Daily Work Report

Reconstructs what was actually worked on by digesting every Claude Code session transcript active in a time window, then cross-referencing git commits so the report reflects real output rather than just conversation.

Arguments

$ARGUMENTS — optional window. Accepted forms:

Map that onto the helper's -Hours / -Since parameters. Never pass both.

Step 1 — Digest the sessions

& "$env:USERPROFILE\.claude\skills\daily\Get-DailyWork.ps1"            # auto window
& "$env:USERPROFILE\.claude\skills\daily\Get-DailyWork.ps1" -Hours 48
& "$env:USERPROFILE\.claude\skills\daily\Get-DailyWork.ps1" -Since '2026-08-01'

Per session it reports: launch dir, branch, work item, active span, whether the session was resumed, the user's prompts, slash commands invoked, files written/edited, PR/WI numbers referenced, and — most importantly — the Did: lines from the mandated end-of-turn orientation footer. Those are self-authored one-line summaries of each real-work turn and are the backbone of the report; prefer them over re-deriving intent from prompts.

Filtering is per entry, not per file: a session resumed today but started weeks ago contributes only its in-window slice. Sessions marked (resumed) began before the window — do not describe their earlier work as today's.

For a stand-up the Did: lines carry almost all the signal, so keep the prompt and file noise down: -MaxPrompts 6 -MaxFiles 6. Raise them only for a session that has no Did: lines and needs its topic inferred from what was asked. -MaxDids and -Json are also available. Expect ~5s for a 24h window (it streams 100+ MB of transcripts).

Step 2 — Cross-reference git

Sessions show intent; commits show what landed. All C:\work\<ID>_<Title> worktrees share the main repo, so one --all log covers every worktree:

git -C /c/work/Payments log --all --no-merges --since="<window start>" \
    --author="$(git -C /c/work/Payments config user.name)" \
    --date=format:'%m-%d %H:%M' --pretty=format:'%ad %h %d %s'

If the digest shows sessions in launch dirs outside C:\work\Payments (e.g. C:\work\other-repo), run the same log in those repos too. Also list uncommitted work-in-progress where it matters: git -C <dir> status --short.

Step 3 — Resolve PR state

Each line has to say whether the work is merged or still open, so look up every PR the digest referenced — one call each, cheap:

for id in <ids from the digest>; do
  curl -s -L --negotiate -u : \
    "https://devops.example.com/DefaultCollection/MyProject/_apis/git/pullrequests/$id?api-version=7.0" \
    | grep -oE '"(status|title)":"[^"]*"' | head -2
done

active = open, completed = merged. The PR title is also the most reliable source for an item's name and work item number.

For a window ≥48h the merged-prs skill gives the broader shipped-to-main view, but skip it when the git log and PR states already answer the question.

Step 4 — Render the report

This is a stand-up list. One line per item, three fields, nothing else:

<WI/PR ID>, <Name>, <one line summary>

Output exactly this shape — a flat bullet list, most substantial work first:

## Stand-up — Fri 2026-08-07 (last 24h)

- **#12345**, Auto-payout customer block, PR 4321 open; 6-agent review closed, all fixes in, both release tasks done.
- **#12346**, Auto-payout processing + shadow report, built end to end, 21 tests green, committed but not pushed — no PR yet.
- **PR 4310**, Deposit limits review, reviewed on request; findings shared in-session only, not posted.
- **#12290**, Routing report data gap, not a bug — localhost reads staging, whose prod snapshot ends 08-02.

Hard rules:

• One line per item. No sub-bullets, no per-item detail block, no session IDs, no file counts, no timestamps, no branch names. If a line needs a second line, cut it down instead.
• Group by work item / topic, not by session. One topic often spans several sessions, and one long session often spans several topics. Merge them into one line.
• The ID comes from the work, not the branch. Take it from #NNNNN in commit subjects, the digest's WIs referenced / PRs referenced lines, or the PR title. Do not use the digest's branch-derived WI field — when sessions are launched from C:\work\Payments rather than a per-WI worktree, every one of them inherits that checkout's unrelated work item. Use PR <id> as the ID when there is no work item (e.g. reviewing someone else's PR).
• State the true shipped state in the summary clause, in the fewest words: merged / PR open / committed, not pushed / local only / analysis only. A commit is not a PR; a PR is not merged; SQL written is not SQL run.
• Do not invent progress. A session with a few prompts, no Did: lines, no commits and no files is not an accomplishment — either give it an honest "explored X, no conclusion" line or leave it out.
• Omit noise entirely: q: answer-only sessions, aborted sessions, and sessions with no meaningful activity. No "N sessions skipped" reconciliation line — this is a stand-up, not an audit.
• Include the current session only if actual work happened in it.
• Add a Blocked: line at the end only if something genuinely cannot proceed without someone else. Otherwise end after the list.

Notes

• Everything here is read-only — no repo, DB or DevOps state is modified.
• The Did: extraction depends on the end-of-turn footer convention in the user's global CLAUDE.md (see docs/conventions.md in this repo). Answer-only (q:) turns and pure Q&A turns have no footer by design, so a session can be legitimately Did:-less; fall back to prompts and files for those.
• Subagent sidechain transcripts (agent-*.jsonl) are excluded — their work shows up under the parent session.