[gmickel]

Why

The tier call that already runs before each worker spawn now sets the worker's reasoning effort as well as its model: one level below the host default for a confident mechanical answer, one above for a confident intelligent one, the default otherwise and without a judge. No new judge question or call; requested, routing-block and role-pinned efforts still win.

What changes for a user or operator

Workers on routine tasks spend less thinking time and workers on hard tasks more, at the same model. The Tier: line in each done summary names the effort applied, or why it was not applied. Runs without a TypeSafe key behave exactly as before.

Scope

23 files changed; +452/-35 lines; 8 generated, 3 mechanical.

1. Effort never followed the tier

Check the spec's precedence and fallback rules: the tier answer already existed but only chose the model.

2. flowctl computes the worker effort

Check judge_tier_effort's ladder step, bounds, floors and precedence, and that scouts (no --task) are unaffected.

• plugins/flow-next/scripts/flowctl.py : judge_tier_effort and spawn_effort on the tier result; four new judge flags; bridged implementers get '<model> at <effort>'. [R1, R2, R3, R4, R5]
• plugins/flow-next/tests/test_judge_consumers.py : Covers bounds and floors, precedence, no-judge default, not-applied lines, bridge transport and the CLI path without a key. [R1, R2, R3, R4, R5]

3. The conductor carries the effort to the spawn

Check that spawn_effort reaches the host's effort parameter, the bridge gets it on the IMPLEMENTER line, and a rejected level retries once at the default.

• plugins/flow-next/skills/flow-next-work/references/judge-tier.md : Effort flags to pass, spawn_effort into the spawn args, the one-retry rule. [R4, R5]
• plugins/flow-next/skills/flow-next-work/references/multi-task.md : Section 3c passes model and effort; the bridged IMPLEMENTER carries 'at <effort>'. [R4, R5]
• plugins/flow-next/skills/flow-next-work/references/worker-bridge.md : The stage line records the effort passed on the bridge command line. [R4, R5]
• scripts/sync-codex.sh : The Codex SECTION3C copy of multi-task 3c matches the new wording. [R4, R5]

4. Docs and reach pages

Check the judge and flowctl docs describe the new flags, and the reach pages state which host versions take a spawn effort.

• plugins/flow-next/docs/judge.md : Tier preset documents effort selection and its fallback. [R4]
• plugins/flow-next/docs/flowctl.md : The four new judge flags. [R4]
• plugins/flow-next/docs/reach/claude-code.md : Spawn effort exists on 2.1.296, absent on 2.1.283. [R4]
• plugins/flow-next/docs/reach/codex.md : reasoning_effort takes the model's own levels; a spawn without one runs at the session effort. [R4]
• CHANGELOG.md : Unreleased entry for 9.0.0. [R4]

8 generated files

5. Flow records

Spec, task and receipt bookkeeping, plus one bug memory entry on dispatch settings reaching bridged children.

• added .flow/memory/bug/integration/dispatch-settings-must-reach-the-2026-10-17.md : Dispatch settings must reach the bridged child, not only the native spawn. [requirement undeclared]

3 mechanical files

6. R6 timed evidence

Check the open item: pass rate held, the mechanical wall-time saving is directional but not confirmed, and the end-to-end control never exercised the feature.

Coverage: R1 → group 2; R2 → group 2; R3 → group 2; R4 → groups 2, 3, 4; R5 → groups 2, 3; R6 → group 6

Blast radius

Read judge_tier_effort in flowctl.py first (the whole decision), then judge-tier.md for how the conductor passes spawn_effort. Scouts and reviewers are untouched. Unproven: the wall-time saving on mechanical fixtures (R6, below), and effort on hosts other than Claude Code 2.1.296+ and Codex 0.146.0+.

Verification

• Implementation review: codex, 3 rounds: NEEDS_WORK, NEEDS_WORK (R6 only), SHIP
• Full suite: 218 files, 4055 ran, 0 failures
• Ruff, Codex sync check, doc anchors: clean
• Live tier call with judge key: Tier: mechanical (jev 0.99) -> claude-sonnet-5-5 at low effort (host default medium)
• Tier call without key: at default effort
• R6 replay pass rate: 15/15 at low vs 15/15 at medium (opus-5.5, 3 mechanical fixtures x 5 reps)
• unverified: R6 mechanical wall time: medians 55/43/131 s vs 76/67/138 s; pre-registered verdict INCONCLUSIVE (gap 10.0 s vs 9.7 s floor)
• unverified: R6 end-to-end control: review rounds 2 vs 2, wall 1,886 vs 1,992 s; no draw spawned a worker, so null by construction

Tradeoffs

Decisions: (1) the 0.8 confidence floor that gates the model tier gates effort in both directions, so an intelligent answer below 0.8 keeps the default. (2) --explicit-effort carries the invocation's effort, else the routing block's implementer line; naming only a model keeps the tier effort (R2, for callers that always pin the implementer model). (3) A role-pinned effort applies only to an in-host spawn, never to a bridged implementer. (4) A rejected effort level is retried once without effort; that rule lives in conductor prose because flowctl cannot observe the rejection. (5) Review round 1 found two paths where a bridged implementer never received the effort; fixed in 3fca663.

Open items

R6 is not met as written. Worker replay (claude-opus-5-5, 3 judge-confident mechanical fixtures x 5 reps per arm): pass rate 15/15 at low vs 15/15 at medium, medians 55.0/43.0/130.9 s vs 76.1/67.2/138.0 s, but the pre-registered paired verdict is INCONCLUSIVE (pooled gap 10.0 s against a 9.7 s floor, within-cell spread 29-88 s). The end-to-end large-feature control (median 2 vs 2 review rounds, 1,886 vs 1,992 s) is null by construction: no draw spawned a worker or made a tier call. About 1% of historical tasks get a confident mechanical answer and 2% a confident intelligent one. Accepting this, or running a multi-task --auto study on Claude Code 2.1.296+, is the maintainer's call. The fn271 harness pins Claude Code 2.1.283 (no spawn effort); an opt-in FN271_CLAUDE_DIR override sits uncommitted in agent-evals. flow-next.dev docs for this change ship with the 9.0.0 downstream walk.

https://claude.ai/code/session_01LbHwMnkD6ovBw6n8cnD8hQ

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 93e6d63. Configure here.}