Log Session

Capture the why and the what-next of a working session — the context that git history alone never records. A commit says what changed; a session log says what you were trying to do, what you decided, and where to resume.

Deliberately cheap (qwen3.6/reasoning_config: none — unlimited; the NaN cluster models carry no token counter, so this skill pays in bytes, not quota): structured summarization, not judgment. It must never reach for an expensive model.

Turn contract — verify before ending the turn

✓ The entry was WRITTEN with accurate git facts: appended to docs/LOGS.md, or — when the fragment carrier `.agentic-workflow/journal/` exists (issue #405) — written as a new fragment there (the aggregator owns the journal; a direct append is refused)
✓ `bun scripts/session-close.mjs close` RAN and exited 0 (or 2, with the leftover paths named) — the entry is committed, not left riding along
✓ No past entry was edited
✓ Artifact language: explicit user instruction > the project's declared docs language > English. The CONVERSATION language never decides — a Spanish prompt still produces English PRs/issues/commits/SPECs unless one of the first two says otherwise
✓ The closing `→ Next:` block is printed as the ABSOLUTE last output

About to end the turn with any box unchecked? The turn is NOT done — complete the missing box first (weak models drop end-of-document duties; this list is first on purpose).

Two doors, one behaviour (#429)

With @gtrabanco/pi-agentic-workflow loaded, /log-session is a command: the handler records the invocation (.pi/.session-<id>.log-session), runs the writer below and names the entry's exact target before this file reaches you; a settled turn with no entry there is reported. The record is code's fact, the prose is yours. Without the extension, or via /skill:log-session, every step below is yours.

When to use

• Before /clear (or your agent's context-reset equivalent) — you're about to wipe context; capture it first.
• Before closing your agent for the day.
• At a natural stopping point — a feature paused mid-way, a thread you want to be able to resume cold.
• After a long session with several decisions worth remembering.

This is the manual, rich counterpart to the lightweight auto-logging hooks the template/ ships (see template/.claude/). The hooks capture the mechanical facts for free on every exit; this skill writes the thoughtful entry when you want one.

Step 0 — Discover the project (always first)

Per the agent guide's Workflow conventions + documentation map, locate the session log. Default path: docs/LOGS.md. If the documentation map names a different location, use that. If no log file exists yet, create it from the shape in template/docs/LOGS.md (a short header + the entry format below) and say so.

If a session-start marker exists (.claude/.session-*.start, written by the template's SessionStart hook), read it to bound the session precisely — it holds the HEAD sha and start time at session open.

Process

1. Establish the session boundary.
   • If a marker is present, the session spans marker_sha..HEAD plus any uncommitted work.
   • Otherwise, use your own conversation context as the source of truth for what happened this session, and corroborate with git log of the recent commits and git status / git diff --stat for uncommitted changes.
2. Gather the mechanical facts (cheaply, with git):
   • branch (git branch --show-current),
   • commits this session (git log --oneline <since>..HEAD),
   • files touched (git diff --stat <since>..HEAD + uncommitted).
3. Write the narrative — the part only you can add:
   • Summary: 1–3 sentences on what this session set out to do and what actually got done.
   • Decisions: the non-obvious choices made and why (the rationale that would otherwise be lost). Link related docs/issues.
   • Next: the concrete next step(s) — the command to run, the unfinished thread, the open question. Write it so a cold reader knows exactly where to resume.
   • Status words are forge-verified. merged, closed, approved — any claim about a PR/issue state is checked with the forge in the same turn (gh pr view <N> --json state,mergedAt) before it is written. A log that records a state the forge contradicts is a false record that the next review must chase (fix #157's F14). Write the actual state: "PR #N open, awaiting the human merge".

4. Render and write the entry — the mechanical facts (timestamp, branch, commit count and range, files touched) are computed for you:

   bun scripts/session-close.mjs render --summary "<what this session did>" \
     --decisions "<key choices + why>" --next "<the concrete next step>"

   It writes nothing: apply its output where the project opted in —

   • No fragment carrier (the default): append to the end of docs/LOGS.md (append-only, chronological) with your editing tool.

   • Fragment carrier (.agentic-workflow/journal/ exists — issue #405): hand the entry to its writer, never to docs/LOGS.md (the aggregator owns that file):

     agentic-workflow journal write --kind narrative --summary "<summary>" \
       --decisions "<choices + why>" --next "<next step>" [--session <id>]

     It computes the timestamp, branch and name, renders the block, refuses bytes the aggregator could not read, and never overwrites — the filename is never composed here. Grammar: journal-fragment-grammar@1 + journal-write-grammar@1 in @gtrabanco/agentic-workflow's README; long prose: … render … | … journal write --kind narrative --content -.

   Omit --decisions / --next when there is nothing real to record — the script drops the line rather than emitting an empty one. A user-supplied note is prepended to the Summary.

5. Close the session — commit the entry and prove the tree is clean:

   bun scripts/session-close.mjs close

   It refuses unless the entry was written the way the project opted in — docs/LOGS.md appended to (a rewritten or truncated log is not a session entry), or, in fragment mode, new fragments added under .agentic-workflow/journal/ with docs/LOGS.md untouched — commits that surface alone as docs(log): <the appended summary>, and reports every path still uncommitted: exit 0 clean, 2 the named leftovers are still dirty, 1 refused. Do not open a PR just for a log entry, and never git add -A here — a concurrent conversation's work must not be swept into a log commit.

The canonical split — this skill's half (feature 66, D2)

Two writers exist on purpose; neither mirrors the other:

• docs/LOGS.md (this skill) is the human/team journal — what a session set out to do, what it decided, where to resume. It travels with git, so a teammate or a fresh clone reads it with no local state.
• Engram (mem_session_summary, and the lane's unit/<NN-slug>/doc topic) is local decision and gotcha recall keyed by unit id, written by unit-lane and execute-phase at unit boundaries. It stays on the machine.

Engram never mirrors this log, and this log never mirrors Engram: team memory travels with git, agent recall stays local. Write the journal half here; the protocol half (boundaries, recall order, degradation) lives in the lane's memory reference, ../unit-lane/references/MEMORY.md.

Guardrails

• Never use an expensive model for this. It's summarization; qwen3.6 (reasoning_config: none → zero thinking tokens) is the natural ceiling on the cluster plan, and the hooks do the free mechanical version.
• Don't invent facts. Decisions and next steps come from the actual session (your context + git), not plausible-sounding filler. If a section has nothing real, omit its line rather than padding.
• Append, never rewrite. Past entries are a historical record — don't edit or "tidy" them. Same reasoning as not rewriting a delivered feature's planning artifacts.
• One file, append-only. Don't fan session logs across files.

Portability (agents other than Claude Code)

The workflow is the contract; Claude Code features are conveniences. On an agent that lacks one, apply the fallback — never skip the step the feature enables:

• No Claude Code hooks — the template's auto-logging hooks (SessionStart marker, SessionEnd entry) don't run on other agents, so there is no marker file and no free mechanical entry: this skill is the only journal writer. Run it before ending every session, and bound the session with git alone (Process step 1's no-marker path).
• No /clear — read it as your agent's context-reset / new-conversation equivalent, in the triggers and in the closing block alike.
• No per-skill model:/effort: — the intent stands: use the cheapest available model for this. It's summarization, never judgment. On NaN the cluster models (qwen3.6, gemma4) carry no token counter — set reasoning_config: none so the append pays zero thinking tokens.

Relationship to other skills

• Complements the template/'s SessionEnd hook (free, mechanical, automatic on /clear and exit) and the opt-in SessionStart restore hook (re-injects the last entry to resume context). This skill is the rich, manual entry.
• Pairs with execute-phase's unit doc ## Progress log: that tracks progress within a unit; the session log is cross-cutting — whatever you touched this session, across units.
• Not part of the PR gate — it's a working-memory aid, not a merge artifact.

Done when

• A new entry is written the way the project opted in — appended to docs/LOGS.md, or added as a fragment under .agentic-workflow/journal/ — with at least Summary and Next.
• The mechanical facts (branch, commits, files) are accurate to git.

• The closing → Next: block is printed:

  → Next: /clear (or your agent's context reset) is now safe — the session is captured
    · resume later → the command named in the entry's **Next** line