Standup Digest

Standup Digest

Overview

Surveys Claude Code, Codex, and OpenCode session transcripts and produces a short standup report: three or four bullets to read aloud, backed by a fuller digest to drill into when someone asks a follow-up.

A companion to gws-workflow-standup-report, which covers meetings and tasks from Google Workspace. This skill covers the work itself — what got typed, committed, opened, and merged — and does not touch Calendar or Tasks.

Run the extractor

python3 ~/devel/lmorchard-agent-skills/standup-digest/scripts/standup_digest.py --format llm

Defaults to the previous workday through today, so a Monday run sweeps Friday through Sunday. Add --date YYYY-MM-DD for one specific day, or --since/--until for a range. Add --no-verify when offline.

The script defaults to writing JSON to stdout. Because this JSON can be extremely large (often over 2MB) and exceed context limits, you must use --format llm to output a condensed summary instead.

Read the output of the script with --format llm and compose from it; do not re-read the raw transcripts yourself.

The digest contract

The digest output (originally derived from JSON with schema_version 3) presents: schema_version, generated_at, window (since/until/rule), stats (sessions/projects/malformed_lines/prompt_chars_dropped/gh_calls), warnings, notes, sessions[], commits[], and working_state[].

Each session carries session_id, transcript, title, project, cwds, repo, branches, launch, started_at, ended_at, prompt_count, prompt_chars_dropped, prompts, assistant_notes, and refs[]. assistant_notes is a bounded, truncated list of the closing prose of assistant turns — a sense of what was discussed, not a transcript (see the conversation register-firewall rule below). Each ref carries kind, repo, number, source, url, verification, state, title, merged_at, closed_at. Each commit carries repo, path, sha, subject, committed_at — repo is null for a local repo with no GitHub remote (use path's directory name).

Each working_state entry carries repo, path, branch, dirty_files, and last_commit (sha/subject/committed_at, or null). One entry per worktree touched, whether or not it produced a commit — path is the worktree, so two linked worktrees of one repository appear separately even though their commits are pooled under the main checkout. branch is null on a detached HEAD; dirty_files is null if git status couldn't be read. last_commit is deliberately not windowed: it dates a branch whose work predates the window, which is how you tell parked work from work never started.

warnings and notes are different in kind. warnings is degradation — something that should have been checkable wasn't. notes is bookkeeping: a cwd that isn't a checkout, a container directory expanded to the checkouts beneath it. Notes are not degradation and must never be rendered as such.

Do not invent fields, and do not read anything from the digest beyond this list.

Language rules

The digest separates what is true from what is interesting. You choose what is interesting. You do not get to upgrade what is true.

unavailable means the claim could not be checked, not that it succeeded or failed. Never describe an unavailable ref with success language ("shipped", "merged", "landed", "closed") or failure language ("abandoned", "rejected") — "worked on" or "looked at" is the ceiling regardless of what a prompt or commit subject implied.

Further binding rules:

• Claims drawn from prompts alone — with no corroborating commits[] entry or a confirmed ref — license only intent language ("started on", "looked at", "picked up"). Outcome language ("landed", "shipped", "stood up", "cleared", "fixed") requires a commit or a confirmed ref behind it; a prompt records what someone said they meant to do, not that it happened. This is the same trap prompts[] created before the extractor learned to strip harness-injected completion claims out of it — the renderer is the layer that still has to enforce it for everything else in that array.
• Conversation is topic signal, not proof. prompts and assistant_notes say what was discussed, debated, and decided — render them in a descriptive register ("worked through", "went back and forth on", "landed on X in discussion", "explored"). They never, on their own, license outcome language ("shipped/landed/merged/fixed/closed"), which still requires a commits[] entry or a confirmed ref. assistant_notes is assistant prose and may carry aspirational completion claims ("now it works", "all green") — treat it as a record of what was said, never as evidence that it happened. This is the firewall that lets the narrative get vivid without the accomplishments getting loose.
• working_state is evidence of work, never of shipping. Uncommitted changes and a pre-window last_commit prove someone was in that tree; they prove nothing landed. "Has uncommitted changes on <branch>", "parked since Saturday", "left mid-change" are all fair. "Shipped", "landed", "finished", "implemented" are not, no matter how confidently assistant_notes describes the work as done. This closes a specific gap: before it, a session that spent hours in a checkout and committed nothing inside the window had no backing evidence at all, so the honest rendering was silence — which read as if the day hadn't happened. working_state gives that day a floor, at the "worked on" tier and no higher.
• A working_state entry on its own is thin. Pair it with the session that touched that path (via cwds) so the reader gets what was being worked on, not just that a tree is dirty. An entry with dirty_files: 0 and a last_commit inside the window is already covered by commits[] — don't report it twice.
• Anything absent from the digest never appears as an accomplishment.
• Sessions with launch: "driver" were kicked off or delegated, not hand-worked. Say so.
• Sessions with launch: "unknown" get neutral phrasing — describe what happened (the repos and branches involved, plus any refs and commits) without characterizing how the session started. Do not call it hand-worked and do not call it delegated; "unknown" means the transcript had no human prompts to judge from, not that the answer is ambiguous between the two.
• When warnings is non-empty, add one short line saying the run was degraded. A partial digest must never read as a clean one. notes is not a warning — never mention a note in the degraded line, and never let one turn a clean run into a hedged one. Notes exist so you can tell "nothing was collected here" from "something broke"; a run whose only complaint is a plain-directory cwd is a clean run.
• refs with source: "prose" are the looser signal — Les may have only mentioned a number while working on something else. Weight them below pr-link refs and below commits. A session with a long list of prose refs (e.g. five or more) is usually one piece of work that happened to touch many issue numbers, not that many separate accomplishments — describe it as one item ("triaged a batch of evals-judge issues") rather than enumerating every number in the headline. The full list still belongs in the detail file.

Missing titles

title is null when the transcript had no ai-title record — this happens. Never print "None" or leave the line blank. Fall back to the first entry in prompts (trimmed to a short phrase) for a working label, and use project plus branches to give it context, e.g. "pilo (feat/retry-backoff): fix the flaky retry test — ...". If prompts is also empty (pairs with launch: "unknown" — no human prompts recorded at all), drop the prompt fragment and label the session from project and branches alone, e.g. "pilo (feat/retry-backoff): session with no recorded prompts".

Output

Two artifacts: the terminal bullets (the tl;dr), and the full detail file — which now opens with those same bullets.

Terminal — the tl;dr

Three or four bullets ordered by significance, not chronology. Annotate each (in your head, against the ref-state table above) before you print it:

## Tue Jul 28

- Refreshed PR #446 (pilo), still open, up for review     (confirmed, OPEN — not landed)
- Cleared a batch of evals-judge issues (#96, #97, #99, #101)   (confirmed, CLOSED)
- Started standing up a zoo eval sandbox; CDP flakiness still blocking clean runs   (no ref, no commit — from a prompt only, intent language)

_threads today: secret-management migration, eval date-robustness, a lot of triage_
_full digest: ~/Documents/Obsidian/main/standup-digests/2026-07-28.md_

The first bullet is the trap to avoid: PR #446 has genuinely been refreshed but is still OPEN — "landed" or "merged" would be the fabrication this whole layer exists to prevent. The second line's "cleared" is only honest because the issue refs are confirmed + CLOSED; if any come back unavailable, drop to "worked on" for those numbers — one confirmed closure doesn't license the whole batch. The third has no ref and no commit, only a prompt describing intent, so it stays in intent language. The optional _threads today:_ line is conversational texture from prompts/assistant_notes — descriptive only, never an accomplishment (see the register firewall).

Detail file — ~/Documents/Obsidian/main/standup-digests/YYYY-MM-DD.md

Create the directory if needed; name the file for the window's start date. Structure:

1. ## tl;dr at the very top — the same significance-ranked bullets shown in the terminal (plus the one-line degraded-run note when warnings is non-empty). Someone should be able to get the day from these alone.
2. One section per project, each session under it (title, or the fallback label, plus branches) with two clearly separated parts:
   • what you worked through — a few descriptive bullets synthesizing prompts + assistant_notes: the questions asked, the debates, what got decided. Descriptive register only (see the firewall) — this is where the day gets its texture.
   • what shipped — only commit- or ref-backed items, phrased as the ref-state table permits. If nothing shipped, say so; never promote discussion into shipment. When nothing shipped but working_state shows a dirty tree or a pre-window last_commit for a path this session was in, say that instead of just "nothing" — e.g. "nothing landed; wt-bidi-webext-commands has 2 uncommitted files, last commit Aug 1." That is the difference between a day that produced no artifact and a day the digest simply couldn't see.

Group commits[] by repo. A session's own repo says which project section it belongs to, but a commit is never attributable to one session within that project — so list a project's commits once at the section level, not split under individual sessions. Commits with repo: null are a local repo with no GitHub remote: group them under a heading from the path's directory basename, e.g. zoo-service (local, no remote).

Note once in the file that prompts/notes are truncated when prompt_chars_dropped is nonzero (for a session or in stats) — the detail file is not a verbatim transcript.

When the day was quiet

If stats.sessions is 0, say the window held no recorded sessions. Do not fill an empty template, and do not pad with anything from the warnings.