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

cli/CHANGELOG.md

@@ -11,6 +11,20 @@ a candidate do not establish public availability or authorize publication.
 
 ## [Unreleased]
 
+### Hosted service
+
+- Hold quotes follow the service's per-run holds. `run quote` and the
+  `run submit` plan send the given `--model`, `--reasoning-effort` and
+  `--environment`, plus `repositories=1` for any `--repo` or
+  `--commit-output`; the `job create` plan sends the spec's `model`,
+  `reasoning_effort` and `environment`, plus `repositories=1` for a
+  `repository_url` or `context_repository_url`. Nothing is defaulted, so a
+  flagless quote is unchanged. `holdBasis`, the human quote and the help no
+  longer call the hold flat.
+- `job create` webhook specs accept `model`, `reasoning_effort`,
+  `repository_url`, `repository_branch` and `output` for the connected
+  contract (the service requires `program_ref` with them).
+
 ## [0.15.0-rc.2] — 2026-10-02
 
 These are accumulated changes in the 0.15 candidate train; some capabilities

cli/SPEC.md

@@ -2469,14 +2469,23 @@ sorted keys of `inputs` as `inputKeys`, `interval_seconds`, `delivery_mode`,
 `visibility`, `role`, `amount_cents`), omitted when none is present; secrets,
 program text and input values never appear. Human previews print it as
 `Summary: field=value; ...` (cents also in dollars) and `CONFIRMATION_REQUIRED`
-as `Planned summary:`; a plan with a quote prints `Hold: $X (flat hold,
-independent of program and model)`. `run submit` and `program draft` check a
+as `Planned summary:`; a plan with a quote prints `Hold: $X (set aside from
+the wallet while the run is live; not its price)`. `run submit` and `program draft` check a
 `--model` against `GET /models` (manifest request `when: "--model"`) before the
 confirmation gate: a name that is neither offered nor the default is
 `INVOCATION_INVALID` naming the three nearest offered
 models, with `suggestedArgv` the same command using the nearest; a failed
 lookup is advisory (the service decides) except an interrupt. `program draft`
-quotes the flat run hold (`GET /run/quote`, advisory) into its plan. `org
+quotes the run hold (`GET /run/quote` with no parameters, advisory) into its
+plan. The hold depends on model, reasoning effort, environment, declared tools
+and bound repositories. `run quote` and the `run submit` plan send only the
+`model`, `reasoning_effort` and `environment` the user gave, plus
+`repositories=1` for any `--repo` or `--commit-output`; the `job create` plan
+sends the spec's `model`, `reasoning_effort` and `environment`, plus
+`repositories=1` for a `repository_url` or `context_repository_url`. Nothing
+is defaulted, the program is not inspected for the quote and tools are never
+sent, so a quote does not cover the program's own run settings or declared
+tools. `org
 create` checks the service slug rule (1-63 lowercase letters, digits or
 interior hyphens; not `openprose`, `system`, `default`, `invitations` or
 UUID-shaped) before any request, suggesting a derived slug as `suggestedArgv`
@@ -2495,8 +2504,9 @@ the `--yes` help line reads `required because <confirmReason>.` and
 `CONFIRMATION_REQUIRED` carries `details.reason` `--yes is required because
 <confirmReason>`; `render_service_help.py --check` and the manifest schema
 refuse a missing, misplaced or circular reason. `run quote` results carry
-`holdBasis` and its human output says the hold is flat and the price is known
-only after settlement.
+`holdBasis` and its human output says what the hold depends on, that the
+quote covers only the options given, and that the price is known only after
+settlement.
 
 **Transport.** Requests never follow redirects and are never retried, not even
 GETs. Each operation has a transport class with bounded time and size:

cli/bun/src/core/service/http.ts

@@ -52,6 +52,29 @@ export function requestFor(operation: ManifestOperation, index: number, path: st
   };
 }
 
+/** The options a `GET /run/quote` hold depends on, as the caller gave them. */
+export interface HoldOptions {
+  model?: string | undefined;
+  reasoningEffort?: string | undefined;
+  environment?: string | undefined;
+  repositoriesBound: boolean;
+}
+
+/**
+ * The `GET /run/quote` query (manifest order): only what was given, never a
+ * default, so a quote with no such option stays parameter-free. Any bound
+ * repository sends `repositories=1`; declared tools are never sent, because
+ * the CLI does not read the program.
+ */
+export function holdQuery(options: HoldOptions): Array<[string, string]> {
+  const query: Array<[string, string]> = [];
+  if (options.model !== undefined) query.push(["model", options.model]);
+  if (options.reasoningEffort !== undefined) query.push(["reasoning_effort", options.reasoningEffort]);
+  if (options.environment !== undefined) query.push(["environment", options.environment]);
+  if (options.repositoriesBound) query.push(["repositories", "1"]);
+  return query;
+}
+
 /** Sets a compact JSON body with sorted keys (both products serialize identically). */
 export function withJsonBody(request: Request, value: Json): Request {
   return { ...request, body: new TextEncoder().encode(canonicalJson(value)) };

cli/bun/src/core/service/jobs.ts

@@ -31,7 +31,7 @@ import { failure, invocationFailure } from "../errors";
 import { humanSafeScalar, quote as quoteText } from "../output";
 import { RunnerFailure } from "../types";
 import { readSource } from "./fs";
-import { encodeSegment, jsonObject, parseJson, requestFor, type Request } from "./http";
+import { encodeSegment, holdQuery, jsonObject, parseJson, requestFor, type Request } from "./http";
 import type { Context } from "./index";
 import { didYouMean, type Environment, type Json, type JsonObject } from "./manifest";
 import { parseOwnAllowed, parseProgramRef, pinned, resolveToRun, validSlug } from "./program-ref";
@@ -584,9 +584,27 @@ async function show(context: Context): Promise<Json> {
   return result;
 }
 
-/** The anonymous GET /run/quote hold for the confirmation plan (default environment); the price policy reference stays internal. */
-async function quote(context: Context): Promise<JsonObject> {
-  const body = await getJson(context, 0, "/run/quote");
+/** A spec key's value when it is a nonempty string (a hold option the spec gives). */
+function specString(spec: JsonObject, key: string): string | undefined {
+  const value = spec[key];
+  return typeof value === "string" && value.length > 0 ? value : undefined;
+}
+
+/**
+ * The anonymous GET /run/quote hold for the confirmation plan, quoted from the
+ * spec's model, reasoning effort, environment and bound repositories (only
+ * those the spec gives); the price policy reference stays internal.
+ */
+async function quote(context: Context, spec: JsonObject): Promise<JsonObject> {
+  const repository = (key: string): boolean => (specString(spec, key) ?? "").length > 0;
+  const request = requestFor(context.operation, 0, "/run/quote");
+  request.query.push(...holdQuery({
+    model: specString(spec, "model"),
+    reasoningEffort: specString(spec, "reasoning_effort"),
+    environment: specString(spec, "environment"),
+    repositoriesBound: repository("repository_url") || repository("context_repository_url"),
+  }));
+  const body = jsonObject(await context.send(request));
   const hold = object(body.hold ?? null, "quote.hold");
   const holdUsd = hold.hold_usd;
   if (typeof holdUsd !== "string" || !/^-?[0-9]+\.[0-9]{2}$/u.test(holdUsd)) throw protocol("quote.hold.hold_usd");
@@ -617,7 +635,7 @@ async function create(context: Context): Promise<Json> {
   const planned = context.planned(1, "/triggers", [], body);
   // A webhook with no program starts no runs: nothing is held.
   if (unpaid(context, spec)) planned.effect = "write";
-  else if (context.invocation.preview || !context.invocation.yes) planned.quote = await quote(context);
+  else if (context.invocation.preview || !context.invocation.yes) planned.quote = await quote(context, spec);
   const gate = context.gate(planned);
   if (gate.kind === "preview") return gate.result;
   const request: Request = { ...requestFor(context.operation, 1, "/triggers"), body };

cli/bun/src/core/service/programs.ts

@@ -437,7 +437,7 @@ async function draft(context: Context): Promise<Json> {
   const request: Request = { ...base, headers: [...base.headers, ["Accept", "text/event-stream"]] };
   const planned = context.planned(0, "/write", [], request.body);
   if (context.invocation.preview || !context.invocation.yes) {
-    // A draft reserves the same flat hold as a run (request 2, advisory).
+    // A draft reserves a hold as a run does (request 2, advisory; parameter-free).
     const quote = await context.advisoryQuote(2);
     if (quote !== undefined) planned.quote = quote;
   }

cli/bun/src/core/service/render.ts

@@ -238,7 +238,7 @@ function firstRunLine(planned: JsonObject): string | undefined {
   return `First run: about 1 second after the job is created, then every ${durationText(interval)}`;
 }
 
-/** The human hold line of a plan with a quote (The hold is flat, not an estimate of this run's price). */
+/** The human hold line of a plan with a quote (the hold is a reservation, not an estimate of this run's price). */
 function holdLine(planned: JsonObject): string | undefined {
   const hold = ((planned.quote as JsonObject | undefined)?.hold as JsonObject | undefined)?.hold_usd;
   return typeof hold === "string" ? `Hold: $${humanSafeScalar(hold)} (set aside from the wallet while the run is live; not its price)` : undefined;

cli/bun/src/core/service/runs.ts

@@ -18,7 +18,7 @@ import { failure, hostedRunFailed, invocationFailure } from "../errors";
 import { humanSafeMultiline, humanSafeScalar, quote as quoteText } from "../output";
 import { RunnerFailure } from "../types";
 import { readText, validRelativePath } from "./fs";
-import { encodeSegment, jsonObject, parseJson, requestFor, type Request, type Response } from "./http";
+import { encodeSegment, holdQuery, jsonObject, parseJson, requestFor, type Request, type Response } from "./http";
 import type { Context } from "./index";
 import { validSession, type JournalEntry } from "./journal";
 import { didYouMean, manifest, type Environment, type Json, type JsonObject } from "./manifest";
@@ -255,8 +255,15 @@ async function quote(context: Context): Promise<Json> {
     if (fallback === undefined || !validText(fallback, 64)) throw protocol("service status environments.default is missing or malformed");
     environment = fallback;
   }
+  // The hold depends on these options; only those given are sent. The
+  // program, inputs and runtime are accepted but never sent.
   const request = requestFor(context.operation, 1, "/run/quote");
-  if (requested !== undefined) request.query.push(["environment", requested]);
+  request.query.push(...holdQuery({
+    model: context.option("--model"),
+    reasoningEffort: context.option("--reasoning-effort"),
+    environment: requested,
+    repositoriesBound: context.optionValues("--repo").length > 0 || context.option("--commit-output") !== undefined,
+  }));
   const body = jsonObject(await context.send(request));
   const { hold } = quoteFields(body);
   let note = "";
@@ -265,15 +272,20 @@ async function quote(context: Context): Promise<Json> {
     note = cleanLine(body.note, 512);
   }
   let text = `Environment: ${humanSafeScalar(environment)}\n`;
-  text += `Hold: $${humanSafeScalar(String(hold.hold_usd))}, set aside from the wallet while a run is live; not its price. The same for every program and model; released within ${String(hold.ttl_seconds)} s when unused\n`;
+  text += `Hold: $${humanSafeScalar(String(hold.hold_usd))}, set aside from the wallet while a run is live; not its price. Released within ${String(hold.ttl_seconds)} s when unused\n`;
+  text += `Depends on: ${HOLD_DEPENDS_ON}; ${HOLD_COVERAGE}\n`;
   text += `Price: known only after a run settles; read it with \`${context.command("run show RUN_ID")}\`\n`;
   if (note.length > 0) text += `Note: ${humanSafeScalar(note)}\n`;
   context.human = text;
   return { environment, hold, holdBasis: HOLD_BASIS, note };
 }
 
-/** `run quote` holdBasis: the hold is not a price estimate. */
-const HOLD_BASIS = "flat hold, independent of program and model; a run's price is known only after it settles";
+/** What the hold depends on (the service prices the hold from these). */
+const HOLD_DEPENDS_ON = "model, reasoning effort, environment, declared tools and repositories";
+/** What a CLI quote covers: it never reads the program. */
+const HOLD_COVERAGE = "quoted from the options given, without the program's own run settings or declared tools";
+/** `run quote` holdBasis: what the hold depends on; it is not a price estimate. */
+const HOLD_BASIS = `depends on ${HOLD_DEPENDS_ON}; ${HOLD_COVERAGE}; a run's price is known only after it settles`;
 
 // ---------------------------------------------------------------------------
 // Event projection (decision 5) and terminal mapping (decision 9).
@@ -876,7 +888,8 @@ interface Submission {
   sourceSha256: string | null;
   session: string | undefined;
   waitMs: number;
-  environment: string | undefined;
+  /** The `GET /run/quote` query of the plan: the hold options given. */
+  quoteQuery: Array<[string, string]>;
 }
 
 async function prepareSubmission(context: Context): Promise<Submission> {
@@ -955,7 +968,8 @@ async function prepareSubmission(context: Context): Promise<Submission> {
   if (bytes.length > MAX_BODY_BYTES) {
     throw invocationFailure(`the submission is ${bytes.length} bytes, above the ${MAX_BODY_BYTES}-byte limit; shrink the program or inputs`);
   }
-  return { body: bytes, extraQuery, sourceSha256, session, waitMs, environment };
+  const quoteQuery = holdQuery({ model, reasoningEffort: effort, environment, repositoriesBound: repositories.length > 0 || commit !== undefined });
+  return { body: bytes, extraQuery, sourceSha256, session, waitMs, quoteQuery };
 }
 
 /**
@@ -1048,7 +1062,7 @@ async function submit(context: Context): Promise<Json> {
     const query = submitQuery(submission.session ?? "{session}", submission.extraQuery);
     const planned = context.planned(2, "/run", query, submission.body);
     const quoteRequest: Request = { ...requestFor(context.operation, 1, "/run/quote"), class: "control" };
-    if (submission.environment !== undefined) quoteRequest.query.push(["environment", submission.environment]);
+    quoteRequest.query.push(...submission.quoteQuery);
     // The quote is advisory: a failed quote never hides the plan.
     try {
       const { hold } = quoteFields(jsonObject(await context.send(quoteRequest)));

cli/conformance/cases/service/framework/help-json-group-records.json

@@ -30,7 +30,7 @@
           {
             "arguments": [
               {
-                "description": "Program file, or - for standard input, as `cli run submit` takes it. Accepted so a quote names the same command as `cli run submit`; the hold does not depend on it.",
+                "description": "Program file, or - for standard input, as `cli run submit` takes it. Accepted so a quote names the same command as `cli run submit`; the quote does not read the program, so the program's own run settings and declared tools are not included.",
                 "name": "FILE",
                 "required": false,
                 "variadic": false
@@ -74,7 +74,7 @@
             "mutation": false,
             "options": [
               {
-                "description": "Run a saved program; a bare SLUG is your own program. @REV is a rev_id, or @N (@revN) for revision N of your own program; without @REV the latest revision is pinned. Accepted so a quote names the same command as `cli run submit`; the hold does not depend on it.",
+                "description": "Run a saved program; a bare SLUG is your own program. @REV is a rev_id, or @N (@revN) for revision N of your own program; without @REV the latest revision is pinned. Accepted so a quote names the same command as `cli run submit`; the quote does not read the program, so the program's own run settings and declared tools are not included.",
                 "name": "--from",
                 "repeatable": false,
                 "required": false,
@@ -96,7 +96,7 @@
               },
               {
                 "default": "the service's default_model (`cli model list`)",
-                "description": "Hosted model id (see `cli model list`). Accepted so a quote names the same command as `cli run submit`; the hold does not depend on it.",
+                "description": "Hosted model id (see `cli model list`). Sent with the quote: the hold depends on it.",
                 "name": "--model",
                 "repeatable": false,
                 "required": false,
@@ -110,15 +110,15 @@
                   "high",
                   "xhigh"
                 ],
-                "description": "Reasoning effort supported by the model. Accepted so a quote names the same command as `cli run submit`; the hold does not depend on it.",
+                "description": "Reasoning effort supported by the model. Sent with the quote: the hold depends on it.",
                 "name": "--reasoning-effort",
                 "repeatable": false,
                 "required": false,
                 "value": "EFFORT"
               },
               {
                 "default": "the service's default environment",
-                "description": "Execution environment offered by the service (for example builtin or linux).",
+                "description": "Execution environment offered by the service (for example builtin or linux). Sent with the quote: the hold depends on it.",
                 "name": "--environment",
                 "repeatable": false,
                 "required": false,
@@ -133,14 +133,14 @@
                 "value": "RUNTIME"
               },
               {
-                "description": "Read-only repository context. Accepted so a quote names the same command as `cli run submit`; the hold does not depend on it.",
+                "description": "Read-only repository context. Any --repo or --commit-output sends repositories=1 with the quote: a bound repository adds to the hold.",
                 "name": "--repo",
                 "repeatable": true,
                 "required": false,
                 "value": "OWNER/NAME[@BRANCH]"
               },
               {
-                "description": "Writable repository that receives the run's commit. Accepted so a quote names the same command as `cli run submit`; the hold does not depend on it.",
+                "description": "Writable repository that receives the run's commit. Any --repo or --commit-output sends repositories=1 with the quote: a bound repository adds to the hold.",
                 "name": "--commit-output",
                 "repeatable": false,
                 "required": false,
@@ -153,7 +153,7 @@
               "stream": false
             },
             "preview": false,
-            "summary": "Report the hold a run reserves: money set aside from the wallet while the run is live. It is not the price. The environment is checked first.\nThe hold is flat, the same for every program and model, and what the run does not use is released when it settles. A run's price is known only after it settles (`cli run show RUN_ID`, price_cents). The program, inputs and model `cli run submit` takes are accepted, so you can quote the exact command you will submit."
+            "summary": "Report the hold a run reserves: money set aside from the wallet while the run is live. It is not the price. The environment is checked first.\nThe hold depends on the model, reasoning effort, environment, declared tools and bound repositories, and what the run does not use is released when it settles. The quote sends only the --model, --reasoning-effort, --environment, --repo and --commit-output you give. It does not read the program, so a program that sets its own model, effort or environment, or declares tools, can hold a different amount. A run's price is known only after it settles (`cli run show RUN_ID`, price_cents). The program, inputs and options `cli run submit` takes are accepted, so you can quote the exact command you will submit."
           },
           {
             "arguments": [