Decision Replay Query

This skill answers targeted project-history questions from Tracework's local decision memory. It is a thin wrapper around the deterministic decision replay helper: the script narrows evidence, and the host agent writes the final answer.

Use query when the current task touches an existing decision, architecture boundary, rejected option, long-running risk, or product tradeoff. Use recall for session-start orientation; use query for a specific follow-up question. Git-only report fallback is not decision evidence here. If the raw record does not support the question, return the evidence gap and suggest targeted capture after the user clarifies the decision.

Scope

• Intra-project by default.
• Raw entries remain the source of truth.
• {vault}/raw/decisions/{project-slug}.json is a derived retrieval index.
• The helper may rebuild an in-memory index from raw entries when no saved index

exists, but this skill does not write new memory.

• Do not use git history, external docs, or model assumptions to fill missing

decision evidence.

Workflow

1. Identify the user's decision question.
2. Pick the mode:

Questions asking for evidence, proof, verification, or whether a claimed result is supported use the existing modes rather than inventing an evidence mode: try impact first to retrieve the claimed effect, then why if the impact query is not answerable. The agent performs this routing; the helper continues to expose only its four concrete modes.

Free-Form Mode

When the user asks a natural-language question that doesn't clearly map to why / alternatives / revisit / impact, use free mode. The agent tries multiple modes sequentially and returns the first result with answerable=true.

Fallback order: why → revisit → impact. Stop at the first answerable=true result. Do not try alternatives in the fallback chain — it answers a different question shape.

If all three modes return answerable=false, say the vault does not contain enough evidence for this question. Suggest capturing the context with /tracework:capture if the user can clarify the decision.

Free-form queries still pass a concrete --mode to the helper on each attempt. The agent handles the multi-mode orchestration; the helper script is not modified.

1. Run the helper:

python <this-skill>/scripts/decision_graph.py query "<question>" --cwd "$PWD" --mode why --limit 5

Use --mode alternatives, --mode revisit, --mode impact, or follow the free-form fallback order when the question calls for it. If the user names a project slug, pass --slug <slug>. If the user or config provides a vault path, pass --vault <path>.

1. Read the JSON evidence pack.
2. Answer from the evidence pack only.

Output Contract

Return a concise answer with:

• A direct answer when answerable=true.
• The matched decision node ids and source timestamps.
• matchedterms, evidencestrength, and answerability_reason.
• sourceentryrefs for every cited decision.
• Direct evidence fields (evidencerefs, sourcerefs, and

directartifactrefs source-of-truth evidence) when present on a cited decision.

• Rejected alternatives when relevant.
• Open questions when relevant.
• confidence and inference_notes when a node is inferred.
• Suggested repo-local docs from suggested_docs, if any.

If answerable=false, say the current Tracework records do not contain enough evidence. Include missing_evidence and suggest capturing the decision with /tracework:capture after the user clarifies it. Do not invent an answer.

Evidence Rules

• Treat confidence=explicit as directly recorded raw-entry evidence.
• Treat confidence=inferred as useful navigation, not a proven fact.
• Treat sourceentryrefs as provenance: they prove where Tracework recorded a

claim, not that the claim's outcome was independently verified.

• Treat evidencerefs, typed sourcerefs, and directartifactrefs as

direct evidence. General artifact_refs are navigation hints unless the raw entry explicitly marked them source-of-truth. Preserve both kinds so readers can drill down without overstating verification.

• Treat artifact dossier fields from {vault}/raw/artifacts/{slug}.json as

navigation plus recorded context. artifactsummary.keyclaims can explain what a linked artifact was believed to cover, but only evidenceboundary: directevidence, raw artifactcontext.sourceoftruth, evidencerefs, or typed source_refs can strengthen verification.

• Treat evidence_strength=weak as a prompt to hedge or ask for more evidence

even when answerable=true.

• An explicit, well-matched raw entry without direct evidence remains

answerable, but it must not be called strong; include its missing_evidence instead.

• Preserve uncertainty in the wording when inference_notes are present.
• Do not quote or cite a decision without its sourceentryrefs.
• Edges and supporting nodes are context expansion, not standalone proof.

Storage

This skill reads:

• {vault}/raw/decisions/{project-slug}.json when present
• {vault}/raw/weeks/{YYYY-WNN}/{project-slug}.json as rebuild fallback
• {vault}/raw/artifacts/{project-slug}.json as optional dossier navigation and recorded context

This skill writes no files.