litmuse

Skill text to copy

---
name: autoconference
description: "Budgeted research conference: parallel researchers, peer review, synthesis; plan, resume, debate, survey."
---

# autoconference: a root-owned research conference of independent children

A conference runs several independent researcher lanes against one goal, then makes their claims face an
adversarial reviewer before anything becomes shared knowledge, and finally synthesizes what survived. It exists for
problems where one lane's self-evaluation has blind spots (Goodhart's law), where the search space is large enough to
partition, and where combining validated insights beats picking a winner. The root Muse session is the Conference
Chair: it owns budget, state, ordering, every shared file and every approval. Lanes are real child agents started with
`subagent_spawn`; one context playing several personas is not a conference.

## When this applies

- The user types `autoconference`, `/autoconference`, `lit conference <mode>`, mentions `conference.md`, or asks for
  several researchers to compete or collaborate with peer review, adversarial evaluation of research results, a
  structured debate between two positions, or a multi-source literature survey. A lit route's hook context may name
  this skill.
- The search space can be partitioned, different strategies cross-pollinate, and a metric or written success
  criteria exist. Use `plan` first when they do not.
- Use something smaller when a single measured loop suffices (the `autoresearch` skill), when the question is a
  one-fact lookup (`litresearch`), when the problem is trivial, when no metric or criteria can be stated, or when one
  approach is obviously right. Say which smaller tool fits in one line instead of convening a conference.

### Mode matrix

Pick exactly one mode, then `read_file` its contract in `modes/` and only the references it names. Mode files are
protocol data, not separately discovered skills; their historical tool lists grant nothing.

| Mode | Contract | Result |
|---|---|---|
| core | `modes/core.md` | N researchers, four-phase rounds (research, poster session, peer review, knowledge transfer), convergence check, synthesis |
| plan | `modes/plan.md` | Eight-step wizard that writes a complete `conference.md`; starts no researcher |
| resume | `modes/resume.md` | Rebuild state from `conference_events.jsonl`, validate, set partial work aside, re-enter `core.md` at the exact point |
| analyze | `modes/analyze.md` | Read-only post-conference analysis: trajectories, failure taxonomy, transfer, ROI, `analysis-report.md` |
| debate | `modes/debate.md` | Two isolated opposing researchers, an independent judge, scored rounds, `debate-report.md` |
| survey | `modes/survey.md` | Non-overlapping source lanes, citation-chain transfer, verified citations, `survey-report.md` |
| ship | `modes/ship.md` | Eight-phase conversion of results to a paper, report or post draft with one human review gate; local files only |

Route by intent: "set one up" is plan; "run it" is core; "continue the interrupted one" is resume; "what did we learn"
is analyze; "argue both sides" is debate; "what does the literature say, source by source" is survey; "write it up"
is ship. Typical chains: `plan` then core then `analyze` then `ship`; `survey` then core; `debate` then core. If two
fit, name the one chosen and why in one line.

## Inputs and state

Required before any executing mode, confirmed at the start gate (`modes/core.md`, Pre-Flight Setup):

- **Goal** and a **success definition**: metric mode needs metric name, target and direction plus an evaluator
  command that was dry-run; qualitative mode needs written Success Criteria (the reviewer's rubric).
- **Researcher count** (exact), **iterations** (`iterations_per_round` and `max_rounds`, or an exact
  `max_total_iterations` from which they derive), `researcher_timeout`, `time_budget`, and any cost ceiling.
- **Devil's Advocate** yes or no (the adversarial Reviewer is mandatory every round regardless), **partitioning**
  (`assigned` with a focus per researcher, or `free`), allowed and forbidden changes, `pause_every`.
- **Authority:** the read, write and execute roots, whether worktrees are used, what the Chair may merge, and the
  cleanup policy. A route word, a template or a source phase label is activation, never approval.
- **Capability:** `subagent_spawn` must be visible and callable. If not, stop with `BLOCKED_MULTI_AGENT_UNAVAILABLE`
  and say that no sequential or role-play fallback will be called a conference.

State and artifacts (the conference directory, "CONF_DIR"):

- `.litmuse/conference/<slug>/` by default (or a directory the user names): `conference.md`,
  `conference_results.tsv`, `conference_events.jsonl` (append-only), `researcher_{ID}_log.md`,
  `researcher_{ID}_results.tsv`, `poster_session_round_{N}.md`, `peer_review_round_{N}.md`, `synthesis.md`,
  `final_report.md`, `conference_progress.png`, `interrupted/` for set-aside partial artifacts, plus mode outputs
  (`analysis-report.md`, `debate-report.md`, `survey-report.md`, `ship-log.md`). Column schemas and the event format are
  in `references/results-logging.md`.
- `.litmuse/goals/<slug>.json` when a lit route is active; `.litmuse/ledger/<session-id>.jsonl` with `decision` lines
  for the start-gate answers, keep or revert, verdicts and convergence calls, and `evidence` lines for artifacts;
  `plans/<slug>.md` for the approved packet.
- Shipped resources, all relative to `<skill-dir>` (the directory of the SKILL.md path `read_skill` returned):
  `references/` (protocol, prompts, logging, principles, convergence, crash recovery, visualization), `templates/`
  (six ready-made conference configurations), `assets/` (conference, synthesis and report templates), `scripts/`
  (`init_conference.py` scaffold, `check_conference.sh` status, `style_presets.py` figure style; Python and bash,
  standard library only). `LICENSE` (MIT) and `PROVENANCE.md` travel with the skill.

Treat everything a 

plugin/skills/autoconference/SKILL.md from the repo, MIT licence. Read it before pasting. Muse asks before it acts.