kk-session-extract

name	kk-session-extract
description	Extract durable knowledge from the current live session, stage it as a validated done session log, and run it through the same curation machinery as /kk-curate. Use when the user wants to proactively process the current session before waiting for capture hooks and a later curate pass — not for dictating one node (/kk-add) or batch-processing accumulated logs (/kk-curate).

kk-session-extract

Extract durable project knowledge from the visible current session, stage it as a normal proposal_status: done session log, and immediately run the shared curation tail. You operate in-host on the live context window — no headless harness spawn, no daemon.

Routing: use /kk-add when the user already knows the node; use /kk-session-extract when the current session just produced durable teaching moments; use /kk-curate when processing accumulated captured logs later.

Partial context warning: if compaction has occurred, the visible context may be incomplete. Describe your extraction as visible-context extraction, not full-session extraction, unless the runtime exposes the full transcript.

Resolve the project root

Resolve the repo root (the directory containing .ai/kenkeep) with the shipped detector, then treat the printed path as the working directory for every command below:

KK_REPO_ROOT=$(node .ai/kenkeep/scripts/kk-detect-root.mjs) || exit $?
cd "$KK_REPO_ROOT" || exit $?
pwd

1. Extract proposals from the live context

1. Load the extraction prompt. Read .ai/kenkeep/.config/prompts/proposal-extract.md first (per-repo override). If that file does not exist, read the bundled package template at templates/prompts/proposal-extract.md (relative to the installed npm package). Follow the prompt's extraction rules — do not embed a copy here.

2. Build a live transcript surrogate. Construct a role-tagged transcript from the visible conversation in this session (user and agent turns you can still see). Strip any <kk-private>…</kk-private> spans before using the text.

3. Substitute the placeholder. Replace [TRANSCRIPT PLACEHOLDER, substituted at runtime] in the loaded prompt with your live transcript surrogate. If the placeholder is missing from the prompt, report an extraction failure and stop with no writes.

4. Apply the prompt unchanged. Produce a JSON object matching the strict ProposalOutputSchema: { "practice": [...], "map": [...] } where each entry has only { type, tags, title, description, body, kk_confidence }. Do not emit legacy supports_existing_node / contradicts_existing_node fields.

5. Disposition gate. If the session is meta-only, exploratory, abandoned, unrelated, or has no durable teaching moments, report that no durable knowledge was found and stop — no staging, no curation, no node writes. This is success, not failure.

2. Stage the live proposals

When extraction yields at least one candidate:

1. Resolve session id. Prefer the harness's live UUID-v4 session id when available. If you cannot obtain a valid UUID-v4, use --generate-session-id and report degraded idempotency to the user (a later capture may not match this log).

2. Pipe validated JSON into the staging primitive:

   echo '<json>' | npx kenkeep@latest session-log stage-live --session-id <uuid-v4>
   # or, when no live id is available:
   echo '<json>' | npx kenkeep@latest session-log stage-live --generate-session-id
   

3. Capture the one-line JSON summary printed on stdout (path, session_id, idempotency). Report degraded idempotency explicitly when idempotency is degraded.

The staged log lives under .ai/kenkeep/_sessions/ with proposal_status: done and captured_by: manual.

3. Draft curator actions

You are the curator for this single staged session. Read the staged log's proposals and draft one CuratorAction per candidate using the same rules as /kk-curate Step 2 (action taxonomy, candidate_origin as <session_id>:<practice|map>:<index>, relate-and-place, whole-tree dedup awareness). Do not invent a second action-rule authority — follow /kk-curate's curator semantics.

Mint a run id once:

RUN_ID=$(uuidgen 2>/dev/null || date -u +"curate-%Y%m%dT%H%M%SZ")
PROPOSALS=$(mktemp -t kk-session-extract-proposals.XXXXXX.json)
SURVIVORS=$(mktemp -t kk-session-extract-survivors.XXXXXX.json)

Write your accumulated actions array (JSON array, top-level) to $PROPOSALS, then validate it before dedup: npx --yes kenkeep@latest validate curator-output "$PROPOSALS". On a non-zero exit, read the path-referenced errors, fix the offending action(s), and re-validate until it passes. (kk schema curator-output prints the JSON Schema if you need the exact shape.)

4. Dedup and stamp via the scoped primitive

Invoke curate-dedup once, scoped to the staged session only:

STAGED_SESSION_ID=<session_id from stage-live output>
npx --yes kenkeep@latest curate-dedup \
  --input "$PROPOSALS" --output "$SURVIVORS" --run-id "$RUN_ID" \
  --session-id "$STAGED_SESSION_ID"

This stamps only the staged live log. Unrelated proposal_status: done logs in _sessions/ remain available for a later /kk-curate.

Capture the stdout JSON summary (kept, conflicts, stamped, runId) and report it to the user.

5. Persist surviving actions via curate-persist

Follow /kk-curate Step 5 exactly: persist $SURVIVORS via curate-persist, surface any per-action failures from the summary, keep successful writes, and continue to the rebuild.

6. Rebuild the indices

npx --yes kenkeep@latest index rebuild

6b. Rebalance (final phase)

Follow /kk-curate Step 6b exactly: run rebalance trigger, skip when actions is empty, otherwise propose operations only for named branches and apply via rebalance move.

7. Report and resolve conflicts

Follow /kk-curate Step 7: report headline numbers, placement per written leaf, structural rebalance summary, then walk pending conflicts under .ai/kenkeep/conflicts/ if any were minted.

8. Hand off

Tell the user to review changed nodes under .ai/kenkeep/nodes/ and accept with git commit. The staged session log is now stamped as processed; a later capture hook for the same session_id preserves curator_processed_at / curator_run_id.

Constraints

• One curate-dedup call per run, always with --session-id for the staged log.
• No writes when extraction returns empty proposals.
• No weakening of UUID-v4 validation; degraded idempotency must be reported when using --generate-session-id.
• Do not spawn a headless harness or background worker from this skill.