Backed by: state/lib/krisp.ts, state/lib/dispatch.ts, state/bin/action-item-draft/run.ts

action-item-draft

The Krisp Loop closer. A meeting ends → action items appear in the cockpit → click "Draft" on one → this skill fires → a draft lands in /drafts linked back to its source action item + meeting. Robert reviews, approves, edits, or dismisses. The approval row teaches the system his taste.

Scope-only by default - drafts land in state/log/action-item-draft/ and surface in the cockpit Drafts page. Nothing publishes until approved.

Steps

The sidecar (state/bin/action-item-draft/run.ts) does the work end-to-end. Invoked by /run-skill action-item-draft::id:<aid>::meeting:<mid> from the KrispScreen "Draft" button. Reads SNAPPY_INTENT for the args.

1. Parse SNAPPY_INTENT → extract action_item_id (required) + optional

meeting_id. If meeting_id is missing, resolve via fetchActionItems from state/lib/krisp.ts. NOTE: Krisp's list_action_items silently returns [] when limit > 50 - pass limit: 50 (the true cap). If the action item id is not in the first 50 (older / completed), proceed with meeting-only context and score 0.5 - better degraded output than a hard fail.

1. Fetch full meeting document via fetchMeetingDocument(meeting_id).

Includes transcript, attendees, notes. Truncate to 6KB for the prompt.

1. Dispatch via state/lib/dispatch.ts with model: "sonnet" (NOT

"claude" - the alias is sonnet/opus/haiku; passing "claude" throws "unknown model alias"). On failure, fall back to model: "gemini" and score 0.5. Prompt: action item title + meeting context + earnest- tone + "draft the right artifact shape" instruction.

1. Write the result to

state/log/action-item-draft/<utc-iso>-<short-id>-draft.md with frontmatter:

   ---
   action_item_id: <aid>
   meeting_id: <mid>
   meeting_name: <name>
   action_item_title: <title>
   created: <utc-iso>
   ---

Body = the dispatched draft text.

1. Call score("action-item-draft", run_id, { score: 1, ... }) with

distinct actor + auditor session ids. Include artifact_path so the eval row points back to the file it produced.

Eval

auto. Shape gate: artifact file exists, frontmatter has both action_item_id and meeting_id, body is non-empty. Score 1 when all true; 0 when any fail; 0.5 when the dispatch fell back to a non-Claude model (recoverable but degraded).

Why

This is the missing seam. Before this skill existed, a Krisp action item sat read-only in the cockpit and Robert had to context-switch to compose anything against it. With this skill + the KrispScreen Draft button + the DraftsScreen approval surface, the loop closes inside the cockpit: meeting → action item → draft → approve/edit/dismiss → done.

action-item-draft - loader

Per-turn rules. Full reference: state/skills/action-item-draft/SKILL.md. Do not skip these.

Critical Rules

• SCOPE-ONLY. Drafts land in state/log/action-item-draft/<utc-iso>-<short-id>-draft.md. Nothing publishes without a row in state/log/approvals.ndjson.
• Frontmatter MUST carry BOTH action_item_id AND meeting_id - cockpit links draft → source via these. Missing either ⇒ shape gate fails ⇒ score 0.
• Pick artifact shape (LinkedIn post / email / message) from the action item content. No default shape. Inference is the actor's job, not a fallback.
• Truncate the meeting document to 6KB before passing to the dispatch prompt. Anything bigger blows the context budget.
• Dispatch via state/lib/dispatch.ts with model: "sonnet" - NOT "claude" (no such alias; throws "unknown model alias"). Valid aliases: sonnet/opus/haiku/gemini. NEVER shell out to claude -p directly.
• Earnest voice. No "holler if anything is off", no "shoot me a note", no folksy closers. End with a concrete next step or "Looking forward to our call."
• Krisp's list_action_items silently returns [] when limit > 50. Use limit: 50. If the action item isn't in the first 50 (old/completed), proceed with meeting-only context and score 0.5 - degraded > fail.
• SNAPPY_INTENT parse: split on :: first, THEN split each token on the FIRST : only. Meeting ids contain colons - naive global : split corrupts them.
• Actor ≠ auditor (CONSTITUTION #3). actor_session_id = sessionId(); auditor_session_id = "action-item-draft-aud-<ts>-<hex>". Build independently.
• Inject the latest state/log/taste-hints.ndjson line for (action-item-draft, <chosen-shape>) into the dispatch prompt when one exists - that's the closed taste loop with distill-approvals.
• A/B control arm: when a hint IS available, randomize 50/50 whether to inject it. The control arm (no hint) is what makes the loop measurable - without it we can't tell if hints are doing anything vs being placebo. Track is_control: bool, active_taste_hint: string|null, and inferred_shape: DraftShape in eval row extra so distill can later answer "does the treatment arm beat control by ≥10pp at n≥30?" Honest acknowledgment: this is randomized A/B, not RL.
• Post-dispatch shape inference uses state/lib/draft-shape.ts::inferShapeFromBody. Same lib distill-approvals uses - drift here silently breaks the loop (writer says "linkedin-post", reader buckets "email", hint never matches the shape that produced it). Don't inline a parallel heuristic.

Commands

| ui dashboard | state/skills/action-item-draft/resources/ui.openui |

Mandatory draft frontmatter

---
action_item_id: <aid>
meeting_id: <mid>
meeting_name: <name>
action_item_title: <title>
created: <utc-iso>
---

Artifact shape selection

Recoverable conditions (exit 0, not 1)

Hard fails (artifact missing / frontmatter incomplete / body empty / wrong model alias) ⇒ score 0. Never crash the run - the cockpit retries on the next tick.

Self-Test

Reading this loader, an agent should:

1. [ ] Parse SNAPPY_INTENT="id:<aid>::meeting:<mid>" without splitting meeting-id colons.
2. [ ] Resolve meeting_id from the action item if absent in intent (with limit: 50).
3. [ ] Truncate the meeting document to 6KB before dispatch.
4. [ ] Pick shape from action-item content (no default), inject taste hint if present.
5. [ ] Dispatch with model: "sonnet" (NOT "claude"); fall back to "gemini" and score 0.5 on failure.
6. [ ] Write the draft with action_item_id AND meeting_id in frontmatter.
7. [ ] Call score() with distinct actor + auditor session ids and artifact_path.
8. [ ] Exit 0 with score: 0.5 + primary_issue on recoverable failures (table above).

Self-report

If this loader fell short:

echo "[$(date -u +%FT%TZ)] action-item-draft: <what was missing>" >> state/log/loader-feedback.log

<!-- kernel-ok: historical attribution for the PID-loader pattern. -->

Self-correcting loader (PID feedback)

This is the rule that makes loaders learn. Ported from snappy-kernel, where it demonstrably worked: the snappy-dom-cartographer skill mapped Skool admin UI once, every later ship-agent inherited the full map for free.

Before you finish, do two things:

1. Fix gaps (P - proportional). MANDATORY. If this loader didn't cover your case - if you had to read another file, run an undocumented command, work around a wrong selector, or discover a quirk - you MUST attempt an Edit to this AGENTS.md before you log.

• Default action: edit this file. Surgical. Add the missing line, table

row, or code example where it belongs. No restructuring.

• Only LOGGED is allowed when: the fix needs >10 lines, spans multiple

files, or requires a structural rewrite. In that case the state/regen/drain.sh queue picks it up asynchronously.

• "I didn't have time" / "it's minor" / "the next agent will figure it out"

are NOT valid reasons. Minor gaps compound - that's the whole point of this loop.

• Concrete example of a valid P-fix: discover agent-browser screenshot

takes path positionally not as --output <path>. One line in a table. Edit it in. Do not just log it.

• The goal: the next agent never has to leave the loader.

2. Log the result.

echo "[$(date -u +%FT%TZ)] action-item-draft: <what was missing or fixed> [FIXED|LOGGED]" >> state/log/loader-feedback.log

• FIXED = you patched this loader inline (P-fix).
• LOGGED = too large for inline; the PostToolUse enqueue + Stop-hook drain

will rewrite the loader from scratch on next session-end.

Do not skip this. Every agent run must leave the system better than it found it. The loader is the setpoint; you are the sensor; the gap is the error signal; closing the gap is the correction.

OpenUI Resource

• Skill-owned OpenUI Lang resource: state/skills/action-item-draft/resources/ui.openui. Read it before rendering or editing this skill's generated component surface.
• Treat this resource as a first-class artifact of the skill, not a generic chat response. Improve it when the skill's user-facing output needs to become richer.
• System resources compose OpenUI primitives and inherit SnappyChat tokens. Use ui_contract: branded in SKILL.md only for deliberate platform or client visuals.