brainstorm / in-progress / success / fail) before creating research packages, bring a project under the binding HTML rules R1-R18 and Trustworthy Research rules T1-T24, or set up the global research-system protocol for autonomous research agents. Project-agnostic: works in any repository the user opens. The dashboard chrome is overview-only — claims, evidence, and stage transitions live on package pages, not on the dashboard.
Use when bridging a raw or existing workspace into the Scope SSOT with /research-onboard: detect/scaffold or analyze the workspace, write prior knowledge, draft a Project objective, and create a PM-ratifiable Project Scope proposal. Always show the proposed Scope clearly to the PM with the next ratification step. The skill only proposes through Triage; it never commits the SSOT or creates packages.
Use when defining or revising Project, Direction, or Task Scope intent, or when execution discovers a needed scope change. The agent only proposes through Triage and must show every proposed Scope clearly to the PM with the next ratification step; accepted Scope writes route through research-op's scope-transition op. Never invokes git.
Use when the user invokes /research-package or asks to create, initialize, draft, scaffold, or materially restructure a research package under research_html/packages, including materializing accepted Scope SSOT Directions, adding dashboard package entries, or large structural results.html edits.
| name | research-op |
| description | Use when the user invokes /research-op or needs guarded changes to an existing research package or Scope SSOT: row/card/section edits, inventory field updates, checks/lints, artifact event fanout, registry additions, scope transitions, or natural-language status fixes. |
| allowed-tools | Bash(python3 *), Read, Edit, Write, Grep, Glob |
| context | fork |
| disable-model-invocation | false |
The single mutation surface for existing research packages. Replaces direct Edit/Write on package files. Enforces the (category, status, op, target) legality matrix and per-target invariants before any byte hits disk. Appends one audit line per op to a local jsonl log.
Two shapes — structured (autonomous) and natural-language (user).
# Structured (agent + workflow.ts + scan-events callers)
python skills/research-op/scripts/research_op.py --pkg <id> --op insert --target methodsTried --payload '{...}'
python skills/research-op/scripts/research_op.py --pkg <id> --event CHAIN_DONE --payload '{"artifact": "..."}'
python skills/research-op/scripts/research_op.py --pkg <id> --op check --scope all
python skills/research-op/scripts/research_op.py --pkg <id> --op check --scope scope-alignment
python skills/research-op/scripts/research_op.py --pkg <id> --op scan-events
python skills/research-op/scripts/research_op.py --pkg _scope --op scope-transition --from-triage <triage-id>
python skills/research-op/scripts/research_op.py --pkg _scope --op scope-transition \
--payload '{"id":"dir/<id>","level":"direction","parents":["project/main"],"version":1,"status":"ACTIVE","spec":{...},"source":"txn-0","op":"create","gate":"USER_CROSS_MODEL_AUDIT"}'
# Project-level knowledge registries (papers / edges / gaps) — durable cross-package stores
python skills/research-op/scripts/research_op.py --pkg <id> --op registry-add --target paper \
--payload '{"id":"dpr2020","title":"Dense Passage Retrieval","url":"https://arxiv.org/abs/2004.04906"}'
python skills/research-op/scripts/research_op.py --pkg <id> --op registry-add --target edge \
--payload '{"from":"paper:dpr2020","to":"paper:ours","type":"extends","evidence":"we adapt its dual-encoder"}'
python skills/research-op/scripts/research_op.py --pkg <id> --op registry-add --target gap \
--payload '{"id":"G1","summary":"no zero-shot benchmark for this domain"}'
# Unified rules registry (data/rules.js) — package rules via the package path,
# project rules via the synthetic _project context (needs a human ack), check anywhere
python skills/research-op/scripts/research_op.py --pkg <id> --op insert --target rule \
--payload '{"level":"package","kind":"binding","slug":"one-notebook","title":"One notebook per figure","text":"Every figure gets its own notebook.","rationale":"reproducibility","addedAt":"2026-06-11"}'
python skills/research-op/scripts/research_op.py --pkg _project --op insert --target rule \
--payload '{"level":"project","kind":"constraint","slug":"no-eval-leak","title":"No eval leakage","text":"Never train on the eval split.","rationale":"validity","addedAt":"2026-06-11","ack":"<verbatim human approval>"}'
python skills/research-op/scripts/research_op.py --pkg _project --op check --target rule
# Natural-language (user manual fixes)
python skills/research-op/scripts/research_op.py --nl 'update: set status of 2026-05-15-panda-baselines to BLOCKED, reason: GPU contention'
--nl is an escape hatch: the CLI does not parse prose — it returns exit 4 and asks for the structured form. Parsing is the agent's job, done here in the body, so the agent stays the one translator (no brittle regex in the script). To handle an NL fix:
set/update → update, add/insert → insert, remove/delete → delete, check/lint → check.YYYY-MM-DD-slug token in the prose).status, payload {"to":"BLOCKED"}; a trailing reason: … clause becomes an extra payload field.→ --pkg <id> --op <op> --target <target> --payload <json>.| Precondition | Check |
|---|---|
| Package exists | test -f research_html/packages/<id>/index.html |
| Inventory entry exists | grep -q "id: '<id>'" research_html/data/research-packages.js |
| Runtime root resolved | RESEARCH_RUNTIME_ROOT env or default outputs/<id>/ exists |
These package preconditions do not apply to --pkg _scope --op scope-transition; Scope transitions
are project-level and can create the first Project node before any package exists.
Primitives: insert · update · delete · check. Composite events: CHAIN_DONE · CHECKPOINT_SAVED · SENTINEL_WRITE · PHASE_MARKER · CANDIDATE_SUBMITTED. Full legality matrix in references/matrix.md. Per-event surface map in references/composite-events.md.
Three ops live outside the (category, status) matrix (they are project-level, not package surfaces):
scan-events — read-only artifact scan (no state-gate, no validation) that lists newly-locked facts for the per-turn propagation cycle.scope-transition — the one gated writer for the Scope SSOT, used by research-scope after human ratification. Invoke it with --pkg _scope, preferably --from-triage <triage-id> so the committed payload is the accepted proposal snapshot. It is gated by the node level (project / direction / task), not the package state machine, and appends one transition to outputs/_scope/transitions.jsonl. When research_html/data/ exists, it also refreshes and checks research_html/data/scope-projection.json/js. The explicit payload form carries the node fields (id, level, parents, version, status, spec, source) plus the transition meta (op, gate, trigger, cause, invalidates, reopens, dial_revert).registry-add — the gated writer for the project-level knowledge registries (--target paper | edge | gap), the durable cross-package stores the Context Pack reads. Gated by per-target reject-before-write validators (registry.py), not the package state machine; dedups and appends one line to research_html/data/{papers,edges,gaps}.jsonl. Payloads: paper = {id|arxiv|source_id (≥1 required), title (required), url, pkg}; edge = {from (required), to (required), type ∈ extends|contradicts|addresses_gap|invalidates (required), evidence}; gap = {id (required), summary (required), status}. A duplicate is a silent idempotent skip (still audited). --pkg is the adding context (must exist) and is recorded on the audit line.--target rule with --pkg _project — the project-level half of the unified rules registry (data/rules.js). Package-level rule rows flow the normal state-gated path (--pkg <pkg-id>, matrix rows I12/U14/D9); project-level rows use the synthetic _project context and require a non-empty payload.ack (the distinct human action supplied by the governed caller). level=universal is write-locked everywhere (the R/T mirror); origin ∈ {mirror, selfevolve} rows are export-owned. --op check --target rule wraps learnings_lint.py lint-rules. Retired targets package-invariant / analysis-rule reject with a pointer to this target.Every Insert / Update / Delete passes Phase 1 (state gate) and Phase 2 (invariant check) before bytes hit disk. Phase 1 looks up (category, status, op, target) in transitions.py; Phase 2 runs per-target rules from references/validate-rules.md. Reject envelope: {rejected: true, phase, rule, file, anchor, field, expected, actual, suggested_fix, op, target}. On reject, the agent retries with the rule visible.
For Scope-materialized packages, run --op check --scope scope-alignment before status transitions,
terminal routing, package result closeout, or any rewrite that depends on sourceDirection /
sourceTasks. This check wraps learnings_lint.py lint-status --pkg <id> and verifies package
provenance against the current Scope SSOT projection.
Read the structured envelope. The suggested_fix field tells you how to adjust the payload. Re-invoke with the corrected payload. Do not bypass — every rejection traces to a real spec invariant.
Path: outputs/<pkg>/_actions.jsonl. One JSONL line per op invocation (success or reject). Verbatim payload included. Never tracked in git. tail -f is the live-observability surface; grep '"validation": "OP_REJECTED"' is the agent-stuck debug surface.
methodsTried row cannot be deleted once status is in success/* or fail/* (enforced by rule_methodstried_terminal_frozen); appending rows and updating terminationMessage stay legal there. verdict and evidencePath are not independent --targets — they are fields inside a methodsTried / result-gate row, mutated through those targets, never directly.scripts/research_op.py — CLI entry.scripts/rules_store.py — data/rules.js load/save + row validation (the registry writer the rule ops use).scripts/transitions.py — the legality matrix as Python dicts (per-op target maps).scripts/events.py — 5 composite events.scripts/validate.py — Phase 2 rules.scripts/router.py — dispatcher.scripts/audit.py — jsonl writer.scripts/scan_events.py — artifact mtime scanner.scripts/ops/{insert,update,delete,check}.py — per-op handlers.scripts/ops/_pkg_block.py — HTML package-block parser shared by the insert/update/delete handlers.