autolit

name	autolit
description	Manage a repo-local academic literature workspace backed by one Zotero collection. Use when Codex needs to initialize `literature/`, list Zotero collections, regenerate `references.bib`, sync PDFs into markdown under `sources/markdown/`, distill pending papers into wiki pages and evidence rows, update cross-paper syntheses, or report workflow status.

Autolit

Use this skill for the full repo-local literature workflow. Treat the current git repo as the project root and <repo>/literature as the workspace.

Phase Selection

Choose the narrowest phase that matches the request:

• CLI plumbing: use autolit for list-collections, init, sync, status, and doctor
• Distill pending papers: use autolit sync first, then use autolit distill for the batch Codex-worker flow or coordinate workers manually inside Codex when you want interactive supervision
• Synthesize across papers: update the wiki directly after reading references/synthesis-contract.md

Do not jump to a broader phase unless the user asks.

Initialization

Treat the repo root as the parent of literature/. If literature/autolit.toml is missing and the user wants setup, use autolit list-collections only if needed, then run autolit init <collection-name>.

Sync And Status

Use the CLI for sync/status plumbing.

1. For deterministic workspace operations, use autolit:
   • autolit sync
   • autolit status
   • autolit doctor
   • autolit list-collections
2. sync is responsible for:
   • regenerating references.bib
   • refreshing sources/markdown/
   • archiving papers removed from the Zotero collection
   • emitting pending bundle files for paper distillation
3. Do not hand-run PDF conversion or manually rebuild workspace state unless you are debugging the skill implementation.
4. Do not write wiki/papers/ or wiki/syntheses/ during sync.
5. status and doctor are reporting phases only.
6. Synthesize remains skill-only. distill is available both as a CLI batch step and as an interactive skill workflow.

Distill Pending Papers

Use this phase after sync has produced bundle files under literature/state/paper-bundles/.

Coordinator Mode

Use coordinator mode when the request is to distill pending papers in general, not a specific bundle.

1. Run autolit sync first so the queue is current.
2. Prefer autolit distill when the user wants a non-interactive batch run over pending bundles.
3. If coordinating manually, act as a coordinator only. Do not interpret multiple papers in the coordinator context.
4. When multiple papers are pending, use one fresh worker per bundle.
5. For paper-distill workers, use:
   • agent_type="worker"
   • model="gpt-5.4"
   • reasoning_effort="xhigh"
6. Each worker should read exactly one bundle and the files it points to, write only the result JSON, and then stop.
7. autolit distill should use gpt-5.4 with xhigh reasoning effort and launch one isolated codex exec worker per bundle.
8. If sub-agents are unavailable, process exactly one paper bundle in the current invocation and then stop.
9. Paper distillation should not create or update wiki/concepts/*.md.
   • Use the paper page and evidence rows to name concepts.
   • Leave concept-page creation and cross-paper updates for the synthesize phase.

Paper Distiller Mode

Enter paper-distiller mode when the prompt names a specific bundle path or says to process exactly one bundle.

1. Read only the bundle JSON plus the files listed inside it.
2. Read references/paper-distill-contract.md before producing output.
3. Write exactly one result JSON at the bundle's result_path, do not edit shared wiki files directly, and stop.

Synthesize

Use this phase for cross-paper reasoning only.

1. Assume sync and paper distillation have already populated wiki/papers/ and wiki/evidence-matrix.md.
2. Read references/synthesis-contract.md before editing.
3. Synthesize owns concept pages.
   • Create new wiki/concepts/*.md pages when recurring concepts emerge across papers.
   • Update existing concept pages to reflect agreements, contradictions, scope conditions, and uncertainty.
4. Update only the cross-paper layer:
   • wiki/concepts/*.md
   • wiki/syntheses/overview.md
   • wiki/syntheses/comparisons.md
   • wiki/syntheses/open_questions.md
   • lightweight wiki/index.md or wiki/log.md touch-ups when they directly support synthesis

Boundaries

• Do not hand-edit sources/markdown/; update it only through sync scripts.
• Do not manually rewrite references.bib; regenerate it through sync.
• Never distill two papers in the same active LLM context.
• Keep concept and synthesis edits grounded, bounded, and reviewable.