crew: route the worker, planner and checker per task (Pareto crew) by santoshkumarradha · Pull Request #1436 · Agent-Field/CodeAF

santoshkumarradha · 2026-09-24T16:39:23Z

Summary

A task's crew (worker, planner, checker) is now picked for each task by a router (internal/crewroute), instead of being fixed by a preset. By default all three seats are on auto, nothing is pinned, every model a connected provider serves is a candidate, and each task is held to a $5 limit. /crew is now an interactive panel for the few things that persist: pins, allowed models, providers and limits.

dev is merged into the branch (Merge dev into feat/pareto-crew); the PR's net change is taken from its fork point, so dev's own work is kept as dev has it. Three commits follow the merge: a fresh-profile test, the docs, and a timing-test fix. dev's OpenRouter-credits work (#1440) is adapted to the routed crew (see Routing).

User-facing behaviour

/crew panel with these rows:
- the three seats: auto · usually <model>, taken from the recent tasks (likely … before there is any history), or ⌖ <model>[@provider] when pinned;
- models: the allowed rule (all, open, a price ceiling, or a custom list);
- providers: one on/off chip per connected provider, plus the free-routes switch in the providers list;
- the cap row: per task $5 · daily none;
- today: what crews spent today and how many tasks ran.
Every change saves at once, ticks its row, and can be undone with z for a few seconds.
Shortcuts make the same writes and then open the panel:
- /crew pin <worker|planner|checker> <model[@provider]>
- /crew unpin <seat|all>
- /crew models <rule|+x|-x>
- /crew cap <$|off>
- /crew cap task <$>
A pin outside the allowed models is refused.
One-task effort: --best / --cheap. It can be given as the first word of /task, as the hand-off's effort field, or as a codeaf do flag. It moves that task only.
codeaf do also takes --pin seat=model[@provider] (repeatable). -json adds class, crew (per seat: model, provider, kind, pinned, est_usd), est_usd, effort, check_model and check_model_source. -yes-spend gets past the daily cap and the plan-price question. It does not lift the per-task limit.
Crew line: each task prints one line, updated in place. It shows the class, the seats and the estimate, then the actual cost beside the estimate when the task ends.
/redo stronger re-runs the last task one rung stronger. It also records that this class of work in this repo was under-served, so the next such task starts one step higher (at most three steps). Each accepted task takes a step back off.
/settings → Providers: the three seat rows are replaced by one seats row that opens the panel. The onboarding screen asks for the daily limit and the chat model only, never a crew.
The model you talk to (/model) is unaffected.

Routing (scoring, classes, fallback, limits)

Classes: bugfix, complex fix (a bugfix whose report shows reach in two or more ways: multiple files, API/protocol, language rules, long report or several repros, a security defect, existing tests that must keep passing), open-ended, and other. A complex fix moves its worker up one rung. On open-ended and other work, upgrades go to the checker first. A task that can't be read confidently counts as open-ended.
Scoring: learned weights (internal/crewroute/prior.json) turn each model's catalog metadata into a quality per seat and class, with a variance:
- a model that publishes indexes or ratings is scored on those alone;
- a model that publishes none is scored from context, release date, licence and family, and never above average;
- price is never read as ability;
- this install's own task outcomes (accepted, kept, redone, failed) move a model's score in a seat by small, bounded steps.
Seat choice: each seat goes to the model where quality minus λ times cost is highest, with λ at the knee of the curve, over the allowed models on connected, enabled providers. Plans and local models cost nothing to route to. Worker and checker must reach an ability floor whenever some allowed model does.
Limits:
- Per task: $5 unless set. Every priced call of the task (seats and helpers) counts against one total. A call that would pass the limit is not made, and the task stops with this task reached its $5 limit · raise it in /crew.
- Daily crew cap: optional. Crews pace toward it. At the cap a task does not start, and codeaf do refuses unless given -yes-spend.
- Checker: its own ceiling, 3× its estimate and at least $0.05.
Fallback: a seat whose first call is refused moves, inside the task, down this ladder:
1. the same model on another route;
2. a model at a similar cost;
3. the last crew that worked in this repo;
4. the chat model.
The line reads running on fallback crew · <seat> <from> → <to> (<cause>). When nothing paid is reachable, free routes are used with free routes in use (may log prompts). With nothing left, the task stops on a single action: add credit on <provider> to continue, reconnect <provider> with /connect, or waiting for the limit to reset.
Route health is kept per route:
- a route that refused a model is left out for a week;
- a rate-limited route rests until its reset;
- accounts out of credit and refused keys are avoided by helpers and re-probed by the next task's first seat call.
chat: know the OpenRouter balance, start a near-zero account on free models, keep the whole 402 sentence #1440 (low OpenRouter balance): the reflex and small-work rows still take their free defaults. For crew seats, a balance read as low marks OpenRouter as out of credit before any call, so seats go to free routes with the notice instead of spending a refused call. A top-up is picked up at the next launch.

Removed

The crew presets (frugal / balanced / max), the model-family row (models.crew.source) and the pick row (models.crew.pick: table / catalog / learn).
internal/crewpick and config/auto.go.
The built-in default models for crew seats. No seat falls back to a model this build chose.
The crew step in onboarding, and the crew, model family and picked from settings rows.
The check seat inheriting the planner's model.
dev's low-balance "free crew" table for the three crew seats (see Routing).

Migration

On the first launch, a profile carrying preset, pick or auto rows is migrated once and told so in one line. Preset, pick and auto rows become auto. Any model id written on a seat stays pinned (kept your pins: …). The open family becomes the allowed rule open. A profile with none of these rows is left untouched. Remote protocol moves from 17 to 18 (Task.Start carries effort, and Task.RedoStronger is new), and an older engine is refused at the handshake.

Tests

make check (with OPENROUTER_API_KEY unset, see #1489) at e5ffb1666: passed — vet, fmt-check, the full test suite (136 packages ok, 0 failures), test-packed-manual, size and build-cross. manual-gates, test-laws and changelog-check also pass.

New tests:

a fresh profile (empty HOME and CODEAF_HOME) starts with auto on all three seats, no pins, rule all, a $5 per-task limit and no daily cap;
no crew seat falls back to a built-in model;
a low OpenRouter balance routes the crew to free pools with the notice, and a healthy balance does not.

How to try it

make build
codeaf                      # /crew opens the panel
/crew pin checker moonshotai/kimi-k3
/crew unpin checker
/crew cap task 10
/task --cheap fix: <a small bug>
/redo stronger
codeaf do --best --json "fix: …"
codeaf do --pin checker=moonshotai/kimi-k3@openrouter "…"

🤖 Generated with Claude Code

A config file whose keys the reader does not know, including a nested object written where a dotted key was meant, read the same as no config at all, with no word to the person. Loading a profile now names every top-level key that nothing consumes, once per load, as a log line and never as an error, so a session still starts. A correctly shaped config says nothing.

Asked about work it had just handed off on the new road, the conversation's tasks tool answered that no task had run, because it read only the old task graph. It now lists a run's tasks from the run's own store, named the way the rail names them, a task by its number and a part by its place under it, each with its state and the first line of its result, and an id reads one of them. The shipped listing of earlier tasks is kept beside it. The manual no longer describes an ask box that no screen opens.

…ased (#1225)

…eanup (#1228)

…e walk (#1227)

…apsed slack (#1229)

… open their pages (#1231)

…ds (#1233)

…operties, with the verdict basis recorded (#1220) The worker declares its checks, and a holds verdict rests on a recorded, auditable, zero-exit run of every declared command. The worker records each command's own exit status, an absent status reads as unknown and earns nothing, and a record opens with a line that marks it as one that records exits, so a declared check that was never observed is refused, including when the record is missing or empty. A record written before exits were recorded still holds by reading and names the checks it did not observe. A does-not-hold verdict is ungated, a task with no declaration gets a reading verdict, and the verdict basis is persisted. One quote-aware reader of "is this one command" lives in internal/approval and is shared by the proposal door, the checker's runner gate and the store.

… callback ports (#1237)

…en it ends (#1232)

…cadence, and the harness that measured it (#1235)

…are PID (#1238)

…er returns (#1243)

…ver the conversation (#1244)

…g the ordering (#1246)

…ntroducing itself again (#1236)

…#1247)

…her's keys (#1248) A profile write now takes a cross-process lock on a stable config.json.lock for the whole read, copy and rename, inside the in-process write mutex. The lock rides internal/filelock, so it holds on every platform the repo builds for. A write waits at most two seconds for another process and then fails with a plain timeout error. The lock file is never unlinked: a crashed holder is released by the operating system closing its file, so there is no stale lock to reclaim. Tests run two real processes forced to contend and show both keys survive, a bounded timeout, and recovery after a holder exits without unlocking. Those forced tests are unix-only; their clocks are sanity bounds, not load bounds.

…ump (#1249) Both tests dumped the whole environment through the capped output collector and read variables out of it. On a machine with a large ambient environment the dump is cut before the TMUX_TMPDIR line, so the test read it absent while the job shell had it, and the two checks for stripped variables passed without proving anything. The shell now prints the three variables on one short line with an explicit word for unset, through the same login shell road the product uses, and the test fails if that line is missing from the captured output. Test only; the product was already correct on every platform.

…llows on a beat (#1251) The run's summary is a read nobody pressed for, and its refresh can wait on a model for seconds. It was asked through the one ordered line that carries a person's gestures to the engine, so a press on a run's row, or a stop, waited behind it: 2.4 to 10.1 seconds on a real screen, 0.09 to 0.18 after. The summary now goes beside the line. What may leave the line is decided by property: the ask was not a gesture and nothing a person does next depends on the engine having seen it first. A law lists every door that stands in the line and fails when one is added without a reason. An open page on a task that can still move was re-read on every paint tick once the last read answered, 509 wire reads in ninety seconds for one page. It now follows on the rail's own beat, counted from the last read for any reason, and a page on an ended or held task is never read again. The follow stays in the line because its fold replaces the page and must not overtake a note.

A run under the task belt was started on a context nothing could cut, and no cancel was kept. Its rows wear task numbers, but a stop by number went only to the task graph, which has never held a run's rows. Measured on the real binary in hosted mode: the stop card answered that there was no such task, the run's own page answered with the store's sentence about who owns the root, ctrl+c and /quit closed the window, and the run carried on to its own landing every time. The id a surface already sends is now resolved to whoever owns the row, so an older window stops a run through a newer engine. A stop of a run ends the run's task and everything open under it in one store write, then cuts the run's context so workers and their calls end at once, then commits what was made on the run's own branch and gives the copy back. Nothing goes into the person's folder, no model turn is bought, and the person is told where the work is and how to bring it in or drop it. A hand-off that joined a run is stopped alone. A stopped run buys no further summary. On the surface: x on the run's own page raises the same stop card, x typed while that page is still loading reaches it too, the page no longer offers a hold the store would refuse, and ctrl+c is read above both the loading page and the open page. A law lists every place that publishes a running row and fails until each has a test that a stop on that row ends its owner. What ctrl+c, /quit and a closed window do to a live run is unchanged here.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…ction on a cut-off failure A free rescue takes a general or code model before one whose name marks it tuned for a single domain (finance, medicine, law and the like). The free notice is one clause. A failed task that saw an account out of credit or a refused key ends on that action rather than offering a stronger redo. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

WithoutPatientRateLimits takes patience off a call and hands a 429 back where the loop would otherwise wait on the same machine; a free move to another machine is still made at once. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…-move line Tiny reads a model size tag from its name; more domain tags are known. A seat ladder puts qualified rivals within 3x the pick cost ahead of dearer ones. A crew line says each seat net move with the last reason and how many were tried between, not every rung. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

… pinned id The free rescue drops domain-tuned and tiny models. A model indexes are filled from any catalog row of the same model and from the evidence snapshot, so the undated row does not lose them. The crew migration retires only preset and pick words and auto rows; any model id in a seat row stays a pin, and the notice names the kept pins. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

… errands guarded Crew seat calls hand a 429 back and move on; a seat tries at most three free pools in a task, so every route down stops on its action in seconds. Only the person own turn on their own model skips the route-health guard, so an errand whose floor is the conversation model is not sent to a quarantined route. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…r rung - Labels are read word by word with emoji shortcodes folded, so a tracker's `bug 🐛` is a bug label; a bug label with no open-ended label is a bugfix instead of the uncertain open-ended default. - Title signals: what the code does not do (never, does not, fails to, should ...) and a call written out in the title; "should support" and kin weigh back on the open-ended side. - A bugfix is read again for reach: more than one source file, an API or protocol, language rules, a long report or several repros, existing tests that must keep passing. Two or more make it complex; the worker then takes one rung up its front (nextRung), planner and checker unchanged, pins never overruled. Decision.Subclass/Reach and CrewRecord task_subclass/task_reach carry it to the log; lines still say bugfix. - The net-move line names the first cause: "(credit unavailable on openrouter; +3 tried)". - Open-ended checker cost priors move to p75 (kimi-k3 $0.0882). - Tests: kornia-4814 as filed and title-only read bugfix; hermes-7680 reads complex and a one-line typo simple; a complex fix gets a stronger worker at the same lambda; a pinned worker stays. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

- CrewSpendCap: the crew cap, and the daily limit where the run is bound by it, whichever is lower, with the action a stopped call ends on. - CrewSeatCeilings: the checker's ceiling on one task, max(3x its estimate, $0.05); none when another seat shares the checker's model id, since the guard keeps spend per model. - A catalog row with no published price, or a non-:free id priced at zero (stealth/preview, meta-routers), is never a candidate or a free rescue; only an explicit-zero :free pool is free. Pins still name such ids. - CrewRouteAnswers for the provider route gate, with a fast path when nothing is unhealthy; a logged route failure drops the health cache so the next call sees it. - The migration notice lists kept pins worker, planner, checker. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

- SpendGuard prices each seat call before it is made (prompt bytes/4 at the prompt or cache-read price, plus the completion ceiling) against the day's cap and the checker's own ceiling; a call that would cross either is not made and the crew stops on the one action. Actual cost is booked after. TUI crews use the crew cap; codeaf do adds the daily limit unless -yes-spend lifted it. - A send that failed its first call on this task is not asked again by any seat: the seat walks on, or ends on its action, without a request. The failure row is written once. - provider.Config.RouteGate is asked before every body the client sends; the session refuses a model whose route health says it will not answer, on every road, except a crew seat's own call and the person's turn. - Tests: checker stops at its ceiling and the day at its cap; every route down asks each free pool at most once across all seats and a retry; with a 403 home, no road reaches the refused route after the seat's own call. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

- The /crew panel, its notes and the crew line draw pin, tick, ring, cross and diamond from the plain tier whatever the icon setting says; a private-use codepoint is a blank cell without the font. - Manual: a call that would cross the cap is not made; the checker's per-task ceiling; complex fixes; label spellings; pools rested for the task; unpriced models never picked; the first cause on the net line. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

- RouteCrew classified first (for the learned offset) and handed Decide only the class, so Decide built Reading{Class, "given"} with no reach and every complex fix ran on the simple fix's worker. Request.Reading now carries the whole reading, RouteCrew passes it and the task, and a class given without a reading still has its reach read off the task's words. - An unmeasured model must publish at least one index the seat weighs to be a pick or a rung; one with none (upstage/solar-mini4 on the planner) was ranked on price alone. The last-rung rescue still takes it. - Unmeasured q stays capped under the measured median (gpt-5.6-luna 3.973, qwen3.8-27b 3.975 are their own published indexes, not backfill) and their weighed cost is floored at the cheapest measured cost, so a measured model wins at equal price; tested. - Tests: Decide with Class given and with the reading; RouteCrew and ResolveSeats (the codeaf do road) route the hermes-7680 report complex with a stronger worker. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…p ends the turn Reproduced in the real TUI against a stub with a stack dump at the provider client: every crew guard bailed on ProfileDir == "", which is what every ordinary launch has (config.ProfileDir is the override variable; "" is the home profile). healthyModel, the route gate, the spend guard and CrewHealthySend/CrewRouteAnswers were all off, so the run summary (RefreshRunSummary, callRole worker, 320 widened to 6400 by the thinking share) asked the refused glm route, and the cap was never checked in a task. - The empty profile is the default profile on every crew path. - A crew's last word (crewStopped, ErrSpendStopped, errRouteResting) ends the turn at once. Weighed as the 429 underneath it, the turn sat out about a minute of backoff and asked the crew again: the all-down idle. - Helper calls in a conversation with crews are priced against the crew's day cap like seat calls, never wait out a rate limit, and are not made for a task that already stopped on its action. - The day holds every in-flight call's estimate and estimates a call at no less than the model's last actual cost today, so concurrent seats and a price sheet far under the real charge cannot pass the cap. - The task's crew line and log row carry helper spend: the helpers made for it, or the day's guarded spend since the task was routed when more. - Tests on the default profile through the real session: a 403 home asks the refused route only on the seat's first call (helper and a direct client refused); a $0.01 cap at $0.004 a call makes two calls and refuses the third and the line carries the helper; a crew stop wrapping a 429 with an hour's Retry-After ends the turn at once, asking once. A law that every provider client in the package is built through clientConfig, which sets the gate. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Remove the experiment data under docs/design/model-pool/data and the scripts that produced the figures and tables (figures.py, paperstyle.py). The paper source, its figures, tables and the built pdf stay, and the Makefile now only builds the pdf. data/seed-cells.csv stays: the pool index seed generator reads it. The router prior (internal/crewroute/prior.json) is unchanged apart from its source notes, which now point at the paper. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Comments in internal/crewroute (route.go, classify.go, prior.go, route_test.go) now state the rule the router follows, not the figures behind it: narrow fixes use the cheapest qualified crew, and open-ended work gets a strong checker. The knee comment says what the knee does (a step up is bought only above [Knee] points per dollar) and no longer lists per-crew deltas and costs. Four test failure messages in route_test.go now say what was expected instead of citing evidence. prior.json keeps all of its data; its source and costs_source notes now read "measured crews per class of work; see the Pareto Crewing paper". The spend guard's comment and its test lead-in describe an overrun without the incident's figures. The manual's crew estimate paragraph and the #1436 change entry refer to the router's table and the design paper instead of a trial. No behaviour or user-facing string changes; gofmt and go vet are clean. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The paper's LaTeX source, bibliography, generated tables (tab-*.tex), figure PDFs (fig-*.pdf) and the Makefile that built them are removed. docs/design/model-pool keeps pareto-crewing.pdf, RUNBOOK.md and data/seed-cells.csv, which internal/pool/index/cmd/seedgen reads. Nothing else in the tree referenced the removed sources; only the removed files referenced each other. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…5 limit Replace the shipped per-model table and hand-set scoring in internal/crewroute with a learned prior read from catalog metadata. crewroute: - prior.json now holds fitted weights only: a joint Gaussian over ability and catalog features (log prices, log context, release date, open weights, the three published indexes, arena Elo), family offsets, a per-(class, seat) linear link from ability to quality, per-seat token shapes and cost scales. No per-model rows ship. - quality() conditions on the fields a row has and returns a mean and a standard deviation; missing fields widen the variance. credible() needs a finite-variance score (evidence ratio) and an upper bound over the floor. - Cost comes from catalog prices x seat shape x class scale x the install factor. costFloor is the cheapest credible priced model; unpriced models are weighed at a priced model of equal ability. - Remove Snapshot/Measured/IsMeasured, seatWeight, unseenMargin, unseenShrink, the measured tie-break and the price ceiling. Every catalog model is a candidate. - Request.Learned applies per-install offsets; Request.TaskCap keeps the pick under the per-task limit (estimate x install factor), --best included. Note reads "held under the $X task limit". - Model gains ArenaElo and Released; ability memoised per model. router/config: - CrewRecord.Learned and CrewLog.Quality: accepted, kept, redo and failed outcomes move a per-(class, seat, model) offset by a bounded step (rate 0.1, bound 1 quality point), recorded on the decision row. - catalog reads `created`; crewModelOf sets Released from it or the slug date. Rows with indexes added are re-scored on the next refresh. - Per-task spend limit models.crew.task_cap (default $5): pre-call guard on every task call, stop line "this task reached its $5 limit · raise it in /crew", /crew cap row "per task $5 · daily $X", `/crew cap task <$>`, `codeaf do` holds it and -yes-spend does not lift it. tui3: Spending tab's per task row reads the /crew limit. Crew panel fixtures carry release dates and Elo; snapshots follow the new picks. Tests: replace the table-reproduction and effort-evidence tests with behaviour tests (stronger-indexed similar-priced model wins a seat, no metadata is not picked unless pinned, install evidence moves but does not freeze a score, bounded outcome learning, --best under the task limit, decision under 2 ms, weights under 2 MB). Manual and changelog updated. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…ach on one-paragraph asks crewroute ability: - A row that publishes any index (AA intelligence/coding/agentic, arena Elo) is conditioned on those indexes only; price, context, date, licence and family no longer move it. A row with none is conditioned on context, date and licence plus the family offset, and its mean is capped at the population mean with the excess added to the variance. Price is not an ability feature. prior.json v2 carries index_features/base_features. - Seats score the solve rate at theta - 1 sd (riskKappa); quality is no longer clamped to [0, 10], so ordering is preserved. - The checker needs mean u >= u_floor whenever any candidate reaches it (seatCredible); Gaps uses u + sd against the open-ended u_ref. - Link refitted with seat pooling within a class (support slopes shrunk toward the worker's), so no seat has a zero slope; --best buys the top model in every seat under the task cap. classify: - Reach signals read the title when a task has no body. - New reach signal: security fixes (traversal, injection, bypass, ...); "N repros" counts; keepPassing allows words between "existing" and "tests". - Title failure words add wrongly/rejected/instead of; a rename or version bump reads as a simple fix. Tests: dominance property test (no dearer and >= on every shared index never loses, 400 random pairs plus the v4.1-flash/glm-5.3-flash case), fix checker not the cheapest thin model, checker floor, --best strongest per seat, index-less row capped at the population mean, one-paragraph reach (T03/T11 texts), defect words and renames. Knee test now asserts open-ended buys a stronger checker at prices a fix does not. Config and tui3 expectations follow the new picks. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…and cheap tolerance - checkerFirst: on open-ended and other work at λ > 0, when the planner and checker slopes are within one SD of each other, the stronger of the two support picks sits the checker seat and the weaker the planner. Pins are never moved; a model that cannot sit the other seat stays. --best (λ = 0) is unchanged. - seatCredible: the worker, like the checker, needs mean u >= u_floor whenever any candidate reaches it (abilityFloor). Fallback ladders are exempt, so a seat that cannot start still moves to a similar cost. - --cheap: strongerWithin trades the worker pick for the strongest eligible model costing at most 1.5x (cheapTolerance). The knee comparison for the effort note runs without it. Tests: the checker is never weaker than the planner at the knee on open-ended and other work, and a pinned planner stays; a cheap worker is credible (glm-5.3-flash over a coding-only model) and the tolerance buys a stronger worker at 1.2x cost but not at 2.4x. Manual and changelog updated. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

dev now carries santos/dev2 (#1410) and #1388, #1426, #1440 and others. The PR's net change is taken from its fork point bdb08cf; the earlier dev2 merge and its revert are left out, so dev's skills (#1396) and worker step-boundary work stay as dev has them. Conflicts: config/crew.go, config/seats.go and tui3/crew.go take the routed-crew versions; /crew keeps dev's notice event. do.go keeps both the kept-record line and the crew outcome log. task_run_belt.go keeps both the machine-hold rail and the crew completer. home.md keeps dev's rows without the preset forms and adds the crew-panel row. truth_test.go keeps dev's new facts and counts the crew as three seats. back_test.go stays deleted as on dev. CHANGELOG.md keeps dev's released history. #1440 on the routed crew: reflex and small work keep their free defaults on a balance read as low; the three crew seats see the OpenRouter account as out of credit before any call (not re-probed per task) and route to free pools with the free-routes notice. Tests follow. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

With HOME and CODEAF_HOME empty, all three seats are routed, nothing is pinned, the allowed rule is all, a task is held to $5 and there is no daily crew cap. With nothing routable, no crew seat falls back to a built-in model and resolving a crew is an error. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Names the four kinds of work (bugfix, complex fix, open-ended, other), the candidates (every model a connected provider serves) and learned weights; draws the panel's cap row as "per task $5 · daily none" and says the seats' "usually" reads the last eight tasks; adds a route health section, the /crew cap task row, the per-task limit on codeaf do (-yes-spend does not lift it), and the crew's limits in LIMITS.md. openrouter-credits.md says how a low balance reaches the routed crew. The changelog entry names the classes, the #1440 change and the redo decay as the code has it. Onboarding comments no longer mention a crew step. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The two-millisecond budget was checked against one batch's mean, which on a loaded shared machine also counts scheduler preemption. It now takes the fastest of five batches of forty. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

santoshkumarradha · 2026-09-25T14:28:17Z

@AbirAbbas this is up to date with dev (merged in, no rebase or force-push; head e5ffb1666), and make check is green at that head. Ready for review.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01V7ShhY74oyWjYGB3SougdE

santoshkumarradha added 30 commits September 18, 2026 21:16

fix(resident): a dispatch pass does not reclaim the leaf it just rele…

ac84ab1

…ased (#1225)

test(session): wait helpers check the condition before the clock (#1226)

3a45c3f

test(session): the pin-the-wire test joins the usage writer before cl…

e88d353

…eanup (#1228)

fix(chatv3): seal the place sweep note at close instead of joining th…

78e8146

…e walk (#1227)

test(provider): order the late-first-token rescue by a signal, not el…

fe9eb62

…apsed slack (#1229)

remote, tui3: a hosted conversation reads and steers its run (#1230)

ccfef48

tui3: a task's page opens with the task, and a run's rows on the rail…

efa5326

… open their pages (#1231)

fix(run): a self-finished codeaf do run seats its review before it en…

4e176a5

…ds (#1233)

test(connect): the second-door tests use dynamic ports, not the fixed…

b208ccc

… callback ports (#1237)

session: a run works in a copy of its own, and its work comes home wh…

98cc254

…en it ends (#1232)

tui3: a wait whose stream has stopped arriving paints at the spinner …

692d4a1

…cadence, and the harness that measured it (#1235)

tui3: a task's page says only what is true of it (#1240)

608c459

chat: name unread config.json keys once before defaults apply (#1239)

2f6559a

test(exec): identify a leaf-end survivor by process identity, not a b…

1cbcb4b

…are PID (#1238)

config: concurrent profile writes in one process compose (#1242)

8f990d5

run: regression test for seating the check after a self-finished work…

5d9c47b

…er returns (#1243)

session: every run a conversation has made stays readable (#1234)

72bbd34

scripts: a repeatable hosted drive of a run on the real binary (#1241)

12840ee

config: unread notice skips keys the product itself retired (#1245)

7eab247

tui3: what is typed after a press on a run's row goes to its page, ne…

5a7748c

…ver the conversation (#1244)

session: make the claim-unspent revision test deterministic by forcin…

1e332c3

…g the ordering (#1246)

fix(connect): a reconnect reuses its registered identity instead of i…

45137b0

…ntroducing itself again (#1236)

test(connect): stop the second-door Slack test racing on a freed port (…

443c661

…#1247)

santoshkumarradha and others added 7 commits September 24, 2026 20:58

manual: how the crew estimate is made and what a pin sends

d886abb

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

manual: bounded free walk, net-move line, pins kept through migration

b737313

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Base automatically changed from santos/dev2 to dev September 25, 2026 02:20

santoshkumarradha and others added 7 commits September 24, 2026 23:11

santoshkumarradha force-pushed the feat/pareto-crew branch from ff1adb4 to cce59b0 Compare September 25, 2026 05:16

santoshkumarradha and others added 10 commits September 25, 2026 01:22

crewroute: drop the test pinned to measured per-crew costs

49b65ed

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

santoshkumarradha mentioned this pull request Sep 25, 2026

crew: count a pinned checker in the panel's coverage warning #1501

Open

AbirAbbas mentioned this pull request Sep 25, 2026

tui3: team delegation #1494

Merged

README: link the Pareto Crewing paper from the crew section

805aba0

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01V7ShhY74oyWjYGB3SougdE

Navigation Menu

crew: route the worker, planner and checker per task (Pareto crew) - #1436

crew: route the worker, planner and checker per task (Pareto crew)#1436
santoshkumarradha wants to merge 923 commits into
devfrom
feat/pareto-crew

santoshkumarradha commented Sep 24, 2026 •

edited

Loading

Uh oh!

santoshkumarradha commented Sep 25, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

santoshkumarradha commented Sep 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

User-facing behaviour

Routing (scoring, classes, fallback, limits)

Removed

Migration

Tests

How to try it

Uh oh!

santoshkumarradha commented Sep 25, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

santoshkumarradha commented Sep 24, 2026 •

edited

Loading