research-mentor

name	research-mentor
description	Long-running research mentor + boss-prep (科研导师 / 发展规划 / 老板汇报). Tracks self-trajectory (stated goals, weekly check-ins, decisions, drift) AND boss context (profile, meeting log, rehearsal) in AgentDB `project/` namespace. Use when the user says "check in", "what should I be working on this week", "am I on track", "回顾本周进度", or "/mentor boss …" / "/boss …" for boss-prep flows.

research-mentor

STATUS: active. mentor.diff_goals + mentor.weekly_checkin_template are real; the project-level research-notes triplet (state.yaml + findings.md + log.md) is implemented in research_assistant.mentor.research_notes; AgentDB writes go through the MCP memory_store tool. Boss subcommands delegate to the boss-historian agent at .claude/agents/boss-historian.md.

When to use

• Weekly / biweekly check-ins.
• User asks "what should I be working on", "is my project on track", "should I pivot".
• User wants a retro at the end of a sprint / paper / quarter.
• User runs /mentor boss … (or the /boss … alias) for boss-prep flows.

Boss flow (delegation)

When $ARGUMENTS starts with boss, strip the leading boss token and hand off to the boss-historian agent at .claude/agents/boss-historian.md. That agent owns the workflows for show, edit, meeting, rehearse, sync, and reads/writes inputs/boss-profile/{profile,meetings,reports,rehearsals} plus AgentDB namespace project/boss/.

This is a delegation, not a re-implementation — keep all the boss-specific prompt logic in the agent file so /boss (the alias) and /mentor boss land in exactly the same place.

Workflow

State model (AgentDB namespace project/)

project/goals               → list of long-term goals + their priorities
project/checkins/YYYY-MM-DD → weekly snapshot (done / blocked / decisions / mood)
project/decisions/<slug>    → architecture-style records for project decisions
project/backlog             → idea backlog, ranked

Check-in flow

1. Read project/goals + the last 3–4 project/checkins/*.
2. Read recent activity signals: new files in outputs/summaries/, new outputs/drafts/<slug>/..., new ideas/* entries.
3. Call research_assistant.mentor.diff_goals(goals, recent_activity).
4. Call research_assistant.mentor.stale_experiments(min_age_days=14) — returns active experiments whose latest versions/<vN.M>.md (or manifest.md if no version exists yet) hasn't been touched in 14+ days. Surface them under a ⏳ "Stale experiments" bucket alongside the goal-vs-activity diff. Each row: slug · latest_version · N days idle. An experiment can look on-track via the goals diff while quietly stalling here — that's the drift signal this catches.
5. Surface: ✅ on-track / ⚠️ drifting / ❌ stalled / ⏳ stale-experiments. Suggest 1–3 path corrections.
6. Write a check-in entry: outputs/mentor/checkin-YYYY-MM-DD.md AND mcp__claude-flow__memory_store into project/checkins/YYYY-MM-DD.

Goal-setting flow

1. Interactive: ask user for 1–3 long-term goals + their definitions of success.
2. Persist to project/goals.

Quick capture flow (/mentor add-past-work [<title>])

Optimized for "I just remembered an old project, log it before I forget." Stays inside the mentor turn — minimal Q&A, smart defaults, return to the check-in.

1. If no <title> in args, ask once (plain text, no AskUserQuestion per the feedback_decision_ui memory): "Title for this past-work entry?"
2. Compute defaults: d = mentor.past_work.quick_capture_defaults(title). This already runs next_available_slug, so the slug is collision-safe.
3. Plain-text optional probes — ask only one round, accept blank to skip:
   • "GitHub repo URL (optional, blank to skip):" — if given, validate it looks like a git URL and stash it for step 5.
   • "One-line abstract (optional):" — into d["abstract"].
   • "Status [in-progress / published / unpublished / abandoned, default in-progress]:" — override d["status"] if non-blank.
4. path = mentor.past_work.compose_past_work_entry(**d) writes the entry to inputs/past-work/<slug>.md with _TODO_ placeholders for anything still empty so the user knows what's left to fill in.
5. If a GitHub URL was given in step 3:
   • mentor.past_work.bind_repo(slug, url) writes the repo: frontmatter.
   • Plain-text Y/N: "Clone the repo locally now? [y/N]". On y, call mentor.past_work.clone_repo(slug). On N, leave it tracked.
6. Index it — to_agentdb_payload(entry) is the metadata=, not the whole call (it has no namespace/key/value fields). Match migrate/reindex.py's shape so a later reindex overwrites the same record:

   entry = parse_entry(path)
   mcp__claude-flow__memory_store(namespace="project/past-work", key=entry.slug,
       value=f"{entry.title}. " + "; ".join(entry.what_i_learned),
       metadata=mentor.past_work.to_agentdb_payload(entry))
   

   so the past-work-historian agent can recall it during /paper direction etc.
7. Print one line: "captured: <slug> → inputs/past-work/<slug>.md". Hand back to the prior mentor flow if one was in progress.

Project-level research notes (/mentor project ...)

A project sits one level above individual experiments and papers — it's "the research thread the user is pulling on for the next 3–18 months." Adapted from Orchestra-Research/AI-Research-SKILLs (MIT) 0-autoresearch-skill.

Each project lives at outputs/research-notes/<slug>/ and owns three files (all helpers in research_assistant.mentor.research_notes):

• state.yaml — ResearchNotesState (project meta, literature shortlist, hypothesis tree, experiments registry, outer-loop cycle counter).
• findings.md — running synthesis. Read this at the start of every session.
• log.md — append-only chronological decision timeline.

The "current project" is a slug stored in AgentDB project/research-notes-context.current (parallel to paper-context and experiment-context). Set on init and show.

Subcommand router:

Subcommand	Action
project init <title>	slug = research_notes.slugify(title); ask for a 1-sentence research question (plain text); research_notes.init_project(slug, title, question); set the cursor.
project list	research_notes.list_projects() → table.
project show [<slug>]	Resolve slug (arg → cursor); print state.yaml summary + findings.md "Current understanding" section + last 5 log rows via research_notes.parse_log(slug).
project log <kind>: <summary>	research_notes.append_log(slug, kind, summary). <kind> ∈ bootstrap / inner-loop / outer-loop / pivot / report / conclude.
project finding <section>: <body>	research_notes.append_finding(slug, section, body). Section is one of Current understanding / Key results / Patterns and insights / Lessons and constraints / Open questions, but any heading is accepted.
project sync	Re-read every outputs/research-notes/*/state.yaml; mcp__claude-flow__memory_store each via research_notes.to_agentdb_payload(state) into namespace project/research-notes.

When this surfaces during a check-in. During the weekly check-in flow, after the goals-vs-activity diff, also:

1. Read project/research-notes-context.current. If unset, skip.
2. Read findings.md (last "Current understanding" + "Open questions" sections).
3. Surface in the check-in: "since last week, has anything in your current understanding changed? Any open questions resolved?". Encourage the user to drop a finding via /mentor project finding ....
4. The Orchestra "outer loop direction" decision (deepen / broaden / pivot / conclude) is what to ask the user when findings.md looks ready.

The project-notes triplet is human-in-loop; lucky-research does NOT run Orchestra's /loop 20m heartbeat. The synthesis cadence is the weekly check-in, not a 20-minute timer.

Outputs

• outputs/mentor/checkin-YYYY-MM-DD.md
• outputs/research-notes/<slug>/{state.yaml, findings.md, log.md}
• AgentDB entries under namespaces project/ and project/research-notes/.

Memory keys touched

• project/goals — read/write
• project/checkins/* — read/write
• project/decisions/* — write
• project/research-notes-context.current — read/write (cursor {slug})
• project/research-notes/<slug> — write (on init and sync)
• papers/*, ideas/*, drafts/* — read (recent-activity signal)

Conventions (settled defaults)

• Check-in cadence: every 7 days. The skill never auto-fires; /mentor is user-initiated. The cadence is what the check-in template promises ("last week's") and what stale_experiments(min_age_days=14) uses as a sensible multiple.
• Activity signal: file mtimes under outputs/summaries/, outputs/drafts/, outputs/papers/, outputs/idea-checks/, and outputs/experiments/*/versions/*.md (used by stale_experiments). AgentDB write timestamps are NOT used — the on-disk truth is the canonical record and survives migrations / AgentDB rebuilds.