Skip to content

t1k:estimate

FieldValue
Modulet1k-delegation
Version2.12.24
Effortlow
Tools—

Keywords: duration, estimate, eta, finish time, forecast, how long, plan eta

/t1k:estimate
<plan-dir>

Gives a plan an ETA from recorded spawn history: the model decides what each phase maps to, the CLI does every number. Runs inline in the main session (the planner); only the per-unit lookups are fanned out, to t1k-estimator.

The CLI is called automatically: at plan finalization (plan-eta <plan-dir> --auto --record, the planner), at cook start and at team-cook start (--auto, print only), and by hand on request. It is never on the development critical path. The t1k-estimator hop is optional and explicit. features.spawnEta: false is the kill switch (plan-eta prints disabled; --auto prints nothing); planEta.autoShow: false silences only the automatic call.

The pure CLI path is the default. bun .claude/scripts/t1k-spawn-eta.cjs plan-eta <plan-dir> (no project copy → bun ~/.claude/scripts/t1k-spawn-eta.cjs) maps every phase itself from the plan’s stated agent, model and brief text, costs zero tokens and needs no agent hop. The t1k-estimator hop (step 2) is OPTIONAL and runs only on an explicit user request. There is no enforced per-hop timeout (the Agent tool has no timeout parameter): a unit that has not returned by the time the CLI result is ready is treated as unknown, and the plan-level figure is reported without it.

  1. Read the plan (plans/<dir>/plan.md and its phase files). For each phase note the stated agent, model and Lane: label. If every phase states agent and model and the brief text is in the phase file, skip step 2 for it: the CLI can map it itself, zero tokens.
  2. Fan out ONE t1k-estimator unit per phase (or lane) in ONE message. Read-only, so one tree is safe. Cap 8 per message at depth 0 (forkDepth.fanOutCaps); a plan with more phases goes in a second wave. Each brief names the phase file, unit id, the stated agent and model if any, and the delivery channel (final message). Estimators are cheap-tier and routable; the brief needs no transcript.
  3. Handle each return. A well-formed ## Estimate line is used as is. A failed, empty, malformed or not-yet-returned hop makes that unit unknown; re-spawn it ONCE natively, and if it still fails leave it unknown. Never fill in a number by hand.
  4. Run the combiner: bun .claude/scripts/t1k-spawn-eta.cjs plan-eta <plan-dir> (no project copy → bun ~/.claude/scripts/t1k-spawn-eta.cjs). It prints ONE headline done by HH:MM <tz> [schedule; method; gates; delivery] n=<n> basis=<basis> validated=<evidence> (human gates and delivery are counted inside), then the sequential and parallel schedules compared, the critical path and the per-phase rows; or unknown naming the thin phase.
  5. Report the CLI output verbatim with its n, basis, validated= (or unvalidated) and any thin/unknown label. Never add a figure the CLI did not print, never drop the validated= tail, never turn an unknown into 0. Flags: --schedule sequential|parallel, --max-parallel <n>, --record, --auto, --no-gh.

No quantile, percentile, critical-path or finish-time code lives here or in the agent. The single source is .claude/hooks/lib/spawn-eta.cjs (per-unit estimate()) and .claude/hooks/lib/spawn-eta-plan.cjs (the combiner), reached through t1k-spawn-eta.cjs. Design: the CLI emits facts, the model decides which phase maps to which key.