Backed by: state/lib/krisp.ts, state/lib/email.ts, state/lib/drive.ts, state/lib/knowledge.ts, state/lib/dispatch.ts, state/bin/email-send-tick.ts

meeting-followup

Fetches recent meetings from Krisp (headless HTTP, no MCP needed), filters to 1:1 and small-group calls (skips masterminds/community calls), composes a follow-up email per meeting, creates a Gmail draft, and - when the eval gate passes at score=1.0 - schedules the draft for auto-send after a 30-minute veto window so Robert can delete or edit it before it ships.

This is the trust feature: the system either sends the right email or doesn't act. Both must be safe. Partial-pass drafts (eval < 1.0) stay in the inbox for manual review and are never scheduled.

Steps

0. Pre-flight -- verify Krisp OAuth token

Before fetching anything, check that a valid Krisp token exists:

import { env } from "../lib/env.ts";
import * as fs from "fs";

const credsPath = `${process.env.HOME}/.claude/.credentials.json`;
let krispTokenPresent = false;
try {
  const creds = JSON.parse(fs.readFileSync(credsPath, "utf8"));
  krispTokenPresent = Object.keys(creds).some((k) => k.startsWith("mcpOAuth[\"krisp|"));
} catch (_) {}

if (!krispTokenPresent) {
  // Graceful fallback: log a warning and exit without hard-failing.
  console.warn("Krisp OAuth token missing from .credentials.json. Run `/mcp krisp` in Claude Code to re-authenticate (~30s browser flow). Skipping meeting-followup.");
  append("chain", { run_id, skill: "meeting-followup", action: "skipped", reason: "no-krisp-token" });
  score("meeting-followup", run_id, { score: 0.0, primary_issue: "no-krisp-token", drafts_created: 0, drafts_skipped: 0 });
  process.exit(0);
}

Token is stored under the key pattern mcpOAuth["krisp|<id>"] in ~/.claude/.credentials.json. It expires approximately every 24 hours. The Claude Code MCP client purges the key on a 401 - headless krisp.ts cannot recover; only the browser OAuth flow restores it.

Heal: run /mcp krisp in any Claude Code session (~30s browser OAuth). Until healed, this skill exits gracefully rather than throwing.

1. Scope -- fetch and filter meetings

import { fetchMeetings, fetchMeetingDocument, type KrispMeeting } from "../lib/krisp.ts";

• Fetch meetings from last hours_back hours (default 48).
• If fetchMeetings throws HTTP 502, treat it as a transient Krisp outage

(Cloudflare bad-gateway - the host mcp.krisp.ai is temporarily unreachable). Log the error, emit score: 0.0, primary_issue: "krisp-502-transient", and exit gracefully. Do NOT mark this as a recipe failure; retry on next scheduled run.

• Skip meetings matching group/community patterns: mastermind, SCC, State Change,

Agentic Building, Build with AI, workshop, webinar, course, training, Claude Meeting (solo Krisp captures), Google Chrome Meeting.

• Extract recipient name from attendees/speakers (excluding Robert). Fall back to

parsing meeting title ("X <> Y", "X and Y", "Name - Session Type").

• Only process meetings with 1-3 other people (not solo or large groups).

2. Gate

• At least one eligible meeting found (otherwise skip with reason).
• Each eligible meeting must have a parseable recipient name.

3. Act -- compose and draft

For each eligible meeting:

1. Fetch full meeting document via fetchMeetingDocument(meeting_id).
2. Extract key points and action items from the document markdown.
3. If no key points found, extract raw transcript and dispatch to gemini

for summarization (via state/lib/dispatch.ts).

1. Look up recipient email via resolvePerson({ name }) from state/lib/knowledge.ts.
2. Search Google Drive for recording + transcript links via state/lib/drive.ts.
3. Compose HTML email with: greeting, key points list, next steps list,

Drive links (recording/transcript), sign-off.

1. Create Gmail draft via createEmailDraft() from state/lib/email.ts.

If a draft already exists for this meeting (tracked in sent-records), check if it still exists with getEmailDraft() and update with updateEmailDraft() instead of creating a duplicate.

3a. Schedule auto-send (only when eval = 1.0)

After all drafts are created/updated AND the shape gate (Step 4 below) has passed at score=1.0 - every required AND non-required criterion green - schedule each successful draft for future send via scheduleSendDraft() from state/lib/email.ts:

import { scheduleSendDraft } from "../lib/email.ts";

// Default 30-minute veto window. Override per-run via SNAPPY_EMAIL_VETO_MINUTES.
const vetoMinutes = Number(process.env.SNAPPY_EMAIL_VETO_MINUTES ?? "30");
const sendAt = new Date(Date.now() + vetoMinutes * 60 * 1000);

await scheduleSendDraft({
  draftId,
  sendAt,
  account: "work",
  recipient: recipientEmail,
  subject,
  meta: { source: "meeting-followup", meeting_id: meeting.meeting_id },
});

Hard rules - do not violate:

• Only schedule send when eval = 1.0, never on partial success. If any

check failed (speaker labels leaked, missing key points, no recipient, malformed HTML, etc.), leave the draft for Robert to review manually.

• Don't double-schedule. scheduleSendDraft() is idempotent on

(draftId, status="scheduled") - calling it twice for the same draft while a row is still scheduled returns the existing row without appending a duplicate. Reschedules require cancelScheduledSend() first.

• Veto path is Robert deleting the draft in Gmail. The consumer at

state/bin/email-send-tick.ts calls sendEmailDraft(); if Gmail returns 404 (draft gone), the row is marked status="vetoed" and no error escalates. Robert never has to remember to cancel - deleting is the cancel.

• Kill switch: SNAPPY_EMAIL_AUTO_SEND_DISABLED=1 halts ALL scheduled

sends globally (consumer-side). Use this if you ever need to pause the pipeline without ripping out the agent.

4. Log + eval

import { append } from "../lib/log.ts";
import { score } from "../lib/eval.ts";
append("chain", { run_id, skill: "meeting-followup", action: "drafted", count, skipped });
score("meeting-followup", run_id, {
  score: has_key_points && no_speaker_labels ? 1.0 : 0.5,
  primary_issue: !has_key_points ? "empty-key-points" : "speaker-label-leak",
  drafts_created: count,
  drafts_skipped: skipped,
});

Eval

Actor: Krisp HTTP client (fetches meeting data) + dispatch model (summarizes transcripts) + Gmail API (creates drafts). Auditor: Shape gate on the composed email: key points present, no "Speaker 2" labels leaked, recipient resolved, subject non-empty, HTML well-formed, Drive links are clickable <a> tags.

Score convention:

Eval criteria (from kernel recipe):

Gotchas

• Missing Krisp OAuth token (P0). mcp.krisp.ai requires a browser OAuth token

stored under mcpOAuth["krisp|<id>"] in ~/.claude/.credentials.json. The token expires ~24h and is silently purged by the MCP client on a 401. Symptom: recipe rejected: No Krisp OAuth token in .credentials.json. Fix: run /mcp krisp in Claude Code (~30s). The skill detects this in Step 0 and exits gracefully rather than throwing.

• Krisp HTTP 502 (transient). Krisp's Cloudflare-fronted API occasionally returns

502 Bad Gateway when the backend mcp.krisp.ai host is temporarily unreachable. This is not a skill bug - it resolves on its own within minutes. Log it as primary_issue: krisp-502-transient and rely on the next scheduled run.

• Speaker labels in transcripts. Krisp uses "Speaker 2", "Speaker 3" for

non-primary participants. The summarization step replaces these with the recipient's actual name before dispatching, but verify in the eval.

• Transcript truncation. Long transcripts are truncated to ~8000 chars

before dispatch to keep summarization fast and cheap.

• Drive time-window matching. Drive files are matched to meetings by

embedded timestamp (+/-30 min window). Falls back to name matching if time matching finds nothing.

• Duplicate draft prevention. Tracks sent records in ndjson. If a draft

already exists for a meeting_id, checks if it still exists (not already sent/deleted) before updating. Sent drafts are skipped, not re-created.

• Group call filtering. Patterns are conservative -- some 1:1s with

unusual names may slip through. The recipient-count gate (1-3 others) is the second defense.

• Veto window timing. Default 30 min. Robert deleting the draft in

Gmail before the window closes cancels the send (404-on-send → status vetoed, not error). The consumer (LaunchAgent every 60s) only acts when sendAt <= now, so a draft scheduled at T+30min is safe to delete any time before that.

• Audit trail. Every scheduled send → row in

state/log/email-send-queue.ndjson (status: scheduled / sent / vetoed / failed / cancelled). Every actual send → row in state/log/email-sent.ndjson with the Gmail message id. Transient send errors → row in state/log/email-send-errors.ndjson with retry_count; the consumer retries up to 3 times before marking the queue row failed.

• Kill switch. SNAPPY_EMAIL_AUTO_SEND_DISABLED=1 in the consumer's

environment halts ALL scheduled sends globally; queue rows stay scheduled and resume the moment the env var is unset.

Graduation

Prose until the gather/compose/deliver pipeline has been executed twice without edits. Then graduate by extracting the runner into a script under the existing state/lib/krisp.ts + state/lib/email.ts surfaces.

Rubric

criteria:
  - name: krisp_token_present
    kind: deterministic
    check: "The log output or `~/.claude/.credentials.json` confirms the presence of a valid Krisp OAuth token or a graceful skip due to its absence."
  - name: eligible_meetings_processed
    kind: deterministic
    check: "The skill log indicates that eligible meetings were identified and attempted to be processed, or a valid reason for skipping was logged."
  - name: drafts_created_or_updated
    kind: deterministic
    check: "The `state/lib/log.ts` output for the skill run contains `action: \"drafted\"` with `drafts_created` > 0 or `updateEmailDraft()` was called."
  - name: no_speaker_labels_leaked
    kind: judge
    check: "No Gmail draft created by the skill contains 'Speaker 1', 'Speaker 2', etc., identifiable as Krisp speaker labels."
  - name: key_points_present_and_relevant
    kind: judge
    check: "Each Gmail draft contains a bulleted list of key points that accurately summarize the meeting's content."

meeting-followup - loader

Per-turn rules for the meeting-followup skill. Full reference: state/skills/meeting-followup/SKILL.md. Do not skip these.

Backed by: state/lib/krisp.ts, state/lib/email.ts, state/lib/drive.ts, state/lib/knowledge.ts, state/lib/dispatch.ts, state/bin/email-send-tick.ts, state/bin/meeting-followup/run.ts, state/bin/meeting-followup-run.ts (cron variant).

Critical Rules

1. AUTO-SEND ONLY ON EVAL = 1.0. Schedule via scheduleSendDraft() only when every shape-gate criterion (required AND non-required) is green. Partial-pass drafts (eval < 1.0) stay in the inbox for manual review, never scheduled.
2. DEFAULT 30-MINUTE VETO WINDOW. sendAt = now + (SNAPPY_EMAIL_VETO_MINUTES ?? 30) * 60_000. Robert deleting the draft in Gmail before the window closes cancels the send - the consumer marks it vetoed (Gmail 404), no error escalates.
3. KILL SWITCH: SNAPPY_EMAIL_AUTO_SEND_DISABLED=1 halts ALL scheduled sends globally. Set on consumer (state/bin/email-send-tick.ts) to pause without unwiring the agent.
4. AUDIT TRAIL is non-optional. Every scheduled send → state/log/email-send-queue.ndjson (status: scheduled / sent / vetoed / failed / cancelled). Every actual send → state/log/email-sent.ndjson with the Gmail sent_message_id.
5. ALWAYS replace Krisp Speaker 2 / Speaker 3 labels with the recipient's actual name BEFORE the dispatch summarization step - leaked speaker labels fail the no_speaker_2 shape gate (which would block auto-send).
6. ALWAYS skip group/community patterns: mastermind, SCC, State Change, Agentic Building, Build with AI, workshop, webinar, course, training, Claude Meeting (solo Krisp captures), Google Chrome Meeting.
7. Only process meetings with 1-3 other people - solo or large group → skip.
8. Idempotent draft handling: check sent-records for an existing draft per meeting_id; if getEmailDraft() confirms it still exists, updateEmailDraft() rather than creating a duplicate. scheduleSendDraft() is idempotent on (draftId, status="scheduled") - calling twice returns the existing row without duplicating.
9. ALWAYS resolve recipient via resolvePerson({ name }) from state/lib/knowledge.ts before composing.

Commands

| ui dashboard | state/skills/meeting-followup/resources/ui.openui | |invoke (cron): npx tsx state/bin/meeting-followup-run.ts |invoke (manual): npx tsx state/bin/meeting-followup/run.ts |fetch meetings: import { fetchMeetings, fetchMeetingDocument } from "../lib/krisp.ts" |compose+draft: import { createEmailDraft, getEmailDraft, updateEmailDraft } from "../lib/email.ts" |schedule send: import { scheduleSendDraft, cancelScheduledSend, sendEmailDraft } from "../lib/email.ts" |drive links: import * as drive from "../lib/drive.ts" |summarize transcript: state/lib/dispatch.ts → gemini |consumer (cron): npx tsx state/bin/email-send-tick.ts # LaunchAgent every 60s |cancel a queued send: npx tsx state/lib/email.ts cancel-send <draftId> |send a draft NOW: npx tsx state/lib/email.ts send-draft <draftId> [account] |eval log: state/log/evals.ndjson (skill: "meeting-followup") |queue: state/log/email-send-queue.ndjson (status: scheduled/sent/vetoed/failed/cancelled) |sent log: state/log/email-sent.ndjson (one row per actual send with sent_message_id)

Pre-flight (Step 0)

Before fetching anything, verify Krisp OAuth state:

• Check ~/.claude/.credentials.json for mcpOAuth["krisp|<id>"]. The token expires ~24h and Claude Code's MCP client purges the entry on a 401.
• Token missing/expired → log a warn line, write an eval row with score: 0.0 and primary_issue: "no-krisp-token", exit gracefully. Heal: run /mcp krisp in any Claude Code session (~30s browser flow). Headless krisp.ts cannot recover on its own.
• Krisp HTTP 502 → Cloudflare bad-gateway transient. Score 0.0 with primary_issue: "krisp-502-transient", retry on next cron tick. Do not escalate to friction.

Install (LaunchAgent - once per machine)

cp state/launchd/com.snappy.email-send-tick.plist ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.snappy.email-send-tick.plist
# Verify it's running every 60s:
launchctl list | grep email-send-tick
tail -f state/log/email-send-tick.launchd.log

OpenUI Resource

• Skill-owned OpenUI Lang resource: state/skills/meeting-followup/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.

Known Pitfalls

• Long transcripts truncate to ~8000 chars before dispatch - keep summarization fast/cheap.
• Drive matches by embedded timestamp (±30 min); falls back to name match if no time hit.
• Recipient parsing falls back to title patterns "X <> Y", "X and Y", "Name - Session Type" if attendees lack a non-Robert name.
• Group-call filters are conservative - the 1-3-others recipient-count gate is the second defense.
• Total <li> per email caps at 17 (7 key points + 10 action items); use clickable <a href> tags for Drive links, not raw URLs.
• state/bin/meeting-followup/run.ts calls scheduleSendDraft ONLY when finalScore === 1.0 AND created+updated > 0. Eval < 1.0 → no schedule, draft stays in inbox.
• scheduleSendDraft() is idempotent on (draftId, status="scheduled"). To reschedule, call cancelScheduledSend(draftId) first, then schedule again.

Self-Test

An agent reading this should correctly:

1. [ ] Schedule auto-send only when eval === 1.0 (never on partial pass)?
2. [ ] Use a 30-minute veto window by default (override via SNAPPY_EMAIL_VETO_MINUTES)?
3. [ ] Replace Speaker 2 / Speaker 3 with the resolved recipient name before summarization?
4. [ ] Skip a meeting matching mastermind or Build with AI in the title?
5. [ ] Treat a Gmail 404 from the consumer as a successful veto (status vetoed), not an error?

Self-report

If this loader fell short, append a line:

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

Known Pitfall (manual append 2026-04-19, area=krisp-token-expired)

Claude Code's MCP client purges mcpOAuth["krisp|<id>"] from ~/.claude/.credentials.json on a 401, ~24h cycle. Headless krisp.ts cannot recover - only browser OAuth restores it. Symptom: Krisp HTTP 401 in recipe stderr, friction area=krisp-token-expired in breakage-report. Heal: run /mcp krisp in any Claude Code session (~30s browser flow). Structural fix needed: detached cron-side token store OR pre-expiry warning at expiresAt - 6h.

<!-- 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)] meeting-followup: <what was missing or fixed> [FIXED|LOGGED] action_kind=<kind>" >> state/log/loader-feedback.log

• <slug> MUST be the literal folder name of this loader

(state/skills/<slug>/AGENTS.md). The class token between [ts] and : is the producer slug, the writeback class, AND the grade class - they must be equal so state/lib/controller-tune.ts can pair the brief.

• 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.

• action_kind is the SECOND pairing predicate (added 2026-04-27, task #327).

Pick the value that describes what you actually did - same slug, different action_kind means the writeback satisfies a different brief layer:

• shape-ok - only frontmatter-shape verification passed (rare from

a human; usually emitted by the lint, not a loader echo)

• skill-ran - the skill ran end-to-end and an eval row landed

in state/log/evals.ndjson

• loader-rewritten - you EDITED this AGENTS.md inline (the FIXED case),

OR the regen drain rewrote it

• pattern-elevated - you promoted a recurring failure to a Critical Rule

(rule fix or new-skill scaffold) If you LOGGED (couldn't fix inline), omit action_kind - the inferrer will pick it up from your body keywords.

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.