trace
/trace
Odtwarza przebieg zapisanej sesji Claude Code po UUID, nazwie lub dacie i streszcza, co zrobiono i jak wyglądał wątek.
- Autor
- proxify-dev ↗
- Repozytorium
- agentic-kit ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 26.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 13
- Skrypty
- 7
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 26.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 7 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: bin/__init__.py, bin/fetch.py, bin/status.py, lib/__init__.py, lib/gen_globals.py, lib/roster.py, lib/sources.py
Co robi i kiedy
Odtwarza przebieg zapisanej sesji Claude Code po UUID, nazwie lub dacie i streszcza, co zrobiono i jak wyglądał wątek.
Wyzwalacze z opisu
Argumenty
/trace <uuid | slug | phrase | date>
Oryginalny opis (EN)
Trace and recap a past Claude Code session by UUID, slug, date, or phrase. Prefetches dialogue + structure + delegated work + the observer's rows for grounded summaries. Use when user asks "what did we do in that session", "trace this work thread", "recap that session", "show session arc", or "link prior session context".
Jak działa
- Step 1 — Interpret the pre-fetched block
- Step 2 — Present recap
- Rules the last seventeen traces taught
- Step 3 — Offer deeper tracing on request
- Anti-patterns
Podgląd SKILL.md
---
name: trace
description: >-
Trace and recap a past Claude Code session by UUID, slug, date, or phrase.
Prefetches dialogue + structure + delegated work + the observer's rows for
grounded summaries. Use when user asks "what did we do in that session", "trace this
work thread", "recap that session", "show session arc", or "link prior
session context".
argument-hint: "<uuid | slug | phrase | date>"
---
# Trace — Session and Thread Recap
Target: `$ARGUMENTS`
The block below is pre-fetched at skill load (bash injection). Synthesize from
this block instead of re-querying tools.
<TRACE_CONTEXT>
!`[ -x "${CLAUDE_PLUGIN_ROOT}/.venv/bin/python" ] || rm -rf "${CLAUDE_PLUGIN_ROOT}/.venv"; uv run --project "${CLAUDE_PLUGIN_ROOT}" python "${CLAUDE_SKILL_DIR}/bin/fetch.py" "$ARGUMENTS"`
</TRACE_CONTEXT>
Template controls live in [`references/template-editing.md`](references/template-editing.md).
## Step 1 — Interpret the pre-fetched block
| Prefix in output | Meaning | Your action |
|---|---|---|
| `RESOLVE_NEEDED` | No session specified | Ask for UUID, slug, date, or topic phrase |
| `RESOLVE_QUERY=` + `CANDIDATES` | Phrase/slug search results | Show top matches, ask user to pick a UUID and re-run `/tracer:trace` |
| `SESSION_UUID=` (+ `TITLE=`, `INDEX=`, `CWD=`) | Session resolved | Build recap from Dialogue + Structure (+ Delegated, Artifacts, Observer when present) |
| `INDEX=no transcript on disk…` | The UUID has no transcript on this machine | Say so. Do not go looking in the indexer's source, the DB, or other project dirs — the check already covered every project dir |
| `ERROR` | Indexer unavailable | Report issue and provide manual fallback command |
## Step 2 — Present recap
Use this shape:
```
Conversation: <TITLE= if present, else the topic>
Date: <from dialogue timestamps>
Project: <cwd>
Branch: <from structure if present>
UUID: <session uuid>
Summary: 2–4 sentences about what was asked and what was decided.
Key tools/friction: notable tools, errors, or blockers from structure.
Next: what the session left open, if anything.
```
## Rules the last seventeen traces taught
- **Recap first.** The user asked what happened in *that* session. Do not
debug, build, or copy files unless their question asks for it.
- **Quote the block, or say it is not there.** Never attribute to the trace a
fact the block does not contain. No invented quotes, commands, or flags.
- **A missing section is not a mystery.** `INDEX=` says whether the transcript
exists and was indexed just now. `sessions replay <uuid>` or
`sessions show <uuid>` before any jq on a `.jsonl`.
What each section is good for:
- **Dialogue** — the last `ASST:` is the final answer *whole*; earlier ones are
500-char previews and end in ` […]` when cut. Quote the final answer's
conclusion rather than re-deriving it. `sessions replay <uuid> --full` for all.
A turn that gave no answer says so and how: `ASST: — no final text;
interrupted by the user after step 6 (Bash). last thinking: …` — that last
thinking is where the turn was heading. A local command the user ran
(`/export`, `/copy`) is one indented line under its turn, not a turn.
- **Structure** — per interaction: tools, `dur=` (how long the turn took —
the answer to "why did this take so long"), `end=interrupted@N` /
`end=tool@N` (no closing answer), `files:` (what it wrote and read, writes
first), `err[…]:` (what the errors actually were).
- **Delegated** — subagents with their final result text, workflows with
their journal path, background tasks (Bash `run_in_background`, Monitor)
with what launched them and the tail of their output. A large foreground
result Claude Code saved to disk is not a task and is not listed. This is
where the results of work the session handed off live.
- **Observer** — the observer's live rows for this session,
`#id · type · title` (forgotten and folded rows left out). "no record"
is normal for a session that has not ended: the observer writes at
session end. Supplementary — the dialogue outranks it.
- **Artifacts** — files the session wrote or edited, whether they still exist,
and the first lines of each markdown deliverable. Read the rest of a file
only if the question needs it.
## Step 3 — Offer deeper tracing on request
> Want deeper detail?
> - `sessions show <uuid>` — how the session went: time,
> tokens, cost, signals (interrupted, errors, repeats, raw reads), and each step with a
> one-line gist of what it returned
> - `trace turn <uuid>-NNN --open 1,4,5` — steps whole, in one call
> - `sessions search "<phrase>"` — sessions ranked by hits (add `--under <path>` to scope
> to a project subtree)
> - raw JSONL read only when verbatim output is explicitly requested
## Anti-patterns
- Do not call mem-search tools for UUID recap when dialogue is already injected.
- Do not read raw JSONL just to summarize.
- Do not drill every interaction when dialogue section already provides the arc.
- Do not open a step to learn what the drill line already says (size, status, timing,
the `↳` gist, a JSON result's keys). Open several at once when you must.
Pliki w paczce (13)
- SKILL.md 5 KB
- bin/__init__.py 1 KB
- bin/fetch.py 6 KB
- bin/status.py 7 KB
- lib/__init__.py 1 KB
- lib/gen_globals.py 1 KB
- lib/roster.py 14 KB
- lib/sources.py 24 KB
- references/status-module.md 2 KB
- references/template-editing.md 1 KB
- schema/report.globals.json 1 KB
- templates/report.md.j2 1 KB
- templates/report.yaml 1 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · proxify-dev/agentic-kit | 100 | MIT |
| jeremylongshore/tons-of-skills-marketplace | 100 | MIT |
| r3dlex/oh-my-auggie | 100 | Apache-2.0 |
| zereight/gitlab-mcp | 100 | MIT |
| biliboss/skills | 100 | MIT |
| oliver-kriska/claude-elixir-phoenix | 100 | MIT |
| moonklabs/aiwf | 100 | MIT |
| Soul-Brews-Studio/arra-oracle-skills-cli | 100 | MIT |
| ccplugins/awesome-claude-code-plugins | 100 | Apache-2.0 |
Migawka bazy z 13.11.2026 · wersja paczki a87d7b9f27a6 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100recall-session
Odnajduje wybraną wcześniejszą sesję Codex po nazwie lub UUID i streszcza jej przebieg. - 100resume-claude-session
Przejmuje pracę z sesji Claude Code po nazwie lub identyfikatorze, gdy użytkownik prosi Codexa o kontynuację. - 100session-recap
Podsumowuje sesję Claude Code na życzenie: co się wydarzyło, jakie były wyniki i co dalej. - 100recap
Odtwarza, co agent Claude Code lub Codex faktycznie zrobił w jednej sesji: cele, decyzje, zmiany w plikach, testy, błędy i stan końcowy. - 100find-session
Znajduje dawną sesję Claude Code po zgłoszeniu, PR, pliku lub temacie i pokazuje, gdzie praca faktycznie się odbyła. - 100dig
Przeszukuje historię sesji Claude Code i pokazuje oś czasu, luki, przypisanie sesji do repozytoriów oraz to, nad czym pracowano.