Evening

/evening

Produce an honest end-of-day accomplishment log. The sibling of /morning: morning plans the day forward, evening proves the day backward. The user often invokes this feeling like the day was wasted — the job is to show the receipts, not to flatter.

Mode Detection

Unlike /morning, global is the default — the end-of-day question is usually "what did I do overall". In global mode, still pass the current project's stream first if it had activity.

Workflow

Step 1: Gather

Reuses the /morning gather script (sibling skill in the same repo) with the today range:

node <skills-repo>/morning/scripts/gather-context.mjs --mode <repo|global> --range today [--project <cwd>]

Supplement with anything visible in the current conversation (work done in this very session won't be in the session index yet — include it from context).

Each CC session reports a Tools line (e.g. Edit×12, Write×3, Bash×40, Agent×2, Skill×1) — read it as receipts:

• Edit/Write/Bash counts → Shipped (files changed, commands run) even when the prose is thin
• Agent/Task counts → directed-agent work (directing is doing)
• Skill/mcp__* counts → tooling/infra/automation, the Invisible work the "did nothing" feeling erases This is the hard signal that turns a "felt unproductive" day into a list of artifacts.

Step 1b: The hours

node <skills-repo>/evening/scripts/compute-hours.mjs [--project <cwd>] [--exclude <this-session-id>] [--json]

Always pass --exclude with the current session's id — writing the recap is not work the day should be billed for. Get the id from the transcript path in your own context; skip the flag only if you genuinely can't find it, and say so.

The script slices the day into quarter-hours, counts a quarter-hour as active for a session when it emitted at least one record inside it, then shares each one out among whatever was live in it. Two splits come back: equal shares, and shares weighted by how busy each session was. They should land close together — a line where they diverge sharply is one to describe rather than quote. Both sum to the day's real active wall-clock, so they can be logged as-is.

Per-session output gives you the active windows, event count, cwd branch, and how much of the time was background agents. Sessions are not streams: fold several sessions into one stream when they're the same piece of work, and sum their split hours. Do not sum the solo column — it double-counts overlap on purpose.

Read the numbers as a floor. Time on a phone, a build the user ran themselves, and thinking away from the keyboard are all invisible to this. Say so once rather than presenting the total as measured fact.

Step 2: Synthesize the receipts

Group into a timeline of work streams, then for each stream pull out:

• Shipped: commits, closed beads issues, merged branches — the hard receipts
• Decided: decisions made, approaches settled, things de-risked (a settled architecture question is output even with zero diff)
• Invisible work: triage, planning, issue filing, tooling/skills/infra, seeds, env setup, reviews of agent output — name it explicitly; this is the work the "did nothing" feeling erases
• Open loops: what's mid-flight, parked, or handed to an AFK instance — with where it stands

Step 3: The verdict

End with 2-4 sentences of honest day assessment:

• Count the receipts ("4 commits, 3 issues closed, 2 filed, 1 skill shipped")
• Use the hours to explain the day's feel, not to grade it — several threads sharing the same quarter-hours is why a full day can read as a blur, and that's a scheduling fact, not a failure
• Name the day's shape ("scattered morning, locked-in afternoon" is a pattern, not a failure)
• If the day genuinely was thin, say so plainly and point at the single most useful thing it produced — don't inflate
• If meta/tooling work dominated, say why it counts (it compounds; tomorrow is faster)

Rules

• Receipts over vibes: every claim links to an artifact (commit hash, issue id, file, doc)
• Chronological-ish, but group by stream, not by hour — nobody needs minute-level accounting
• Don't moralize about distraction gaps; mention them only if the user asks where time went
• Include work done by directed agents/instances today — directing is doing
• If context is thin, say so honestly rather than padding
• Keep it scannable: short sections, bold the artifacts

Output Format

See references/output-format.md.