--- name: morning-digest description: Synthesize the operator's overnight work (cse-forecast, cse-intake, cse-engagement-finder runs; war-room, forecast, and 1:1 prep outputs; any [ATTN]/[ERR] lines from the skill logs) into one Slack self-DM written in the operator's voice. Use when the user asks for the morning digest or overnight recap, or when an operator-installed external scheduler invokes it at 06:30 Pacific local time on weekdays. This repository does not install a scheduler. argument-hint: "[date or 'yesterday']" --- # Morning Digest ## Compact MCP routing Follow the shared [compact MCP routing contract](../../shared/compact-mcp-routing.md). The facade tools are `cse_capabilities`, `cse_read`, `cse_apply`, `context_assemble`, and `cse_session_info`; named operations are capability ids. On interactive `/mcp/full`, direct named reads and each named write's per-tool guarded preview/apply flow are allowed; `cse_read`/`cse_apply` are also available facades. On facade-only `/mcp`, route reads through `cse_read` and mutations through `cse_apply` twice: preview first, then the identical capability and arguments with underlying `execute:true`, justification, and the returned `preview_digest`. Call `cse_session_info` directly for auth recovery. Do not call `context_assemble` or run `cse-sweep` for this digest. Worked ticket read: ```json jira_get_issue({"key": "CSE-123", "fields": ["summary", "status"]}) ``` Read the operator's skill logs from the past ~24h, sparsely verify surfaced Jira keys and unanswered Slack messages, and post a single self-DM summarizing what needs attention vs. what was handled autonomously overnight. Written in the operator's voice per `.agents/shared/prose-voice.md`. ## Domain Configuration No domain discovery is needed. Ticket verification requests only standard `jira_get_issue` fields; customer names come from the log line being verified. ## Voice This skill writes in the authenticated operator's voice. Follow the register rubric `.agents/shared/prose-voice.md` (sentence-case opens, contractions, dash palette, declarative cadence, two-facts-max, no em-dash) plus the `writing` skill for the patterns to favor and the AI-slop tells to avoid. `$CSE_OPERATOR` is only a bootstrap-resolved voice selector; if it is unset or ambiguous, fail closed. What this means concretely: - Declaratives are the default closer. Use "Parked." "Tracking." "Should be all you need." Avoid "gut check?" / "thoughts?" on every section. - Two facts max per bullet, connected with `--` / `;` / `,`. Never `and ... and ...` stacking. - No Unicode em-dash (`—`). Use `--` for dashed asides. - Contractions on. "It's", "don't", "we've", "I'll". Never "do not" / "cannot". - First-person for writes I performed; passive for state I'm just reporting. ## Setup Before running, `source` `.agents/shared/skill-bootstrap.sh` so `CSE_TOOLD_BIN` and `CSE_OPERATOR` resolve. The daemon keeps the self-DM write path's keychain auth fresh on its own; no mint step is needed: ```sh # shellcheck source=.agents/shared/skill-bootstrap.sh source "${PLUGIN_ROOT:-${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$PWD/plugins/cse-tools}}}/.agents/shared/skill-bootstrap.sh" ``` After that: `$CSE_OPERATOR` and `$CSE_TOOLD_BIN` are available; Jira and Slack auth are read from daemon keychain entries. On a 401, request `cse_session_info({"force_refresh": true})` once and retry once. A 403 means access is blocked or revoked: surface the named operator action and stop without retrying. Direct-tool guardrails for this skill: - For Slack direct search prefer `q`; `query` remains a compatibility alias. - Jira reads use `jira_get_issue` only for keys already surfaced by logs; request explicit `fields`. - Slack reads stay bounded to one overnight search plus the candidate threads needed to determine whether the operator replied; on `/mcp/full`, use the named tools directly, and on `/mcp`, route their capability ids through `cse_read`. Required surfaces: - **Read strategy:** Follow [`.agents/shared/read-strategy.md`](../../shared/read-strategy.md). This is a log-centric digest, not an entity-retrieval skill. Do not route it through `context_assemble` or `cse-sweep`. Re-resolve only specific logged ticket keys and unanswered Slack candidates with capability reads. - **Tool frugality:** Follow [`.agents/shared/tool-frugality.md`](../../shared/tool-frugality.md). Use compact defaults, explicit Jira fields, bounded Slack results, and no speculative source fan-out. - **Skill logs** -- `~/.config/cse-tools/logs/*-YYYYMMDD-*.log` (each skill run writes one). Inside the log, grep for `[ATTN]`, `[ERR]`, and the "ACTIONS EXECUTED" / "HELD" blocks that `cse-forecast` and `cse-engagement-finder` emit. Repo-relative `logs/` is also accepted on a dry-run from a checkout. - **Meeting / forecast artifacts** -- confirm expected overnight files under `war-rooms/`, `forecasts/`, and `1on1s/` when overnight logs name them. - **Slack reads** -- `slack_search`, `slack_read_thread`, and `slack_users`: call named tools directly on `/mcp/full`, or route their capability ids through `cse_read` on `/mcp`. Use only for unanswered mentions and DMs received overnight. - **Slack writes (self-DM)** -- `slack_open_dm` and `slack_post_message` follow the two-phase write flow: use each named tool's guarded preview/apply on `/mcp/full`, or `cse_apply` on `/mcp`; prose goes through a `/tmp` `text_path`. - **Jira** -- `jira_get_issue`: call the named tool directly on `/mcp/full`, or route its capability id through `cse_read` on `/mcp`, only for ticket keys surfaced by `[ATTN]`, `[ERR]`, or `HELD` lines. Do not run a broad `jira_search`. If a surface is unavailable, skip that section and note it in the digest body as `DATA UNAVAILABLE -- `. Never fabricate activity. ## When to use this skill - External schedule (not installed by this repository): the operator may add `CRON_TZ=America/Los_Angeles` followed by `30 6 * * 1-5 ` to their crontab. This runs at 06:30 Pacific local time with daylight-saving changes handled by the timezone. On-demand invocation remains fully supported. - On-demand: operator types "morning digest" / "overnight recap" / "what happened while I was asleep" in chat. - Argument: optional date (default = today; operator can pass `yesterday` to re-compute for the prior day's logs if the cron missed). ## Inputs - `$1` (optional) -- date slug: `today`, `yesterday`, or `YYYY-MM-DD`. Default `today`. - Implicit: authenticated operator identity from `cse_session_info` (routes the self-DM to the right user); `$CSE_OPERATOR` is a display/voice selector, never an email derivation source. ## Process ### Phase 1: Resolve operator identity 1. Call `cse_session_info` and read `operator_identity` from the daemon-owned authenticated operator identity. Require `operator_identity.available === true` and a non-empty `operator_identity.email`. Never construct an email from the `$CSE_OPERATOR` slug. Stop if the session does not expose that email. `cse_session_info` does not supply a Slack user ID. 2. Resolve that email with capability `slack_users` and arguments `{action:"search", query:}`. Keep only results whose profile email equals the daemon email (case-insensitive exact match). Fail loudly if zero or multiple exact matches are returned. Profile the single match (`action:"profile"`) and store its ID as `$OPERATOR_SLACK_UID`. Abort if the profile email does not equal `operator_identity.email`. 3. Dry-run `slack_open_dm({users:$OPERATOR_SLACK_UID, dry_run:true})`, then execute with identical arguments plus `execute:true`, a justification of at least 16 trimmed characters, and the returned `preview_digest`. Before accepting the returned DM channel, profile `$OPERATOR_SLACK_UID` again and hard-verify both: (a) the profile ID equals `$OPERATOR_SLACK_UID`, and (b) the profile email equals `operator_identity.email` (case-insensitive). **ABORT without posting** if either check fails; never fall back to a username, hard-coded email, `$CSE_OPERATOR`, or default person. ### Phase 2: Gather the bounded Pacific-local interval Resolve `$1` (`today`, `yesterday`, or `YYYY-MM-DD`) as a calendar date in `America/Los_Angeles`; reject invalid dates. The digest window is `[06:30 on the preceding Pacific date, 06:30 on the target Pacific date)`. Convert both endpoints to instants for timestamp filtering. Enumerate candidate filenames for both Pacific calendar dates touched by the interval (and both filename-date prefixes when the UTC conversion crosses midnight), then retain only records whose opening timestamp is inside the half-open interval. Do not use the host timezone or a UTC-date wildcard as the filter. **Skill logs:** 1. List candidate `$LOG_DIR/*-YYYYMMDD-*.log` for every filename date identified by the bounded interval, where `$LOG_DIR = ~/.config/cse-tools/logs`, or the resolved repository root's `logs/` on a repo-relative dry-run. Filter by each log's opening timestamp against the interval before consuming it. 2. Per-log, extract: - Opening timestamp and real completion receipt. For `cse-tools-run-skill` logs, parse its final stdout JSON object: schema `cse_tools.skill_run_result.v1`, `status:"pass"`, plus process exit code 0 is success; `status:"blocked"` or exit code 2 is blocked, and a thrown runner error exits 1. Its JSONL transcript's `context_pack` record is retrieval evidence, not a generic completion trailer. For other producers, consume an explicit supervisor/process exit receipt when present; if none exists, mark completion `UNKNOWN` rather than inventing failure or success. - `[ATTN]` lines (soft-skip items the skill surfaced). - `[ERR]` lines (hard failures with a `ecs-claude --resume ` hint). - `ACTIONS EXECUTED` block (`cse-forecast` format): transitions, field updates, comments when execute mode ran. - `HELD` block: intended actions blocked by H1/H2/H3 hold conditions. 3. De-duplicate: if both `[ATTN]` and `HELD` reference the same ticket + field, collapse to one digest line with the more specific reason. **Slack overnight activity:** 1. After parsing the logs, run one bounded `slack_search` with `after: from: (channel_has: OR dm)` scoped to `#cse-requests`, `#help-cse`, `#gold-accounts`, plus the operator's DMs. Use the capability's `q` argument. 2. Filter to messages that contain a direct @-mention of the operator OR a DM to the operator. 3. Use `slack_read_thread` only for those candidates to see if the operator already replied. If not: candidate for "Needs your attention". **Jira verification:** 1. Extract unique ticket keys from `[ATTN]`, `[ERR]`, and `HELD` lines. 2. Call `jira_get_issue` once per unique key, in parallel, requesting only `summary` and `status`. 3. Do not enumerate Jira or infer new overnight activity outside the logs. **Meeting and forecast coverage from local artifacts:** Use only war-room / forecast / 1:1 output paths named by overnight logs. Confirm whether each expected file exists under `war-rooms/`, `forecasts/`, or `1on1s/`; do not call another evidence source to discover meetings or forecasts. ### Phase 3: Classify and prioritize Sort findings into three buckets: **Needs your attention** (put at the TOP of the Slack DM): - Any `[ERR]` line with a session-resume handle. Include `ecs-claude --resume ` verbatim. - Any `[ATTN]` tagged to a specific ticket + missing field that the skill attempted to resolve and couldn't. - Held H1 (missing prerequisite field that auto-research failed to fill). - Held H3 (customer unresponsive 10+ days where the ticket still sits in `Customer Validation` or later -- those need a strategic call the skill shouldn't make alone). - Any Slack DM or direct @-mention the operator hasn't replied to. **Handled autonomously**: - Successful transitions (group by target stage, one line per stage: "3 moved from Pilot Scoping to Pilot Validation"). - Successful field resolutions (count + one example). - Tagged escalation comments posted. - Intake tickets created by cse-engagement-finder Bucket A. - Ruled-out leads appended by cse-engagement-finder. **Meeting and forecast coverage**: - Yesterday's meetings and forecasts with files on disk (`war-rooms/*`, `forecasts/*`, `1on1s/*`) -> one-line "Prep landed" or "Forecast landed" summary. - Logged prep/forecast runs whose expected output file is missing -> "No prep artifact for " or "No forecast artifact for " (gap signal). Drop anything that doesn't fit a bucket. The digest is a synthesis. ### Phase 4: Compose the DM (operator voice) The DM is a single file-backed `slack_post_message`. The Slack adapter accepts `blocks` only as inline arguments, not by file path, so do not send blocks; put the complete fallback-formatted Slack text in `text_path`. Structure: ``` *Morning digest -- * *Needs your attention* - - *Handled* - *Meetings* - _Full logs: _ ``` ### Voice rules for the body Apply `.agents/shared/prose-voice.md` rubrics: - Open with sentence-case. Never "Auto-digest:" / "Here is your..." / "AI summary". - Declaratives beat elicitations. Use "3 moved to Pilot Validation." Avoid "Should I move these?" - One connective of any type per bullet. No `and ... and ...` chains. - 60-140 chars per bullet is the sweet spot. Two sentences only when the caveat requires it. - Hedge palette: `kind of`, `basically`, `my read is`, `might be worth`. One per bullet max. - Contractions required. - No emojis. Period. - For `Needs your attention` bullets, include the ticket key first (e.g. `CSE-94 Hatch Bank`) so the operator can scan; the operator reads left-to-right and wants the ticket identity before the issue category. - For `[ERR]` lines, lead with the skill + time + one-line cause: `cse-forecast 06:12 -- 502 from Jira during transition sweep. ecs-claude --resume ` - Hedges on judgment calls only: `My read is the Hatch exec sponsor is still Brad but he was quiet last week` is in-voice; `My read is the transition succeeded` is not (that's a fact, state it). ### Phase 5: Post 1. Compose the message body in a `/tmp` file so the operator can inspect it before the write (set `DIGEST_DRY_RUN=1` to skip the write and print to stdout; respect this in cron invocations during the first week). 2. Dry-run `slack_post_message({channel:$OPERATOR_DM_CHANNEL, text_path:, dry_run:true})`. Re-call with the same arguments plus the returned `preview_digest`, `execute:true`, and a specific justification of at least 16 trimmed characters. Do not pass inline prose or `blocks`. 3. Capture the returned `ts` so the digest can be edited later. For an edit, write the full replacement text to a `/tmp` file, dry-run `slack_update_message` with arguments `{channel, ts, text_path, dry_run:true}`, then execute with identical arguments plus the returned `preview_digest`, outer `execute:true`, and a justification of at least 16 trimmed characters. 4. Record the channel and `ts` in the secondary digest file only. Do not append to a producer log; `~/.config/cse-tools/logs/morning-digest-YYYY-MM-DD.md` is the canonical writable receipt target for digest updates. ## Output - Primary output: one Slack DM in the operator's self-conversation. - Secondary output and canonical digest-update receipt: digest body written to `~/.config/cse-tools/logs/morning-digest-YYYY-MM-DD.md` for grep-friendly replay, including the channel, `ts`, and raw inputs hash (sha256 of the concatenated log filenames + Slack activity IDs) so a deterministic re-run can detect when nothing changed. - Exit codes: - `0` -- digest posted cleanly (or successful dry-run). - `2` -- auth preflight failed. - `3` -- Slack post failed; digest body IS on disk at the secondary-output path, operator can copy-paste from there. ## Validation checklist (run before posting) - [ ] `$OPERATOR_SLACK_UID` was resolved from an exact Slack profile email match to `cse_session_info.operator_identity.email`; re-profile before post confirms the same ID and email; `$CSE_OPERATOR` is not used as an identity source - [ ] "Needs your attention" bullets each carry a concrete handle (ticket key, Slack permalink, or `ecs-claude --resume `) - [ ] No bullet starts with "Auto-", "AI-", "I have ", "I was unable to" - [ ] No Unicode em-dash in the body (grep the composed text for U+2014) - [ ] No emojis anywhere - [ ] At least one declarative closer (not every bullet ends with "?" or "thoughts?") - [ ] "Handled" section is counts + one example - [ ] If `[ATTN]`/`[ERR]` is zero and "Handled" is empty, the DM opens with "Quiet overnight -- nothing to flag." and exits ## Behavioral rules - Never fabricate. If cse-forecast didn't run, say so: `cse-forecast 06:07 -- no log on disk. Check supervisor on the instance.` - Never page the operator twice for the same issue. If the last digest's `ts` referenced the same ticket + error, skip re-surfacing it; the previous message is still in the DM. - Never dump the raw skill log into the DM. The digest is a synthesis; the log is a link at the bottom. - Never include cleartext customer-sensitive info (churn signals, exec feedback) in the digest body. Surface the ticket key + "needs a read" instead; the operator can open the ticket to read the raw context. - Never skip the voice loader. A digest in wrong-operator voice is worse than no digest. - If the daemon connection is refused, preserve/write the secondary digest file with all locally gathered content, point the operator to `cse-mcp-setup`, and stop. Do not attempt `slack_open_dm`, `slack_post_message`, or `slack_update_message` while the daemon is unreachable. - If the voice loader cannot resolve the operator profile or `prose-voice.md` is unreadable, abort and post a single line: `Morning digest offline -- voice profile unavailable. Check ~/.config/cse-tools/ or re-source skill-bootstrap.sh.`. Missing raw corpus or examples alone is not a reason to use generic prose; continue from `prose-voice.md`.