Repository navigation
Releases
Releases · archid21/rightsize
rightsize 0.14.0
Cache and compacts
- Keep warm only while it pays. The default of Before the cache goes cold is now "while it pays" (it was keeping warm until your usual start). rightsize records every idle break, whether you came back before the cache expired, after it, or not at all. At each point a keep-warm request is due, it sends one only when your own chance of coming back before the next expiry covers its cost, at most 2 per break; until it has enough breaks on record it sends one request at the first expiry on a 1-hour cache and none on a 5-minute cache. With no request due it compacts when that pays, else saves a whatsnext handoff note. On a 1-hour cache the band shows the countdown with an Extend 1 hour button. If the rule's own runs cost more than they saved, it pauses itself and says so. A value you set in /config still behaves as before.
- Keep warm stops on a cache miss. After each keep-warm request, rightsize checks that it read the conversation from the cache. If it read under 80%, the cache missed: keep warm ends, says why in the log, and sends no more requests.
- Wrap up, and the handoff text for every bare compact. A
/compactwith nothing after it, and Claude Code's automatic compact, now use rightsize's handoff text; text you type after/compact, whatsnext's text and another plugin's are never replaced. The text comes from a PreCompact hook, so it also works where Claude Code keeps plugin modules away from compaction (Team and Enterprise logins); it needspython3. On the countdown row of keep warm while it pays, a Wrap up button ends keep warm and saves a whatsnext handoff note, or compacts when whatsnext is not set up in the project. Typingwrap upin a project where whatsnext is not set up now compacts too, instead of sending those words to the model. When rightsize compacts where a note would have done, a hint says so: to install whatsnext (at most once in 30 days; Don't show again in the ⋯ menu hides it for good), or to runwhatsnext init(once per project). On its first start rightsize also turns your earlier runs into idle breaks, so keep warm has a record to judge from straight away. - rightsize's own compact writes a handoff. When whatsnext is not installed or cannot compact, rightsize now asks Claude Code's compaction for a handoff under eight headings: goal, next action, state, files, decisions, your instructions, open questions and dead ends. It quotes your words where they are unclear and drops tool output and file contents. The text was chosen by testing it against the old one on saved sessions. With whatsnext, its own compact text is used as before.
Cost warnings and estimates
- A warning for unusually expensive turns (opt-in). After a turn that cost more than 99% of your own turns, the band asks once whether to warn you before a prompt that looks that expensive (Advanced · Big-turn warning: off, auto, $1, $2 or $5; off by default). With it on, a prompt whose estimate reaches the amount is held for the Question wait with Run it, Plan first (the prompt goes on with a request to propose a short plan and wait for your go-ahead) and Lower effort (only where rightsize would lower that kind of work now). A turn that is heading for the amount shows a band row with the estimate, what it has spent so far, Keep going and Finish up; Finish up puts a message in your prompt box asking Claude to finish the step and report, for you to send, and nothing is ever sent for you or stopped. If you stop that turn yourself, a request for a report is put in your prompt box too. rightsize learns as you go: early warnings can be false alarms. Non-interactive runs (-p, including --max-budget-usd) are left alone.
- Estimates learn in dollars. Calibration now weighs each turn by what it cost, instead of averaging how far off each turn was. A cheap turn that ran ten times its estimate used to pull every estimate up; now an expensive turn counts for more, which is what a route choice compares. Your existing calibration carries over and shifts as new turns come in. After a turn's first 5 requests, rightsize also projects its cost again from its own pace and logs it beside the first estimate.
- Fast mode is priced. rightsize knows fast mode's prices and keeps them current with the price check. It reads Claude Code's
fastModesetting and checks each request's real cost, so it sees fast mode even when the setting is not the whole story. Estimates, keep warm and the report use fast rates while fast mode is on. When you turn fast mode on or off, or a model switch does, the band says what reading the conversation again at the new speed costs. On a subscription, the report says fast mode spend is paid from your usage credits.
Report
- How each task ended. Every finished task now has an outcome: tests passed last, 👍, corrected, undone, 👎, tests failed last, or unknown. Your own signals count first. The Whole tasks section in
/rightsize report detailscounts the outcomes, with unknown shown as unknown and never as success, and shows for each route (unchanged, lowered, raised, handed off) how many tasks were clean and the median cost of a clean task. - The rater can say it is not sure. When it cannot tell how hard a request is (too vague, or it depends on something it cannot see), rightsize leaves the prompt as it is: no lower or higher effort, no hand-off, no switch. The log says so, and the report's details count how many prompts it left alone this way.
- Why lowered turns stepped back up. When a turn rightsize lowered goes back to your effort after a tool error, the log and the run log now say which tool failed and what kind of error it was: a failed test, a missing file, an edit that failed, another failed command, or other.
/rightsize report detailshas a new LOWERING table: for each kind of work and model, the lowered turns, how many stepped back up and the most common cause, the real failures (you corrected, undid or 👎 the turn, or its last test failed), the miss rate the pause uses, and whether lowering is on or paused. When rightsize lowers or steps up has not changed.
Helpers
- A busy helper is not stopped at a fixed count. A helper on Haiku or Sonnet with no estimate for its task used to stop at 25 or 40 tool calls however well it was going. Past that count it now goes on while it is making progress (a file change or a passing test run in its last 10 calls) and stops when it is not, or at 4 times the count. A helper that repeats itself is still caught at once by rightsize's stuck signs.
rightsize 0.13.0
Report
- The report leads with a checked saving.
/rightsize reportopens with the overall saving as a range and a share of your cost. It counts the prompts rightsize left alone too, and checks the saving against what staying really costs. Prompts you overrode (advice you declined, or a change you undid with Retry at original effort or Run on your model) are left out. - A saving check on the prompts rightsize left alone. The details show how far its "what staying would cost" prediction is off, per kind of work, with Estimated and Checked columns.
- A shorter report. Fixed context is one line that points to
/usageand/trim:context. The raw savings estimate and the 30-day summary moved to/rightsize report details. - Where output cost goes. One line gives output's share of cost, split into thinking, text and code written. When thinking is the largest part, it says the Lowest effort and Highest effort rows control it.
- How rightsize tunes its estimates.
/rightsize report detailsshows when the calibration was last saved, how many runs it has measured and a factor per kind of work. - Did it work? After you change a setting that alters what rightsize does, the report compares your prompts after the change with the days before it: cost, requests, tool errors, failed tests and time per prompt.
Cache and watch only
- Cache-life advice. Once a month, at a turn end, the band prices your last 30 days on a 5-minute cache with keep warm and on a 1-hour cache. If the cheaper one is not the life Claude Code uses now, it asks once whether to switch. It shows only on an API key, a cloud provider or gateway, or usage credits, and writes only on your answer.
- Watch only: a row after 7 days and a summary. After 7 full days of watch only, one band row shows what the Saving check says rightsize would have saved, with Turn on, Ask before changing and Keep watching. When you turn watch only off, one row sums up the same range.
Helpers
- A helper has a spend limit, not a call count. A cheaper helper may spend 1.5, 2 or 3 times what its task is estimated to cost on your model. At 80% of the limit, while it is still finding new things, the new Advanced · Long helpers row decides: ask (the default), continue while making progress, or stop at the limit. A stopped helper cannot be resumed; the main chat starts a new one with the list of what is left.
- An unanswered helper question never loses the work. With no answer, the helper goes on while it makes progress, up to 10 times the estimate, then stops and sends its report back.
- A stopped helper can hand back its report. The last call a stopped helper makes is allowed, so its report reaches the main chat and the work is not redone.
Settings
- Six plain rows. Mode, Routing, Hand-offs, Ask before changing, Cache and Band come first. Question wait, Price check and Rater are now fixed values (30 seconds, monthly, Sonnet) with no row; a value an earlier release stored is kept and used.
Full list: CHANGELOG
rightsize 0.12.0
- Start fresh, next to Compact now. With whatsnext installed, the context-size warning offers Start fresh beside Compact now when at least 40K tokens of the window are left. Start fresh has whatsnext save a handoff note (it skips that when the note is already fresh), then the row says
Handoff saved. Keep going here, or start a new session from it.with the new-session prompt, and three buttons: Compact and continue here, Copy for new session and Not now. In the terminal, the row gives theclaude "..."command to run in a new window instead of Copy. Typing your own prompt clears the row. A plugin cannot start a new session as you, so rightsize gives you the prompt to paste and starts nothing itself. - Compact now says what it did. After a compact from the warning, the row reads
Compacted. Each turn now costs less. - A Concise output suggestion. When Claude's written output is a large share of what a session costs (at least a quarter, after 20 requests and $1), the band offers Claude Code's Concise output style: This session (set back when the session ends), This project, or Not now, with a hint that
/configsets it everywhere. It writes only the project's local Output style setting. - trim's savings in
/rightsize report. With trim installed, the report adds TRIM SAVINGS (tokens, estimated): tokens saved per trim feature, all time and the last 30 days, with fixed estimates marked "assumed". rightsize only reads trim's savings files; it never writes them. - A model switch fits the session's real window. A per-turn switch is capped at the window Claude Code reports for the session when that is smaller than the model's listed window, and no switch happens when the conversation is larger than the target model's window.
- Fewer false "stuck" raises. A tool call refused by another plugin's hook is steering, not a sign Claude is stuck, and bookkeeping tool calls (task lists and the like) are left out of the stuck signs.
- Asks give one reason and four answers. Each ask (effort, stronger model, subagent, stuck) gives one sentence on why, then once, for this session, keep, or always. Remember choices is gone; earlier saved choices move to the matching rows once.
- SessionStart hook output is measured from the messages. The report's fixed context counts it from the session's messages, and says "not measured" when it cannot read them.
- The cache offer counts down with a colored bar. The offer before the cache goes cold now shows how much of the cache life is left:
Goes cold in 42 min ▰▰▰▰▰▰▰▱▱▱ (14:55) · coming back after costs about $0.63. The minutes and bar are green with over half the life left, yellow down to a fifth, and red under it; the words read the same without color. - A shorter keep-warm row. It drops "so far":
Warm until 8:30 · ~0.1% used($0.04on the API). It shows no countdown: keep warm refreshes the cache before it goes cold, so the time that matters is when keep warm stops.
rightsize 0.11.0
- A welcome table after install. The first session after install opens a pane with what rightsize does: the short version in three steps, then a table of what you get and how it works, with the five features to learn first. Esc closes it. Claude Code places a pane nobody asked for only in a terminal of 144 columns or more; narrower, and in the mobile app, one line says rightsize is on and names
/rightsize welcome, which shows the same table in the transcript, drawn as a Markdown table in the terminal and the desktop app. - Tips on the band. At the end of a turn, when the band has nothing else to say, one tip shows on its own Tip row:
/rightsize reportonce rightsize has saved you something, Retry at original effort once a turn was lowered, then marking a call 👍 or 👎 and trying watch only. Next tip shows the next one, Stop tips turns them off, × hides one, and/rightsize tipsbrings them back. No tip shows while a turn runs, next to a question or a cache offer, or with Band set to Never. - One ask for a GitHub star, shared with whatsnext. The last tip asks you to star rightsize. It is offered on its own once, at a turn end after rightsize has saved you something, when your own
ghsays the repo is not starred. Star on GitHub stars it withgh; if that fails, the repo page opens. No thanks ends it for good. With whatsnext installed, one ask names both repos and one answer covers both, through~/.config/archid21/stars.json. Withoutgh, signed out, offline, or withCLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFICset, rightsize does not ask. This is rightsize's only contact with a third party; PRIVACY.md and WHAT-IT-TOUCHES.md list it. /rightsize reportis shorter and says what the numbers mean. The report now shows three things: what you gained, what rightsize did, and what to change. On a subscription the headline is hours of Claude's work, for exampleYou got about 22–23 more hours of Claude's work from your plan., then the 5-hour windows that stands for and its value at API prices (under one window, the share of one). On the API it is dollars and the share of your cost. YOUR SAVINGS is one table with Today, 7 days, 30 days and Lifetime: extra work time, 5-hour windows and weekly limit on a subscription, then the value, the share of cost saved and the extra on hard prompts. It fits in 80 columns. The Session column is gone; the band still shows the session total. 30-DAY SUMMARY lists prompts rated, left as they were, effort lowered, sent to a cheaper model, effort raised, sent to a stronger model, subagents moved, suggestions you declined and keep warm, one row each. RECOMMENDATIONS replaces Next steps: numbered actions (cut fixed context, with its cost and largest parts; a cheaper start; keep warm, when it cost more than it saved), then a line for a start that needs no change and one for a paused hand-down. A short footer says the figures are estimates and points to the details./rightsize report detailsholds the sections behind the numbers: Starting point, Fixed context, Accuracy (now with close calls), Not counted as savings, Time and Whole tasks, with the full footer./rightsize helplists it.- The report's tables line up in the desktop app.
/rightsize reportand/rightsize report detailsnow pick their layout when the row is drawn: off the terminal the tables go in code blocks, so Today, 7 days, 30 days and Lifetime stay in their columns. Before, the desktop app could get the terminal layout, whose columns need a fixed-width font. - Not counted as savings stays within 80 columns. A turn raised because Claude looked stuck lists each sign on its own line, and a long reason or sign wraps under itself instead of running past the edge.
rightsize 0.10.2
- Fable 5.2 prices. Fable is now priced as Fable 5.2: $8 input, $40 output, $0.20 cache read, $10 and $16 for 5-minute and 1-hour cache writes (USD per million tokens). The automatic price check reads the Fable 5.2 row.
- Older models get their own prices. rightsize priced each family at its newest model, and older versions cost more, so its costs and keep-warm and compact choices were wrong there.
prices.jsonnow has aversionslist with a row for each older version the pricing page lists (for exampleclaude-opus-4-8at $5 input and $25 output), and the price check keeps it current. A request is priced by its version's row wherever the version sits in the model ID, so Bedrock and Vertex IDs match too. On an older model an effort change writes the whole conversation again, so rightsize prices each switch that way. A main chat on an older version with no row pauses rightsize, as if you turned it off: the band says so once, the status line says why, and/rightsize pricesfetches the current list. Switching to a priced or current model turns it back on. Helpers are not checked. - The cache's life is read, not assumed. rightsize took the prompt cache's life from
prices.json(1 hour unless edited), so on an API key, a cloud provider or a subscription drawing on usage credits, where Claude Code writes a 5-minute cache, keep warm sent its first request at 57 minutes to a cache that had died at 5, and the idle warning waited an hour. It now reads the life the way Claude Code decides it:FORCE_PROMPT_CACHING_5M,CLAUDE_CODE_PROMPT_CACHE_TTL, thepromptCacheTtlsetting,ENABLE_PROMPT_CACHING_1H, acacheTtlyou put inprices.json, then the billing the rate limits show (1 hour on a subscription within its plan, 5 minutes otherwise; a subscription window at 100% means usage credits). Helpers, Claude's subagents and forks are their own bucket, 5 minutes unlesssubagentPromptCacheTtlor its variable says 1 hour, and a helper's cache writes are now priced at 1.25x input, not 2x. The life is read at start, when the billing is learned and after each turn, logged when it changes, and shown by/rightsize status. The shippedprices.jsonno longer carriescacheTtl; one you add still speaks for a gateway rightsize cannot read. - Keep warm inside a turn, on a 5-minute cache. A tool call that runs longer than the cache lives (a test run, a wait for CI, a subagent's work, a question left open) let the conversation go cold in the middle of the turn, and the request after it wrote the whole conversation again. When the cache lives under 1 hour, rightsize now sends its keep-warm request 4 minutes after the main chat's last request and every 4 minutes while the call runs. It follows the same bars as keep warm between turns: Cache on, Before the cache goes cold not off, a cold return that costs at least the Idle warning amount, and never more requests than the cold request would cost. It starts from a turn's second request on, not on the first tool call after your prompt and not after you type into a running turn, because there the request would miss the cache. When the call outlasted the cache and the next request read the conversation from it, the saving is counted like any other keep-warm return. Nothing shows on the band; the log and the report's Keep warm line carry it. A 1-hour cache is left alone.
- A helper's long foreground wait is refused. rightsize cannot keep a helper's (subagent's or fork's) cache warm, so a helper that waits on a foreground Bash call longer than its 5-minute cache lives lets the cache die, and its next request writes its whole context again. A helper's foreground Bash call with a timeout over 270 seconds (4.5 minutes, 30 seconds under the cache's life) is now refused with a hint to run it in the background or use a shorter timeout. A call with no timeout (Claude Code's default is 2 minutes), a background call, a helper on a 1-hour cache (
subagentPromptCacheTtlor its variable) and the main chat are left alone; the main chat has in-turn keep warm. - The report names the memory files and counts SessionStart hook output. Fixed context's Memory row now names your largest memory files (
CLAUDE.md,MEMORY.md, rules files) with their sizes, the way the MCP tools row names servers. A new SessionStart hooks row counts the text that your settings' and plugins' SessionStart hooks add to every new session. It sits in the conversation, so every request reads it again from the cache, but/contextshows it in no row of its own. The row gives the total, not the source: Claude Code does not say which plugin wrote it, so look in/plugin. A memory file or the SessionStart hooks of 10,000 tokens or more (about the 40,000 characters at which Claude Code itself warns about a large CLAUDE.md) is flagged with aLarge:line. The Memory advice now says to trimCLAUDE.mdandMEMORY.md(/memory). - An automatic permission denial is not a tool error. In a session with nobody to answer a permission prompt (
claude -p, CI), Claude Code denies a call that needs approval by itself. rightsize counted that as the work failing, so a turn it had lowered went back to the session's effort level. It now treats the denial like your own "no" at the prompt: the lowered level stays, and the stuck signs do not count it. The same goes for a write your permission settings deny ("File is in a directory that is denied by your permission settings"). A real failure after it still brings the level back. - A lookup that misses a file is not a tool error. Claude often reads around with one line of several lookups and guesses a path that is not there (
grep … | head; head CHANGES.rstin a repo that hasCHANGES.md). The line ends with an error, and one such miss sent a lowered turn back to the session's effort level. A Bash line made only of read-only lookups (grep,cat,head,ls,sed -n,find, read-onlygitand the like, with no output written to a file) that failed on "No such file or directory", or that sent its errors to/dev/nulland ended with exit code 1 or 2, now counts like a search that found nothing. A grep-style search in a;list is now judged by the list's last command, whose status the shell returns, not its first. - The hand-off brief is smaller and tells the helper when to stop. The hidden note that tells Claude how to brief a cheaper subagent now asks for what is out of scope, and for code as
file:lineranges or short excerpts, never whole files pasted in: a smaller brief costs less to cache, and every later helper step reads it again. The note also tells the helper that when it meets a choice the brief does not settle, or something that contradicts the brief, it stops and reports the question with its options instead of guessing, and Claude decides and continues it. - Keep warm or compact is priced over the time keep warm would run. Before the cache goes cold, rightsize compares keeping the cache warm over your usual idle time with compacting. It priced the requests over your usual idle time even when keep warm would stop sooner (its hours, or your usual start time), so it could choose a compact that keep warm would never have cost. It now prices the shorter of the two.
rightsize 0.10.1
Measured on our benchmark of everyday tasks from an Opus start (lookups, searches, a bulk edit, debugging, a refactor, features, a review, writing): the same model at low effort passes as often as at medium and costs about a fifth less, with no mid-turn raise needed. Long prompts that state a limiting instruction late, after text that contradicts it, are the one case still under test.
- Effort changes are priced as free on the current models. Anthropic's docs say an effort change keeps the prompt cache on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 with the Claude API or a Claude subscription (Claude Code 2.1.260 or later). rightsize assumed a small share of switches miss until 20 of your own had been measured, and that assumed cost blocked the lower-to-low rule on every small task: a lookup rated low on every model ran at medium. A switch is now free from the first prompt, so on Opus or Fable every task but a review runs at low from the start. Measured misses still price it again where a change does rebuild the conversation (a gateway, Amazon Bedrock, Google Cloud's Agent Platform, an older model). Under the Opus and Fable rule a lowering also pays for one switch at most, since the next prompt lowers too.
- The "tool calls failed" stuck sign waits for a full window with no progress. It fired on 3 failed tool calls anywhere in a turn, which is what the first steps of a debug turn look like (a repro script that fails is the work), and raised effort for nothing. It now fires only when the last 10 tool calls hold 3 or more failed calls and none of the 10 was a file edit that worked or a test run that passed. Test runs do not count as failed calls, as before.
- A search that finds nothing is not a failed tool call.
grep,rg,git grep,diff,testand the like exit 1 when they find nothing or the files differ, and Claude Code hands that back as an error. rightsize counted it: three misses fed the "tool calls failed" sign, the same search missing twice looked like the same call failing twice, and one miss stepped an auto-lowered turn back up. A Bash result that reports exit code 1 from one of those programs (the command's first program after anycd … &&, or the last stage of a pipeline) now counts as a success for every stuck sign and never steps a turn back up. Exit code 2 or more, and any other program, still count as errors. - The Haiku-start switch always asks. On a Haiku session, debugging, features, refactors and writing switch the main chat to Sonnet for the turn, and under the Stronger model rule "always send" that happened without a question. On our benchmark Haiku alone usually managed this work and switching every turn cost several times more, so this one move now asks on the band with both costs, and nothing switches until you answer. Review switches follow your rule as before.
rightsize 0.10.0
- 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 (…)orAsk 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) andidleCompactionis 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'swrapup.autoIdlesaves 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
longsign 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).
rightsize 0.9.0
- New tagline and README. "Make every Claude token count." The README opens with the benefits (same outcome at up to 30% lower cost from an Opus start; stronger reviews, possibly at a higher cost, from a Sonnet start; less manual switching; lower cache and compact costs), a feature table with how to use each feature, and a note that rightsize works better with whatsnext. The sections on how it works, what it runs and its limits now cover compacting, the context warning and whatsnext.
- Haiku is priced by prompt length everywhere. Over 100,000 tokens a Haiku request pays $0.50 input, $0.625 / $1.00 cache write (5-minute / 1-hour), $0.05 cache read and $2.50 output, against $0.10, $0.125 / $0.20, $0.01 and $0.50 below it; the prompt counts cache reads and writes. rightsize priced only a Haiku compact figure this way. It now prices every Haiku request by its own prompt: the cost of real responses, a Haiku helper whose prompt grows past 100K, and a cold return or keep-warm ping on a Haiku session.
prices.jsonhaiku.over100know holds all four prices, and the monthly price check compares all of them. Helper work that stays under 100K is unchanged. - whatsnext decides "installed". rightsize reads
projectfromcompact-instructions --jsoninstead of looking for.whatsnext/config.json. - One fallback compact for the limit button and the idle compact, with whatsnext's own keep/drop text.
- Wrap-up ordering. After
wrap up, at each turn end rightsize asks whatsnext again and compacts only when no wrap-up is missing; after 2 turns it gives up with a band note. With under 40,000 tokens of room it skipswrap upand says "no fresh handoff". - Keep warm vs compact by expected idle time. When Before the cache goes cold is the profile's keep warm, rightsize compacts instead if keeping warm over the usual idle time (median of the last 20 gaps in the project, at most 8 hours) costs at least 3 times a compaction and its summary.
- Context-size warning. At the end of a turn, when the conversation first reaches 40, 50, 60, 70, 80 or 90% of the window, the cache row says so with a Compact now button, once per step. Steps at or above Claude Code's auto-compact point are skipped. It compacts only when you press the button; ⋯ has Why? and Don't show again.
- No 200K ceiling on the idle note or the cache row.
- Plainer paused-hand-down text. When the pause is for prompts the rater could not sort into a kind, the report says "unclassified work" instead of "other work".
- Removed the last mentions of the old handoff plugin from the docs.
rightsize 0.8.4
- Every use of
$is now plain, ahead of the directory check. The check reports three lines per scan, so this release sweeps the whole plugin instead of waiting for the next three:retriesUsedis replaced by a plain state read and a pure helper,useChanceuses a lock instead of a closure, theconfig.sethook is one registration instead of a loop, six nested$uses are hoisted into their own lines, and the test kit no longer castsonor names a non-hook valuectx. Theconfig.setandcommand.runhooks pass errors on with.catch. No change in behaviour.scripts/check-plain-dollar.mjsnow scans the test files too and flags loop-registered hooks, aliases ofon, and any hook parameter name used elsewhere. - Plainer test text and README. Test fixtures no longer quote
rm -rf, a registry host or a sign-in URL; the README says the mod fetches no code, names theHOMEread, and says that theconfig.setandagent.spawnhooks pass every event on withnext(e).docs/WHAT-IT-TOUCHES.mdmatches.
rightsize 0.8.3
- The directory check can read every use of
$. Three places that dropped a field from the band destructured it into a variable named_, which the directory read as an unclear use of$. They now use plain names. No change in behaviour.scripts/check-plain-dollar.mjsnow flags destructuring into_. - Plugin manifest: the
typesfield is removed fromplugin.json(Claude Code ignores it, and the directory reported it as unrecognized).types/index.d.tsstays. - Test text: three tests that fake a failed network call no longer quote
curl, so the directory does not read them as download-and-run commands. The tests check the same things.