You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The per-start rules of the effort policy are now code. On Opus or Fable every task but a review runs at low effort, whatever the rating says (a start at high or xhigh lowers to low, not to the rated level). On Sonnet a rated raise is taken only for a task rated high or more with 20 steps or more, and never for debugging. A Haiku helper takes mechanical work only (lookups, searches, bulk edits, extraction, summaries); a helper on any other model needs a task of 20 rated steps or more. On a Haiku session, debugging, features, refactors and writing switch the main chat to Sonnet for the turn under the Stronger model rule, as a review switches to Opus. Claude's own subagents are not held to the type or size rules.
Six Advanced rows removed. Hand-off threshold (70%), Lowering threshold (90%), Explore close calls (off), Pause after misses (20%), Helper call cap (25 Haiku / 40 Sonnet calls) and Rater check (every 20th Haiku rating) are now fixed values; a stored value for any of them is ignored. Budget pacing still moves the two bars when a usage window runs hot.
Stuck sensitivity row removed. The stuck signs use one set of limits (the former normal values); Raise when stuck turns them off.
One rule set, two switches; the Profile row is gone. Quality, Balanced and Economy differed in a hand-off threshold, an evidence rule, retry counts and which cases ask, none of which the benchmark could tell apart. The Balanced values are now fixed (hand-off threshold 70%, pause at 20% misses with no minimum, Sonnet checks every 20th Haiku rating, 2 chances per task, keep warm until your usual start, reviews at up to Highest effort), each with its Advanced row. New plain row Ask before changing (off by default): on, every move rightsize would make on its own asks first; a rule row you set still wins for its case. A stored Quality profile reads as Ask before changing on until you set the row; Balanced and Economy read as the defaults. The band now says The default decided (…) or Ask before changing decided (…) where it said which profile decided.
No effort raise at prompt time on Opus or Fable. On our benchmark the rated level did not predict which tasks failed, and Opus at low effort passed as often as at medium or high. A rating that asks for more effort on Opus or Fable now changes nothing at prompt time; reviews still run at high or more, and the mid-turn stuck raise still fires. Sonnet keeps its rated raises except for debugging, as before.
A lowered turn stays lowered. An automatically lowered turn used to step back to the session's level after 5 tool calls, which gave the saving back on every task longer than that. It now stays at the lower level for the whole turn. A tool error still brings the session's level back, and the misses log still counts that.
The rater rates difficulty, not size. Its levels no longer map "multi-file changes" to high or "long multi-step plans" to xhigh: a change across many files or steps is rated by its hardest step, and the size goes in the step count. High needs a reason the rater can name.
Fixes that keep leaving the tests red raise effort. New stuck sign failed-check: a test run fails, Claude edits, and a test run fails again is one failed fix attempt; two in a row raise effort one level for the rest of the turn, under the Raising rule and the per-turn cap. How many calls a fix takes does not matter, and a red run repeated without an edit between is not an attempt. A task whose turn ended on a failing test run is not lowered or handed down on its next prompt.
Rightsize steps aside for Claude Code's idle compact. Claude compacts an idle session 54 minutes after its last request when the context is at or above 200,000 tokens (CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, at least 100,000) and idleCompaction is on, its default. Keep-warm pings do not reset that timer, and after the compact each one pays for the uncached summary. In that state, on a 1-hour cache, rightsize sends no pings, makes no offer and compacts nothing before the cache goes cold; whatsnext's wrapup.autoIdle saves the handoff. When a session has been compacted while you were away, the band says so on your return (Claude compacted this while you were away (210K to 18K tokens)., with the handoff path when whatsnext has one), and the return is never counted as a cache miss. Below the threshold, or with idle compaction off, nothing changes.
README: keep warm does draw a little usage on a subscription. It said the rebuild did not eat into your window. A keep-warm request draws a small share, far less than rebuilding the cache, and the band shows the running share.
A tool-call cap for cheaper helpers. A Haiku helper may make 25 tool calls for a task and a Sonnet helper 40. At the cap its next call is refused and it reports back, and the main chat finishes the work, as for a helper that is struggling. New row Helper call cap (tight, normal, loose, off). Opus and Fable helpers and the main chat have none.
The report adds Today and Lifetime. A second table, in the same rows as the first, shows today (since local midnight) and lifetime. Runs older than 30 days are folded into totals in the plugin's own store at session start, so the Lifetime column outlives the logs. Both are estimates, and the report says so and says the date lifetime counts from.
A test that keeps text off the cached prefix. A hand-off adds its note to the prompt's context only; the rater's system text and every keep-warm request are fixed. No behavior changed.
The context-size warning shows while keep warm runs. Keep warm starts at the end of a turn, just before the warning is checked, and the warning used to wait while the cache row was in use. On a profile that keeps warm, it was almost never said. It now joins the keep-warm row (warning first, then "Warm until …"), with Compact now beside Stop keeping warm. The usage-limit advice works the same way. Compact now also stops keep warm, since the summary replaces the conversation it was keeping.
A long turn raises effort only when it stops making progress. The long sign used to fire on step count alone, so a healthy turn working through a big plan (editing files, passing tests) was raised and then ran slower and cost more per step. Past 3x the rated steps (at least 15), rightsize now also checks the last 10 calls: it raises only when none was a file edit or write that worked or a test run that passed, and fewer than half were calls not made earlier in the turn. Reading many new files counts as progress. The check repeats on later steps, so a turn that stalls late is still caught. Relaxed looks at the last 12 calls (under 40% new) and Sensitive at the last 8 (under 60% new).