Capture

Adaptive-Depth Session Recap

Capture session-end or checkpoint signals into Tracework raw entries. The goal is not a chronological diary. Preserve the lightest durable work signal that can improve Daily, Weekly, and Monthly first, then support evidence drill-down, decision replay, roadmap synthesis, and session-start recall when needed.

Tracework is reporting-first: Daily, Weekly, and Monthly can still produce direct value without a fresh capture. Capture improves their evidence boundary and keeps lower-frequency Query, Recall, and Roadmap available when needed.

Progressive References

Keep this file focused on routing and workflow. Read these references only when needed:

• references/capture-routing.md: choose capture_depth and handle explicit lite/standard/deep overrides. Read before finalizing entries.
• references/capture-entry-writing.md: write raw entries, artifact context, reporting metadata, Markdown fallback, and quality checks.
• references/capture-operations.md: resolve config, call tracework_raw.py, write artifact dossiers, and format quiet receipts.
• references/capture-day.md: scan indexed Codex and Claude sessions for one date without mixing reporting groups. Read for day mode only.
• references/tracework-storage-convention.md: shared schema and storage rules.

Zero-Config Mode

This skill works with or without a configured knowledge vault:

• With a vault: write entries to {vault}/raw/weeks/{ISO-week}/{project-slug}.json; write artifact dossiers to {vault}/raw/artifacts/{project-slug}.json only when the routed capture depth and session signal justify it. Keep output quiet by default.
• Without a vault: return structured Markdown in the conversation. Do not write files or create directories. Add a one-line setup hint.

Zero-config mode is best-effort. Fill fields that are supported by the conversation and local evidence, but do not interrogate the user just to enrich optional fields.

Workflow

1. Detect mode and depth.

   • Default mode is session-end capture.
   • Use checkpoint mode for "checkpoint", "记一下当前进展", or /tracework:capture checkpoint.
   • Use day mode for /tracework:capture day [YYYY-MM-DD] [scope] or its natural-language triggers. Read references/capture-day.md and follow it instead of extracting only the current conversation.
   • Read references/capture-routing.md, then choose one session-level capture_depth: lite, standard, or deep.
   • Respect explicit depth overrides such as /tracework:capture deep or "简单记一下". If no override exists, route from the session evidence.

2. Extract signals.

   • Read references/capture-entry-writing.md.
   • Classify the primary archetype: decision, repair, investigation, build, or maintenance.
   • Extract only durable signals: goal, state change, decision/trade-off, rejected path, risk, evidence boundary, artifact context, and next useful action.
   • Group related work into 1-3 entries. Use at most 5 entries for unusually broad sessions.

3. Generate raw entries.

   • Include capture_depth on every new session-recap entry.
   • Add the minimal reporting object only when the report boundary is clear. Keep work_stream at the top level and do not pre-write channel-specific Daily/Weekly/Monthly carry-forward prose.
   • Add artifact_context and artifact dossier side effects only when the session materially changed durable artifacts or routed to deep.
   • Do not assign report-local O#, W#, D#, or E# labels.

4. Write or fallback.

   • Read references/capture-operations.md.
   • Use tracework_raw.py append-entry as the canonical write path.
   • Upsert artifact dossiers when warranted and enabled.
   • Register the project as a best-effort side effect.
   • If config or helper calls fail, return the Markdown recap instead of blocking the user.

Output Policy

Vault mode returns a concise confirmation plus a capture receipt from references/capture-operations.md. Include:

• chosen depth and one short routing reason;
• the top durable summary;
• non-zero risk / open-question / abandoned-alternative counts only;
• one forward-looking line that connects this capture to later Daily/Weekly reports (for example that it can enter this week's report candidates).

Keep the default vault reply about six lines or fewer. Do not dump transcript bodies or long recaps unless verbose mode is requested.

Day mode returns one aggregate scan receipt. Never echo transcript bodies or paths. List only counts, project names, and explicit skip reasons.

Verbose output is opt-in. Treat "verbose", "show recap", "展开总结", or an explicit request to inspect the recap as permission to include the Markdown recap after the receipt.

Zero-config mode is always verbose because the conversation output is the only deliverable. It does not need the short vault receipt; keep the existing one-line setup hint after the Markdown recap.

Quality Gate

Before finalizing each entry, check:

• Does the entry explain why the work mattered, not just what files changed?
• Is capture_depth no heavier than the signal deserves?
• Does status bound the claim as done, ongoing, risk, or decision?
• Does reporting.outcome_candidate.kind, when present, avoid overstating the evidence?
• Can Daily, Weekly, and Monthly recover the state change and next gate from factual fields without relying on pre-written channel prose?
• Are git, files, docs, and logs treated as evidence or coverage rather than outcomes by themselves?
• Are durable artifacts represented through artifact_context or dossier metadata only when they are material?
• For checkpoint mode, is this a durable stage signal rather than a progress log?

Before finalizing vault-mode user output, also check:

• Does the receipt include a top summary grounded in the written entries?
• Are zero counts omitted rather than shown as "0 个"?
• Is there exactly one forward-looking report line (candidate for weekly/daily, or checkpoint "已记入本周证据")?
• Is the default reply short and quiet unless verbose was requested?

Anti-Patterns

• Do not fabricate fields to satisfy the schema.
• Do not ask confirmation questions just to enrich optional fields.
• Do not preserve process noise that would never appear in a daily or weekly review.
• Do not use checkpoint mode to record command-by-command progress.
• Do not scan host-wide transcript directories directly; day mode reads only manifests created by the opt-in Tracework hook.
• Do not read a transcript until its project and reporting group pass scope classification.
• Do not split one coherent feature across many entries.
• Do not write new entries with source: "arch-doc"; that source is legacy-only.
• Do not generate formal Stage/Pipeline architecture documents from this skill.