Recall

Session Start Recall

This skill prepares a bounded project context at the start of a session so the developer and AI do not restart from zero. It is useful when durable Tracework memory exists; it is not required before daily, weekly, monthly, or query can produce their best available output.

Recall is a durable-memory surface, not a report fallback. Unlike daily or weekly reports, it must not use git history to invent missing memory.

Scope

• Intra-project only.
• Raw entries are the semantic source of truth.
• Artifact dossier entries are optional source navigation plus recorded context.
• Decision context entries are optional synthesized decision-replay hints from {vault}/raw/decisions/{project-slug}.json.
• If the saved decision index is missing, invalid, empty, or older than the raw entries, the helper refreshes the derived decision index from raw entries as a best-effort side effect, then uses that fresh Decision Context.
• Do not use git commits as fallback in v1.
• Do not invent history when no vault data exists.

Workflow

Pre-Check: Vault Health

Before running the recall helper, do a quick health check:

1. Resolve vault path from config
2. Check if {vault}/raw/weeks/ has any {slug}.json files
3. If no raw entries exist at all: suggest /tracework:capture and stop
4. If raw entries exist but are older than 14 days: note staleness in output
5. If recall_context.py is missing: warn and suggest re-install

If the pre-check passes, continue to the main workflow:

1. Resolve the current project context.
2. Run the helper:

python <this-skill>/scripts/recall_context.py --cwd "$PWD" --limit 12

If the user explicitly names a project slug, pass --slug <slug>. If the user or config provides a vault path, pass --vault <path>.

3. Read references/recall-output-template.md.
4. Produce Markdown in the conversation.

Output Contract

The response must include:

• Recent progress
• Relevant decisions
• Decision Context, when {vault}/raw/decisions/{project-slug}.json exists or can be rebuilt in memory from raw entries
• Abandoned alternatives that may affect the current task
• Open questions
• Risks to check before implementation
• Repo-local docs worth reading
• Potentially stale intent artifacts
• Suggested entry point for the current session

If there is no vault or no raw data, say that Tracework has no durable memory for the current project yet, suggest using /tracework:capture after this session, and suggest /tracework:cold-start-interview if no config exists. Do not fill the gap with git history or assumptions.

Evidence Rules

• Cite raw entry timestamps for decisions, risks, abandoned alternatives, and open questions.
• Cite artifact ids for docs worth reading.
• Cite decision context source_entry_refs when using decision_context; preserve confidence and inference_notes instead of presenting inferred nodes as fact.
• Treat artifact dossiers as independently readable but not independently authoritative. artifact_summary can explain scope and recorded context; direct evidence still comes from raw entries, evidence_refs, typed source_refs, or source-of-truth artifact references.
• Include decision_context_source when present, especially if rebuilt=true, so the user can see whether the section came from a saved index or an in-memory freshness rebuild.
• Treat intent_artifact_flags as read-only staleness hints. Phrase them as "may need review" rather than facts. Do not write or propose diffs from this skill.
• Use "Inferred" wording when the helper returns weak or old-schema entries.
• Missing artifact dossier index is acceptable and must not block output.

Storage

This skill reads:

• {vault}/raw/weeks/{YYYY-WNN}/{project-slug}.json
• {vault}/raw/artifacts/{project-slug}.json when present
• {vault}/raw/decisions/{project-slug}.json when present

This skill writes {vault}/raw/decisions/{project-slug}.json only when the derived decision index is missing, invalid, empty, or older than raw entries. Raw entries remain the source of truth.

Artifact Discovery

The artifact dossier index at {vault}/raw/artifacts/{slug}.json is a structured catalog of durable project documents plus recorded context. When present, use it to enrich the recall output:

1. Find related docs: When raw entries mention related_docs or artifact_context, cross-reference the artifact dossier index for topic tags, decision threads, and artifact_summary.scope to find adjacent documents.
2. Detect stale documents: Artifacts with status: superseded or a superseded_by field indicate documents that have been replaced — flag these so the next session doesn't rely on outdated references.
3. Navigate by topic: Artifact topics, decision_threads, and artifact_summary.key_claims provide stable retrieval terms for finding documents relevant to the current session.
4. Respect availability: Use source_availability, last_seen, and deletion_behavior to show whether the source document can still be opened and whether the dossier remains useful if it cannot.

In the "Docs worth reading" section, include for each artifact:

• Why the artifact matters (from artifact_summary.scope, topics, or recorded raw artifact_context.scope)
• Whether it is still active (from status)
• Source availability and deletion behavior when present
• Evidence boundary for any key claim when present

Missing artifact dossier index is acceptable — the recall output must not depend on it. When absent, derive doc references from raw entry fields alone.

Absorbed Intent-Artifact Flagging

The old standalone intent-sync behavior is now a lightweight recall section. When recent raw entries include sync_suggestions, or mention intent artifacts and contract-like terms such as design, plan, architecture, prompt, schema, contract, migration, or config, include a "Potentially stale intent artifacts" section.

This section is read-only. Its job is to tell the next session what to inspect, not to decide that a document is stale and not to modify the document.