Reuse
Detects sessions that re-plan the same work and exports reviewable skeleton templates.
Reuse detects clusters of sessions that share a planning skeleton, meaning the structural shape of the agent’s opening plan. When the same “patch release” or “triage” workflow gets re-planned dozens of times, each run re-buys a plan whose tool sequence was identical and only a version string changed. Reuse names that waste and exports each skeleton as a reviewable template.
tj optimize reuseReuse never reuses a plan for you. It reports “these plans look structurally identical.” You review the skeleton and decide whether to template it.
How it works
For each session, Reuse finds the first LLM call that precedes any tool call, the model’s opening plan. It then clusters sessions by a signature, and the signature sharpens as more capture is available:
- Mode 1 (always on). The ordered tuple of tool names following the planning call. Works against any telemetry, including raw Claude Code backfills with no capture toggles set.
- Mode 2 (
[capture] prompts = true). Also hashes a variable-stripped prefix of the planning prompt. Digits, ISO dates, and paths are normalized first, so “release v0.3.4 on 2026-06-15” and “release v0.3.5 on 2026-06-17” hash identically. The cluster key becomes the intersection of tool sequence and prompt prefix, which splits apart unrelated tasks that happen to share a tool sequence.
A cluster surfaces only when it clears all three thresholds: at least 3 sessions share the signature, average planning output is at least 200 tokens, and at least $0.01 is recoverable.
Two recoverable numbers per cluster
Each cluster carries two framings side by side:
| Field | Heuristic | Meaning |
|---|---|---|
| cache-reuse | avg_planning_cost × (repetitions − 1) | Recoverable going forward by reusing the existing skeleton. Conservative — you already paid once. |
| script-replacement | avg_planning_cost × repetitions | Upper bound — replacing every planning call with a template eliminates all of it. |
The aggregate figure that the Lens Overview reads uses the conservative cache-reuse number. All dollar figures flow through the shared framing layer, so subscription users see token-share framing and local users see token counts.
Report
tj report --reuse # all agents, 30d window
tj report --reuse my-agent # scope to one agent
tj report --reuse --since 7d # custom window
tj optimize reuse --export-templates # Markdown skeletons only, no browserThe HTML report renders one section per cluster with the tool signature, both recoverable numbers, and the skeleton with variable regions highlighted as {{slot_1}}, {{slot_2}}, and so on. Each cluster also gets a Markdown sidecar you can paste straight into a slash command or saved prompt. Rendering the skeleton text needs [capture] completions = true; without it the clusters and numbers still show but the skeleton is replaced by a hint.
Honesty caveat
Reuse confidence is heuristic, and every figure is estimated recoverable. It measures that planning calls repeat, not that the plans were interchangeable. The variable-slot highlighting is itself an honesty surface: when slots cover most of the skeleton, you can see the match is mostly placeholders. The caveat appears on every surface (CLI, HTML, Markdown):
Structural skeleton match, not a guarantee the plans were interchangeable. Review the templates before reusing them.