weightclass 0.1.0

Local, policy-driven routing for Codex and Claude workflows.

weightclass

weightclass is a local, policy-driven router for Codex and Claude Code workflows. It classifies a task in memory as low, standard, or high, chooses a deterministic model-and-effort route, and can start one selected vendor process in the foreground.

By default, a request stays with its explicit source vendor. Cross-vendor routing is available only through a reviewed policy opt-in. An optional V2 route can start a separately installed API runtime after explicit review and egress acknowledgement; weightclass never reads API credentials or makes provider network requests itself.

Install

weightclass has no runtime dependencies beyond Python 3.10 or later.

uv tool install weightclass      # or: pipx install weightclass
brew install ictechgy/tap/weightclass

Or from a local checkout:

git clone https://github.com/ictechgy/weightclass.git
cd weightclass
python3 -m pip install .

All three install the wclass command. The native Codex and Claude CLIs must already be installed and authenticated; weightclass never reads or changes their authentication or subscription state.

Releases are cut by pushing a tag; see RELEASING.md.

For reviewable native Codex and Claude Code invocation examples, see Native integrations.

Run locally

wclass --help lists the whole surface:

wclass [-h] [--version] {classify,route,run,render,v2} ...

classify, route, and run read the task from standard input. render prints the command of a policy route named by a workflow descriptor and never reads a task. v2 selects a declarative API route; see V2 API routing.

Every malformed invocation — an unknown subcommand, a missing argument, a bad policy — exits 2 with {"error": "invalid_input"} on standard error and nothing else, so a caller can parse the failure without scraping usage text. Flag names are never abbreviated: --confirm-api-egress cannot be shortened.

Exit codes are weightclass's own; a selected command's status never overwrites them:

Code	Meaning
0	Success. For run and v2 run, the selected command exited 0.
2	invalid_task or invalid_input.
3	unsupported_route — no policy route matched.
4	executor_unavailable — the command could not be started.
5	api_confirmation_required — V2 without --confirm-api-egress.
6	route_fingerprint_mismatch — the reviewed route changed.
7	executor_failed — the command started and exited non-zero.

Code 1 is not weightclass's; it means the interpreter died on an unhandled exception, which is a bug worth reporting.

Code 7 carries the real status in its diagnostic, as {"error": "executor_failed", "executor_exit_code": N} or, for a command killed by a signal, {"error": "executor_failed", "executor_signal": N}. A selected command inherits standard error, so this diagnostic is always written on a fresh line and is the last line of standard error — parse that line, not the whole stream, which also holds whatever the command itself printed.

A vendor CLI that reports success while declining to do the work still exits 0; weightclass cannot detect that and does not claim to.

Inspect a route before running it:

printf '%s' 'Fix a spelling typo in the README.' | wclass route --source-vendor codex
printf '%s' 'Fix a spelling typo in the README.' | wclass run --source-vendor codex

The built-in routes are intentionally conservative:

• Codex: low, standard, and high use an ephemeral exec session in a workspace-write sandbox with model_reasoning_effort set to low, medium, and high. Codex has no dedicated effort flag, so the effort is passed as a -c configuration override for that one invocation.
• Claude: low, standard, and high use print mode, no session persistence, and efforts low, medium, and high. Permissions are acceptEdits, because print mode is non-interactive: a permission mode that asks a human has nobody to ask, so every edit is refused while claude still exits 0 — the router would report success having changed nothing. acceptEdits auto-accepts file edits only, which lets the Claude route change files as the Codex route already could. It does not make the two identical: Codex's workspace-write also runs commands, while under acceptEdits a non-edit tool still goes to a prompt that print mode cannot answer.

Neither default route pins a model. Model selection stays your reviewed policy's decision, expressed inside that policy's command; see Override the routes.

--source-vendor is required when weightclass is called from a Codex or Claude integration. With the default policy, --source-vendor codex selects only Codex routes and --source-vendor claude selects only Claude routes. weightclass is a standalone process, so it does not try to infer its parent application.

When --source-vendor is omitted, weightclass still pins every tier to a single vendor: the vendor of the first route declared in the policy (codex for the built-in routes). A tier is never silently served by a second vendor — that requires "allow_mixed_vendors": true. The vendor field is always present in wclass route output.

Classification is local and deterministic. Security, authentication, authorization, data, migration, concurrency, performance, production, and architecture signals route to high. Short typo, spelling, formatting, and rename tasks route to low; other valid tasks route to standard. Unknown or oversized task input fails closed.

Three rules make the outcome predictable:

• Signals are matched on whole words, so reproduction does not count as production. Korean has no word boundaries, so Korean signals are matched by containment and a compound word that embeds a signal may over-escalate.
• When both a high and a low signal are present, high wins. Under-rating a task is the more expensive mistake.
• A task of 1,200 characters or more is treated as high on length alone, so pasting a large context escalates the tier regardless of wording.

Override the routes

Use wclass route --policy policy.json or wclass run --policy policy.json to use a reviewed local policy. Routes are considered in listed order, so the first matching tier is selected. Add --source-vendor codex or --source-vendor claude when invoking it from that vendor. Configure model labels and vendor-specific effort arguments only with labels you know are available to you.

{
  "allow_mixed_vendors": false,
  "routes": [
    {
      "id": "codex-low",
      "vendor": "codex",
      "tier": "low",
      "command": ["codex", "exec", "--model", "your-low-model-label", "-"]
    },
    {
      "id": "claude-high",
      "vendor": "claude",
      "tier": "high",
      "command": ["claude", "--print", "--model", "your-high-model-label", "--effort", "high"]
    }
  ]
}

The command tokens are opaque policy values. weightclass validates their shape but does not assert vendor CLI semantics or subscription access. Always run wclass route with a representative non-sensitive task to inspect a policy before using wclass run.

A token is passed to the selected program as one argv entry, without a shell, so a token may contain spaces — an install path such as /Users/me/My Tools/claude, or a multi-word flag value.

A token may not contain a character that a reviewer would not see, since review is the whole point of rendering the command. Rejected are every Unicode C category — control characters, format characters such as zero-width space and the bidirectional overrides, surrogates, private-use and unassigned code points — along with any whitespace other than the ASCII space, and leading or trailing whitespace. The same rule applies to V2's model and effort labels.

Bind a run to the selection you reviewed

wclass route prints a route_fingerprint over the selected route id, vendor, command, tier, and the policy's allow_mixed_vendors setting — every field the descriptor itself shows, so you can recompute it from what you read. Pass it back to bind the run to that selection:

task='Review this authorization change.'
fingerprint="$(printf '%s' "$task" | wclass route --policy policy.json \
  | python3 -c 'import json,sys; print(json.load(sys.stdin)["route_fingerprint"])')"
printf '%s' "$task" | wclass run --policy policy.json \
  --ack-route-fingerprint "$fingerprint"

If the policy, the selected route, or the task's tier changed since the review, the run stops with exit 6 and {"error": "route_fingerprint_mismatch"} rather than executing an unreviewed command.

Three limits are worth stating plainly:

• The flag is optional, and omitting it binds nothing. wclass run without --ack-route-fingerprint re-selects from the policy as it finds it. This differs from wclass v2 run, where the acknowledgement is mandatory because that path can send your task to a paid API.
• The task is not bound, only its tier. A fingerprint reviewed for one low task will run any other low task that selects the same route. Binding the task would mean retaining a hash of it, and weightclass does not hash task content.
• The argv is bound, not the program. If the command names a path whose contents are replaced between review and run, the fingerprint still matches. It binds the policy's selection, not the identity of the executable — the same limit V2 has for --api-runtime.

A route has no separate model field, and a policy that declares one is rejected. Only command is ever executed, and weightclass cannot verify that a label matches the model a command actually selects without asserting vendor CLI semantics it deliberately does not assert. A label it cannot verify would let a reviewed descriptor advertise one model while another runs, so the model is declared once, inside command, where wclass route prints it in full.

Set "allow_mixed_vendors": true only when you intentionally want a Codex request to select a Claude route, or the reverse. When it is false or absent, the vendor filter is applied before tier selection — including when --source-vendor is omitted, in which case the vendor of the first declared tier route is used.

V2 API routing through an external runtime

V2 adds declarative API-route selection without turning weightclass into an API client. weightclass does not read API keys, inspect authentication, or make network requests. Instead, you provide an already-installed, trusted runtime at an absolute, executable path. That runtime is responsible for provider credentials, HTTP, billing, and any provider output.

Use a V2 policy only for API routes; unlike the V1 legacy policy, it cannot contain arbitrary command arrays. A route is eligible only for its declared source vendors. codex maps to the OpenAI provider family and claude maps to the Anthropic provider family; allow_cross_provider must be true before a route can cross those families.

{
  "schema_version": 2,
  "allow_cross_provider": false,
  "allow_api": true,
  "routes": [
    {
      "id": "openai-high-api",
      "tier": "high",
      "eligible_source_vendors": ["codex"],
      "provider": "openai",
      "transport": "api",
      "model": "your-openai-model-label",
      "effort": "high",
      "intended_recipient": "OpenAI API",
      "intended_billing_boundary": "your OpenAI API account"
    }
  ]
}

First review the selected destination and copy the returned fingerprint. weightclass reports the intended recipient and billing boundary from the reviewed policy; it does not verify either claim.

printf '%s' 'Review this authorization change.' | \
  wclass v2 route \
  --policy api-policy.json --source-vendor codex \
  --api-runtime /absolute/path/to/weightclass-runtime

Starting an API route requires both an explicit egress confirmation and the exact fingerprint from that review. weightclass recomputes the route before spawning the runtime, so a change to the selected model, effort, source, destination, runtime path, or API/cross-provider permission invalidates the acknowledgement.

printf '%s' 'Review this authorization change.' | \
  wclass v2 run \
  --policy api-policy.json --source-vendor codex \
  --api-runtime /absolute/path/to/weightclass-runtime \
  --confirm-api-egress --ack-route-fingerprint 'sha256:copied-from-route'

For a selected V2 route, weightclass invokes exactly this fixed protocol, without a shell, and passes the task only on standard input:

/absolute/path/to/weightclass-runtime --provider PROVIDER --model MODEL --effort EFFORT

Do not put API keys, tokens, task text, or personal information in the policy, route metadata, or command line. V2 does not provide retries, failover, credential management, background execution, or a bundled provider runtime.

Security boundary and non-goals

• No persistence: weightclass writes no router artifacts or vendor configuration.
• Task text is read only from standard input, held in memory to classify and pass to the selected child process, then discarded. weightclass never logs, stores, echoes, or places it in diagnostics.
• weightclass never reads credentials, subscription balances, pricing, cookies, or vendor configuration. It does not capture or process vendor output. V2 does not issue provider HTTP requests; a separately installed runtime may do so only after the explicit acknowledgement described above.
• Route selection is deterministic. Unsupported, malformed, or unsafe input fails closed with a redacted JSON diagnostic.
• weightclass does not infer source vendor, model availability, subscription tier, or remaining usage. Supply the source vendor explicitly and put model arguments in a reviewed policy's command when model routing is required.
• wclass run starts exactly one configured command in the foreground without a shell, retry, backgrounding, recovery, or process supervision.
• weightclass is not an API proxy, credential manager, cloud service, subscription checker, bundled provider runtime, or unattended multi-agent supervisor.
• Policies must be reviewed before use. Do not place secrets in a policy.
• wclass route binds a later wclass run only when you pass the fingerprint it prints, and only to the policy's selection — see Bind a run to the selection you reviewed for what that does and does not cover. Your control of the policy file, plus the built-in routes that live in code and cannot be swapped, is the boundary that always applies. Treat a policy file the way you treat a shell script.
• A selected command receives the task on standard input and inherits standard output and error. Whatever it does with the task — including writing it somewhere — is outside weightclass's control, and its exit status is its own.

Development verification

weightclass has no runtime dependencies. These development tools are not required to use it, only to reproduce what CI checks:

PYTHONPATH=src python3 -m unittest discover -s tests
PYTHONPATH=src python3 -m compileall -q src

python3 -m pip install ruff mypy build twine
ruff check src tests
ruff format --check src tests
mypy
python3 -m build && twine check dist/*

License

MIT