FE-1763: Let Brunch build Petrinaut nets as capably as the existing assistant

.changeset/brunch-in-band-browser-tools.md

@@ -0,0 +1,5 @@
+---
+"@hashintel/petrinaut": patch
+---
+
+Allow a host to run named AI tool calls and return their results itself while the assistant response is still streaming, through `aiAssistant.inBandBrowserTools: { has(toolName), run(call, execute) }`: Petrinaut calls `run` when the call's turn in same-document order starts, with a `signal` that aborts on Stop, and the host calls `execute(input)` at most once and reports the resolved output or the failure; stock hosts retain their existing tool path. `executePetrinautAiMutation` lets a host run one canonical AI edit with the assistant's own no-op detection and output.

.changeset/petrinaut-ai-model.md

@@ -0,0 +1,5 @@
+---
+"@hashintel/petrinaut-core": patch
+---
+
+`petrinautAiModel` names the model and reasoning effort Petrinaut assistants run by default: `gpt-5.5-2026-04-23` at `medium`. `@hashintel/petrinaut-core/ai` also exports the stock prompt's behavioural frame and capability guidance separately, as `petrinautAiStockBehavioralFrame` and `petrinautAiCapabilityGuidance`, and composes `petrinautAiPrompt` from them.

AGENTS.md

@@ -86,7 +86,7 @@ The compose stack and the graph take these, so a new dev server stays off them a
 | 8200       | Vault                       | `HASH_VAULT_PORT`              |
 | 9000, 9001 | MinIO API and console       |                                |
 
-The Brunch agent's `start` and `start:test` scripts and its production image listen on 3002, where the integration tests expect it; the dev pair uses 4321 and 4915.
+The Brunch agent's `start` script and its production image listen on 3002; the dev pair uses 4321 and 4915.
 
 ### Starting Services
 

apps/brunch-agent/.pi/extensions/brunch-persona-testing/README.md

@@ -109,7 +109,7 @@ For Anthropic schema acceptance after a tool, schema, or adapter change, run `ya
 
 Add `--openai` to `yarn workspace @apps/brunch-agent test:persona` for the same proof through the registered OpenAI provider at low effort. The native Responses serializer and SSE parser remain real; only HTTP responses are synthetic. Each request checks the mounted tools' schemas/descriptions, `strict: false`, model and effort; captured `openai-requests.json` includes browser-result history. Passing is synthetic wiring evidence, not OpenAI server acceptance or a live-model result.
 
-The construction proof holds the recording pause and checks that no submission occurs before release. `test/persona-extension-lifecycle.test.ts`, enabled with `PI_PERSONA_CLI=$(command -v pi)`, crosses the installed Pi's flag hydration and tool-registration boundary with a synthetic socket reply and no inference. After building Brunch, `node --experimental-strip-types test/provider-accounting.integration.ts --disabled` checks that native requests proceed with an unusable historical ledger, preserve it untouched and retain usage in the original database. These checks do not prove live-model fidelity or successful generation with the operator's credential.
+The construction proof holds the recording pause and checks that no submission occurs before release. `test/persona-extension-lifecycle.test.ts`, enabled with `PI_PERSONA_CLI=$(command -v pi)`, crosses the installed Pi's flag hydration and tool-registration boundary with a synthetic socket reply and no inference. These checks do not prove live-model fidelity or successful generation with the operator's credential.
 
 From the HASH root, build and run the synthetic browser proof:
 

apps/brunch-agent/README.md

@@ -2,15 +2,15 @@
 
 ## Run the Petrinaut panel locally
 
-From the repository root, make `ANTHROPIC_API_KEY` available in the environment and run:
+Brunch and the stock chat route both run `petrinautAiModel` from `@hashintel/petrinaut-core` unless a deployment overrides it (`BRUNCH_CHAT_MODEL` and `BRUNCH_CHAT_THINKING` for Brunch, `PETRINAUT_AI_MODEL` and `PETRINAUT_AI_REASONING_EFFORT` for the stock route). An overriding `BRUNCH_CHAT_MODEL` runs without a thinking level unless `BRUNCH_CHAT_THINKING` sets one. From the repository root, make `OPENAI_API_KEY` available and run:
 
 ```sh
 yarn dev:brunch
 ```
 
-The first step builds the Petrinaut libraries the panel imports (`dist/` and design-system codegen). Then it starts the Brunch server at `http://127.0.0.1:4321` and the real Petrinaut website at `http://127.0.0.1:4915`. The website proxies `/agents/chat/*` to Brunch without changing the request origin or Flue protocol. The typed panel and Voice mode talk to one Flue chat agent composed from the context-independent core prompt in `@hashintel/brunch-agent/flue`, the SDCPN/Petrinaut instructions, modelling runbook skill, SDCPN plugin tools in `@hashintel/brunch-agent-plugin-sdcpn`, and app-owned deployment material.
+The first step builds the Petrinaut libraries the panel imports (`dist/` and design-system codegen). Then it starts the Brunch server at `http://127.0.0.1:4321` and the real Petrinaut website at `http://127.0.0.1:4915`. The website proxies `/agents/chat/*` to Brunch without changing the request origin or Flue protocol. Local development loads `apps/brunch-agent/.env.development`, selecting `openai/gpt-5.6-sol` with low reasoning by default, like the persona launcher. An explicit process environment or app-local `.env.local` can override these values; deployment model settings are separate and not established by this dev default. A configured credential does not authorize a paid run. When Brunch is selected, the typed panel and Voice mode bind the conversation to the open document: Brunch composes its prompt, SDCPN skill, elicitation, Ledger, document-revision attribution, and explanation with Petrinaut-owned capability guidance and the complete canonical `petrinautAiTools` catalogue. Petrinaut retains canonical schemas, execution, and model-visible outputs.
 
-The ordinary browser tools are `read_petrinaut_docs`, `read_petrinaut_net`, `read_petrinaut_diagnostics`, `mutate_petrinaut_net` (one ordered batch that adds, removes, or edits existing parts of the root net by ID), and `layout_petrinaut_net`, whose browser result carries a separately recorded `layoutRecord` of observed pre/post hashes and position effects. The [tool catalogue](src/agents/chat-agent/tool-catalogue.ts) records the mounted names and their definition/execution owners. The skill is activated via `activate_skill`, with supporting resources disclosed via `read_skill_resource`; the app-owned deployment diagnostic is `ping`. There is no generalized elicitation loop, sweep tool, or `brunch_ask` on this path. Brunch settles workpiece revisions through the server-side `mutate_workpiece` tool into Flue persistent state and retrieves them with `read_workpiece`; evidence exports are derived from the retained records, not a separate authoritative capture store.
+Native Stock remains a separate assistant choice; `VITE_PETRINAUT_DEFAULT_ASSISTANT` selects native Stock or Brunch.
 
 For browser-visible persona testing, use the [persona launcher and operator guide](.pi/extensions/brunch-persona-testing/README.md):
 
@@ -19,19 +19,27 @@ yarn brunch:persona --list-cases
 yarn brunch:persona --case inventory-purchasing
 ```
 
-The launcher opens a dedicated Chrome window, pauses for recording readiness, then drives the real panel with a private Pi persona. Brunch's own tool calls update the visible net and workpiece. Select any listed case or a directory containing `situation-pack.md` and `opening-message.md`. There is no automatic budget cutoff; native usage is retained. The guide owns prerequisites, stop/resume and evidence instructions; consult it before paid execution.
+The launcher opens a dedicated Chrome window, pauses for recording readiness, then drives the real panel with a private Pi persona. Petrinaut's stock tool calls update the visible net. Select any listed case or a directory containing `situation-pack.md` and `opening-message.md`. There is no automatic budget cutoff; native usage is retained. The guide owns prerequisites, stop/resume and evidence instructions; consult it before paid execution.
 
-By default outside production, conversations persist in SQLite at `apps/brunch-agent/.data-wipe-me/conversations.db`. `BRUNCH_DEV_DB_PATH` overrides that local path. The hermetic browser-transport test uses `BRUNCH_CHAT_DB_PATH` to point at its own sqlite file. Flue history is the conversation log. The panel rehydrates from the SDK's canonical conversation observation and does not resubmit or replay settled turns.
+By default outside production, conversations persist in SQLite at `apps/brunch-agent/.data-wipe-me/conversations.db`. `BRUNCH_DEV_DB_PATH` overrides that local path. Flue history is the conversation log. The panel rehydrates from the SDK's canonical conversation observation and does not resubmit or replay settled turns.
 
 The mounted Flue URL `/agents/chat/:instanceId` requires the principal and logical conversation identity in `x-brunch-principal` and `x-brunch-conversation`. The path id is the hash of those values, not a bearer token or trusted authentication.
 
+Provider admission bounds a dispatch with no model event at 60 seconds and bounds both newly opened and active reasoning silence at 120 seconds. These phase-specific limits tolerate supported reasoning models that legitimately pause for tens of seconds: an observed valid continuation exceeded the former 15-second first-reasoning-delta limit, then its sole retry exceeded the former 10-second between-delta limit. On expiry Brunch cancels the invocation, requires cancellation acknowledgement within two seconds, and permits at most one retry only before any tool call completed; completed tool work is never replayed. Tests may shorten these bounds with `BRUNCH_MODEL_STREAM_FIRST_EVENT_TIMEOUT_MS`, `BRUNCH_MODEL_STREAM_REASONING_START_TIMEOUT_MS`, `BRUNCH_MODEL_STREAM_IDLE_TIMEOUT_MS`, and `BRUNCH_MODEL_STREAM_CANCELLATION_TIMEOUT_MS`; non-test deployments always use the production policy.
+
 Print a human-readable transcript of one conversation from that same Flue history (server already running):
 
 ```sh
 yarn workspace @apps/brunch-agent transcript -- --principal <key> --id <conversationId>
 ```
 
-## Local Postgres for fixture-producing development
+## Tests
+
+Neither `test:unit` nor `test:integration` needs anything running. `test:integration` checks the built app, starting it as a server and loading it in process with scripted OpenAI responses on the default model.
+
+The `test:manual:*` scripts are not part of either suite and do not run in CI. `test:manual:browser` builds the website and drives it in the local macOS Chrome against the built app.
+
+## Local Postgres
 
 Set `BRUNCH_DB_KIND=postgres` explicitly to use the existing Postgres adapter and migrations locally. An unset selector defaults to SQLite outside production; `BRUNCH_DB_KIND=sqlite` also selects that lightweight path. Production always requires Postgres (selector unset or `postgres`), and rejects `sqlite`. Selector values are exact and case-sensitive; blank or unknown values fail.
 
@@ -130,7 +138,7 @@ Petrinaut `/api/chat` stays on the website; the accepted later website path is `
 Releasing `/agents/*` to production browser ingress requires separate authentication,
 authorization, ingress, and rate/spend gates. CORS, caller-supplied principals, and conversation
 hashes are not authentication. Desired count remains one until same-conversation ownership
-across replicas is separately proven.
+across replicas is separately proven. The direct browser-result handoff is ephemeral in that single owner: its one-use issued-call capability and existing conversation ownership headers do not authenticate a user or route callbacks across replicas. A lost result after a possible document effect is unknown; Stop aborts an active wait, while silent browser disappearance is detected only after the renewable 25-second lease expires. Production release still requires the external ingress authentication, authorization, routing and spend gates described above.
 
 The deployed chat path stores Flue conversations, submissions, compaction records, attachments,
 claims, leases, and settlement state in Postgres.
@@ -155,12 +163,12 @@ against this branch's image until a Mission 8 successor retargets it to `/agents
 
 Voice is a second input modality over the panel's conversation. It is not a Voice route and does not own provider audio or durable conversation state.
 
-|                       |                                                                                                                                   |
-| --------------------- | --------------------------------------------------------------------------------------------------------------------------------- |
-| URL                   | `/agents/chat/:instanceId`, called through the public Flue browser client and the same-origin local proxy                         |
-| Identity              | `x-brunch-principal` plus `x-brunch-conversation`; the server verifies that their hash matches the mounted instance id            |
-| Initial turn          | One `FlueClient.send()` carrying `{ kind: "user", body }`                                                                         |
-| Client-tool follow-up | One `FlueClient.send()` carrying the `client-tool-result` signal for completed client-tool parts, correlated by `toolCallId`      |
-| Response              | `FlueClient.wait()` chunks projected into one finite AI SDK UI-message stream; observation/history provides canonical rehydration |
+|               |                                                                                                                                   |
+| ------------- | --------------------------------------------------------------------------------------------------------------------------------- |
+| URL           | `/agents/chat/:instanceId`, called through the public Flue browser client and the same-origin local proxy                         |
+| Identity      | `x-brunch-principal` plus `x-brunch-conversation`; the server verifies that their hash matches the mounted instance id            |
+| Initial turn  | One `FlueClient.send()` carrying `{ kind: "user", body }`                                                                         |
+| Browser tools | A document-bound conversation awaits browser results over HTTP within the server tool call; no follow-up submission is sent       |
+| Response      | `FlueClient.wait()` chunks projected into one finite AI SDK UI-message stream; observation/history provides canonical rehydration |
 
 Typed and finalized spoken turns use this same route. The panel's explicit **Stop** requests a conversation-wide Flue abort before cancelling its local stream. Local Voice interruption stops playback only and leaves canonical history unchanged.

apps/brunch-agent/docs/task-dependencies.json

@@ -235,32 +235,6 @@
       "dependsOn": [],
       "cache": false
     },
-    "start:test": {
-      "dependsOn": [
-        "build"
-      ],
-      "cache": false,
-      "persistent": true,
-      "affectedBy": [
-        "error-stack",
-        "error-stack-macros",
-        "harpc-types",
-        "harpc-wire-protocol",
-        "hash-codec",
-        "hash-codegen",
-        "hash-graph-api",
-        "hash-graph-authorization",
-        "hash-graph-store",
-        "hash-graph-temporal-versioning",
-        "hash-graph-types",
-        "hash-temporal-client",
-        "type-system"
-      ]
-    },
-    "start:test:healthcheck": {
-      "dependsOn": [],
-      "cache": false
-    },
     "test:docker": {
       "dependsOn": [
         "build:docker"

apps/brunch-agent/package.json

@@ -10,7 +10,6 @@
     "build:docker": "docker buildx build --tag brunch-agent --file docker/Dockerfile ../../ --load",
     "dev": "vite dev",
     "fix:eslint": "oxlint --fix --type-aware --type-check --report-unused-disable-directives-severity=error .",
-    "fixture:worked-model": "node --experimental-strip-types src/evaluations/persona/create-worked-model-fixture.ts",
     "lint:eslint": "oxlint --type-aware --type-check --report-unused-disable-directives-severity=error .",
     "lint:tsc": "tsgo --noEmit",
     "measure:context-replay": "node --experimental-strip-types src/diagnostics/context-replay-measurement.ts",
@@ -21,19 +20,10 @@
     "smoke:deployment": "node --experimental-strip-types src/deployment-smoke.ts",
     "start": "PORT=3002 node dist/server.mjs",
     "start:healthcheck": "wait-on --timeout 1200000 http-get://localhost:3002/health",
-    "start:test": "NODE_ENV=test PORT=3002 node dist/server.mjs",
-    "start:test:healthcheck": "wait-on --timeout 600000 http-get://localhost:3002/health",
-    "test:anthropic-tools": "turbo run build --filter '@apps/brunch-agent...' && node --experimental-strip-types test/anthropic-tool-preflight.ts",
-    "test:compiler-feedback": "VITE_BRUNCH_CHAT_ENDPOINT=/agents/chat VITE_PETRINAUT_DEFAULT_ASSISTANT=brunch turbo run build --filter '@apps/brunch-agent...' --filter '@apps/petrinaut-website...' --env-mode=loose && node --experimental-strip-types test/compiler-feedback.integration.ts",
     "test:docker": "node --experimental-strip-types test/container-smoke.ts",
     "test:integration": "vitest run --config vitest.integration.config.ts",
-    "test:native-schema": "node --experimental-strip-types test/integration/native-schema-carriage.integration.ts",
-    "test:passage-policy": "node --experimental-strip-types test/passage-policy.integration.ts",
-    "test:persona": "VITE_BRUNCH_CHAT_ENDPOINT=/agents/chat VITE_PETRINAUT_DEFAULT_ASSISTANT=brunch turbo run build --filter '@apps/brunch-agent...' --filter '@apps/petrinaut-website...' --env-mode=loose && env -u BRUNCH_STEP_A_ACCOUNTING -u HASH_OTLP_ENDPOINT node --experimental-transform-types test/persona-construction.integration.ts",
+    "test:manual:browser": "VITE_BRUNCH_CHAT_ENDPOINT=/agents/chat VITE_PETRINAUT_DEFAULT_ASSISTANT=brunch turbo run build --filter '@apps/brunch-agent...' --filter '@apps/petrinaut-website...' --env-mode=loose && node --experimental-strip-types test/browser-witness/petrinaut-panel.ts",
     "test:unit": "vitest run --config vitest.config.ts",
-    "test:worked-model-bundle-copy": "vitest run --config vitest.config.ts --reporter=verbose test/worked-model-bundle-copy.contract.test.ts",
-    "test:worked-model-net-projection": "VITE_BRUNCH_CHAT_ENDPOINT=/agents/chat turbo run build --filter '@apps/brunch-agent...' --filter '@apps/petrinaut-website...' --env-mode=loose && node --experimental-strip-types test/worked-model-net-projection.integration.ts",
-    "test:workpiece-evidence": "node --experimental-strip-types test/workpiece-evidence.integration.ts",
     "transcript": "node --experimental-strip-types src/diagnostics/transcript-cli.ts"
   },
   "dependencies": {
@@ -57,7 +47,6 @@
     "valibot": "1.4.2"
   },
   "devDependencies": {
-    "@anthropic-ai/sdk": "0.74.0",
     "@earendil-works/pi-tui": "0.84.3",
     "@flue/vite": "2.0.3",
     "@playwright/test": "1.58.2",

apps/brunch-agent/petrinaut-local.vite.config.ts

@@ -16,6 +16,8 @@ import {
   type UserConfig,
 } from "vite";
 
+import { brunchEnv } from "@hashintel/brunch-agent/constants";
+
 import {
   defaultChatOrigin,
   petrinautLocalServer,
@@ -43,7 +45,7 @@ export default defineConfig(async (environment) => {
     throw new Error("PETRINAUT_WEBSITE_ROOT is required.");
   }
   const root = resolve(websiteRoot);
-  process.env.VITE_BRUNCH_CHAT_ENDPOINT ??= "/agents/chat";
+  process.env[brunchEnv.viteChatEndpoint] ??= "/agents/chat";
   process.env.VITE_PETRINAUT_DEFAULT_ASSISTANT ??= "brunch";
   // Babel resolves the React compiler plugin from the launched project's cwd,
   // not from the imported config file. Match a native hash launch before the
@@ -57,7 +59,7 @@ export default defineConfig(async (environment) => {
   if (!loaded)
     throw new Error(`Could not load Petrinaut's Vite config from ${root}.`);
 
-  const chatOrigin = process.env.BRUNCH_CHAT_ORIGIN ?? defaultChatOrigin;
+  const chatOrigin = process.env[brunchEnv.chatOrigin] ?? defaultChatOrigin;
   return mergePetrinautPanelConfig({
     chatOrigin,
     loadedConfig: loaded.config,