Task-Scoped Context

codedecay context retrieves a bounded, inspectable packet of engineering context for one task. Use it when a user-owned coding agent needs the relevant routes, files, tests, requirements, memory, docs, owners, and proof references without receiving a whole-repository dump.

npx codedecay context \
  --task "Allow finance admins to retry failed payouts" \
  --format markdown

The command refreshes local analysis artifacts, reads repo-local context, ranks nodes with deterministic lexical and graph-neighbor matching, writes .codedecay/local/task-context.json, and prints Markdown or JSON.

What It Reads

Context retrieval uses existing CodeDecay evidence as source-of-truth input:

• requirements from --task and optional --requirements
• changed files, impacted areas, route/API impacts, symbol impacts, and test proof from the current analysis report
• normalized impact graph artifacts from .codedecay/local/impact-graph.json
• local .codedecay/memory.json
• .codedecay/config.yml design contracts, product endpoints, commands, and probes
• repository docs and ADR/RFC-style markdown
• CODEOWNERS entries
• package manifests
• latest local product/runtime verification evidence when analysis includes it

It does not execute configured checks, call models, generate embeddings, use network access, or send telemetry.

Output Shape

JSON output includes:

• query: task text, deterministic tokens, source revision, and node limit
• graph.nodes: selected context nodes with stable IDs, kind, trust class, confidence, provenance, limitations, and source revision
• graph.edges: selected evidence links such as serves, tests, depends-on, owns, and observed-by
• selected: ranked relevance explanations and evidence references
• rejected: near-miss decoys that scored below the selected bound
• safety: proof that no model, command, network, telemetry, or hosted service path was used

Trust classes distinguish current facts from context:

• current-revision-fact: current code, config, analysis, or runtime/tool evidence
• memory: repo-local memory that is useful but not proof
• historical-context: docs, ADRs, and similar context that may inform review
• stale-context: deprecated, superseded, outdated, or conflicting context
• ai-suggestion: untrusted agent text when present in upstream evidence

Memory and docs can guide investigation, but they cannot prove current behavior or merge safety without corroborating current-revision evidence.

Examples

Retrieve context before editing:

npx codedecay context \
  --task "Add a dashboard filter for reviewed uploads" \
  --max-nodes 16 \
  --format markdown

Retrieve context for an active PR diff:

npx codedecay context \
  --task "Change payout retry formatting" \
  --base main \
  --head HEAD \
  --format json

Use structured acceptance criteria:

npx codedecay context \
  --task "Add billing export" \
  --requirements .codedecay/requirements.yml \
  --format markdown

MCP

The MCP server exposes the same retrieval boundary as task_context.

Example input:

{
  "task": "Allow finance admins to retry failed payouts",
  "requirements": {
    "acceptanceCriteria": [
      {
        "id": "AC-1",
        "text": "A retry request enqueues exactly one worker job.",
        "requiredProof": ["API integration test", "worker idempotency test"]
      }
    ]
  },
  "format": "json",
  "maxNodes": 16
}

The tool is report-only. It writes the same local artifact and uses the same trust, provenance, and limitation fields as the CLI.