razbakov/skills/daily-digest

Daily Digest

Chief-of-Staff orchestrates the daily digest: collects review-ready slots from the top-manager agents, consolidates into one message, sends to the Commander via Telegram, saves to sessions.

Implements ~/Projects/ikigai-team/rules/agent-proactivity.md. Read that for the philosophy (cadence triggers, review-ready threshold, focus signal, digest discipline). This skill is the executable contract.

Trigger

• Cron / launchd fires this skill at the configured cadence (typically weekday 06:30).
• User says "run daily digest", "send digest", or invokes /daily-digest.

Inputs from project context

Read these from the project's CLAUDE.md and agent files:

• Manager roster — agents in .claude/agents/<name>.md. The Chief-of-Staff role is whichever agent has "Daily Digest Protocol" in their file.
• Each manager's Cadence section — answers when they participate (some daily, some weekly, some opted-out).
• Telegram bot token — for sending. Look up via the project's documented sender (commonly .bin/telegram-send.py --agent <chief>).
• Focus signal file — typically ops/focus.md. Read most-recent entry.
• Sessions directory — typically ops/sessions/. Save digest there.

If any of these are missing, halt and tell the user what's needed — do not guess.

Execution

Phase 1 — Read context

1. Read the focus signal file. Most-recent line is current. If open / clear / absent, treat as no-focus.
2. Determine today's day-of-week. Some managers participate only on certain days (e.g., a Strategy role may only submit Mondays, a Personal-Coach role may opt out entirely).
3. List participating managers for this run by reading each manager's Cadence section.
   • "submits to daily digest" + day matches → in
   • "out-of-cycle" / opted out → skip
   • "weekly on <day>" + day matches → in this run only

Phase 2 — Collect slots

For each participating manager:

1. Dispatch a slot-request to that manager's tmux session. Format:

Daily digest slot — <date> (<dow>)
Focus: <current focus or "open">

Per your Cadence section, scan your domain and reply with ONE of:
- A review-ready slot (≤ 5 lines) in the format from agent-protocols.md:
    Why: <KR>
    What: <one or two lines>
    Media: <link · screenshot · doc>
    Asking: <decision>. Default: <action> by <time> unless objection.
- "nothing to surface" if no decision/outcome/drift to report

Reply by <deadline, typically T-15 min before send>.

2. Track replies. After deadline, any silent manager → record "nothing to surface (no reply)".

Phase 3 — Assemble + send

Suppression gate — check this FIRST. If every collected slot is "nothing to surface" (including "no reply") — i.e. no manager reported a pending decision, a delivered outcome, or a drift — then there is nothing review-ready. Do NOT send anything to Telegram. Skip straight to Phase 4, record the silent cycle, and exit. A digest that is only "nothing to surface" lines is exactly the noise this protocol exists to prevent (agent-proactivity.md: "a daily silent cycle is a healthy cycle"). Only proceed to formatting and sending when at least one slot carries real review-ready content.

Format the digest:

Daily digest — <date> · focus: <current focus or "open">

<Chief role> (<role-tag>): <slot>
<Manager 1> (<role-tag>): <slot>
<Manager 2> (<role-tag>): <slot>
...

• Each slot is what the manager replied with — do not paraphrase, do not edit beyond trimming whitespace.
• The Chief writes their own slot first, before assembling.

Send via the project's documented sender (e.g., .bin/telegram-send.py --agent <chief> --file <digest-file>).

Phase 4 — Save

Save to <sessions-dir>/<date>-daily-digest.md. Include:

• The sent message verbatim — or, if the Phase 3 suppression gate fired, a one-line "silent cycle — nothing to surface, no message sent" note instead
• A list of who participated and who was silent
• The current focus signal at send time

The cycle is always logged, even when nothing is sent — silence is recorded, not invisible.

Phase 5 — Track replies (async)

After send, watch for Commander replies. Parse per agent-proactivity.md:

• ok / ship it / yes → consent now, route to originating manager
• no / change X / do Y instead → revise request, route to originating manager
• pause / hold → mark the open ask paused, log to sessions
• focus: <theme> → append to ops/focus.md with timestamp

If the Commander replies to a specific slot only, route that decision to that manager without disturbing other open asks (their time-boxes age normally).

Out-of-cycle bypass

This skill handles the routine daily flow only. Out-of-cycle event triggers (urgent items: prod incident, festival contact mid-day, partnership emergency) bypass the digest — the originating manager sends directly. Do NOT include those in the digest after the fact; they had their own thread.

Failure modes

• Tmux session for a manager not running — log and treat as "nothing to surface (no reply)". Do not block the digest.
• Telegram send fails — retry once. If still failing, save the digest to sessions and log the failure; do not silently drop.
• No managers participating today (e.g. weekend) — skip the run. Log "no participants today, skipping" to sessions and exit.

Why a skill, not prose

Putting the protocol in a skill (instead of inline prose in the Chief's agent file) makes the flow:

• Deterministic — the same steps fire every day, same shape every day
• Testable — manual /daily-digest invocation runs the same path as cron
• Portable — any project that imports the framework can adopt the skill
• Revisable — protocol updates land in this skill, not in N agent files