version: "1.0.1" name: analyze-project description: Read project assets (analysis.json) together with surrounding context (intake.json, edit.json, planning/, README, and past projects), create the interpretation layer interpretation.json, and render a read-only analysis report limited to facts and readings of the material. Use to understand a multi-asset project across sources, integrate completed analyze-footage results, or identify missing primary information for interviews before choosing a direction. edit-plan reads this output before deciding direction.

Analyze a project

Language: Respond in the user's language — conversation, questions, approval prompts, and the reports and records you write (including decision-log rows) follow the language the user writes in.

Do not Read all of edit.json / captions.json; grep by ID and read only matching lines (Reading edit.json). In other stages that need writes, Edit the relevant lines or use edit-store's script API.

Hard rules

• Do not rerun pass 1. Delegate per-asset fact extraction (heavy processing such as STT/frame analysis) to existing

analyze-footage analysis.json files. Run only for missing assets, 1 asset = 1 execution. Use existing analysis.json without modifying it.

• Display absent ledger chapters as “unobserved”. Do not imply they did not exist or fill them by inference

from assets/other chapters. Keep those report chapters visible and marked unobserved; use observations[] execution history as evidence when present.

• Do not rewatch footage in pass 2. Limit it to text reasoning over pass-1 outputs (transcript, keyframe notes, events).

Return to footage only at pinpointed locations pass 2 determines are necessary.

• Pass-2 inputs are not limited to analysis.json files. Read surrounding project context (intake.json,

planning/, README, and referenced past projects), and grep edit.json by ID to read relevant lines only. Field validation (internal multiasset-dogfood run, 2026-07-22) confirmed cases where analysis.json alone could not establish asset provenance (orphan/unclear, etc.) confidently.

• Include `memory` connections when reading surrounding project context. If .akari/connections.json

declares memory (contract-2026-07-25-memory-connection-v0.md, Japanese), start at entry (default INDEX.md) and read only the include/exclude scope. Record referenced file paths as sources in interpretation.json's inputs.context. Full-text ingestion is forbidden. Do nothing if undeclared; this is not an error.

• Every interpretation claim requires evidence: all assets[].summary/role/relations/flags and arc[].

Evidence is limited to quoted transcript ranges or keyframe-note references. Fabricate nothing. Do not write unsupported claims; send remaining gaps to open_questions.

• open_questions are interview questions only. Do not ask users to make decisions. Ask only for primary information

absent from both footage and the internet. See interpretation.md and references/open-questions-examples.md.

• Stop condition (80-point line): can the proposed arc form a coherent whole without fabricated facts?

If so, end interviewing. Send later questions to post-edit refinement (edit-plan).

• Do not display arc in the report. Keep it as schema data in interpretation.json (source of interview

questions and seeds for AI proposals during dialogue). Removal from display was settled in the 2026-07-22 revision.

• Do not add helper fields absent from interpretation.schema.json.

Finalize only JSON that passes validate-interpretation.mjs.

• No decision UI in reports (choices, knobs, or finalize buttons). Decide direction in chat.

The AI proposes with evidence instead of asking “What should we do?”

• Connect decision records to existing decision_log conventions (edit-plan). Do not invent a format.

See decision-record.md.

• Use OpenMontage only as structural inspiration; do not copy AGPL text or code.

Execution order and contents

1. Read collect.md, list project assets, and check for analysis.json.

Run analyze-footage only for missing assets, 1 asset = 1 execution.

1. Read interpretation.md; read all analysis.json files with surrounding project context,

then create interpretation.json with mandatory evidence and zero fabrication.

1. Read validate-and-render.md, confirm validate-interpretation.mjs PASS,

generate a report using renderer --analysis <ref>=<path>, and open it. Verify absent analysis.json chapters display as unobserved.

1. Only if open_questions remain, read interview.md and interview in chat

(no decision delegation). Promote answers to interview context and regenerate interpretation/report (return to steps 2–3).

1. Read decision-record.md; record interview Q&A and major pass-2 decisions in the decision log.

Read only the file for the current stage, without reading ahead.