session-ingest

AKB Session Ingest

Multi-agent workflow that analyzes a session JSONL file and writes structured notes to an AKB vault.

Workflow classification. LLM-wiki Ingest workflow. Cognitive-debt layer — it captures the in-flight reasoning of a coding session (how the team arrived at an understanding, the decisions and dead ends), turning ephemeral working context into persistent compiled-truth notes (TIL / decision / idea / task) before it evaporates. The four drafter subagents share this classification.

Prerequisites

• AKB MCP server is accessible.
• --vault {name} provided.

Execution Flow

Phase 0 — Resolve       parse args, resolve JSONL path (auto or explicit), maybe delegate & exit
Phase 1 — Prepare       detect duplicates, fetch open tasks
Phase 2 — Draft         read JSONL, identify commits/repo, draft session note
Phase 3 — Sub-Draft     launch til / task / idea / decision drafters; triage ideas
Phase 4 — Write         dedup ↔ finalize ↔ write session, sub-notes, resolve tasks

Phase 0 — Resolve & Maybe Delegate

Resolve invocation arguments and the JSONL path. Optionally fire-and-forget to a different target's CLI and exit early. All Phase 1+ work assumes $VAULT and $JSONL_PATH are set and the JSONL file exists.

0-1. Parse Arguments

Parse positional [jsonl_path] (optional) and flags:

• --vault {name} (required) — hard-fail with --vault {name} required. if absent
• --lang {language} (optional) — preferred language for synthesized note body
• --delegate {claude|codex} (optional) — fire-and-forget to a different target's CLI and exit early

Store as $JSONL_PATH, $VAULT, $LANG_FLAG, $DELEGATE.

0-2. Reject Self-Delegate

The current target was baked in at build time. If --delegate matches it, abort to prevent infinite recursion.

SELF="claude"
if [ -n "$DELEGATE" ] && [ "$DELEGATE" = "$SELF" ]; then
  echo "Refusing $SELF -> $SELF delegate (would recurse)." >&2
  exit 2
fi

0-3. Resolve JSONL Path

If $JSONL_PATH was supplied positionally, use it. Otherwise discover the current session's JSONL automatically — discovery is target-specific.

Claude target — ${CLAUDE_SESSION_ID} template substitution + $PWD encoding.

The Claude Code skill runtime substitutes ${CLAUDE_SESSION_ID} with the current session's UUID at render time (see the substitution table at code.claude.com/docs/en/skills.md) — it is NOT a shell environment variable. Do not gate the path on an env-var check; the substitution has already happened by the time bash runs.

if [ -z "$JSONL_PATH" ]; then
  ENCODED=$(printf '%s' "$PWD" | tr '/.' '--')   # both / and . map to -
  JSONL_PATH="$HOME/.claude/projects/${ENCODED}/${CLAUDE_SESSION_ID}.jsonl"
fi

If the substitution did not happen (e.g., this skill was somehow invoked outside a Claude session), the resulting $JSONL_PATH will contain the literal text ${CLAUDE_SESSION_ID} and the file-exists check below will catch it.

Verify the resolved path exists before proceeding:

if [ ! -f "$JSONL_PATH" ]; then
  echo "JSONL path not readable: $JSONL_PATH" >&2
  exit 2
fi

0-4. Delegate Fire-and-Forget (Optional)

If $DELEGATE is set, hand the ingest work to a different target and exit. The current session does no ingest work — the receiver runs Phase 1+ on the same JSONL. claude / codex spawn the target's CLI in the background.

Each target runs the receiver unattended while limiting blast radius if the JSONL contains injected content:

• claude: --permission-mode acceptEdits auto-approves file edits; Bash/MCP rely on the skill's allowed-tools.
• codex: -s workspace-write -c approval_policy="never" — confines file writes to the workspace AND skips approval pauses. --skip-git-repo-check lets the receiver start in a non-git pwd (delegate does not require the JSONL's project to be checked out). Avoid --dangerously-bypass-approvals-and-sandbox: that drops the sandbox entirely, letting prompt-injection in the JSONL escalate to arbitrary shell.

LOG="${TMPDIR:-/tmp}/session-ingest-delegate-$(date +%s)-$$.log"
ARGS="\"$JSONL_PATH\" --vault \"$VAULT\""
[ -n "$LANG_FLAG" ] && ARGS="$ARGS --lang \"$LANG_FLAG\""

case "$DELEGATE" in
  claude)
    nohup claude -p "/session-ingest $ARGS" --permission-mode acceptEdits \
      >"$LOG" 2>&1 &
    echo "Delegated to claude (pid $!, log $LOG)."
    ;;
  codex)
    # $session-ingest must NOT shell-expand. Single-quote the literal,
    # then concatenate the double-quoted $ARGS so the file path stays quoted.
    # --skip-git-repo-check: the receiver may run in a non-git pwd (delegate
    # does not require the JSONL's project to be checked out); codex exec
    # otherwise aborts before starting.
    PROMPT='$session-ingest '"$ARGS"
    nohup codex exec --skip-git-repo-check -s workspace-write -c 'approval_policy="never"' "$PROMPT" \
      >"$LOG" 2>&1 &
    echo "Delegated to codex (pid $!, log $LOG)."
    ;;
  "")
    : ;;
  *)
    echo "Unknown --delegate target: $DELEGATE (expected: claude, codex)" >&2
    exit 2
    ;;
esac

if [ -n "$DELEGATE" ]; then
  exit 0
fi

Phase 1 — Prepare

1-1. Detect Duplicates

Before reading the full JSONL, check whether this session was already ingested.

Detect JSONL Schema

Read the first ~10 lines of the JSONL (Read with limit: 10). The opening line may be a noise event without identifying fields (CC sessions sometimes start with permission-mode or file-history-snapshot, neither of which carries sessionId); scan all 10 to find the first line that decisively identifies the schema:

• If any of those lines has a top-level sessionId field (e.g., {"type": "worktree-state", "sessionId": "..."} or {"type": "user", "sessionId": "..."}), the JSONL is CC schema (flat per-line events). Set jsonl_schema = "cc".
• If any line has a top-level payload object containing id (e.g., {"timestamp": "...", "type": "session_meta", "payload": {"id": "...", "cwd": "..."}}), the JSONL is Codex schema (event-wrapped). Set jsonl_schema = "codex".
• If none of the first 10 lines yields either signal, hard-fail with "Unrecognized JSONL schema: $JSONL_PATH" — do not guess.

Schema detection is independent of runtime.target: a Codex JSONL ingested from a Claude harness still uses Codex parsing, and vice versa.

Extract Session ID, Timestamps, and Source CWD

Reuse the first-10-lines window from schema detection and read the last few lines of the JSONL. Capture everything the head window carries in one pass — later phases reuse these values rather than re-scanning. Field lookup is schema-aware:

For jsonl_schema == "cc":

• Store the first encountered sessionId field as session_id. (Lines without sessionId such as file-history-snapshot are skipped — they precede the first content event in some sessions.)
• If a worktree-state entry is present, store worktreeSession.originalCwd as session_cwd; otherwise leave session_cwd unset.

For jsonl_schema == "codex":

• Find the entry with type == "session_meta" (typically the first line). Store payload.id as session_id and payload.cwd as session_cwd.

For both schemas:

• Last entry's top-level timestamp → jsonl_last_timestamp.

Store:

• session_id — the resolved session/thread id
• jsonl_last_timestamp — the ISO 8601 timestamp of the final entry
• session_cwd — the source session's working directory (or unset)

Search Vault for Existing Session

mcp__akb__akb_search(
  query="{session_id}",
  vault="{vault_name}",
  collection="sessions",
  tags=["session:{session_id}"],
  type="session",
  limit=1
)

No match → sync_mode = "new", proceed to Phase 2.

Match found → read the existing session note:

mcp__akb__akb_get(vault="{vault_name}", doc_id="{matched_doc_id}")
→ existing_session_doc_id, existing_tags, existing_session_content

Extract the last-synced:{timestamp} tag value from the existing note's tags. Store as last_synced_timestamp. Preserve existing_session_content (the full markdown body) for Phase 2-3 and Phase 4-1.

Compare Timestamps

• jsonl_last_timestamp <= last_synced_timestamp → no new content since last sync → stop: "Session {session_id} already synced up to {last_synced_timestamp}. No new content. Skipping."
• jsonl_last_timestamp > last_synced_timestamp → proceed to the new-content assessment below.

Assess New Content

Read the JSONL entries with timestamps after last_synced_timestamp. Filter out noise. The filter list is schema-aware:

For jsonl_schema == "cc":

• type: "system" messages
• type: "permission-mode" entries
• type: "file-history-snapshot" entries
• User messages containing only /exit, /quit, or session-end signals
• isMeta: true entries

For jsonl_schema == "codex":

• type: "turn_context" entries (per-turn configuration)
• type: "event_msg" with payload.type in {task_started, token_count} (system markers and telemetry)
• type: "response_item" with payload.type == "reasoning" (model-internal thinking, analogous to CC thinking blocks)
• User messages (type: "event_msg" with payload.type == "user_message") containing only /exit, /quit, or session-end signals
• type: "session_meta" (metadata-only, already consumed for session_id extraction)

After filtering, judge whether the remaining entries represent meaningful work — substantive user messages, assistant responses with real content, tool calls that changed files, or decisions made.

Not meaningful (e.g., resume then immediately exit, only system noise) → stop: "Session {session_id} has entries after {last_synced_timestamp} but no meaningful new work. Skipping."

Meaningful → set sync_mode = "update", store existing_session_doc_id and last_synced_timestamp, proceed to Phase 1-2.

1-2. Retrieve Open Tasks

Before reading the full JSONL, retrieve project-scoped open tasks from the vault so the session drafter can identify which tasks this session resolved.

Parallel execution with Phase 1-1: the project-name parse step below runs immediately (no MCP), and the akb_search for open tasks is independent of Phase 1-1's duplicate-detection search. Issue Phase 1-1's existing-session search and Phase 1-2's open-task search in a single parallel tool-use block. Act on Phase 1-2 results only after Phase 1-1 greenlights continuation (otherwise discard).

Derive Project Name

Project name is the basename of the source session's working directory. Use session_cwd (captured in Phase 1-1) when set — it is authoritative metadata from the JSONL, and cross-target delegate (e.g. Claude → Codex) puts the receiver in a different pwd than where the session originated, so pwd is not a reliable fallback.

• session_cwd set → project_name is its basename. (For CC worktree sessions this is originalCwd, the unworktreed project root, not the worktree name.)
• session_cwd unset (CC session with no worktree-state entry) → use the current working directory's basename. The CC JSONL parent directory uses a lossy encoding (/ and . both map to -), so decoding it is unreliable and not used.

If derivation fails entirely, set project_name to the empty string. The downstream open-tasks query becomes a no-op (no project: tag match) — degraded but safe.

Query Open Tasks

mcp__akb__akb_search(
  query="open task",
  vault="{vault_name}",
  collection="sessions/tasks",
  type="task",
  tags=["project:{project_name}"],
  limit=20
)

From the results, exclude any document whose tags contain resolved. Store the remaining results as open_tasks — a list of {doc_id, title, summary} for each match.

If no results or the query fails, set open_tasks to an empty list and continue.

Phase 2 — Draft Session Note

The main agent acts as the session drafter. Read the JSONL file, extract session information, and produce both a session draft (navigator skeleton) and a session context for sub-agents. The session draft is finalized later in Phase 4-0 after sub-drafter results are reconciled.

2-1. Read Session JSONL

Read the JSONL file. It contains the complete conversation history: user messages, assistant responses, tool calls, and results. The field-name layout differs by jsonl_schema (set in Phase 1-1) — apply the analogous mapping when extracting content:

Concept	CC schema	Codex schema
User message	type: "user" entry, message.content[] with type: "text"	type: "event_msg", payload.type: "user_message"
Assistant text response	type: "assistant" entry, message.content[] with type: "text"	type: "event_msg", payload.type: "agent_message" (and/or response_item with payload.type: "message")
Tool call	type: "assistant" entry, message.content[] with type: "tool_use" (nested inside the assistant turn — NOT a top-level event type)	type: "response_item", payload.type: "function_call"
Tool result	type: "user" entry, message.content[] with type: "tool_result" (nested inside a user-role turn that wraps the tool reply)	type: "response_item", payload.type: "function_call_output"

Update mode (sync_mode = "update"): Focus JSONL analysis on entries with timestamps after last_synced_timestamp — this is where the new information lives. The session draft you produce must be a combined compiled truth: read existing_session_content (retrieved in Phase 1-1) and integrate its Summary / Narrative / Artifacts / Knowledge Produced / Open Questions with the new work from the resumed portion. The written draft is the new canonical state of the session, not a delta.

New mode (sync_mode = "new"): Read the full JSONL as normal.

Focus on extracting:

• Work performed (main topics and tasks), as a chronological narrative
• Files and commits touched
• External URLs (articles, repos, docs) referenced in the session
• Unresolved questions (to be promoted in Phase 4-0 where possible)

Skip tool call/result noise — focus on user messages and assistant text responses. Reusable lessons and long-lived decisions will be extracted by the Phase 3 drafters; the session draft itself does not need to generalize.

2-2. Identify Session Commits and Repository

Scan the JSONL for git commit activity. Extract commit hashes to determine the range of changes made during the session.

If commits were found, run:

git log --oneline {earliest_commit_hash}^..{latest_commit_hash}
git diff --stat {earliest_commit_hash}^..{latest_commit_hash}

If no commits were found in the JSONL, skip these commands.

Also collect the repository URL:

git remote get-url origin

Store the result as repo_url. If the command fails (no remote configured), set repo_url to null.

2-3. Draft Session Note

Compose the session draft per references/session-template.md — that file carries the full draft template, section-omission rules, update-mode handling, and authoring guidelines. The session note is a navigator + narrative + timeline; deep knowledge (reusable lessons, long-lived decisions, follow-up tasks, creative ideas) belongs to the Phase 3 sub-notes and must not be restated here.

Determine the session's primary theme (debugging vs. feature build vs. refactor) and let it guide the emphasis in ## Narrative. Number ## Open Questions bullets Q1, Q2, … so Phase 3 sub-drafters can target them via resolves_question: Q{n}.

2-4. Identify Resolved Tasks

Compare the work performed in the session (from Phase 2-1) against the open_tasks list (from Phase 1-2).

A task is "resolved" if the session performed substantial work directly addressing the task's topic. Partial completion counts — any remaining work will be captured as a new task by the task-drafter.

For each resolved task, record:

• doc_id — from the open_tasks entry
• title — the task's title
• path — the task's path (for Knowledge Produced link)
• reason — brief explanation of how it was addressed (e.g., "Implemented the retry logic described in the task")

Store as resolved_tasks list. If no tasks were resolved, set to an empty list.

2-5. Construct Session Context for Sub-Agents

Build a lightweight session context as a roadmap for the Phase 3 drafter agents. This supplements the session draft and JSONL — it helps agents orient quickly.

Session Context:
- Project: {project name}
- Repository: {repo_url from Phase 2-2, or "N/A"}
- Date: {YYYY-MM-DD from JSONL or today}
- Session JSONL: {$JSONL_PATH resolved in Phase 0}
- Work performed: {brief list of main topics/tasks}
- Committed changes: {from Phase 2-2 — commit messages and files changed, or "None"}
- Problems encountered (for possible troubleshooting TILs): {brief list}
- Long-lived decisions considered (for possible Decision notes): {brief list}
- Rejected alternatives (anti-pattern memory for Decisions): {brief list}
- Tools/technologies used: {notable APIs, libraries, frameworks}
- Vault name: {resolved vault_name from --vault}
- Collections: {plugin's default_collections mapping — includes decisions}
- Ingest mode: {new|update}
- Last ingested timestamp: {ISO timestamp from existing session note, or "N/A" if new}
- JSONL analysis start: {last_synced_timestamp if update, or "beginning" if new}
- Open tasks: [{doc_id}] {title} (for each task in open_tasks, or "None")
- Resolved tasks: [{doc_id}] {title} — {reason} (for each task in resolved_tasks, or "None")
- Open Questions (session draft): Q1: {text}, Q2: {text}, ... (each question available for sub-drafters to resolve via `resolves_question: Q{n}`)

Keep this concise — it is a roadmap, not a comprehensive record.

Phase 3 — Parallel Sub-Drafting & Validation

3-1. Launch Drafter Agents

Launch 4 agents simultaneously in a single message. Each agent receives the session context, a session headline (frontmatter + > **Summary:** line + numbered ## Open Questions list only — not the full Narrative), and the JSONL path. The JSONL is the primary source; drafters do not need Narrative prose.

Agent(
    subagent_type="akb-sessions:til-drafter",
    description="TIL learning notes",
    prompt="[Session Context]\n\n## Session Headline\n{frontmatter + Summary + Open Questions}\n\nExtract learnings from the JSONL, validate against vault, and compose TIL notes."
)

Agent(
    subagent_type="akb-sessions:task-drafter",
    description="Follow-up task notes",
    prompt="[Session Context]\n\n## Session Headline\n{frontmatter + Summary + Open Questions}\n\nIdentify follow-up tasks from the JSONL and validate against vault."
)

Agent(
    subagent_type="akb-sessions:idea-drafter",
    description="Creative idea notes",
    prompt="[Session Context]\n\n## Session Headline\n{frontmatter + Summary + Open Questions}\n\nGenerate creative ideas inspired by the JSONL and validate against vault."
)

Agent(
    subagent_type="akb-sessions:decision-drafter",
    description="Long-lived decision notes",
    prompt="[Session Context]\n\n## Session Headline\n{frontmatter + Summary + Open Questions}\n\nExtract long-lived architectural/product decisions (ADR) from the JSONL, validate against vault, and compose Decision notes."
)

Agent Summary

Agent	Model	Purpose	Output
til-drafter	sonnet	Learnings + vault validation	0–N learning notes with dispositions + til_kind tag
task-drafter	sonnet	Follow-up tasks + vault validation	0–N task notes with dispositions
idea-drafter	opus	Creative ideas + vault validation	0–N idea notes with dispositions
decision-drafter	sonnet	Long-lived decisions (ADR) + vault validation	0–N decision notes with dispositions

Each agent receives:

1. Session Context — lightweight roadmap (from Phase 2-5)
2. Session Headline — frontmatter + > **Summary:** line + numbered ## Open Questions. The full Narrative is not sent — drafters read the JSONL directly for domain-specific detail.
3. JSONL path — primary source for each drafter's domain analysis
4. Vault config — vault_name and collections for duplicate checking

Each agent drafts notes, then validates them against the vault using akb_search. Drafts are returned with a disposition (create / append; supersede additionally for decisions), optional resolves_question: Q{n} when a draft addresses a specific session Open Question, and related_to doc IDs. Duplicates are reported as skipped and excluded from the output.

idea-drafter specific. Idea drafts additionally carry impact / effort / confidence enum assessments. The drafter silent-drops any idea it rates impact: low before returning (no report, no skipped block). The main agent re-triages the surviving drafts in Phase 3-2.

Agents that find nothing to write return "No {type} notes for this session."

3-2. Triage Idea Drafts

After the idea-drafter agent returns, the main agent filters the surviving idea drafts using their impact / effort / confidence frontmatter fields. The drafter already silent-dropped any idea it rated impact: low before returning; this phase is a second gate that catches cases the drafter missed and removes high-cost, low-conviction ideas.

Scope. Applies only to idea drafts. TIL / task / decision drafts pass through unchanged.

Inputs. Each idea draft from Phase 3-1 with disposition ∈ {create, append}.

Automatic drop rules — apply in order, drop on first match:

1. impact == low — drop. Safety net for the drafter's self-filter.
2. effort == large AND confidence == low — drop. High-cost, low-conviction ideas clutter the vault.

All remaining idea drafts pass through to Phase 4.

No user prompt. No skipped block written. No vault write. Dropped drafts are discarded in-memory; they do not feed Phase 4-0 or Phase 4-2. If a dropped idea set resolves_question: Q{n}, discard that mapping as well — the Q ID stays in the session's ## Open Questions.

Phase 4 — Write to AKB

Save all drafts with create, append, or (decisions only) supersede disposition directly — no user approval needed. Each drafter sets status: active in its ---DRAFT--- block; Phase 4-1 / 4-2 pass it through verbatim.

4-0. Dedup Drafts & Finalize Session Body

In-memory reconciliation between the session draft and the sub-note drafts before any AKB write. Single in-context pass — no extra LLM call, no MCP call, no per-pair loop. Inputs: session draft body, all sub-note drafts with disposition ∈ {create, append, supersede}, resolved_tasks list.

Step A — Promote Open Questions. For each sub-note draft with resolves_question: normalize the value (strip / upper / ^Q\d+$; non-conforming → log warning + no-match), find the matching Q{n} bullet in the session's ## Open Questions, and remove that bullet entirely — the answer surfaces in ## Knowledge Produced (Step C). Multiple sub-notes can target the same Q{n}; the bullet is removed once. If no session question matches, log a warning and leave the sub-note unchanged.

Step B — Cross-Type Content Overlap Detection. For each sub-note, extract its headline sentence — TIL: > **Key Insight:** …; Task: > **Next Action:** …; Idea: > **Inspiration:** … plus Proposal first sentence; Decision: > **Context:** … plus Decision first sentence. Scan the session Narrative and Artifacts prose for sentences that match a headline. A sentence matches when at least two of these hold: (1) core noun phrases overlap, (2) assertion direction is the same (not negated / qualified differently), (3) reasoning or evidence cited is the same. One criterion alone is not enough — preserve the Narrative sentence if in doubt. On a match, remove the redundant sentence or replace it with a brief pointer ("…the lesson captured in Knowledge Produced"); do not rewrite surrounding prose beyond what grammatical flow needs. Overlap resolution always favors sub-notes — never edit a sub-note body here.

Step C — Finalize Knowledge Produced. Build the block from deduplicated sub-notes plus resolved_tasks, in the order TIL → Idea → Task → Decision → Resolved:

## Knowledge Produced

- 🧠 TIL: [{title}](sessions%2Flearnings%2F{doc_path}.md) — {one-line hook from Key Insight}
- 💡 Idea: [{title}](sessions%2Fideas%2F{doc_path}.md) — {one-line hook from Inspiration}
- ✅ Task: [{title}](sessions%2Ftasks%2F{doc_path}.md) — {one-line hook from Next Action}
- 📐 Decision: [{title}](sessions%2Fdecisions%2F{doc_path}.md) — {one-line hook from Decision first sentence}
- ♻️ Resolved: [{title}](sessions%2Ftasks%2F{path}.md) — {reason from resolved_tasks}

For newly-drafted sub-notes the link's doc_path is substituted at render time from each Phase 4-2 put response. For resolved tasks the link uses the existing task's path captured in Phase 2-4. See references/note-templates.md → "Cross-Note Links" for why this hybrid (body link + depends_on frontmatter) is used.

Step D — Final omissions. Omit ## Knowledge Produced entirely when no sub-notes were produced and no tasks resolved. Omit ## Open Questions if Step A left zero unresolved bullets. The resulting body is what Phase 4-1 writes — do not mutate further.

4-1. Write Session Note

The session note is written first because sub-notes reference it via depends_on. Sub-notes always use the full URI akb://{vault_name}/doc/{session_doc_path} (always-URI cross-ref convention; future-proof against vault tier splits).

New mode (sync_mode = "new"). Extract metadata from the draft's ---DRAFT--- block and akb_put(vault="{vault_name}", ...) with type="session", collection={draft.collection}, status={draft.status}, tags=["session", "claude-code", "session", "session:{session_id}", "project:{project_name}", "last-synced:{jsonl_last_timestamp}", ...topic_tags], and the finalized body from Phase 4-0. Capture session_doc_id and session_doc_path.

Update mode (sync_mode = "update"). The session note already exists. Phase 2-3 + 4-0 produced a full compiled-truth replacement plus exactly one new timeline entry. Apply the timeline merge procedure — Mode A (full-body merge) (see appendix) against existing_session_content. Build the updated tag list (replace the old last-synced:{old_timestamp} with last-synced:{jsonl_last_timestamp}; preserve all other tags; add any new topic tags). Then akb_update(vault="{vault_name}", doc_id=existing_session_doc_id, content={merged}, tags={updated}, summary={draft.summary}, message="Update from resumed session: {brief description}"). Assemble session_uri = akb://{vault_name}/doc/{existing_session_doc_path} for sub-notes' depends_on.

4-2. Write Sub-Notes (Learning, Task, Idea, Decision)

Issue all sub-note akb_put calls in a single parallel tool-use block. Sub-notes are independent of each other; only the session note's order matters (via depends_on).

For each sub-note, extract metadata from the draft and call akb_put(vault="{vault_name}", ...) with collection={draft.collection}, type={draft.type}, status={draft.status}, tags=["session", "claude-code", ...draft_tags], depends_on=["akb://{vault_name}/doc/{session_doc_path}"], related_to=[...draft.related_to] (full URIs as emitted by drafters). Collection per type: TIL → sessions/learnings, Task → sessions/tasks, Idea → sessions/ideas, Decision → sessions/decisions.

Idea notes — append four namespaced tags to draft_tags before put: category:{draft.category}, impact:{draft.impact}, effort:{draft.effort}, confidence:{draft.confidence}. Drop duplicates (the drafter may have already emitted some as tags). These four tags are the durable carriers for the external maintenance layer's consolidation synthesis and staleness checks.

Task notes — after the task document is created, also create a lightweight todo via akb_todo(vault="{vault_name}", title=draft.title, priority={P0→urgent, P1→high, P2→normal, P3→low}, ref_doc=task_doc_id, note=draft.summary).

append disposition — the drafter has already merged compiled truth + one new timeline entry. Read the existing document with akb_get(vault="{vault_name}", doc_id=draft.append_target), apply timeline merge Mode A against draft.body, and akb_update(vault="{vault_name}", doc_id=draft.append_target, content={merged}, tags=[...existing, ...new from draft], summary=draft.summary, message="Appended from session ingest: {brief description}").

supersede disposition (decisions only) — write the new decision via akb_put(vault="{vault_name}", ..., tags=[..., "supersedes:{old_doc_id}"]) (tags use bare doc_id; the : already binds key:value). Capture new_doc_id and new_doc_path; compose new_doc_ref = akb://{vault_name}/doc/{new_doc_path}. Then read the old decision via akb_get(vault="{vault_name}", doc_id=draft.supersedes), apply timeline merge Mode B with new_entry = "- **{today}** | Superseded by {new_title} ({new_doc_ref}). {reason}. [Source: session:{session_id}, {today}]", and akb_update(vault="{vault_name}", doc_id=draft.supersedes, content={merged}, tags=[...old, "superseded-by:{new_doc_id}"], status="superseded", message="Superseded by {new_doc_ref} in session ingest"). Setting status="superseded" removes the old decision from the external maintenance layer's active-document pools.

4-3. Resolve Completed Tasks

If resolved_tasks is non-empty, fetch open todos once: akb_todos(vault="{vault_name}", status="open") → open_todos. For each resolved task:

1. Scan open_todos in memory for a todo whose ref_doc matches the task's doc_id. If found, akb_todo_update(vault="{vault_name}", todo_id={matched}, status="done", note="Resolved in session {session_doc_id}").
2. Read the task document with akb_get(vault="{vault_name}", doc_id={task_doc_id}), then compose the resolution timeline entry - **{today YYYY-MM-DD}** | Task resolved in session:{session_id}. {reason}. [Source: session:{session_id}, {today YYYY-MM-DD}] and apply timeline merge Mode B against task_content (compiled-truth zone preserved as-is).
3. akb_update(vault="{vault_name}", doc_id={task_doc_id}, content={merged}, tags=[...existing, "resolved"], status="archived", message="Resolved in session {session_doc_id}").

If no matching todo is found, still archive the task document. If the up-front akb_todos fails, skip all todo updates this ingest but archive the task documents anyway.

Error Handling

If a drafter agent (Phase 3) fails or returns an error:

• Continue with results from the remaining agents
• Log the failure in the session summary
• Do not block the pipeline for a single agent failure

If an akb_put fails partway through Phase 4:

• Continue writing the remaining notes from this invocation
• The partial result is queryable under the "session:{session_id}" tag

If Phase 4-0 Step A finds a resolves_question that does not match any session Open Question ID:

• Log a warning naming the sub-note and the invalid Q ID
• Leave the sub-note draft as-is, write without session mutation for that question

References

• references/session-template.md — session draft template, section-omission rules, update-mode handling, authoring guidelines.
• references/note-templates.md — sub-note templates (TIL / task / idea / decision).
• Agent prompts in agents/ — full drafter-agent instructions.

Appendix: Timeline Merge Procedure

Shared procedure for every skill that writes back to an existing note. Two modes — every caller picks one explicitly.

Shared helper. split_body(content) — locate the first standalone --- line that is followed (eventually) by a ## Timeline heading. Return (body_above, timeline_entries). If no ## Timeline heading exists, return (content, []).

Mode A: Full-body merge

Used when the caller has produced a rewritten compiled-truth body (the new canonical state) plus exactly one new timeline entry. Callers: session-ingest Phase 4-2 (session update), session-ingest Phase 4-3 (sub-note append), and ingest-jira (issue upsert).

Inputs: draft_body (full new body — includes --- + ## Timeline + one new bullet), existing_content (current doc).

1. (draft_above, draft_entries) = split_body(draft_body). draft_entries must contain exactly one bullet. If it contains zero, treat as drafter error → fall back to akb_update with draft_body unchanged and log a warning.
2. (_, existing_entries) = split_body(existing_content).
3. Compose merged content:

   {draft_above}
   
   ---
   
   ## Timeline
   
   {draft_entries[0]}
   {existing_entries}
   

4. Newest entry first. Never edit or remove past entries.

Mode B: Entry-only append

Used when the caller produced only a new timeline bullet and wants the existing compiled-truth zone preserved verbatim. Callers: session-ingest Phase 4-4 (task resolve) and the decision supersede path, ingest-doc re-ingest, and ingest-confluence re-ingest.

Inputs: new_entry (a single bullet line, possibly with sub-bullets), existing_content.

1. (existing_above, existing_entries) = split_body(existing_content).
2. If existing_content had no ## Timeline section, append "\n\n---\n\n## Timeline\n\n" to existing_above before proceeding.
3. Compose merged content:

   {existing_above}
   
   ---
   
   ## Timeline
   
   {new_entry}
   {existing_entries}
   

4. Never rewrite the compiled-truth zone in this mode. If the caller needs to change compiled truth, it must use Mode A instead.