Standup Report

Standup Report

No em-dashes. Nothing this skill writes may contain an em-dash; use a comma, colon, or parentheses instead. Enforced by a repo hook that flags em-dashes in generated .md. See /flagrare:write-docs.

Generate a daily standup recap that reads like a human wrote it. The reader is either the user (preparing what to say at standup) or their team (skimming a Slack post). Either way, they want the story of yesterday's code work, not a JSON dump of commits.

The single rule that shapes everything below: work is named by what it was, not by its ID. "Fixed the image cache eviction" beats "Merged PR #481" every time. Nobody remembers numbers; everyone remembers the bug.

Setup (first run only)

Config lives at ~/.claude/skills/flagrare/config.json: a single file shared across all flagrare skills, outside the plugin tree so it survives plugin updates and reinstalls. Skill-agnostic keys (GitHub login, display name, repo scope, local roots) sit at the top level; skill-specific keys nest under skills.<name>.

In Bash, expand ~ explicitly: "$HOME/.claude/skills/flagrare/config.json". The directory may not exist yet, mkdir -p "$HOME/.claude/skills/flagrare" before writing.

Step 1: Migrate legacy config (one-time)

If ~/.claude/skills/flagrare/config.json does not exist but a legacy per-skill config does, migrate it:

LEGACY="{skill_directory}/config.json"   # old location, lost on plugin reinstall
NEW="$HOME/.claude/skills/flagrare/config.json"
if [ ! -f "$NEW" ] && [ -f "$LEGACY" ]; then
  mkdir -p "$(dirname "$NEW")"
  # Wrap legacy keys: standup-specific ones (extra_mcps) move under skills.standup-report
  jq '{
    github_login, display_name, first_person, repo_scope, local_repo_roots, tracker_mcp,
    skills: { "standup-report": { extra_mcps: .extra_mcps } }
  } | with_entries(select(.value != null))' "$LEGACY" > "$NEW"
fi

Tell the user once: "Migrated your config from the old per-plugin location to ~/.claude/skills/flagrare/config.json so it survives plugin updates."

Step 2: First-time setup (if no config exists at the new path)

Use AskUserQuestion to collect:

1. GitHub login: needed for author:, commenter:, reviewed-by: searches. If missing, run gh api user --jq '.login' and confirm.
2. Display name: how to refer to the user in the narrative ("Alan reviewed two PRs" vs "you reviewed two PRs"). Default to first-person ("I").
3. Repo scope: one of:
   • org:<name>, all repos in a GitHub org (best for company work)
   • user:<login>, all repos under the user's account (best for personal projects)
   • explicit list of owner/repo strings, when work spans orgs
4. Local repo roots: directories under which to scan for local commits (e.g., ~/Dev, ~/work). The skill walks one level deep looking for .git/ to enumerate repos.
5. Tracker MCP: detect which is installed in the session (Linear, Jira, Notion, Asana, Shortcut, Trello). If multiple, ask which one this user actually files tickets in. If none, skip.
6. Additional MCPs to query: after the above is filled in, list the other MCPs currently available in the session (Slack, Discord, Google Calendar, PostHog, etc.) and ask if any should feed context into the recap. Example: Slack DMs/channels might surface conversations that explain why a PR was opened. Save which ones the user opts in to.

Save the shared keys at the top level and the standup-specific extra_mcps under skills["standup-report"]:

{
  "github_login": "aturing",
  "display_name": "Alan",
  "first_person": true,
  "repo_scope": { "type": "org", "value": "acme-corp" },
  "local_repo_roots": ["~/Dev", "~/work"],
  "tracker_mcp": "linear",
  "skills": {
    "standup-report": {
      "extra_mcps": ["slack"]
    }
  }
}

Confirm with the user before saving. Preserve any pre-existing skills.* blocks from other flagrare skills, merge, don't overwrite.

Returning user

If ~/.claude/skills/flagrare/config.json exists, read the top-level keys plus skills["standup-report"]. If only some top-level keys exist (e.g. brag-doc was set up but not standup-report), reuse what's there and only ask for extra_mcps. If the user says "reconfigure" or "edit setup", re-run the setup flow, but only rewrite the skills["standup-report"] block plus any top-level keys the user changes; leave other skills' blocks untouched.

Resolve the time window

First, determine today's date and current time by running:

date '+%A %Y-%m-%d %H:%M'

Do NOT guess the day of the week from context or the current date string. LLMs are unreliable at day-of-week calculations. Always run the command above and use its output.

"Yesterday" at standup means last working day, not literal yesterday. Apply this logic to the output of the date command:

today = Mon → window starts Friday 00:00
today = Tue-Fri → window starts previous calendar day 00:00
today = Sat/Sun → window starts most recent Friday 00:00

The window always ends at the current time (now), not at midnight of the previous day. Work done earlier today (before the skill is invoked) is part of the standup. For example, if it's Tuesday 9:30am, the window is Monday 00:00 through Tuesday 09:30.

If the user specifies a window in the prompt ("since Friday", "last 3 days", "this week"), honor it. Otherwise apply the rule above. State the resolved window in the report header so the reader knows what's covered.

Data collection

Run the following queries in parallel: they're independent and slow if serialized. Use ISO 8601 dates (YYYY-MM-DD) for the window boundaries. {FROM} is the start of the window (last working day). {TO} is today's date (so that any work done between midnight and now is captured).

1. GitHub: your authored PR activity

gh api "search/issues?q={SCOPE}+is:pr+author:{LOGIN}+updated:{FROM}..{TO}&per_page=100" \
  --jq '.items[] | {number, title, html_url, state, draft, merged_at, created_at, updated_at, body, repo: (.repository_url | split("/") | last), labels: [.labels[].name]}'

For each PR returned, also fetch:

• Commits you authored in the window:

  gh api "repos/{OWNER}/{REPO}/pulls/{NUMBER}/commits" \
    --jq '[.[] | select(.commit.author.email == "{EMAIL}" or .author.login == "{LOGIN}") | {sha, message: .commit.message, date: .commit.author.date}]'
  

• Reviews on the PR (to know if you addressed feedback yesterday):

  gh api "repos/{OWNER}/{REPO}/pulls/{NUMBER}/reviews" \
    --jq '[.[] | select(.user.login | test("\\[bot\\]$") | not) | {user: .user.login, state, submitted_at, body}]'
  

2. GitHub: reviews you gave to others

gh api "search/issues?q={SCOPE}+is:pr+reviewed-by:{LOGIN}+updated:{FROM}..{TO}&per_page=100" \
  --jq '.items[] | {number, title, html_url, repo: (.repository_url | split("/") | last), author: .user.login}'

For each, fetch your reviews specifically to know whether you approved, requested changes, or commented:

gh api "repos/{OWNER}/{REPO}/pulls/{NUMBER}/reviews" \
  --jq '[.[] | select(.user.login == "{LOGIN}") | {state, submitted_at, body}]'

3. GitHub: comments you left

gh api "search/issues?q={SCOPE}+commenter:{LOGIN}+updated:{FROM}..{TO}&per_page=100" \
  --jq '.items[] | {number, title, html_url, repo: (.repository_url | split("/") | last)}'

Deduplicate against the prior two queries, only count this PR if it wasn't surfaced as authored-by or reviewed-by you. The remaining set is PRs where you contributed just a comment.

4. Local git commits

For each configured local repo root, enumerate */.git/ one level deep, then for each repo:

git -C {repo_path} log --author="{LOGIN_OR_EMAIL}" \
  --since="{FROM}" --until="{TO} 23:59:59" \
  --date=format:'%a %Y-%m-%d %H:%M' \
  --pretty=format:'%H|%ad|%s|%b%n---'

%ad with --date=format:'%a ...' makes git print the weekday (Mon, Tue, ...). Use that weekday verbatim. Never compute a weekday from a date yourself.

These catch work that hasn't hit GitHub yet, local WIP, branches not pushed. Deduplicate SHAs against the GitHub PR commits already collected.

5. Deploys / release automation

For each merged PR in step 1, check whether a release workflow fired afterward:

gh api "repos/{OWNER}/{REPO}/actions/runs?created=>={MERGE_DATE}&per_page=20" \
  --jq '.workflow_runs[] | select(.name | test("release|deploy|publish"; "i")) | {name, conclusion, html_url, created_at, head_sha}'

Match on head_sha (the merged PR's merge commit SHA) or by time proximity (workflow created within ~5 min of merge). Record: did it run, did it succeed, when. If the conclusion is failure, surface that explicitly, a "shipped" PR that didn't actually deploy is the kind of detail standup is for.

6. Linked tickets (if tracker MCP configured)

For each PR collected, scan body and title for ticket references:

• Linear: \b[A-Z]{2,5}-\d+\b (e.g., ENG-142)
• Jira: same shape, configured project keys
• Notion: URLs matching notion.so/...
• GitHub issues: #\d+ references resolved within the repo

Use the configured tracker's MCP to fetch each ticket's title and status. The title is what you use in the narrative ("the image cache eviction bug"), the ticket ID stays in the footnote.

7. Optional MCP context

For each opted-in extra MCP (from setup), pull anything from the time window that mentions the user or references the PRs/tickets above. Examples:

• Slack: messages mentioning @{user} or threads where they posted code-related replies during the window. Filter aggressively, birthday wishes and lunch polls are noise.
• Calendar: meetings tagged or named for the repos/projects involved. A "design review for X" meeting is context the standup can mention ("the morning was meetings, design review for X took an hour").

Treat MCP results as narrative seasoning, not primary data. If an MCP errors out or returns nothing, skip silently, the standup still works without it.

8. Today's queue

Run these queries in parallel with steps 1-7. Their output feeds the Today section only, they do not affect the Yesterday narrative.

Carry-over threads (derived from steps 1: 4, no new queries needed)

After collecting yesterday's data, flag any of these as carry-over candidates:

• Authored PRs still open and not merged, especially those with changes_requested reviews or no approvals yet
• Authored PRs still in draft state
• Local branches with commits not yet in any PR (from step 4, commits with no matching SHA in the GitHub results)

Order carry-overs by recency of last push (most recently active first). These are the highest-signal "likely working today" items because work was already in motion.

Assigned ticket queue (tracker MCP, if configured)

Fetch tickets assigned to the user in a not-yet-started state:

• Linear: assignee: me with state type in [triage, backlog, unstarted], ordered by priority ASC
• Jira: assignee = currentUser() AND status in ("To Do", "Open", "Backlog") ORDER BY priority ASC
• Notion: query the tasks database for rows where Assignee = user, Status not in Done/Cancelled, sorted by Priority property

Normalize priority to a consistent scale for ranking: Urgent → High → Medium → Low → No priority (no-priority sorts last). Limit to 10 results; note "and N more" if the queue is longer.

Cross-reference against carry-overs: if an open PR or local branch references a ticket that also appears in the queue, treat them as the same item, list it once, under the carry-over heading, with the ticket title as the label.

If the tracker MCP is unavailable or returns nothing, omit the queue silently. Carry-overs from GitHub/local git are always shown.

Pending review requests (GitHub)

gh api "search/issues?q={SCOPE}+is:pr+is:open+review-requested:{LOGIN}&per_page=50" \
  --jq '[.items[] | select(.draft == false) | {number, title, html_url, repo: (.repository_url | split("/") | last), author: .user.login, updated_at}]'

These are open, non-draft PRs where the user is an explicitly requested reviewer and has not yet submitted a review. Exclude:

• PRs already in step 2 results (reviewed-by, already acted on)
• PRs authored by the user (already in carry-overs)
• Draft PRs (.draft == true), the author isn't ready for review

For each result, note the PR author, "Daniel's auth refactor" is more useful than a bare title in the Today context. If the PR body or title references a linked ticket, resolve its title via the tracker MCP (same logic as step 6).

Naming work in human terms

For every PR, commit cluster, or ticket, derive a human phrase that names the work. Priority order:

1. Linked ticket title, lowercased and trimmed (e.g., "Image cache evicts entries on retry" → "the image cache eviction bug" or just "image cache eviction")
2. PR title with conventional-commit prefix stripped (fix(cache): LRU eviction on retry → "fixing the LRU eviction on retry")
3. Commit subject if the PR has no descriptive title
4. First meaningful file path as a last resort (src/cache/lru.ts → "the LRU cache code")

The phrase goes in the prose; the PR number and ticket ID go in the refs footnote. Inline numbers belong only when the user explicitly asks ("include PR links inline").

Pin every event to a real date before you write

Temporal hallucination, the wrong weekday or work attributed to the wrong day, is this skill's most common failure. Prevent it mechanically: before writing a single sentence, build a dated timeline from the timestamps you already collected, and treat that timeline as the ONLY source for any day or time reference.

• Every collected event carries a real timestamp: commit dates (now printed with their weekday via %ad), PR merged_at / created_at, review submitted_at, deploy created_at. Tag each event with its weekday and date straight from that timestamp.
• Do NOT infer a weekday from a date in your head. If you need the weekday for an ISO date that did not arrive with one, compute it: date -j -f '%Y-%m-%d' '<ISO>' '+%A' (macOS) or date -d '<ISO>' '+%A' (GNU). Never reason it out.
• The date output from "Resolve the time window" is today's anchor. Everything else is positioned relative to it using real timestamps, never guessed.

The hard rule: any specific day or time-of-day claim in the report must trace to a timestamp in this timeline. "Tuesday", "this morning", "late afternoon", "yesterday evening", "first thing" are calendar or clock claims. Use them only when an event's actual timestamp supports them. No timestamp, no temporal claim.

Narrative synthesis

This is where the skill earns its keep. The naive version of this skill enumerates events. The version worth shipping reads like a Staff Engineer's standup: impact-first, root-cause-aware, and honest about judgment calls.

Write like a Staff Engineer, not a junior

The reframe that matters: a standup is not a status report on you. It's a status report on the system: what's better, who's unblocked, what risks were caught, what's still in flight and why. The work is the vehicle; the impact is the cargo.

Apply these shifts when synthesizing:

• Lead with impact, not action. "The X bug that's been paging us all week is fixed and deployed" beats "Fixed bug X". The reader cares that the pages stop, not that you typed.
• Name root causes, not just symptoms. If you debugged something, say what was actually wrong, that's the standup-worthy detail. "Found the LRU was admitting entries faster than it evicted under burst load" is what makes the team smarter; "fixed the cache" doesn't.
• Reviews are judgment calls, not tasks. Recast: "Approved Daniel's PR" → "Unblocked Daniel on the auth refactor, the migration plan was sound." "Requested changes on Carol's PR" → "Pushed back on Carol's caching strategy; I think the eviction model will bite us next quarter." Name what you decided and why, not just the GitHub button you clicked.
• Connect work to systems and people. Which team is unblocked? Which downstream service was at risk? Which on-call rotation just got quieter? Standup-readers want to map your work onto the org.
• Acknowledge what didn't ship and why. Half-done work and deliberate punts are part of the story. "Started the queue-sharding work but parked it after a quick spike showed the bottleneck is upstream, refocusing today" is staff-level. Silence on unfinished work reads as overpromising.
• Forward-looking notes earn their keep. If something you reviewed is going to bite later, say so once, plainly. The standup is the cheapest place to surface a future risk.
• Vary sentence shape. Don't open every beat with a past-tense verb ("Fixed... Reviewed... Addressed..."). Lead with the noun some of the time ("The auth refactor Daniel's been on landed cleanly"), the constraint ("Most of the day went to..."), or the surprise ("One thing that took longer than expected was...").

Before / after

The same events, written two ways. Both are accurate. Only one reads like a Staff Engineer wrote it.

❌ Junior recap (events as a to-do list, dressed up):

Fixed the LRU eviction bug in the image cache and merged the PR. The release workflow deployed it. Reviewed two PRs, approved Daniel's auth refactor and requested changes on Carol's cache benchmarks. Addressed three comments on my queue refactor PR.

✅ Staff recap (impact, causation, judgment):

The image-cache regression that's been paging the on-call rotation for a week is fixed and in prod, root cause was the LRU admitting entries faster than it evicted under burst load, which only surfaced because our retry policy amplifies traffic on the hot path. While that was baking, I unblocked Daniel on the auth-service refactor; the migration plan was sound and we agreed on the rollback path inline. Pushed back on Carol's cache-benchmark methodology, the workload she's measuring doesn't match what production sees, and shipping the conclusions as-is would have driven the wrong tuning decisions next quarter. My own queue refactor is in re-review after addressing the back-pressure concerns from the earlier review pass.

Notice: same five events, but the second version names the consequence of each one. The reader walks away knowing what changed about the system, not what tickets moved columns.

Look for connections

Before writing, scan the collected data for these patterns. They're the raw material for the impact framing above:

• Same files touched in multiple PRs → likely one thread of work; group them and name the thread.
• PR descriptions referencing prior PRs ("follow-up to #478", "addresses feedback from review of #475") → causal link, weave it in.
• Comments you left that were resolved by your own subsequent commits → that's "addressed feedback and pushed back into review" as a single beat.
• A review you gave that triggered changes → that's a judgment call, not a button click. Name what you flagged and what changed.
• A merged PR that deployed cleanly → one beat, framed by impact ("shipped, error rate normalized" / "shipped, no rollback needed").
• A merged PR that didn't deploy, or deployed and failed → its own beat, surface explicitly. A "merged" PR that didn't reach prod is the kind of detail standup exists to catch.
• Tickets that closed without your PR closing them → maybe a review of yours unblocked the close; cross-reference.

What never appears in output

Across all sections, paragraph, recap, bullets, refs, these categories are always excluded:

• AI/skill invocations: "ran debug-hunt", "invoked standup skill", "used the TDD writer", slash commands, skill names, evaluation runs
• Pure process steps that produced no observable outcome: "read the docs", "grepped the codebase", "ran evals", "researched X"
• Ticket-column moves with no associated code or review ("moved ENG-142 to In Review")
• Tooling churn that is self-evident from the outcome ("set up the dev environment", "ran the test suite locally")

If an activity produced a real outcome (a bug found, a risk surfaced, a decision made), write the outcome, not the activity that led to it. A debugging session is worth one standup line: the root cause, not the method.

Tone

Write in past tense. First-person if first_person: true, otherwise use the configured display name in the third person.

Use connective tissue that conveys rhythm: most of the day, while that was running, on the side, one thing that took longer than expected. These convey shape without asserting a clock or calendar position, so they are always safe. Phrases that DO assert a position, late afternoon, between meetings, Tuesday, first thing, yesterday evening, are allowed only when a real timestamp in your timeline supports them (see "Pin every event to a real date"). Order beats by importance rather than by clock, but never invent a time to do it.

Avoid the word "PR" in prose if you can substitute the work-name. "Shipped the image cache fix" beats "Merged the image-cache PR". Avoid "addressed feedback" without saying what the feedback was, "addressed the back-pressure concerns from the earlier review" tells the reader something; "addressed feedback" doesn't. (If you do name a day, it must come from a real timestamp.)

Synthesizing Today

The Today section is the forward-looking half of the standup. Its job is to answer "where are you putting your energy today?" without overpromising.

These are inferences, not commitments. Frame everything as "likely working today", the report is a prediction based on what's in motion, not a schedule the user will be held to. The reader should walk away with a reasonable picture of where the day is headed, not a list they expect to be checked off.

Ordering logic, three tiers:

1. Active carry-overs first: threads already in motion (open PRs awaiting review, PRs with changes-requested, draft PRs close to ready, local branches not yet PR'd). These have the highest prior probability of being worked today. Order by how close they are to done: a PR with one approval short of merge ranks higher than a draft with no reviews yet.
2. Pending review requests second: open PRs where a teammate explicitly requested the user's review and no review has been submitted yet. Someone is blocked waiting; these are the highest-urgency external asks. Order by updated_at descending (most recently active first, the author may be waiting on this right now). Name both the work and the author: "Daniel's auth refactor" not just the PR title.
3. Priority queue items third: assigned tickets in To Do / Backlog / Unstarted state, ordered by tracker priority (Urgent → High → Medium → Low → No priority). If the tracker has no priority field, order by creation date descending.
4. Omit tickets that have a matching carry-over, don't list the same work twice. The carry-over entry wins because it's more specific.

Cap at 5 items total. If there are more, pick the 5 with the strongest signal (active carry-overs first, then highest-priority queue items) and note "queue has N more" in the Refs footnote.

When a carry-over is close to done, say so. "One review away from merge" or "waiting on CI" is more useful than just the work name.

When the queue is empty and no carry-overs exist, omit the Today section entirely rather than speculating. An honest "nothing lined up" is better than invented filler.

Report format

# Standup: {YYYY-MM-DD}
> Covering: {window_human} | {N} PRs touched | {M} reviews given | {K} commits

## Yesterday

{2-3 sentence narrative paragraph. Lead with impact: what changed about
the system, who's unblocked, what risk was caught. Not "I did X and Y", 
"The X problem is fixed and shipped; Y is unblocked; Z is now flagged
for next quarter." If the day was mostly meetings or interrupts, say so
plainly and frame what it enabled.}

## Recap

{Long-form journal section. 2-5 short paragraphs grouping events by
thread, not by source. Each paragraph is one coherent story with its
own impact line. Cross-reference where it helps, "the same auth
refactor I reviewed earlier in the window, Daniel ended up shipping it;
the migration ran clean." (Any day or time reference must come from a
real timestamp, see "Pin every event to a real date".)

Open paragraphs with varied shapes, not every one starts with a verb.
Lead with the noun ("The auth refactor..."), the constraint ("Most of
the morning..."), or the surprise ("One thing that took longer than
expected..."). Pure verb-first openers across every paragraph is the
junior tell.}

## Today

{Inferences about where the day is headed, not a schedule, not a
commitment. Frame each item with hedged language. Three tiers:
carry-overs → pending review requests → priority queue.
Omit this section entirely if all three are empty.}

**Carry-overs:**
- {work name}, {status: "one approval short of merge" / "waiting on CI" / "changes requested"}
- …

**Pending reviews:**
- {author}'s {work name}, {context, e.g. "requested Tuesday, auth service"}
- …

**Up next (from queue):**
- {ticket title}, {priority}
- …

## For the channel

{One opening line, the **big picture** of the day in plain English.
What was the day *about*? Not a list of things done, but the
single sentence a teammate would use to summarize your day to
someone who wasn't there. Examples:
  "Mostly heads-down on the image cache, root cause found, fix shipped, on-call relieved."
  "Split between unblocking the auth refactor and spiking the queue-sharding approach."
  "All reviews day, three PRs, design review for the new API surface."
If the day had no clear theme, lead with the most impactful item.}

**Yesterday:**
{Tight bullet list, 4-8 lines, slack-pasteable. Each bullet still
carries impact framing, what changed, what's unblocked, what's
flagged. Naked "Reviewed X" / "Approved Y" bullets are a regression
to the junior format; even compressed, the bullets should say *why
it mattered*. Names work in human terms. Deploy state inlined only
when it's notable.

**NEVER include in bullets or prose:**
- **Implementation details**: file names, function or symbol names, "refactored the X service", "added a nullable column", "split the handler". The channel reader wants the outcome, not the changeset. This is the most common channel regression.
- Skills invoked, slash commands run, or AI tooling used ("ran debug-hunt", "used standup skill", etc.)
- Process steps ("grepped the codebase", "ran evals", "read the docs")
- Ticket-column moves that produced no code change
- Self-reviews, bot-authored PRs handled individually (collapse to "approved N dependency bumps")

These are process, not outcome. If a debugging session is worth mentioning,
name what was *found* and *fixed*, not the method used.

A channel before/after, since the bullets are what the model copies:

❌ Enumerated, implementation-focused (a changelog, not a standup):
- Refactored DailyHighlight.svelte and extracted IntentionCard component
- Added HighlightOutcomeActions with yes/partial/moved_forward/not_today
- Removed EveningWitness.svelte and its test
- Updated daily-highlight.service.ts to return outcome

✅ Outcome-focused (what changed for people):
- The /today screen redesign is in review, one calm intention instead of the old wall of cards, and the evening check-in is gentler
- Returning after a quiet stretch no longer reads as a failure, the new momentum signal welcomes you back instead of resetting a streak

Same work. The first lists files; the second tells the team what is different. Write the second.}

- Image-cache regression shipped, on-call rotation should stop paging
- Unblocked Daniel on auth refactor (approved after working through the rollback path inline)
- Pushed back on Carol's cache-benchmark methodology, workload doesn't match prod
- Queue refactor back in review after addressing the back-pressure concerns

**Today:**
{1-3 lines max. Priority order: carry-overs first, then pending
review requests, then queue items. Don't list more than 3, the
team needs the headline, not the full picture. Hedged language:
"likely", "plan to", "continuing", "reviewing", "starting on".
If nothing is lined up, omit this block entirely.}

- Continuing queue refactor, one approval away from merge
- Reviewing Daniel's auth service changes (requested Tuesday)
- Starting on auth token expiry edge case (high priority)

## Refs

{Compact footnote with the actual identifiers, for anyone who wants to
chase a link. One line per work item. Include today's queue items
at the end with their ticket IDs.}

- image cache eviction, PR acme-corp/api#481, ENG-142, deployed
- auth refactor (Daniel), PR acme-corp/api#478 (reviewed)
- LRU cache benchmarks, PR acme-corp/api#479 (reviewed, changes requested)
- queue refactor, PR acme-corp/api#483 (carry-over, awaiting review)
- auth service refactor (Daniel), PR acme-corp/api#485 (review requested)
- auth token expiry edge case, ENG-145 (High, in queue)

Length rule of thumb

• Yesterday paragraph: 2-3 sentences, never more than 4. If the day was busy, summarize at a higher level rather than running long.
• Recap section: aim for 3 paragraphs. If you have only one beat for the day, the recap collapses into one paragraph and that's fine.
• Today section: 2-5 items total across carry-overs, pending review requests, and queue. If all three are empty, omit the section.
• For the channel, Yesterday bullets: 4-8 lines. More than 8 means commits are listed individually when they should be grouped under one thread.
• For the channel, Today bullets: 1-3 lines. The team needs the headline, not the full queue.

When there's nothing to report

If the window contains no activity (PTO, sick day, all-meetings day):

# Standup: 2026-05-26
> Covering: Friday 2026-05-22 | no code activity

## Yesterday

Quiet day, nothing landed in code. {If MCPs surfaced context: "Spent the day in meetings, design review for X, planning for Y." Otherwise omit this clause.}

## Today

{Still show the Today section if there are carry-overs or queue items, 
a quiet day doesn't mean there's nothing lined up next.}

## For the channel

No code shipped yesterday.

**Today:** {highest-priority carry-over or queue item, or omit if nothing}

Don't fabricate activity. An honest "quiet day" beats invented bullets.

Edge cases

• PRs that crossed midnight: a PR opened Friday 23:50 and merged Monday 09:15 belongs to whichever side has the user's activity within the window. If both, mention it as a thread that carried over.
• Reviews where you self-approved: skip from the "reviews given" count, not interesting at standup.
• Bot-authored PRs you reviewed (Renovate, Dependabot): surface as a single line ("approved 3 dependency bumps"), not one bullet each.
• Force-pushed commits: GitHub may double-count under your authored work. Dedup by sha.
• Multiple PRs on the same ticket: collapse to one work-item in the prose, but list each PR in Refs.
• MCP unavailable or unauthenticated: skip silently for tracker/Slack; if gh itself is unavailable, tell the user to run gh auth login first since GitHub is the primary source.

Output

Always render the full report inline in the conversation.

If the user asks to "post it" or "send to Slack", offer to copy the "For the channel" section specifically, the prose sections are for them, the bullets are for the team. If they configured a Slack MCP and gave a channel, offer to post directly.

If the user asks to save it, default to ./standup-{YYYY-MM-DD}.md in the current working directory.