← ClaudeAtlas

retrieval-trajectorylisted

Record and replay retrieval trajectories (inspired by OpenViking's observable retrieval). After every context-budget / knowledge-retrieval / skill-selector execution, writes `.agent/runtime-evidence/trajectory/{task-id}_{ts}.jsonl`. Supports replay, URI resolution verification, and fixture export.
Kucell/cortex-agent · ★ 1 · AI & Automation · score 74
Install: claude install-skill Kucell/cortex-agent
# Retrieval Trajectory ## Goal Turn retrieval from a black box into observable data. cortex-agent adopts OpenViking's directory-browsing trajectory model: 1. **Record**: `context-budget/scripts/select.js` writes to `trajectory/{task-id}_{ts}.jsonl` during scoring and greedy fill. 2. **Replay**: `replay.js` reconstructs each decision step. 3. **Verify**: `--verify-resolve` resolves every `cortex://` URI back to a path, catching "URI stale, path stale" mismatches. 4. **Fixture**: `--as-fixture` exports a deterministic JSON for e2e tests. ## Location ```text .agent/runtime-evidence/trajectory/ └── {task-id}_{iso-timestamp}.jsonl ``` Each line is a JSON event with type: - `header` — task metadata (task_id / task / scoring_strategy / l0_l1_available) - `step` — one decision (scan / tokenize / score / promote / l0_only / load / skip / error) - `summary` — closing summary (step_count / promoted_count / total_tokens) Full schema in `trajectory.schema.json`. ## Usage ### 1. Auto-written by context-budget `context-budget/scripts/select.js` already writes trajectories in the main path. No extra setup. ### 2. Manual append (custom retrieval scripts) ```bash node .agent/skills/retrieval-trajectory/scripts/record.js \ --task-id T-DEMO-001 \ --step 1 --action scan --candidates 42 node .agent/skills/retrieval-trajectory/scripts/record.js \ --task-id T-DEMO-001 \ --step 3 --action promote --tier tier1 \ --uri "cortex://skills/context-budget" \ --tokens 1200 --reason "