Quote holds with the run's options and accept webhook binding keys

[rawwerks]

I'm an AI agent (Claude Code) opening this PR on behalf of Ray (@rawwerks). He asked for this fix and approved opening the PR, but he did not write the text below. The service-side facts come from the hosted service's maintainers.

Why

The hosted service's run hold is no longer flat. It depends on the model, reasoning effort, environment, declared tools and whether a repository is bound. GET /run/quote accepts model, reasoning_effort, environment, repositories and tools. With no parameters it quotes the default model at its default effort: builtin, no repositories, no tools.

The CLI still behaved as if the hold were flat:

• Hardcoded wording. It printed holdBasis: "flat hold, independent of program and model", "The same for every program and model", and "the hold does not depend on it" on --model, --reasoning-effort and --repo. The guide, SPEC and docs said the hold is flat and "currently $1.02".
• Ignored options. run quote, the run submit / run create plan and the job create plan quoted without parameters. run quote --model <cheaper model> therefore showed the default model's hold.

Separately, job create refused model, reasoning_effort and repository_url in webhook specs, though the service accepts them on a webhook with a program_ref. A webhook that reviews pull requests on a given model with a read-only repository therefore could not be created from the CLI.

What changes

Quote requests. Both ports send only what the user gave; nothing is defaulted.

• run quote and the run submit / run create plan: --model → model, --reasoning-effort → reasoning_effort, --environment → environment, and repositories=1 for any --repo or --commit-output.
• job create plan: the spec's nonempty model, reasoning_effort and environment, and repositories=1 for a nonempty repository_url or context_repository_url. This applies to every job type.
• What stays the same:
  • A flagless quote stays parameter-free.
  • program draft still quotes without parameters.
  • Tools are never sent, because the CLI does not inspect the program.
  • The plan quote stays advisory for run submit and required for job create, as before.

Wording.

• holdBasis is now depends on model, reasoning effort, environment, declared tools and repositories; quoted from the options given, without the program's own run settings or declared tools; a run's price is known only after it settles.
• The human run quote output gains a Depends on: line.
• The help, guide, SPEC and docs/service/* match. A program that sets its own model, effort or environment, or declares tools, can hold a different amount, and they all say so.

Webhook spec. The job.create webhook variant accepts model, reasoning_effort, repository_url, repository_branch and output, with the same rules as a schedule. Without a program_ref the service refuses these keys, and its message reaches the user as details.serviceMessage.

Files.

• Shared contract:
  • The manifest: quote query templates, the run.quote and job.create text, and the webhook spec keys.
  • The runs.schema.json holdBasis const.
  • The regenerated help.v1.json and service-guide-* cases.
  • The three hand-maintained framework cases with the manifest digest (service-capabilities-json, service-capabilities-jsonl, help-json-group-records), refreshed from identical Rust and Bun output.
• Corpus:
  • Nine edited runs/* cases.
  • Six new cases: runs/quote-sends-hold-options, runs/quote-commit-output-only, runs/submit-preview-repo-quote, jobs/job-create-webhook-binding-quote, jobs/job-create-schedule-environment-quote and jobs/job-create-schedule-empty-strings-quote. Each fixture pins the exact quote query.
• Rust: service/{runs,jobs,render,programs}.rs.
• Bun: service/{http,runs,jobs,render,programs}.ts.
• cli/CHANGELOG.md: an [Unreleased] entry.

Compatibility notes

• holdBasis text. The schema const is a visible contract change.
• run quote --model. It still does no local model check. An invalid model now comes back from the service as SERVICE_REQUEST_REJECTED, where before the option was silently ignored.
• A follow-up is possible. The service maintainers are adding a program-aware quote (program_ref, and a POST /run/quote that takes the run body). Once it lands, the CLI can include a program's own settings and declared tools without parsing it.
• Overlap with Document webhook delivery signing and fix endpoint_url name #49. That PR also edits guide.v1.md and the service-guide-* cases. Whichever merges second needs python3 cli/ci/render_service_help.py --write.

Evidence

Local, on macOS with Python 3.10.21, Rust from rust-toolchain.toml and Bun 1.4.2. These gates pass:

• shared-contracts, service-help, public-surface-files and service-operations-corpus
• differential-conformance and service-coverage
• rust-format and rust-clippy
• service-operations-rust-build and service-operations-rust: 1095 cases, 0 failures
• bun-typecheck, service-operations-bun-build and service-operations-bun: 1095 cases, 0 failures

Rust service tests (prose-runner-core lib, service_runs, service_jobs, service_programs, service_framework) and Bun service tests also pass.

Not passing locally, and not caused by this change:

• rust-tests. The full gate fails on this machine under load, and fails the same way on unmodified main @ 8e2a258. The failures are process-timing tests that time out with exit 124, a different set each run (weave_host, cli.rs adapter tests). The weave_host tests pass 8/8 run alone on this branch.
• Bun dev-endpoint.test.ts and build-identity.test.ts. They fail on unmodified main here too; local Bun is 1.4.2, while the repo pins 1.3.5.

GitHub CI is authoritative for both.

An independent review agent read the diff before this was opened. It found a Rust/Bun divergence on empty spec strings, now fixed and pinned by job-create-schedule-empty-strings-quote, plus wording overclaims, all addressed.

🤖 Generated with Claude Code