ai-session-wiki-ingest

name	ai-session-wiki-ingest
description	Turn one date of local Codex and Claude Code session evidence into one Chinese Daily Wiki page.

AI Session Wiki Ingest

Write Chinese prose while preserving technical names, paths, commands, repo names, and ticket ids in English.

Purpose

The Daily workflow has one job:

session logs -> bounded Evidence Snapshot -> one Daily page -> local verify

Node handles evidence extraction, Snapshot packing, and lint. The root agent receives one Snapshot and writes one summary. Do not delegate routine Daily work or split it into per-card model calls.

The .capture.json is a lossy, regenerable Evidence Snapshot; the raw session logs remain the complete fact source. Priority and noise filtering control evidence density; Snapshot size is not a business skip condition. prepare persists the Snapshot and returns one metadata JSON line containing its path; it never emits Snapshot bytes.

Run

1. As the first tool command after reading this skill, run:

   node scripts/daily-memory-workflow.mjs prepare YYYY-MM-DD
   

   Run it once with a 30-second blocking wait. If the tool still reports running, wait once. Never poll or rerun prepare. Do not inspect the helper source; this skill and the helper output are the complete run contract.

2. Handle the one-line status:

   • ready: read the persisted Evidence Snapshot through the helper's chunk reader:

     node scripts/daily-memory-workflow.mjs read YYYY-MM-DD
     

     Run it repeatedly. Each run prints one chunk — a CHUNK i/N ... header line, the chunk payload, then a trailing END CHUNK i/N line — and records that chunk in the read ledger; verify later refuses a Daily whose ledger is incomplete. The helper owns chunk sizing and offsets; never compute offsets or read the Snapshot file any other way. Stop when the command prints {"done":true,...}. If a chunk is missing its END CHUNK i/N line or the tool reports output truncation, report snapshot_read_truncated; never reread a chunk or continue synthesis.

   • skipped_no_sources: leave the Daily unchanged and report skipped.

   • skipped_with_reason: leave the Daily unchanged and report its reason.

   Prepare output is metadata-only. Never expect Snapshot JSON after the metadata line, and never interpret includedTurns / omittedTurns as transfer counts; they report evidence filtering. If the run cannot continue, report the exact stage and one of these reasons instead of the generic incomplete tool transfer:

   • prepare_failed: ... when the prepare command itself fails;
   • snapshot_missing: ... when evidenceSnapshot does not exist;
   • snapshot_unreadable: ... when the path exists but cannot be read;
   • snapshot_invalid_json: ... when the persisted file is not valid JSON;
   • snapshot_chunk_missing: ... when verify reports the Snapshot read ledger incomplete;
   • snapshot_read_truncated: ... only when a chunk-read tool explicitly reports output truncation.

   Include the Snapshot path, snapshotBytes from prepare when available, and the original tool error text. Never report a business-level blocked state.

3. Use the complete ordered chunks from step 2 as the only synthesis input. Do not read any chunk again or open raw session JSONL, an existing Daily, automation memory, global memory, or prior agent runs. Omitted older turns are an expected cost-control result, not a reason to stop.

4. In one synthesis pass, write exactly wiki/sources/ai-chats/YYYY-MM-DD.md using the schema and template included in the Snapshot. Do not create any other knowledge page.

5. Run once:

   node scripts/daily-memory-workflow.mjs verify YYYY-MM-DD
   

   Use the same blocking-wait rule. If verification reports a mechanical formatting error, make at most one targeted correction and verify once more. If it still fails, leave the run as skipped_with_reason; do not broaden evidence reads or call it blocked. Do not read or edit wiki/log.md; verify owns lint, diff, status, and logging.

Compile The Daily

• Group by workstream and final outcome, not transcript chronology.

• Preserve substantive goals, root causes, decisions, rejected options, verified results, blockers in the underlying work, and unresolved follow-ups.

• Omit routine tool narration, polling, acknowledgements, and repeated status updates.

• Every ### topic under 关键会话 starts with one to three representative links to Evidence Cards from the Snapshot:

  - 证据来源：[Codex · codex-...](../../../.vault-meta/captures/ai-chats/YYYY-MM-DD.capture.json#codex-...)
  

• Never link a card to a claim it does not support. The linked Snapshot card retains the original session path for later audit.

• If evidence is incomplete because older turns were omitted, state only what the Snapshot supports and lower confidence when appropriate. Do not invent missing detail and do not open another source.

• detailed preserves all valuable supported information in the Snapshot. concise shortens prose, not factual accuracy.

• A day with no substantive work may be skipped_with_reason: no high-signal work instead of producing an empty Daily.

Cost Boundary

The normal path is exactly one prepare, the minimum number of non-overlapping Snapshot chunk reads, one synthesis/write, and one verify. Do not issue separate model turns for individual chunks, lint, diff, status, logging, or progress narration; the Node helper owns the mechanical steps.

Report

Report:

• the date and whether the result was written or skipped;
• Evidence Snapshot path and Evidence Card count;
• included and omitted turn counts from prepare;
• prepare status and Snapshot read status, including the exact failure stage and reason when skipped;
• Daily path when written;
• verification result;
• confirmation that no git commit was made.