Skill v1.0.1
currentAutomated scan100/100+4 new
version: "1.0.1" name: ccx-recap description: "Recap what a coding-agent session or time window actually did from ccx evidence: story, decisions, verified vs merely claimed, mistakes, cost. Use for 'what happened', 'what did it do', 'recap', 'summarize the session', 'what did I work on today/this week'. For 'what went wrong', use ccx-retro."
ccx-recap
One question: what did the agent actually do? Answer it with receipts, not vibes. The agent's own final message is a claim, not evidence — the recap is built from the trace.
Scope
| Trigger | Evidence source | |
|---|---|---|
/ccx-recap | Latest session here: ccx trace | |
/ccx-recap <session-id> | That session: ccx trace <id> | |
/ccx-recap today / week / month | The whole fleet: ccx insight --scope <s> --all --json (every workspace's sessions digested), then ccx trace per session that decides the story |
Timezone: preserve the user's. Warn when a --tz offset makes "today" differ from their local day (e.g. --tz +8 from a US machine).
This skill describes ccx >= 0.8. If a command or flag here is missing (unknown flag: --turn), the binary is older than the skill: upgrade ccx, then run ccx skills install — the binary embeds skills matching its own CLI surface.
Process
- Outline first.
ccx trace --json(orccx tracetext) — every
turn and step headline with tools/edits/errors/cost. This always fits; read it whole. Steps are the agent's own narration at the moment it acted — that sequence IS the story skeleton.
- Pick what matters. Turns with edits, errors, linked commits,
high cost, user pushback, or human interventions (interrupts — the human pressed stop; denials — a tool call refused at the permission prompt; both are on turns, steps, and stats). Ignore command noise (is_command).
- Drill only there.
ccx trace --turn Nfor full narration,
mutations, and error attribution. Raw text when needed: ccx view <session> --at <message_id> opens the cited message in context. When the story crosses sessions (a handoff picked up, a fork, a parallel agent), ccx related <session> says which sessions and how, with evidence — cite those anchors, do not guess the link.
- Verify before claiming. "Done" requires evidence: a passing
test step, a linked commit, a verification narration. Tag every important claim:
observed— tool output / git / trace shows itagent-claimed— the agent said it; nothing independent checkedunverified— plausible, not checked
- Write the recap. Conversation reply first; offer an HTML
artifact for sessions worth keeping.
- Land it. A recap that ends at a chat reply evaporates — the
next session re-pays the full archaeology. Offer to write the distillation into whatever durable store the user already runs (devlog, memory file, wiki/vault page; never invent a new store). Keep the claims inline and cite the session as provenance: [ccx:<session-id> #turn.step] — anyone can re-open the receipt with ccx trace <session-id> --turn N. If the user declines or has no store, the reply is the deliverable.
The 60-second test
A reader must know, within a minute: what was asked, what the agent did about it, the 3-5 decisions that shaped the work, what is verified vs claimed, what it cost, and what is still open. If the recap can't produce that paragraph, it is not done.
Format (conversation or HTML):
Recap — <session/scope> | <duration> | $<cost> | <files> files | <errors> tool errorsWhat happened 3-8 sentences, chronological, cited [#turn.step]Decisions the real ones only, with the rejected alternativeVerified vs claimed anything important that lacks evidence, flaggedStill open unfinished work, unverified claims, warningsNumbers small table: turns, steps, edits, errors, cost
Citations are #turn.step (e.g. #133.9). Copy trace warnings into the recap — evidence gaps are part of the answer.
Multi-session (today / week / month)
The subject is everything the human did with agents in the window — every workspace, every provider — not the repo you happen to be in. ccx insight --scope <s> --all --json (ccx.insight.v1) is the whole evidence set and it fits: read it whole before deciding what matters.
metricsandworkspaces[]say where the work went and what it
cost; quote cost with cost_coverage (an unpriced model is not a free one).
sessions[]carry each session's humanprompts[](intent, cited
by session:line), final_answer (agent-claimed), edits, commits, interrupts/denials, cost_usd + cost_status, and relation (a session that started before the window is a long-running container, not this window's work).
- Group sessions into workstreams by workspace and prompt content,
not by session count. Then ccx trace <id> only the 3-5 sessions that decide the story (cost, edits, commits, interventions, or the human's stated interest) and verify their claims.
- Write the recap per workstream: what was asked, what the agent
did, verified vs claimed, still open. Never re-bucket ccx log records yourself; never let one workspace stand in for the fleet.
For a human review surface, write the HTML cockpit described in HTML-REPORT.md: judgment on top of the digest, into the insights dir so it shows at /insights in ccx web.
Hard rules
- No judgment without a citation. No citation, say "unverified".
- Do not paste transcripts; the recap is the distillation.
- Numbers come from the trace/insight, never estimated.
- The final assistant message is input, not truth: check it against
the evidence like any other claim.