polym-slack-digest

name	polym-slack-digest
description	Generate Slack daily digests, highlight customer risks and action items, and publish summaries to Lark or local outputs.
allowed-tools	Bash, Read, Write, Edit
version	1.2.0

Slack Daily Digest Skill

End-to-end Slack-to-Feishu daily intelligence pipeline. The skill handles the full lifecycle on Mira, Codex, and Claude Code:

1. Onboarding — guide the user to create a Slack App, install it to their workspace, and configure both Slack and Lark Bot credentials.
2. Index Setup — create a single Feishu doc that lists every daily digest, newest at top.
3. Schedule — register a recurring task at the user-specified time (default: 11:00 and 16:00 in the configured timezone).
4. Daily Run — pull, synthesize, write per-day doc, pin to index.
5. Watchdog — 20 minutes after run, verify the doc exists and contains expected sections; if not, diagnose and retry until success.
6. Notify — push Highlights + TODO summary to user via Lark IM bot.

---

Platform Matrix

Platform	Lark docs	Scheduling	Durability	Retry/catch-up
Mira	Built-in Lark doc skills	Mira scheduled task	Durable	Mira watchdog + retry
Codex	Local Lark CLI	Durable Codex routines (MCP)	Durable UI	Durable retry routine (+15 min)
Claude Code	Local Lark CLI	Durable scheduled tasks (MCP)	Durable UI	Durable retry routine (+15 min)

For Codex and Claude Code, all scheduled routines are durable host UI routines — they persist across session restarts and are visible/manageable in the host's routine/automation UI. Local OS schedulers (launchd, cron, systemd) are not used.

For one-off E2E runs, the current host agent handles synthesis via the HOST_AGENT_HANDOFF file. Do not look for a codex or claude executable for this path.

0. Conventions

• Working directory: All persistent files live under ~/.slack-daily-digest/. The skill creates this on first run.

• Credential file: ~/.slack-daily-digest/credentials.json (mode 0600). Never echo credentials back to the user.

• State files:

  • ~/.slack-daily-digest/users_cache.json — Slack user ID → name/title/email
  • ~/.slack-daily-digest/channel_owner_cache.json — channel → BytePlus owner
  • ~/.slack-daily-digest/index.json — index doc URL + folder token + schedule
  • ~/.slack-daily-digest/routine.json — local Codex / Claude Code routine configuration
  • ~/.slack-daily-digest/routine_state.json — local routine slot attempts and catch-up state
  • ~/.slack-daily-digest/runs/<YYYY-MM-DD>/ — per-day pull/synth artifacts

• Output language: All user-facing prompts, doc titles, and digest content default to English. The skill may switch to the user's working language only if the user explicitly asks.

---

1. Routing — When to Trigger Each Phase

Read ~/.slack-daily-digest/credentials.json:

• Missing or incomplete → run Phase A: Onboarding (§2)
• Complete but no index.json → run Phase B: Index Setup (§3)
• Codex / Claude Code local mode and no local_platform.json → run scripts/setup_lark_cli.py, then continue setup when ready
• All present, user asks "run digest now" → run Phase D: Daily Run (§5)
• All present, user asks "verify today" / triggered 20min after run → run Phase E: Watchdog (§6)
• User asks "change schedule" / "reconfigure" → re-run the relevant phase

Platform detection is centralized in scripts/platform.py. Honor SLACK_DIGEST_PLATFORM=mira|local_agent when present; otherwise Mira-specific runtime hints select Mira, and all other environments use local-agent mode.

---

2. Phase A — Onboarding (Slack App + Lark Bot)

Show the user two parallel setup tracks in one message. Both must be completed before proceeding.

2.1 Slack App Setup

Display this verbatim (English):

Step 1 — Create your Slack App

1. Go to https://api.slack.com/apps and click Create New App → From scratch.
2. Name it Mira Slack Digest and pick your workspace.
3. In OAuth & Permissions → User Token Scopes, add:
   • channels:history, channels:read
   • groups:history, groups:read
   • im:history, im:read
   • mpim:history, mpim:read
   • users:read, users:read.email
4. Click Install to Workspace and authorize.
5. Copy the User OAuth Token (starts with xoxp-...) and paste it back here.

⚠️ Use the User Token, not the Bot Token — the digest reads channels you joined as a user.

After the user pastes the token, validate it:

curl -s -H "Authorization: Bearer $SLACK_USER_TOKEN" \
  "https://slack.com/api/auth.test" | jq '.ok, .user, .team'

If ok=false, show the error and ask them to retry.

2.2 Lark Bot Setup

Step 2 — Create your Lark Custom Bot

1. Go to https://open.feishu.cn/app and click Create Custom App.
2. Name it Mira Slack Digest Notifier.
3. In Permissions & Scopes, add:
   • im:message (send messages)
   • im:message:send_as_bot
   • contact:user.base:readonly (resolve email → open_id)
   • docx:document (read/write docs)
   • drive:drive (move docs to folder)
4. In Version Management & Release, publish v1.0 and wait for admin approval (or self-approve in trial workspaces).
5. In Credentials & Basic Info, copy App ID and App Secret and paste both here.

Validate by acquiring a tenant access token:

curl -s -X POST "https://open.feishu.cn/open-apis/auth/v3/tenant_access_token/internal" \
  -H "Content-Type: application/json" \
  -d "{\"app_id\":\"$LARK_APP_ID\",\"app_secret\":\"$LARK_APP_SECRET\"}" | jq '.code, .tenant_access_token'

code=0 means success.

2.3 Persist Credentials

Use scripts/save_credentials.py:

python3 ~/.claude/skills/slack-daily-digest/scripts/save_credentials.py \
  --slack-token "xoxp-..." \
  --lark-app-id "cli_..." \
  --lark-app-secret "..." \
  --lark-user-email "<the user's lark email>"

This writes ~/.slack-daily-digest/credentials.json with mode 0600.

Then proceed to Phase B.

---

3. Phase B — Index Setup & Schedule

Use scripts/platform.py to select the setup path.

3.0 Local-Mode Notice

When the platform is Codex / Claude Code local mode, tell the user:

I will create durable host UI routines — one per digest slot plus one retry routine (+15 min) per slot. Routines persist across session restarts and are visible in the Codex / Claude Code routine UI. An optional end-of-day catch-up is also available. All routines run local_routine.py run-due, which deduplicates via routine_state.json so the same slot is never published twice.

No local OS schedulers (launchd, cron, systemd) will be used.

Ask in one prompt — do not ask for times/timezone without also collecting working hours, retry, and catch-up:

Schedule configuration Please provide:

• Run times — e.g. 10:00, 14:00 (comma-separated HH:MM)
• Timezone — e.g. Europe/London, Asia/Shanghai, Etc/GMT-1 (default: Asia/Shanghai)
• Working hours — e.g. 08:00-18:00 (default: 08:00-19:00)
• Retry routines? — fires 15 min after each slot if main run missed/failed (default: yes)
• End-of-day catch-up? — fires ~30 min after last slot to retry missed runs (default: yes)

3.1 Create the Folder

Ask the user (English): "Should I create a new Feishu folder named Slack Daily Digest to store all daily docs, or do you have an existing folder? If existing, paste the folder URL."

If Mira and new → use lark-drive-wiki-skill create-folder --name "Slack Daily Digest" --folder-token <root> and capture the token.

If Codex / Claude Code and new → use Lark CLI through scripts/lark_doc_client.py / lark-cli drive +create-folder; do not require a custom Lark app for document writing when Lark CLI user auth is available.

3.2 Create the Index Doc

Use scripts/lark_doc_client.py so Mira keeps lark-docs-skill behavior and local mode uses lark-cli.

Title: Slack Daily Digest · Index
Body:
## This document is auto-maintained by Mira. Daily digests are listed below in reverse chronological order (newest on top). Monitoring scope: all Slack channels the configured user has joined; the channel list refreshes on every run.
## Daily Digest List (newest on top)

Move it into the folder through the same document client.

3.3 Ask for Schedule

Use the run times, timezone, working hours, retry, and catch-up preferences collected in §3.0.

Persist to ~/.slack-daily-digest/index.json:

{
  "index_doc_url": "https://www.feishu.cn/docx/...",
  "index_doc_id": "...",
  "folder_token": "...",
  "timezone": "Etc/UTC",
  "schedule": [{"hour": 11, "minute": 0, "timezone": "Etc/UTC", "tz": "+00:00", "window_hours": 19},
               {"hour": 16, "minute": 0, "timezone": "Etc/UTC", "tz": "+00:00", "window_hours": 5}]
}

For local mode, also persist ~/.slack-daily-digest/routine.json:

{
  "timezone": "Etc/UTC",
  "slots": ["11:00", "16:00"],
  "working_hours": {"start": "09:00", "end": "18:00"},
  "catchup_hours": 6,
  "check_interval_minutes": 10,
  "agent": {"type": "auto", "command_template": null}
}

3.4 Register Recurring Task

For Mira, use the loop skill (or platform's scheduled-task mechanism) to register:

• Run slack-daily-digest run at each scheduled time.
• Run slack-daily-digest watchdog 20 minutes after each scheduled time.

For Codex / Claude Code — durable host UI routines only. No local OS scheduler. No session-only routines.

Error codes from setup-scheduler

Exit code	Meaning	Action
host_unknown	Not in a Codex or Claude Code session	Stop. Do not create any scheduler.
session_only_routine_not_allowed	Host detected but does not support durable routines	Stop. Report to user.
host_durable_ui_routine_unavailable	mcp__scheduled-tasks__create_scheduled_task unavailable	Stop. Do not fall back to CronCreate or local OS scheduler.
stable_path_required	SLACK_DIGEST_HOME is under /tmp or a test path	Ask user to confirm stable path (default ~/.slack-daily-digest) before continuing.
scheduler_setup_blocked	Missing credentials, index, or runner config	Fix the listed items, then re-run.

Step 1 — Validate and get setup request

python3 scripts/local_routine.py setup-scheduler \
  --slots "<times>" --timezone "<timezone>" --working-hours "<window>"

Prints a scheduler_setup_request JSON on success. Key fields:

• host_scheduler — "codex" or "claude_code"
• scheduler_durability — always "durable_ui"
• host_instructions — host-specific creation notes (read host_instructions.create_method)
• host_routines — one descriptor per slot (main)
• retry_routines — one descriptor per slot at slot + 15 min
• optional_catchup_routine — end-of-day catch-up descriptor
• local_scheduler_used — always false
• session_only_routine_used — always false

If setup-scheduler exits non-zero, stop and show the error verbatim to the user.

Step 2 — Create durable host UI routines

Read host_instructions.host from the JSON.

---

Step 2A — Codex (host_instructions.host == "codex")

Use mcp__scheduled-tasks__create_scheduled_task. All routines are durable and visible in the Codex automation UI.

Main routines — for each entry in host_routines[]:

mcp__scheduled-tasks__create_scheduled_task(
  name        = entry.name,
  description = entry.description,
  command     = entry.command,    # local_routine.py run-due --slot HH:MM --timezone TZ
  cron        = entry.cron,       # e.g. "0 10 * * *"
  env         = entry.env,        # SLACK_DIGEST_HOME + SLACK_DIGEST_PLATFORM
)

Retry routines — for each entry in retry_routines[]:

mcp__scheduled-tasks__create_scheduled_task(
  name        = entry.name,       # e.g. "slack-digest-1000-retry"
  description = entry.description,
  command     = entry.command,
  cron        = entry.cron,       # e.g. "15 10 * * *"
  env         = entry.env,
)

Example for slots ["10:00", "16:00"] Europe/London:

slack-digest-1000        cron="0 10 * * *"   → main 10:00
slack-digest-1000-retry  cron="15 10 * * *"  → retry 10:00
slack-digest-1600        cron="0 16 * * *"   → main 16:00
slack-digest-1600-retry  cron="15 16 * * *"  → retry 16:00

Optional catch-up (ask user — see below). After creation record IDs (Step 3).

---

Step 2B — Claude Code (host_instructions.host == "claude_code")

Use mcp__scheduled-tasks__create_scheduled_task. Scheduled tasks are durable, persist across session restarts, and are visible in the Claude Code Scheduled/Routines UI.

Do NOT use CronCreate. CronCreate jobs are session-only and auto-expire after 7 days — they are not allowed for production scheduled runs. If mcp__scheduled-tasks__create_scheduled_task is unavailable, stop with host_durable_ui_routine_unavailable — do not fall back to CronCreate, session-only jobs, launchd, or any local OS scheduler.

Stable path check: Before creating any scheduled task, confirm SLACK_DIGEST_HOME is not under /tmp or a test directory. If it is, stop and ask the user to confirm a stable path (default ~/.slack-daily-digest). Do not create durable tasks that point to /tmp.

Main routines — for each entry in host_routines[]:

mcp__scheduled-tasks__create_scheduled_task(
  name        = entry.name,        # e.g. "slack-digest-1000"
  description = entry.description,
  command     = entry.command,     # local_routine.py run-due --slot HH:MM --timezone TZ
  cron        = entry.cron,        # e.g. "0 10 * * *"
  env         = entry.env,         # SLACK_DIGEST_HOME + SLACK_DIGEST_PLATFORM
)

Retry routines — for each entry in retry_routines[]:

mcp__scheduled-tasks__create_scheduled_task(
  name        = entry.name,        # e.g. "slack-digest-1000-retry"
  description = entry.description,
  command     = entry.command,     # same run-due command; skips silently if slot succeeded
  cron        = entry.cron,        # e.g. "15 10 * * *"
  env         = entry.env,
)

Confirm to the user that tasks are durable and visible in the Claude Code Scheduled/Routines UI.

⚠️ One-time permission approval required (Claude Code)

Claude Code scheduled tasks run with permissionMode: default. The first time a task fires, it may pause and prompt for Bash/Write tool approval before executing local_routine.py run-due.

Required action after creating routines:

1. Open the Scheduled section in the Claude Code sidebar.
2. Find the first digest task (e.g. slack-digest-HHmm).
3. Click Run now.
4. When prompted, approve Bash and Write tools.
5. Tool approvals are stored on the task and auto-applied to all future scheduled runs.

Do not call scheduled setup fully verified until a task has actually executed local_routine.py run-due and a Lark digest doc URL appears in routine_state.json.

Alternative (pre-approve via settings): Add the following to the project .claude/settings.local.json inside the permissions.allow array so that scheduled task sessions (which run from cwd ~/.claude) auto-approve without prompting:

"Bash(SLACK_DIGEST_HOME=<STATE_DIR> SLACK_DIGEST_PLATFORM=local_agent python3 <SKILL_DIR>/scripts/local_routine.py run-due *)",
"Bash(<SKILL_DIR>/scripts/local_routine.py *)",
"Write(<STATE_DIR>/**)"

Replace <STATE_DIR> with SLACK_DIGEST_HOME value and <SKILL_DIR> with the skill path.

Jitter and timing: Scheduled tasks apply a small deterministic delay (up to ~10 min) at dispatch time. The retry routine (+15 min) and optional catch-up routine handle any slots that fire late or miss their window. Do not treat a delayed first fire as a failure.

Optional catch-up (ask user — see below). After creation record IDs (Step 3).

---

Catch-up routine (all hosts — create if user opted in during §3.0 schedule configuration)

Create a durable routine from optional_catchup_routine using the same method as above (Step 2A or 2B). Do not ask the user again — catch-up preference was already collected in §3.0.

---

Step 3 — Record routine IDs and report

python3 scripts/local_routine.py set-backend --backend host_routine \
  --durability durable_ui \
  --routine-ids "<main-id-1>,<main-id-2>" \
  --retry-ids "<retry-id-1>,<retry-id-2>" \
  [--catchup-ids "<catchup-id>"]

Do not call set-backend unless durable routine creation succeeded and IDs were returned.

Validation before reporting success

Report scheduler setup success only after all of the following are true:

• mcp__scheduled-tasks__create_scheduled_task (or Codex equivalent) returned task IDs
• set-backend completed without error
• routine.json records scheduler_durability: "durable_ui" and the routine IDs
• No local scheduler (launchd / cron / systemd) was created
• No CronCreate job was created (not even session-only)
• No session-only routine was created
• At least one task has actually executed local_routine.py run-due and produced a Lark doc URL in routine_state.json (confirmed via "Run now" or a scheduled fire)

If a routine was created but has never executed end-to-end, report setup as pending verification, not success. Instruct the user to find the task in the Scheduled/Routines UI, click "Run now", and approve Bash/Write tools if prompted.

Final report (keep short)

host: codex | claude_code
scheduler backend: durable host UI routine
main routine IDs: [...]
retry routine IDs: [...]
catch-up routine ID: <id> | none
timezone: <TZ>
working hours: HH:MM-HH:MM
SLACK_DIGEST_HOME: <path>
local scheduler used: no
CronCreate used: no
session-only routine used: no

Shared routine logic

All routines invoke local_routine.py run-due --slot HH:MM --timezone TZ, which:

• Loads routine_state.json — skips immediately if this date+slot already succeeded
• Acquires routine.lock — prevents overlapping runs
• Runs the full pipeline: pull Slack → build digest JSON → agent synthesizes final_*.md → publish to Lark
• Marks success only after a Lark doc URL is confirmed in result_*.json
• Marks transient_error on network/API failures — retried by the +15 min retry routine
• Marks credential_error on auth failure — stops retrying, notifies user
• Marks skipped_missed_window when the catch-up window expires

---

4. Phase C — User-Cache Refresh (weekly)

Run scripts/refresh_users_cache.py:

• If users_cache.json mtime > 7 days → call users.list paginated (Tier 2, sleep 1s between pages) and rewrite cache.
• Otherwise skip.

This step is automatic at the start of every Daily Run.

---

5. Phase D — Daily Run

Steps 1–8 below mirror the proven pipeline. All scripts live in scripts/.

Step	Action	Script
1	List all conversations the user joined	scripts/list_channels.py
2	Compute time window from schedule	scripts/run_digest.py (inline)
3	Pull conversations.history per channel (sleep 0.3s, Tier 3)	scripts/pull_history.py
4	Resolve user IDs → names; replace <@Uxxx> mentions; identify BytePlus owner per channel	scripts/build_digest.py
5	AI synthesis — Executive Summary, Highlights, TODO, Pipeline, Risks, By-Channel (full message text, never truncate), Statistics	LLM call inside the conversation
6	Create per-day Feishu doc Slack Daily Digest · YYYY-MM-DD; move to folder; if doc already exists for today (16:00 run), edit-inplace replace the placeholder section	scripts/lark_doc_client.py
7	Pin the new entry to the top of the index list (insert_before the current first list item, NOT insert_after the title)	scripts/lark_doc_client.py
8	Lark IM notify the user	scripts/notify_lark.py

5.1 Time-Window Calc

import datetime, calendar
now_utc = datetime.datetime.utcnow()
window_hours = config["window_hours"]  # 19 for 11:00, 5 for 16:00
oldest = calendar.timegm((now_utc - datetime.timedelta(hours=window_hours)).timetuple())
latest = calendar.timegm(now_utc.timetuple())

5.2 v4 Synthesis Spec — Mandatory Sections

The digest MUST contain, in order:

1. 🎯 Executive Summary — 3-5 bullets, each with 🔴/🟡/🟢 + Owner + concrete action
2. 📌 Highlights — 6-10 cross-channel one-liners
3. ✅ TODO Backlog — table: Priority | Owner | Item | Deadline | Source channel
4. 📈 Business Pipeline — table: Partner | Stage/Amount | Status | Owner | Health
5. ⚠️ Risk Signals — bulleted with channel + risk
6. 🗂️ By Channel — for each active channel, four sub-blocks:
   • 📖 Context — 2-4 sentences
   • 💬 Full Messages — every message verbatim, newlines normalized to / (slash-space). No truncation.
   • 🎯 Next Steps — at least one P0-P2 with named Owner + Deadline
   • Risk — one line, 🔴/🟡/🟢 prefix
7. 📝 Statistics — top channels, top contributors, identified BytePlus owners
8. Trailing placeholder for the next intra-day run: ## 🕙 16:00 Update (placeholder, to be filled) — only on the morning run

5.3 Pin-to-Top Rule (CRITICAL)

The natural insert_after against the index title appends to the bottom of the list (because all list items are siblings of the title). The correct anchor is the current first list item:

# Read index doc, find first "- **YYYY-MM-DD**" line
# Use insert_before with that line as the anchor
edits = [{
    "mode": "insert_before",
    "selection_with_ellipsis": "- **<previous-newest-date>** · [Slack Daily Digest · ...]",
    "markdown": "- **<today>** · [Slack Daily Digest · <today>](<new-url>)"
}]

If the list is empty, fall back to insert_after against the title (only happens for the first-ever entry).

---

6. Phase E — Watchdog (T+20min)

Triggered 20 minutes after each scheduled run.

python3 ~/.claude/skills/slack-daily-digest/scripts/watchdog.py --date <today> --slot <11:00|16:00>

Watchdog checks:

1. Doc existence — read the index, look for <today>'s entry. Missing → fail.
2. Doc content — read the doc body. Verify all 7 mandatory sections (§5.2) are present.
3. Section completeness — for the slot in question (11:00 → main body; 16:00 → 16:00 section must not contain the placeholder text "placeholder, to be filled").

On failure:

• Read ~/.slack-daily-digest/runs/<today>/error.log if it exists; otherwise inspect the latest pull/build artifacts.
• Diagnose root cause (network, token expired, rate-limit, doc-edit ambiguity).
• Self-retry: rerun from the failing step. Token expired → notify user via IM and stop. All other failures → up to 3 retries with backoff [60s, 300s, 900s].
• After each retry, re-invoke watchdog. Loop until success or until the token-expired terminal condition.

On success:

• Append a one-line success note to ~/.slack-daily-digest/runs/<today>/watchdog.log.
• Do not notify the user on success (only failure / recovery).

---

6.5 Local Routine Catch-Up

For Codex / Claude Code local mode, scripts/local_routine.py run-due owns schedule catch-up:

• One success per date + slot.
• If the computer was off/asleep at the scheduled time, run automatically on the next active check when current time is inside working hours and within catchup_hours after the scheduled slot.
• If a run fails due to transient errors, retry on future routine ticks while still inside the catch-up window.
• If credentials are invalid/expired, stop retrying and notify the user clearly.
• If the catch-up window expires, mark the slot as skipped_missed_window and do not create a stale digest next day.
• Use ~/.slack-daily-digest/routine.lock to prevent overlapping runs.

The local routine invokes scripts/run_digest.py. One-off local mode pulls Slack history, builds digest_*.json, writes a host-agent handoff for synthesizing final_*.md from references/digest-template.md, publishes via LarkDocClient after final_*.md exists, updates the index newest-on-top, and writes ~/.slack-daily-digest/runs/<date>/result_<slot>.json.

---

7. Phase F — Lark IM Notification

After successful daily run, push to the configured user's Lark IM:

python3 ~/.claude/skills/slack-daily-digest/scripts/notify_lark.py \
  --user-email "<from credentials>" \
  --doc-url "<today's doc URL>" \
  --highlights "@~/.slack-daily-digest/runs/<today>/highlights.md"

The notify script:

1. Resolves email → open_id via Lark Contact API.
2. Sends an interactive card with:
   • Title: 🔔 Slack Daily Digest · <today HH:MM configured timezone>
   • Top 3 Highlights
   • Top 3 TODO items with owner + deadline
   • "Open Full Doc" button → doc URL

---

8. Error Handling Reference

Error	Action
Slack ratelimited	sleep 60s, retry up to 3×
Slack invalid_auth / token_revoked	Append to error.log, IM the user "Slack token expired, please reconfigure", stop.
users.info fails for one user	Fallback to User(<id>), do not block the run.
Lark 99991663/4/5/8 or "token expire"	The agent harness's token-refresh hook handles this — do NOT re-implement here.
Lark edit-inplace validation 1101 (locator not found)	Re-read doc with --output-format markdown, recompute the exact anchor (watch for full-width vs half-width punctuation), retry.
Lark doc create timeout	Save raw markdown to ~/.slack-daily-digest/runs/<today>/backup.md and IM the user.

---

9. Cookbook — Common Operations

Reconfigure schedule

python3 ~/.claude/skills/slack-daily-digest/scripts/save_credentials.py --reschedule "11:00,16:00"

Force a one-off run for an arbitrary window

python3 ~/.claude/skills/slack-daily-digest/scripts/run_digest.py --oldest <epoch> --latest <epoch>

Backfill a missed day

python3 ~/.claude/skills/slack-daily-digest/scripts/run_digest.py --date 2026-05-04 --slot 11:00

---

10. Files in This Skill

slack-daily-digest/
├── SKILL.md                          # this file
├── scripts/
│   ├── save_credentials.py           # encrypt + persist creds
│   ├── refresh_users_cache.py        # weekly users.list refresh
│   ├── list_channels.py              # users.conversations
│   ├── pull_history.py               # conversations.history per channel
│   ├── build_digest.py               # ID resolution + owner detection
│   ├── run_digest.py                 # orchestrator (Steps 2-7)
│   ├── platform.py                   # Mira vs local-agent detection
│   ├── lark_doc_client.py            # Mira docs / local Lark CLI abstraction
│   ├── setup_lark_cli.py             # local Lark CLI readiness helper
│   ├── local_routine.py              # local schedule + missed-run catch-up runner
│   ├── watchdog.py                   # T+20min verification + self-retry
│   ├── notify_lark.py                # Lark IM notification
│   └── lark_helpers.py               # shared: token, email→open_id, doc edit
└── references/
    ├── slack-app-setup.md            # extended Slack onboarding docs
    ├── lark-bot-setup.md             # extended Lark onboarding docs
    └── digest-template.md            # full v4 synthesis template