Back to Gitbutler

GitButler Agentlog Skill

crates/but-agentlog/skill/SKILL.md

0.20.06.3 KB
Original Source

GitButler Agentlog Skill

Use but agentlog to skim prior agent work from captured sessions.

Trigger this skill for plain prompts like "get context for branch", "branch context", "catch up on this branch", "recover context", "what happened on this branch", or "what prior agent work happened here".

If both the general GitButler CLI skill and this skill seem relevant, use this skill for agentlog discovery. Use the general GitButler skill only for version control operations like commit, push, branch, or diff.

Core Workflow

Do not search for or read local SKILL.md files. If this skill is active, use these instructions directly. Do not inspect the general GitButler skill first. Start with but agentlog skim.

Start with skim when prior work could affect your next action. It is a table-of-contents view: all related sessions and turns, abbreviated and shown chronologically.

sh
but agentlog skim
but agentlog skim branch <branch-name-or-ref>
but agentlog skim review <review-id-or-pull-request-key>
but agentlog skim change <change-id-or-key>

Use show only when skim is too thin, ambiguous, or you need exact evidence:

sh
but agentlog show <session-key> --limit 20
but agentlog show <session-key> --turn <turn-key> --limit 20

For normal current-branch recovery, run skim without target arguments first. It discovers the applied GitButler branch itself. Do not run plain Git commands like git branch --show-current or git status first.

sh
but agentlog skim

Pass an explicit target only when the user gives one.

Use JSON only when you need exact handles for drill-down:

sh
but --json agentlog skim
  1. Start with skim for a clean orientation.
  2. Treat skim like compacted turn history: all related sessions and turns are present, but each turn is abbreviated. It is not the full transcript.
  3. If skim is enough for a lightweight status answer, summarize it and stop.
  4. Drill down only when skim is ambiguous, misses the rationale, or you need exact evidence.
  5. To drill down, rerun skim with --json to get the relevant session_key and turn_key, then use show.
  6. Use show <session-key> for turn-level context.
  7. Use show <session-key> --turn <turn-key> only for turns that need exact detail.

Target Discovery

When the user says "current branch" or gives no explicit target, run:

sh
but agentlog skim

skim performs GitButler branch discovery automatically. If target discovery fails, then inspect GitButler state:

sh
but status

Do not use plain Git's gitbutler/workspace branch as an agentlog target.

Rules

  • Prefer orientation before payload.
  • Prefer skim for turn-history recovery.
  • Treat skim as complete but abbreviated. It includes every related session and every turn in those sessions, not every record or the full transcript.
  • Prefer human skim output first. Use --json only for drill-down handles or exact evidence.
  • Run show when skim is thin, ambiguous, missing the why, or when the user is asking you to make or verify a consequential claim.
  • Do not dump records for every candidate session.
  • Do not dump full records or transcripts. If session detail is needed, start with show <session-key> --limit 20 and summarize previews.
  • Treat matches as related evidence, not ownership.
  • Keep session_key and turn_key values internal unless the user asks for evidence handles or they are needed for a follow-up command.
  • Use record snippets only when they justify the next action.
  • Say "session related to branch X via turn Y", not "session for branch X".

Reading skim

Use skim to answer "what has already been discussed, planned, attempted, or changed in this work?"

Human output is intentionally compact but includes every related session and every turn in chronological order. JSON output keeps full session_key and turn_key values so you can drill into show without printing those handles to the user.

The skim is a compressed table of contents. It abbreviates each turn but does not omit related sessions or turns.

Useful fields:

  • target_kind and target_key: target used for discovery.
  • sessions: related sessions in chronological order, with counts and previews.
  • coverage: shown sessions, shown turns, and direct related turn count.
  • sessions[].turns: every turn in order, with labels and previews.

If skim has enough context for a lightweight answer, summarize from it and stop. If it is ambiguous or misses the why, use JSON skim handles and drill down with show.

Reading show

Use show to open a session or one turn.

sh
but agentlog show <session-key> --limit 20
but agentlog show <session-key> --turn <turn-key> --limit 20

Without --turn, show answers what happened in a session at turn granularity.

Useful fields:

  • coverage: returned turns versus total turns.
  • turn_key: handle for show --turn.
  • turn_index: order within the session.
  • capture_kind: backfill or incremental.
  • record_count: turn size before hydrating records.
  • source_record_index_range: provider/source indexes covered by the turn.
  • observed_targets: targets observed in the turn.
  • latest_user_preview and latest_assistant_preview: compact orientation.
  • tool_counts: tool call/result counts and tool names.

Increase --limit only when the current window misses the related turn or setup.

With --turn, show answers what exactly happened inside one turn.

Useful record fields:

  • coverage: returned records versus total records in the turn.
  • turn_record_index: order within the turn.
  • source_record_index: original provider/source index when available.
  • timestamp, kind, role, text: message/tool orientation.
  • tool_name and tool_input: tool details when present.
  • source_record: redacted stored provider envelope for low-level debugging.

Turn records are intentionally bounded. Increase --limit only for a specific turn after session-level show output proves that more detail is needed.

Skim Summary

When summarizing a skim, do not say you recovered "full context." Say you skimmed the full turn history in abbreviated form.

Include:

  • target used for discovery
  • previews explaining intent and outcome
  • minimal record snippets for the next action

Do not lead with raw session or turn identifiers. Mention them only when the user asks for evidence handles or when passing them to another command/tool.