Repository navigation
Releases
weightclass 0.32.0
Highlights
weightclass is now only the launcher it always was underneath: wclass starts one
agent CLI process — Codex, Claude Code, Antigravity (agy), or Grok — with the
exact command you reviewed. Everything that existed to measure or promise a cost
saving is gone, because the measurements this project ran did not support the
promise. The measured record stays in the README and under docs/.
wclass-advisoryand theweightclass.advisorypackage are removed, together
with the packaged Agent Skill bundle, campaigns, verifiers, councils, and the
speculative cheap-first runner. 0.31.1 is the last release that has them;
pin it withuv tool install weightclass==0.31.1or
pipx install weightclass==0.31.1. The Homebrew tap follows the latest release
only.wclass delegate(role delegation, conformance drivers, the qualification
registry,delegate native route|run) andwclass v2(API routing through an
external runtime) are removed.wclass recommend,wclass review-cost-profile,wclass render,
classify --ask-vendor,classify --show-triage-command, and
run --suggest-escalationare removed.- What remains:
discover,classify(local, offline),route,run,usage,
example-policy,review-preset, and the policy generatorsprofileand
select; policy schemas 1, 2, and 3; the aggregate-only usage store; the
terminal review default and fingerprint binding from 0.31.1.
Why
The three-vendor review that shaped 0.31.1 agreed that the one lever with a
large measured effect, model grade, was blocked by a verify gate nobody had
measured. Stage A0 of docs/cheap-first-decision-plan.md measured it: the
default verify command caught 6 of 33 blind-injected defects, and a
class-aware invariant script caught 24 of 33 (Wilson 95% [0.56, 0.85]), below
the pre-registered lower bound of 0.80. The plan closed at A0 with a no-go and
the owner chose to narrow the tool rather than keep a framing the data does not
support. docs/verifier-recall-result.md has the full reading and
docs/launcher-scope.md the removal scope.
Compatibility and safety
Breaking change — removed entry point and commands. wclass-advisory,
wclass delegate, wclass v2, wclass recommend, wclass review-cost-profile,
and wclass render no longer exist. An invocation of a removed subcommand exits
2 with {"error": "invalid_input"}, like any unknown name. The removed flags
on kept commands are rejected the same way. Exit code 8 (triage_unavailable)
is no longer produced.
Breaking change — schema-3 policies that require native_delegation. The
confirmation flag for that requirement was removed with delegate; a route that
lists native_delegation in required_confirmations now exits 3
unsupported_route before the task is read.
Unchanged: every route fingerprint for unchanged inputs (the built-in route
table, the packaged preset files, and the schema-2/3 descriptor goldens are
byte-identical); the one-foreground-child contract; no credential reads, no
network, no persisted task content; the terminal review default and its
non-interactive escape hatches; the usage store format.
The advisory Skill upgrade-ledger step that used to run before the release build
no longer exists, so the "previous bundle" check that failed the v0.31.0
candidate cannot recur. The release workflow's validate jobs no longer smoke the
removed commands.
Documents
Documents that described removed surfaces moved to docs/archive/ with an index;
nothing was deleted. HANDOFF.md was rewritten for the narrowed tool and its
prior contents were appended verbatim to docs/handoff-archive.md. The README
lost the sections for the removed surfaces and kept the measured results,
security boundary, and fingerprint binding sections byte for byte.
weightclass 0.31.1
Highlights
wclass runandwclass routerequire you to say where the tier comes from: exactly one of
--tier(you choose) or--suggest-tier(the local classifier suggests, and prints its own
measured record next to the suggestion). The refuted heuristic classifier is no longer the
default front door, and a suggested tier cannot start a vendor without a review.- When stdout is a terminal,
wclass runshows the task-free route on the controlling terminal
and asks before starting it. No copied fingerprint is needed there.--no-review, an explicit
--jsonbefore the command, or--ack-route-fingerprintselects the non-interactive path, and a
pipe never gains a prompt. - Every
runandrouteflag now has help text.--source-vendorand the tier selection lead
the usage line, and each--helpends with a first command to try. wclass --helplists the seven daily commands and names the seven advanced ones in one line.
profile,select,review-cost-profile,recommend,render,delegate, andv2still
parse, run, and answer their own--help; they are only absent from the listing.- The packaged presets are now
<vendor>-model-override, which is what they do: bind opaque model
and effort labels onto the built-in route shapes. The old<vendor>-cost-focusednames claimed
a saving nothing here has measured. They remain accepted as aliases everywhere a preset is
selected, and the policy files, command bytes, and fingerprints are unchanged. - The README leads with what the tool does, then Install, Quick start, and "What you get, and what
you do not". The measured classifier record and the paired token study were moved, not removed:
see "Why the classifier is opt-in" and "Measured results for tier routing".
Core routing
- The tier requirement is an argparse mutually exclusive group, so it fails before the task is
read, and the diagnostic stays the closedinvalid_inputobject. The usage line is where the
requirement is discoverable, and a test pins that. --suggest-tiercarriesCLASSIFIER_MEASURED_AGREEMENTin the receipt and in the console
review: agreement 10/24 (41.7%), high-tier recall 1/9 (11.1%), over-routing 6/24 (25.0%), and
under-routing 8/9. An explicit tier adds no receipt field, so the frozen schema-1 route bytes are
unchanged and no fingerprint moved.- The terminal review default is keyed to one signal: stdout is a terminal. An explicit
--human
in a pipe does not open a prompt; an explicit--jsonon a terminal does not either.--review
and--no-reviewtogether are rejected rather than resolved in one direction. A pty-backed
harness looks like a terminal, so give it--no-reviewor a fingerprint. review-presetandroutereceipts report the canonical preset name in theirpresetfield. A
script that compares that field to an old-cost-focusedstring needs the new one; nothing else
about the receipt changed.docs/integrations.mdexamples that ranroute/runwithout a tier source were fixed, and a
test now scans every shell block in every markdown file for that defect.
Guidance and documentation
- Project guidance is split by directory scope. The root
AGENTS.mdcarries shared rules and an
index;src/weightclass/AGENTS.md,src/weightclass/advisory/AGENTS.md,tests/AGENTS.md,
.weightclass/AGENTS.md, andpackaging/AGENTS.mdare authoritative inside their subtrees. HANDOFF.mdrecords the direction research done with external reviewers, what survived
checking and what did not, and the usability batch above.
Compatibility and safety
Breaking change — run and route require a tier source. An invocation without --tier or
--suggest-tier now exits 2 with {"error": "invalid_input"} instead of classifying silently.
Automation should pass --tier; interactive use can pass --suggest-tier and review.
Breaking change — terminal runs review by default. A wclass run whose stdout is a terminal
now asks on the controlling terminal before starting a vendor, including for built-in routes that
previously started directly. Scripts driven through a pty should pass --no-review or
--ack-route-fingerprint. Non-terminal behaviour, including the fingerprint requirement for policy
runs, is unchanged.
Breaking change — preset receipt field. The preset field in route and review-preset
output carries the <vendor>-model-override name even when an alias selected the preset.
The one-foreground-child contract, the credential, network, and task-persistence boundaries, the
advisory companion, campaign records, statistical gates, sealed manifests, and every route
fingerprint are unchanged. The core native agy and grok routes still deliver the task in argv;
that boundary is unchanged and still documented in src/weightclass/AGENTS.md. The packaged
advisory Skill bundle is byte-identical to 0.30.0; the 0.30.0 bundle is registered in the upgrade
ledger, which the release workflow requires for the previous tag whether or not the bundle changed.
v0.31.0 was tagged first and failed that ledger check before build. Nothing was published under
it; the tag is retained and not reused, as the release policy requires.
Not in this release
Cheap-first execution with automatic escalation, the only lever the repository has measured as
large, remains unimplemented. It conflicts with the one-foreground-child contract and is a separate
decision, recorded in HANDOFF.md.
weightclass 0.30.0
Highlights
- The built-in
agyadvisory route no longer places task text in command-line arguments. agy reads
the prompt from one NDJSON message on standard input under--input-format stream-json, and the
CLI itself rejects an argv prompt in that mode, so the documented local process-inspection
exposure is closed rather than merely narrowed. No built-in advisory route now carries a task in
argv. wclass-advisory ask --councilstarts its members together instead of one after another. Wall
clock is the slowest member rather than the sum of all members, and the whole-council deadline is
shared by every member instead of starving the ones requested last.- Egress consent can be granted once for a shell with
WCLASS_ADVISORY_EGRESS=sessioninstead of
once per invocation. Nothing is written to disk and the grant dies with the shell. --humanpreview output is an actual summary. It previously printed the raw receipt, whose
inlined JSON schema pushed every reviewable flag off the screen.- Advisory Skill onboarding generation 18 removes the now-false argv-delivery instruction, states
the parallel council contract, and documents the session grant.
Advisory task delivery
- Both
agyroute builders drop--print {{task}}for
--input-format stream-json --output-format stream-json. Verified against agy 1.1.23. - The standard-input payload is wrapped as one NDJSON user message only when a route declares
that input format, judged by argv token position rather than by executable name or substring. - A route that declares NDJSON standard input while also carrying a
{{task}}or{{task_file}}
slot is rejected instead of being resolved in one direction, because that route describes two
destinations for the same task. - Route review and
--previewnow reporttask_stdin_encoding, so the reviewed argv is no longer
the only description of what bytes the child receives. - The
agylocal capability check now requires--input-format. An agy build without stream-json
input fails closed before any task is read rather than during execution.
Council execution
- Members run in one bounded thread pool over the existing per-member runner. Each child keeps its
own process group and is reaped by its own pid, so members cannot take each other's exit status. - Each member's timeout is fixed once at submission, so identical input does not produce different
budgets across runs. - Results are always reported in the requested vendor order, never in completion order.
- A failed peer is not cancelled, matching the existing campaign coordinator: cancelling would
discard work the vendor has already charged for. - The target worktree is compared once before the members start and once after all of them finish.
Consent and output
WCLASS_ADVISORY_EGRESSmust equal exactlysession; any other value, including an empty one,
does not grant. The receipt recordstask_egress_confirmation_sourceasflag,
session_environment, orterminal, and--previewreportssession_egress_grant_active.- Skipping the question does not skip the disclosure. The vendor list, per-member delivery mode,
and the context warning still print on the controlling terminal under a session grant. - The variable is absent from the vendor child environment allowlist, so no vendor CLI receives it.
- Preview rendering summarizes workflow, context, child counts and limits, repository behavior,
consent state, and per-route delivery. Arguments longer than 120 characters render as a length
and a SHA-256 prefix;--jsonstill carries exact bytes.
Compatibility and safety
Breaking change — schema-1 agy advisory routes. The agy argv changed, so every schema-1
agy profile digest and advisory route fingerprint changes with it. A sealed campaign population
bound to an older agy route must complete the documented migrate-routes --vendor agy step
before its next dispatch. No existing record is rewritten or merged. Claude and Codex profile
digests, and Grok's {{task_file}} delivery, are unchanged.
Breaking change — consent becomes expressible as configuration state. Before this release the
egress gate was strictly per action: a flag on the command, or an answer at the terminal. A shell
variable now authorizes every later call in that shell and every descendant process, so a wrapper
script or agent harness inherits it, and exporting it from a shell rc file makes the prompt
permanently unreachable. This is deliberate and is what makes non-interactive use possible, but it
is a change in the nature of the gate, not only in its ergonomics.
Core wclass routing, the one-foreground-child contract, campaign records, statistical gates, and
every sealed manifest are unchanged. The core native agy and grok routes still deliver the task
in argv; that boundary is unchanged and still documented in AGENTS.md.
Advisory Skill onboarding advances from generation 17 to 18. The exact published 0.29.0 four-file
bundle is registered in the upgrade ledger, so an installed 0.29.0 bundle upgrades without a
conflict.
Known limitation
The agy CLI applies its own --print-timeout, which defaults to five minutes, while the advisory
per-child ceiling is 3,600 seconds. A long agy review can therefore be cut off by the vendor CLI
before the reviewed ceiling is reached. This predates the release and is not changed by it.
weightclass 0.29.0
Highlights
- One-shot advice can now skip locally trivial plan/final work before vendor probes, consent, or
repository snapshots. Task-only context no longer snapshots the worktree or performs local
file-and-line triage. - Councils have one bounded whole-council deadline through
--total-timeout-seconds. Members that
cannot start before the deadline are reported separately, and partial councils use exit code 3. - Human review output keeps locally rejected high/critical findings visible for manual inspection
and describes structural triage as locally checkable rather than semantically supported. - Managed campaign verification protects the complete project
.weightclassverifier zone.
doctorchecks every selected workflow route and reports whether the installed managed wrapper
matches the current package. - New campaign gates use population rule v2, excluding infrastructure failures from every
attempted route. Historical sealed v1 gates retain their original population definition. - Advisory Skill onboarding generation 17 shortens the primary workflow, moves campaign detail to
its reference, reports customized installs underconflicts, and adds an automated release gate
preventing omission of the previous exact Skill bundle from the upgrade ledger.
Core routing and diagnostics
- Schema-2 fingerprint mismatches that cannot match any reviewed route now fail before task input.
- Argv-delivered tasks reject NUL before interactive confirmation, and platforms without
os.ctermidfail closed instead of raising an unexpected attribute error. - A schema-3 aggregate usage-store failure after child completion now reports the child's bounded
return code so callers can distinguish child and accounting outcomes. - Legacy escalation suggestions explicitly report that schema-3 rework accounting is unsupported
instead of recommending an unusable--usage-reworkflow.
Compatibility and safety
This release intentionally changes one-shot receipt schema version 1 to 2, changes a partial
council's exit status from 2 to 3, and removes the ineffective wclass-advisory ask --role option.
Consumers must use the receipt's complete and per-member status fields rather than treating all
non-zero exits alike. Task-only receipts report worktree_checked:false,
worktree_unchanged:null, no requested repository access, and no local triage.
Existing schema-1 managed gate populations remain valid and are evaluated with their sealed v1
rule. Existing managed roots whose copied verifier wrapper predates this package remain preserved,
but doctor marks dispatch not ready until the owner explicitly reruns init to install the
current package wrapper. Candidate-controlled dependencies outside the protected .weightclass
zone remain outside the verifier guarantee.
Four scrubbed GLM review packets covered core routing, one-shot behavior, managed campaigns, and
Skill/release operations (172,667, 94,999, 118,240, and 81,920 bytes; SHA-256 prefixes
6bdb523325f7, 6fd9bf00607c, 517eda63f090, and 326d77c9a47c). Provider output was treated as
untrusted and every retained finding was reproduced or validated locally before implementation.
Local validation passed 1,573 tests with 35 skips, Ruff check and format-check over 248 files,
strict mypy over 204 source files, compileall, and whitespace checks. Extracted-sdist isolation
passed 1,566 tests with 80 skips. Implementation PR #173 passed all eight Linux Python 3.10-3.14,
macOS Python 3.10/3.14, lint, type, build, and distribution checks.
Installation
uv tool install --upgrade "weightclass==0.29.0"
brew upgrade ictechgy/tap/weightclassweightclass 0.28.0
Highlights
- Declare when advice is needed with
--stage manual|plan|pivot|final; optionally skip locally
classified trivial plan/final work with--auto-skip-trivial. - Inspect task-free egress before execution with
--preview, including selected vendors, delivery
exposure, context, local probe count, and task-bearing child bound. - Request an explicit two-to-four-vendor council with
--council. Every member runs in a fresh
independent process, partial results are preserved, and descriptive consensus/dissent never
selects a winner or claims quality verification. - Select the smallest useful context with
--context task|diff|files|repo. Diff context is a
bounded trackedHEADdiff; files context uses repeated explicit--filepaths. - Review findings now receive snapshot-bound local
file:linetriage and invocation-only exact
duplicate grouping. Human output folds rejected/duplicate findings while JSON retains every
original finding and annotation. - Install advisory Agent Skill onboarding generation 16 with the new one-shot timing, context,
preview, and explicit-council guidance.
Compatibility and safety
The single-vendor ask --vendor ... path, manual stage, and repo context remain the defaults.
Council routing is never automatic and cross-vendor consent cannot be inferred from risk or
complexity. Call budgets are per invocation and no task-derived state, semantic key, result, or
context is persisted.
Preview reads neither task stdin nor repository content and starts no vendor process. Real runs
still perform two task-free local CLI probes per vendor before task input. A council contains at
most four task-bearing children, runs them sequentially in requested order, and exits 2 with a
machine-readable partial receipt when a member fails.
Explicit file context rejects traversal, every .git component, symlinks, controls, binary or
non-UTF-8 content, files over 32 KiB, and aggregates over 128 KiB. File bytes and local triage are
bound to the pre-execution snapshot. Diff context resolves and observes an absolute
non-repository Git executable before stdin, binds --git-dir and --work-tree to a standard
repository, disables external diff, text conversion, color, hooks, fsmonitor, and optional locks,
and omits untracked files. Linked-worktree .git files fail closed for this narrow selector.
The tool still does not confine a selected CLI from the host filesystem, verify provider recipient
or billing, or check top-level Git metadata. Agy task/context delivery remains visible in local
process arguments and is disclosed before execution. Every model-authored result and local
consensus label remains explicitly untrusted and quality_verified:false.
Two scrubbed GLM reviews used public packets of 81,943 and 87,976 bytes (SHA-256 prefixes
e7c87f987906 and 344e43c60f56). Confirmed consent, Git/worktree binding, executable-path,
snapshot-race, aggregate-I/O, finding-mute, prompt-boundary, citation, partial-receipt, and argv
disclosure issues were fixed and retested. A complete nine-production-file Codex Security diff
review retained no finding; TAC status was not queried. Full validation passed 1,563 tests with 35
skips, Ruff check and format-check over 205 files, strict mypy over 202 source files, compileall,
and the full Python 3.10-3.14 plus macOS Python 3.10/3.14 PR matrix.
Installation
uv tool install --upgrade "weightclass==0.28.0"
brew upgrade ictechgy/tap/weightclassweightclass 0.27.1
Highlights
- Restore automatic Agent Skill upgrades from the exact package-owned bundle shipped through
weightclass 0.26.0. - Retain every 0.27.0 usability feature: same-process
wclass run --review, stateless
wclass-advisory ask, anonymous campaign stdin, grouped command families, verifier scaffolding,
and Agent Skill onboarding generation 15.
Compatibility and safety
The 0.27.0 installer omitted the exact 0.26.0 four-file Skill bundle from its historical
compatibility ledger. Consequently, an unmodified installed 0.26.0 Skill was safely rejected as
skill_conflict instead of being overwritten. This patch adds only those reviewed hashes. Modified,
extra-file, or symlinked Skill destinations still fail closed, and install/uninstall retain their
descriptor-bound exact-content controls.
The regression was reproduced against both local Codex and Claude 0.26.0 Skill installations.
After the fix, skill status --target both reported both exact upgrades as planned; the upgrade
completed and a readback reported both destinations already_installed. The focused Skill and
distribution suite passed, followed by the complete Python 3.10-3.14 and macOS CI matrix.
Installation
uv tool install --upgrade "weightclass==0.27.1"
brew upgrade ictechgy/tap/weightclassweightclass 0.27.0
Highlights
- Run a selected native route with
wclass run --review: the same process displays the
task-free route on the controlling terminal, asks for confirmation, and starts at most one
child without requiring a copied fingerprint. - Use
wclass-advisory askfor a stateless read-only review, research, diagnosis, or design
answer with the selected CLI's configured default model. It needs no initialization, profile,
campaign, project verifier, route digest, or task file. - Use grouped
campaign,skill, andadvancedcommand families. Existing flat commands remain
accepted for automation compatibility. - Send new campaign tasks on standard input. The managed parent forwards bounded task bytes to
package-pinned runners through anonymous pipes without creating a task pathname. - Scaffold a safe reject-all project verifier with
campaign verifier scaffold, check committed
customization and baseline rejection withcampaign verifier check, and inspect or remove only
exact package-owned Agent Skill bundles withskill statusandskill uninstall --confirm. - Install Agent Skill onboarding generation 15, which uses one-shot
askfor ordinary advisory
modes and reserves campaign setup for implementation or explicit measurement.
Compatibility and safety
- Piped output remains JSON. On a terminal, non-execution
wclasscommands and
wclass-advisory askrender human-readable output by default; pass--jsonfor machine output
or--humanexplicitly. A barewclassorwclass-advisoryinvocation now prints help and
exits successfully. - Existing
routeplus--ack-route-fingerprint, flat advisory commands, managed task-file
dispatch, explicit-profile consult, and low-level campaign commands remain available. - Guided review supports schema-1, native schema-2, and observation-bound native schema-3 routes.
Route identifiers now reject terminal control characters, and built-in executable resolution is
checked again after the human confirmation pause.
Security boundaries
Stateless ask performs two task-free bounded local --help/--version probes before reading the
task, then starts one task-bearing child after explicit egress consent. Route permissions request
read-only behavior and a bounded no-follow snapshot rejects non-.git worktree mutations. The
receipt explicitly reports git_metadata_checked:false, quality_verified:false, and
host_filesystem_confined:false; use an external container or jail for a hostile CLI or
repository. Agy retains its documented argv process-inspection exposure, while Grok uses an
inherited anonymous /dev/fd/N pipe.
Exact-bundle uninstall renames the selected Skill to a private tombstone, revalidates the bound
directory descriptor, and deletes only the verified package files. Customized, extra-file, or
symlinked destinations fail closed.
Three scrubbed GLM reviews covered the runtime and Skill/documentation surfaces. Confirmed
lazy-loader, escalation fingerprint, terminal-control, child-cleanup, claim-consistency, and
destructive-race issues were fixed and retested. The final complete Codex Security working-tree
diff scan reviewed all 13 changed production files with complete coverage and retained no finding;
TAC status was unavailable because the connector was not connected.
Full validation passed 1,540 tests with 35 skips; extracted-sdist isolation passed 1,528 tests
with 79 skips. Ruff check and format-check passed 201 files, strict mypy passed 198 source files,
the distribution build and isolation checks passed, and both implementation PRs passed the full
Python 3.10-3.14 and macOS Python 3.10/3.14 CI matrix.
Installation
uv tool install --upgrade "weightclass==0.27.0"
brew upgrade ictechgy/tap/weightclassweightclass 0.26.0
Highlights
- Scrub managed project-verifier environments and terminate their owned process groups on timeout,
including descendants that retain output pipes. - Cache one-shot leader-exit observations and distinguish live process-group permission failures
from the bounded Darwin owned-zombie notification race. - Keep selector-construction failures inside owned cleanup for bounded capture, advisory preflight,
safe Git, and managed verifier Git helpers. - Replace avoidable 100 ms capture and Git polling with kernel event waits bounded by the actual
monotonic deadline. - Share managed parser/service contract constants and remove unreachable path-based Agent Skill
walkers while retaining the complete descriptor-based historical upgrade ledger.
Compatibility and safety
There are no intentional command-line, policy-schema, or exit-code breaking changes in this
release. Managed CLI direct-script behavior and legacy monkeypatch seams remain supported. The
explicit install-skill command is documented as the sole vendor-recognized-path persistence
exception and still publishes only the exact reviewed bundle to selected personal Skill roots.
Verifier environment scrubbing and process-group cleanup are local integrity and availability
controls, not host confinement. Use an external container or jail when vendor or verifier code is
hostile.
The final complete Codex Security diff scan reviewed all nine changed production source files and
retained no finding. Full validation passed 1,514 tests with 35 skips, extracted-sdist isolation,
Ruff and format checks, strict mypy, package builds, and the complete Linux/macOS CI matrix.
Installation
uv tool install --upgrade "weightclass==0.26.0"
brew upgrade ictechgy/tap/weightclassweightclass 0.25.0
Highlights
- Require explicit
--confirm-task-egressacross exact-command, profile, campaign, and
internal-consult advisory execution before task access or child start. - Harden timeout, cancellation, output-overflow, and resistant-descendant cleanup with retained
process-group ownership. - Bound Git and verifier I/O, resolve managed verifier inputs to immutable Git objects, and harden
state-root, workspace-registry, and Agent Skill publication paths. - Make campaign JSONL lookup linear and reduce cold CLI startup through split entrypoints and lazy
imports.
Compatibility and safety
Breaking change: advisory callers using exact-command, profile, campaign, or internal consult
execution must now pass --confirm-task-egress. Core wclass remains a single-foreground-child
router, and the advisory companion remains explicit and experimental.
The final 24-file security diff scan retained no finding. Full validation passed 1,496 tests with
35 skips, Ruff and format checks, strict mypy, package isolation, and Homebrew source installation
tests.
Installation
uv tool install --upgrade "weightclass==0.25.0"
brew upgrade ictechgy/tap/weightclassweightclass 0.14.0
Highlights
- Add task-free
review-cost-profilevalidation and a non-executingrecommendcommand for same-vendor, same-tier expected-completed-cost advice. - Bind every recommendation to exact route, cost-profile, and qualification-card fingerprints with conservative sample, coverage, confidence, quality, and critical-failure gates.
- Add an offline aggregate-only scorer for evaluator-sanitized provider usage, separating metered-cost evidence from fixed-subscription quota capacity.
- Reject byte-identical baseline/candidate commands as
candidate_route_unchanged, and keep unqualified Codex, Antigravity, and Grok standard experiment routes at medium effort.
Compatibility and safety
Built-in routes, source-vendor boundaries, policy schemas, and one-child execution behavior are unchanged. Recommendations never execute a vendor, retry, read task input before their document gates, inspect credentials, infer pricing or entitlement, or claim an actual bill reduction. Cost-focused routes remain explicit opt-ins, and the rejected Codex Luna experiment is not qualified or automatically applied.