Get back into the flow of a session in one glance. A read-only ORIENTATION report — not what-did-we-do-and-was-it-good (that's /summarize) but where do things stand RIGHT NOW: what work was done, where the branch is (git), where the ticket stands (Linear), what's blocked on YOUR decision, and what's next. The orchestrator writes its own live bearings, then a subagent scans the raw transcript JSONL (dropping read-noise, KEEPING every decision/state-bearing call — AskUserQuestion, Skill/Task, plan-accept, commit/ticket/PR), reconstructs the state BLIND, flags stale narrative artifacts, and reconciles against the bearings, transcript-authoritative. Reuses /graph to render the work trajectory (only when it actually branches) — phases, divergences, snapshots, milestones, plan completion, where you are now, what's upcoming — then walks you through every open thread. Saved to builds/ by default; degrades gracefully (no ticket/plan/thin transcript → report whatever exists). Dual audience: a human returning cold, and
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Get back into the flow of a session in one glance. A read-only ORIENTATION report — not what-did-we-do-and-was-it-good (that's /summarize) but where do things stand RIGHT NOW: what work was done, where the branch is (git), where the ticket stands (Linear), what's blocked on YOUR decision, and what's next. The orchestrator writes its own live bearings, then a subagent scans the raw transcript JSONL (dropping read-noise, KEEPING every decision/state-bearing call — AskUserQuestion, Skill/Task, plan-accept, commit/ticket/PR), reconstructs the state BLIND, flags stale narrative artifacts, and reconciles against the bearings, transcript-authoritative. Reuses /graph to render the work trajectory (only when it actually branches) — phases, divergences, snapshots, milestones, plan completion, where you are now, what's upcoming — then walks you through every open thread. Saved to builds/ by default; degrades gracefully (no ticket/plan/thin transcript → report whatever exists). Dual audience: a human returning cold, and the agent itself getting its bearings. A building block: it orients and reports, never touches code/git/ticket content. Triggers: "/report", "where are we", "where do things stand", "what's the state of this session", "get me back in the flow", "orient me", "where did we leave off — on STATE, not on what-we-did", "I'm lost — catch me up on the state".
version
1
tier
lightweight
args
[-- <what you most want re-oriented on>]
Get back into the flow of a session in one glance. /report is the present-tense orientation sibling of the workflow family. Where /summarize reports backward (what did we do, was it good, interrogate it) and /session status prints thin mechanical state (.state.json phase, lifecycle), /report answers the question you actually have when you return to a window cold: "where do things stand right now, what's waiting on me, and what's next?" — with a rendered state graph.
It is read-only (¶INV_REPORT_READ_ONLY) and sessionless, and runs within the active session: locate the transcript → write the orchestrator's own bearings → spawn a transcript-scanning subagent (file handoff to builds/) → reconcile → graph → render → walk the open threads → save. <trailDir> = <sessionDir>/builds/.
Dual audience. The report serves two readers at once: a human returning after a break (or after /clear — the session persists and is reactivated), and the agent itself re-establishing its bearings. Both need the same thing — the current state, grounded in what actually happened, not a hopeful summary.
Hard boundaries. Not /summarize (never reviews work quality, runs the inverse-ask palette, or dwells on the past beyond what establishes the present). Not /snapshot (never commits, posts to a ticket, or re-runs a gate). Not /session (a rich re-orientation, not a state dump). A bug or risk it notices is an open-thread line item, not a repair.
/report Protocol
1. Locate the transcript & resolve targets
Establish the ground-truth sources, degrading gracefully (¶INV_REPORT_GRACEFUL) — a missing source narrows the report, never aborts it.
The active session (<sessionDir>): its *_LOG.md, plan (if present), DIALOGUE.md, and builds/.
The raw transcript JSONL (<transcript>) — the subagent's primary source (§3). Resolve it deterministically from the session id Claude Code exports (¶INV_REPORT_TRANSCRIPT_IDENTITY) — never by newest-mtime, which under multiple open windows returns a sibling window's file:
If ends up empty, the subagent falls back to + (already tool-call-free) — thinner, but never a hard fail, and never the wrong window's transcript. The orchestrator resolves the path and passes the substituted absolute path to the subagent.
# Fallback ONLY if the env var is unset or the file is missing:
if
"$CLAUDE_CODE_SESSION_ID"
"$TRANSCRIPT"
then
ls
"$HOME/.claude/projects/$SLUG"
head
# Mechanical verify: the tail must contain THIS session's dir, else it's the wrong window.
tail
"$TRANSCRIPT"
"$(basename "$sessionDir")"
""
fi
<transcript>
DIALOGUE.md
*_LOG.md
The ticket(s) — derive the Linear key from the session's tickets[] (.state.json) or the branch name (fin-2833-… → FIN-2833). Read-only. Sanity-check relevance: a branch-derived key may be unrelated to this session's actual work (e.g. an engine-tooling session on a feature branch) — flag low-relevance rather than presenting it as the ticket. No key → skip the section.
Git — the repo(s) the work actually landed in (¶INV_REPORT_GIT_FOLLOWS_WORK). Do NOT assume cwd-git is the work's repo. Report cwd git, AND any other repo the session actually modified (e.g. an engine/skill session edits ~/.claude/engine — a separate repo; cwd-git is then unrelated). For each: branch, dirty-file count, last commit; ahead/behind vs base only if an upstream/base is set (guard it). Use git log --oneline -n 10 (safe on young branches — never HEAD~10.. which errors under 10 commits).
Echo one line:Orienting on <session "<slug>"> — transcript <id|fallback>, ticket <KEY(rel?)|none>, work-repo <path>. Building bearings…
2. Orchestrator's own bearings
Write down what you (the orchestrator) currently believe the state to be, from your own in-context memory. Do it now so it's captured — but it is a hypothesis to be checked, not the answer (¶INV_REPORT_RECONCILE). Keep it tight (you hand it to the subagent in §3): the goal as you understand it, what you think was just done, what you believe is pending, and — honestly — where you're unsure or may have lost the thread. On a cold start with little context, say so plainly; the transcript scan then carries the load.
3. Spawn the transcript-scanning subagent (file handoff)
Before dispatching — §CMD_LOG_SKILL_INVOCATION: log this dispatch to the session log (why + context-pack pointer + one-line re-tread) so a restarted session can re-tread it. Fire it as the last step before the Task/Agent handoff.
Spawn one subagent (general-purpose or analyzer) to reconstruct the state from <transcript>. File handoff (¶INV_REPORT_FILE_HANDOFF): it WRITES its full digest to <trailDir>/REPORT_DIGEST.md and RETURNS only a 6–10 line summary — keeping the orchestrator lean and leaving a durable artifact beside the other builds/ files. Run it foreground (§4–§8 need it).
Guard: if neither <transcript> nor DIALOGUE.md/*_LOG.md exists, skip the subagent — the orchestrator does git + ticket directly.
Construct a fully self-contained prompt:
ROLE: You are a session cartographer. Reconstruct the TRUE current state of a work session from its raw transcript. You READ and RECONSTRUCT; you never mutate, commit, post, fix, or re-run a gate.
PRIMARY SOURCE — the raw transcript: Read <transcript> (JSONL, one message/line). Filter as you read (¶INV_REPORT_TRANSCRIPT_FILTER) — the cut line is decision/state-bearing vs read-noise, NOT "only AskUserQuestion":
KEEP every user text message and every assistant text block (reasoning, plans, claims).
KEEP, unfiltered, all decision/state-bearing tool calls + results: AskUserQuestion (both sides — the decision record), Skill/Task invocation headers (name + one-line input + one-line result — these are the major moves: /build, /snapshot, /experiment, /ticket, /pr), ExitPlanMode (plan acceptance), and any commit/ticket/PR-producing call + its result (git commit, linear save_*, gh pr create → the SHAs / keys / URLs the narrating text often omits).
DROP read/inspection noise: Read, Grep, Glob, read-only Bash, and engine log appends. Note that such work happened from surrounding text; don't ingest the payloads.
If <transcript> is absent/empty, fall back to DIALOGUE.md + *_LOG.md and say so.
RECONSTRUCT BLIND FIRST (¶INV_REPORT_RECONCILE): Derive your State line, Work-done, Pending, Open-threads, and Graph-inputs from the transcript before you read the orchestrator's bearings. Do not read the bearings block below until your independent reconstruction is written. This prevents anchoring on a possibly-wrong narrative.
STALENESS CHECK (always run — ¶INV_REPORT_STALENESS): Compare the mtime of every narrative artifact (the debrief and other non-log *.md, the plan, and — if fetched — the ticket/PR body) against the timestamp of the last substantive log entry / commit. Flag any narrative artifact that predates the last real work as stale (stat -f %m on macOS). "Artifact X was written before verdict/work Y" is consistently the single highest-value finding — surface each stale artifact in Open threads, and in Reconciliation deltas when it changes the state read.
ALSO READ (state, not narrative), read-only: git for every repo the work touched (cwd, and e.g. ~/.claude/engine for engine work — label which holds this session's work and whether committed) — git -C <repo> status --porcelain + git -C <repo> log --oneline -n 10; the plan file's [x]/[ ]if one exists (many skills have none — skip silently); the Linear ticket via MCP only if a relevant key is given.
THEN — and only then — read the orchestrator's bearings and diff:<paste the §2 bearings note verbatim>
Where the transcript contradicts or outruns the bearings, the transcript wins; capture each mismatch in the deltas section.
WRITE the full digest to <trailDir>/REPORT_DIGEST.md with these sections, then RETURN a 6–10 line summary (state line + the headline finding + counts):
State line — one sentence: where the work stands right now.
Work done — 3–6 substantive things, most recent first.
Git state — per repo, labeled; which holds this session's work; committed or not.
Ticket state — status + most relevant point, or "none / branch-derived (unrelated)".
Pending on the user — decisions/questions explicitly awaiting the human (mine unresolved AskUserQuestion exchanges + direct asks).
What's next — immediate next actions, ordered.
Open threads / dangling — unanswered questions, #needs-* tags, deferred items, stale artifacts, "come back to X" — each with a one-line disposition hint.
Graph inputs — ordered trajectory nodes (milestones, phase/branch divergences, snapshots=commits, decision points, plan steps) each marked done/current/upcoming; AND a one-word verdict branches | linear on whether the trajectory actually forks (drives §5's gate).
Reconciliation deltas — where your (transcript-first) reconstruction differs from the bearings, including any stale-artifact corrections.
4. Reconcile & assemble
Read <trailDir>/REPORT_DIGEST.md. The subagent's transcript-first reconstruction is authoritative on facts; your §2 bearings contribute only unstated intent (what you were about to do) — never override a transcript fact with a bearing. Surfacing a reconciliation delta (§3.9) is often the single most useful line for an agent that had lost the thread.
5. Render the flowgraph — reuse /graph, gated on real structure
Gate first (¶INV_VISUALIZE_STRUCTURE_WITH_GRAPH): render a flowgraph only when the trajectory actually branches / loops / diverges (digest §8 verdict = branches). If it's linear or thin (degraded path → a flat commit list), skip the graph — the ordered "What's next" list is clearer; a flowgraph of 1→2→3 is noise.
When it does branch, invoke /graphinline via the Skill tool (tell it to render once, no revision prompt) over the digest's graph-inputs, as a trajectory graph in the §CMD_FLOWGRAPH Status & Trajectory vocabulary — never freehand glyphs (§INV_VISUALIZE_STRUCTURE_WITH_GRAPH):
done → ✓; upcoming → ○; dropped/cancelled → ✗; stale → ⚠ (pairs with the staleness check);
current position → ◄ HERE (exactly once); snapshots → ▣ <sha>; pivots → ╰► ⟨branched-off⟩;
decisions → ◆ with the outcome annotated (◆ … → chose X / ◆ (OPEN)).
Wrap in a code fence.
6. Render the report
Render the assembled report in chat (graph included when §5 rendered one), using the shape in assets/TEMPLATE_REPORT.md (the skill's base dir — do not hardcode ~/.claude). Match the project's prose density and stick to the template's plain section headers — no decorative emoji unless the template itself carries them (the project leans terse-and-plain; do not freehand decoration on top of the shape). Lead with the state line and a compact state strip (branch · work-repo committed? · dirty · [plan x/y — omit if no plan] · open-threads), then the reconciliation deltas (near the top — it's the payload), then: work done → git (per repo) → ticket → the flowgraph (or ordered next-list if gated out) → pending-on-you → what's next → open threads.
7. Walk every open thread
Walk the user through each open thread / pending decision (digest §5 + §7) for disposition — the thorough close is what turns a report into being back in the flow. Batch via AskUserQuestion (4 per batch, §INV_BATCH_SIZE_4). Per thread offer disposition: address now / defer (tag) / dismiss / note & move on — and always honor an "explain first" request (a returning human often needs the why before deciding, not just fix/defer/dismiss): explain the thread's stakes, then re-present its disposition. "Address now" = route it (resume the work, or hand to the owning skill) — never fix code inside /report. "Defer (tag)" applies an orientation #needs-* bookkeeping tag (permitted by ¶INV_REPORT_READ_ONLY's scope — it touches no code/git/ticket content). Zero open threads → say so and skip to §8.
8. Save (default) & stop
Save the report by default (¶INV_REPORT_SAVE_BY_DEFAULT) — it's a durable artifact, not an ephemeral glance. Write it to <sessionDir>/builds/REPORT.md, latest-wins (overwrite on re-run). The digest already sits beside it at <trailDir>/REPORT_DIGEST.md. Report the saved path (clickable), then stop.
Constraints
¶INV_REPORT_READ_ONLY — /report never mutates code, git, or ticket content: no commits, no ticket posts, no code edits, no gate/test re-runs. Permitted: writing its own report + digest under <trailDir>, and applying orientation #needs-* bookkeeping tags during the §7 thread-walk (session-state bookkeeping only). A bug or risk it notices is an open-thread line item, not a repair.
¶INV_REPORT_ORIENTATION_NOT_REVIEW — Forward/present-facing: where are we, what's blocked, what's next. Disjoint from /summarize — it does NOT judge work quality, run goal-vs-actual retrospection, or offer the inverse-ask palette. "Was this good / interrogate what we did" → route to /summarize. Past events appear only insofar as they establish the present state.
¶INV_REPORT_TRANSCRIPT_IDENTITY — Resolve the transcript by session identity ($CLAUDE_CODE_SESSION_ID), never by newest-mtime. With multiple windows open, newest-.jsonl is routinely a sibling session; the mtime path is a last-resort fallback and must be verified against this session's dir before use.
¶INV_REPORT_TRANSCRIPT_FILTER — The subagent drops read/inspection noise but retains all decision/state-bearing calls: AskUserQuestion (unfiltered), Skill/Task invocations, ExitPlanMode, and commit/ticket/PR calls (with their SHAs/keys/URLs) — plus all user + assistant text.
¶INV_REPORT_RECONCILE — The subagent reconstructs from the transcript blind first, then reads the orchestrator's bearings only to diff for the deltas section. Transcript-authoritative on facts; bearings add unstated intent only. Reconciliation deltas are surfaced, not hidden.
¶INV_REPORT_STALENESS — The subagent always checks narrative-artifact mtimes against the last substantive work and flags any that predate it (a debrief/ticket/PR body written before the last real work). Stale-narrative is a recurring highest-value finding — name it, don't hope it emerges.
¶INV_REPORT_FILE_HANDOFF — Subagent→orchestrator handoff is via file: the digest is written to <trailDir>/REPORT_DIGEST.md (returning only a short summary), and the final report is saved to <sessionDir>/builds/REPORT.md. Both are durable artifacts beside the other builds/ files.
¶INV_REPORT_SAVE_BY_DEFAULT — The report is saved by default (not offered/ephemeral), latest-wins.
¶INV_REPORT_GIT_FOLLOWS_WORK — Report git for the repo(s) the session actually modified, not cwd by assumption; cwd-git can be entirely unrelated (e.g. engine work under ~/.claude/engine).
¶INV_REPORT_GRACEFUL — Degrade, never hard-fail. No ticket → skip. No plan → omit plan segment. No transcript → DIALOGUE + log; none of those → skip the subagent, orchestrator does git + ticket. Young branch → -n 10, guarded ahead/behind. Always render whatever exists.
Subagent for the transcript read; interactive walk stays with the orchestrator.
