costwright

Static budget certificates for LLM-agent workflows. Point it at your LangGraph / CrewAI / OpenAI-Agents-SDK repo and it tells you — before anything runs — the worst-case budget ceiling of every workflow graph, where that ceiling comes from, and when it is effectively vacuous.

Backed by a machine-checked (Lean 4) cost-soundness theorem: A Potential-Based Calculus for Resource Bounds of LLM-Agent Workflows (repo) — well-typed ⟹ spend ≤ declared budget, on every trace. costwright is the static frontend of that calculus. It never executes your code (pure AST analysis, zero dependencies).

pip install costwright
costwright check .

costwright — budget certificate check  (schema costwright.v1)

  ▲ src/agents/researcher.py :42   default_dependent  ≤1000 supersteps (framework_default)
  ‼ scripts/run_forever.py   :17   runaway            while-true-driver
  ✗ src/agents/fanout.py     :88   non_certifiable    send-fanout

  12 graph units | ✓ 3 certifiable | ▲ 7 default-dependent | ✗ 1 non-certifiable | ‼ 1 runaway

  ⚠ 6 unit(s) rely on a framework default of ≥1000 supersteps (LangGraph ≥1.0.6) —
    that budget ceiling is effectively vacuous. Set recursion_limit explicitly.

Why

We measured 254 real agent workflows from 45 public production-grade repos:

• 88% of cyclic workflows rely on a framework default for their only budget bound — or have none at all.
• LangGraph's modern default (recursion_limit) is 1000 supersteps (it was 25 before v1.0.6). A "protected" workflow can take a thousand turns before the framework stops it.
• Only 12% of LLM calls in real code carry an explicit token cap.

A runtime budget tracker tells you what you spent — after you spent it. A static certificate tells you the worst case before you deploy, rejects the runaway pattern at check time, and the math behind the bound is machine-checked, not vibes.

_{Methodology disclosure: dataset is ~80% LangGraph (204/254 units; CrewAI and Agents-SDK samples are small); units cluster ~5.6 per repo; public-GitHub visibility bias applies; LangGraph version assumed ≥1.0.6 (default 1000) where undeclared; false-reject 0 was measured on N=3 surviving rejects. Full experiment: spec, frozen dataset, and audit trail in this repo's specs/ and results/.}

Commands

costwright check [path]

Maps every workflow graph to the typed-budget calculus and reports, per graph unit:

• --json → stable costwright.v1 output for CI/tooling.
• --fail-on reject|default-dependent|non-certifiable → exit 1 on policy violation (default: never fails — findings are warnings; strictness is opt-in).
• Exit codes: 0 ran fine, 1 policy violated, 2 infrastructure error. Never anything else.

costwright caps [path]

Finds LLM constructors without a token cap and tells you the right parameter for that provider (it is parameter-specific, not provider-specific — verified against primary provider docs):

• OpenAI Responses → max_output_tokens (bounds reasoning+output)
• Azure/OpenAI reasoning (Chat) → max_completion_tokens (reasoning tokens are inside it)
• Anthropic standard → max_tokens (with budget_tokens < max_tokens); interleaved/adaptive thinking degrades the ceiling (budget may exceed max_tokens) — flagged
• Gemini → max_output_tokens plus thinking_budget (thinking is billed as output but NOT bounded by maxOutputTokens) — flagged when missing

--patch out.diff emits a unified diff (git apply out.diff). costwright never edits your files.

costwright fuse (experimental)

Bundles a costwright cost certificate with an eleata-verify risk certificate (a per-output selective-risk SLA) into one tamper-evident costwright.fusion.v1 audit record per agent run — answering both "will it blow my budget?" and "can I trust this output, or send it to a human?"

costwright check . --json > cost.json                 # the cost certificate
# ... your eleata-verify step writes risk.json (VerifyResult.to_dict()) ...
costwright fuse --cost cost.json --risk risk.json --run-id "$RUN_ID"

It is the cartesian product of two independently-scoped certificates — NOT a composed guarantee: it does not assert the agent is "safe", has no aggregate "both passed" boolean, and does not claim a bounded budget preserves the risk SLA. The non-interference theorem (budget ⊥ risk, Hoare-style) that would justify composition is unproven / future work. The bundler is pure stdlib and never imports eleata-verify (costwright's zero-dependency core is intact). See docs/FUSION.md, docs/NON-INTERFERENCE.md, and the runnable examples/fusion_demo.py.

GitHub Action

- uses: hernaninverso/costwright/action@v0.1
  with:
    fail-on: reject        # reject | default-dependent | non-certifiable | none

JSON schema

costwright.v1 is frozen: closed enums (category, provenance, aggregation), stable unit_id, line spans only (no source code in output — CI-log safe). New fields are additive in v2. The signature field is reserved (null in OSS output): signed, independently verifiable certificates are a separate service — contact below.

What costwright does NOT do (exact scope)

• It does not bound Send fan-out, interrupts/human-in-the-loop, CrewAI hierarchical mode, or dynamic goto — those map to non_certifiable, never to a fake bound. Conservative by construction: when in doubt, no certificate.
• Nested subgraphs (graph.add_node("x", inner.compile())) ARE composed since 0.2.0 — but only when the inner graph is fully visible: a straight-line sequence of direct add_node calls on a uniquely-bound, non-imported local StateGraph. Any way a static reader can't pin down the inner graph's identity or node count — built in a loop/comprehension, bound more than once, passed into a helper, imported, from x import *, add_sequence, a RetryPolicy/error_handler — fails closed (non_certifiable), never a smaller composed number. Dynamic rebinding (globals(), setattr, exec, monkeypatching) is unobservable to any static analyzer and voids the certificate, as with every type checker.
• The bound is worst-case (deliberately over-approximate). Tightening it is roadmap.
• A token-level dollar bound additionally needs caps on every call (costwright caps) and a per-provider billing ceiling — see §3.2 of the paper for where that holds and degrades.

Theory

The calculus (7 constructs, affine potential in the style of Hofmann–Jost AARA), its cost-soundness theorem, the relational↔functional equivalence, the termination measure, and the affine no-double-spend layer are machine-checked in Lean 4 (no sorry; axioms [propext, Quot.sound]): hernaninverso/typed-resources.

Certification / enterprise

Signed budget certificates (Ed25519, independently verifiable), a live per-provider billing-ceiling feed (providers change semantics monthly — your certificate can silently invalidate), and org-level CI policy are a paid layer on top of this OSS tool. Contact: hernaninverso@gmail.com.

License

Apache-2.0.