Quality & Governance

Deliverables

What Pulse AI agents produce so task work can be inspected, heard, and approved.

The promise

Deliverables are the review surface for agent work. They translate the work into artifacts a human can inspect without reconstructing the whole terminal session.

Depending on preferences and task type, deliverables can include an HTML presentation, Markdown summary, screenshots, validation logs, structured data, and narration. In this workspace, task deliverables require both HTML and Markdown, and narration is required when voice preferences are enabled.

Deliverable types

TypeWhat it is forWhere it belongs
HTML deckVisual, navigable explanation of the work and evidence..pulse/knowledge/<slug>/artifacts/
Markdown summaryFast scan, changelog-friendly closeout, or review notes.Same KI artifacts folder
Narration WAVSpoken briefing tied to the deck or response.Same KI artifacts folder
Screenshots/mediaVisual proof of product state, KI views, review surfaces, or validation output.Same KI artifacts folder
Validation artifactsLogs, JSON, manifests, source maps, or route proof.Same KI artifacts folder

scripts/finalize-deliverable.js is the preferred path because it attaches the KI to the board item and keeps the artifacts discoverable in review.

How deliverables travel into review

The review path is:

  1. Work starts from a board item.
  2. Phase 1 defines checks and commits the manifest.
  3. The task produces evidence.
  4. scripts/finalize-deliverable.js packages the evidence under .pulse/knowledge/.
  5. The task frontmatter gets a deliverables entry such as ki:<slug> or a presentation reference.
  6. scripts/completion-gate.js validates the package and promotes the task to review.
  7. The reviewer opens the task, the KI, and the artifacts before deciding.

If the board item does not link the KI, the reviewer may never see the work. If the KI exists but the deck is stale, the reviewer should reject or request revision.

Example: what the reviewer opens

For web-fix-reset-password-memory-leak, the reviewer opens:

Review objectWhy it matters
Board taskConfirms the original request, completion notes, rubric, and deliverables link.
KI overviewExplains the closeout and the sibling-scope boundary.
HTML deckPresents the evidence in a reviewable format.
Narration WAVGives the spoken summary required by voice preferences.
Validation checksShows which commands were used to prove the listener cleanup.

This is the expected request-to-evidence path: a reviewer should never have to trust an unlinked chat claim.

Media needs

Deliverables should include media when it improves review quality:

Page or task typeUseful media
Knowledge Item educationScreenshot of the KI artifact list or a deck cover.
Review Queue educationScreenshot of a task in review with findings and attached deliverables.
Completion Gate educationTerminal capture of a passing or failing gate run.
Research educationResearch card or diagram that separates product docs from deeper papers.

Do not include screenshots that expose secrets, private paths beyond what the user already sees, or agent governance material that belongs outside public docs.

Related links