Meeting/transcript summarization

summarize-meeting is the first concrete AI workflow. A manifest such as examples/jobs/meeting-summary.org2-ai.json turns raw meeting or transcript material into a reviewable Org2 artifact:

npm run org2 -- ai run \
  --task summarize-meeting \
  --file raw/meeting.org2 \
  --out views/meeting-summary.org2

npm run org2 -- ai run \
  --job examples/jobs/meeting-summary.org2-ai.json \
  --out views/meeting-summary.org2

The generated draft is safe-by-default: preview-only unless --apply is passed, stored as a generated view artifact, and marked ORG2_REVIEW_STATUS: review-required. The body includes:

• a concise meeting summary

• key decisions with file:path::line citations

• TODO-style action item suggestions

• people/org/project entity candidates

• suggested link candidates that must be reviewed before promotion

• source citation and provenance metadata, including source hashes and adapter/model metadata

The built-in deterministic mock adapter keeps tests and local examples provider-free while exercising the same adapter/request boundary that hosted or local model adapters can implement later.

Draft format

A generated draft starts with an Org2 property drawer:

:PROPERTIES:
:ID: team-sync-summary-team-sync-summary.org2
:ORG2_ARTIFACT_SCHEMA: org2-artifact-metadata/v1
:ORG2_ARTIFACT_ROLE: view
:ORG2_PROVENANCE: file:raw/team-sync.org2
:ORG2_GENERATOR: org2 ai run team-sync-summary adapter=local-default model=fixture-model
:ORG2_GENERATED_AT: 2026-05-20T01:00:00.000Z
:ORG2_AI_JOB_ID: team-sync-summary
:ORG2_AI_TASK: summarize-meeting
:ORG2_PROMPT_TEMPLATE: meeting-summary@v1
:ORG2_AI_ADAPTER: local-default
:ORG2_AI_MODEL: fixture-model
:ORG2_SOURCE_HASHES: file:raw/team-sync.org2=sha256:...
:ORG2_REVIEW_STATUS: review-required
:END:

The body includes a review checklist, job metadata, source list, generated summary or extraction section, and cited source excerpts using file:path::line links.

Review and promotion

Promotion is explicit:

1. Inspect the draft and verify generated claims against cited source lines.

2. Edit the draft until the accepted content is safe for canonical notes.

3. Change ORG2_REVIEW_STATUS from review-required to reviewed.

4. Run org2 ai promote with --apply to append the reviewed body to a canonical note file.

npm run org2 -- ai promote \
  --file views/team-sync-summary.org2 \
  --to-file notes/team-sync.org2 \
  --apply

Promotion refuses drafts that are still generated or review-required. On success it appends only the draft body to the target note, omits the generated artifact drawer from canonical notes, and marks the source draft promoted.

Lint rules

org2 lint flags generated artifacts that still need review with artifact-generated-unreviewed. Existing artifact checks also catch missing provenance, stale provenance mtimes, invalid or mismatched source hashes, and invalid review statuses.

This makes AI output behave like a build artifact: useful, traceable, lintable, and overwriteable until a human explicitly reviews and promotes it.

VS Code commands

The VS Code extension exposes the same safe path through command-palette actions:

• org2.aiRunDraft picks a job manifest, previews org2 ai run, confirms, writes the draft, and opens it for review.

• org2.aiPromoteDraft previews org2 ai promote for the active/selected draft, requires confirmation, and opens the canonical target after promotion.

Both commands use the CLI; the editor does not bypass review status checks or write canonical notes without the explicit promote confirmation.