@amenophis1er/foreman 0.1.6 → 0.1.8

package/src/orchestrator.ts

@@ -90,11 +90,13 @@ class MessageStream implements AsyncIterable<SDKUserMessage> {
   }
 }
 import { WORK_DIR, makePolicy, type PendingPermission } from './policy.js';
+import { PLAYWRIGHT_MCP_CLI } from './browser.js';
 import type { AgentEnv } from './provider.js';
 import { generateRunTitle } from './title.js';
 import { combineBasis, costBasisOf, isPriced, type CostBasis } from './types.js';
 import { priceUsage, type ModelPrice } from './prices.js';
 import { captureBaseline } from './deck.js';
+import { memorySection, readMemory, writeMemory } from './memory.js';
 import type { RunMeta, TokenUsage, WorkerMeta, WorkerProgress } from './types.js';
 
 /** A run's usage before its first `result` message. */
@@ -248,6 +250,32 @@ export async function ensureIgnoreLines(file: string, lines: readonly string[]):
 // a metered one that dollars already stop.
 const DEFAULT_MAX_TURNS = 150;
 
+/**
+ * The same idea counted in tokens, for the same reason the turn cap exists.
+ *
+ * A turn cap bounds how many times the director speaks, not how much it says,
+ * and those come apart badly on a free or unpriced run: 150 turns over a large
+ * context is millions of tokens that nothing here was watching. It binds only
+ * where dollars cannot — a priced run already has a real cap, and this one
+ * must never end it first. Sized from the ledger, not from a hunch: cache
+ * reads are most of a director loop's traffic, and finished missions on this
+ * machine have run to 8–16M tokens total (a 5M figure was first proposed and
+ * would have cut several of them short). 20M is clear of every honest run
+ * seen so far and still well inside what a runaway reaches before the clock.
+ */
+export const DEFAULT_MAX_TOKENS = 20_000_000;
+
+/**
+ * The token cap as a director should read it: "5M tokens", not "5000000".
+ * A budget is only useful if the agent it constrains can hold it in mind, and
+ * a raw seven-digit figure in a prompt is one more thing to misread.
+ */
+export function tokenCapLabel(n: number): string {
+  if (n >= 1e6) return `${Number((n / 1e6).toFixed(1))}M tokens`;
+  if (n >= 1000) return `${Number((n / 1000).toFixed(1))}k tokens`;
+  return `${n} tokens`;
+}
+
 /**
  * How long a worker may say nothing at all before it is treated as stalled.
  *
@@ -675,6 +703,12 @@ directing worker agents. Non-negotiable rules, in priority order:
    failure is an escalation, never a self-repair.
 7. When DONE WHEN is verified, update MISSION.md (all boxes ticked, final log
    entry) and end with a short summary of what was built and how you verified it.
+8. LEAVE NOTES FOR THE NEXT CREW. Before you finish, call mcp__foreman__remember
+   with the whole project memory as it should read now: how to run and test
+   the project, ports and paths that matter, conventions and the reasons
+   behind them, traps you fell into. Facts, one line each, a page at most.
+   Rewrite stale lines rather than appending; drop what no longer holds.
+   Never a secret, never anything outside this project.
 `;
 
 export const WORKER_CHARTER = `
@@ -697,9 +731,6 @@ When finished, end with a concise report of what you did and how you checked it.
 `;
 
 /** Foreman's bundled Playwright MCP server, resolved from this repo. */
-const PLAYWRIGHT_MCP_CLI = fileURLToPath(
-  new URL('../node_modules/@playwright/mcp/cli.js', import.meta.url),
-);
 
 /** Claude subscription/quota exhaustion — an external pause, not a failure. */
 const USAGE_LIMIT_RE = /out of usage credits|usage limit reached|upgrade to increase your usage/i;
@@ -756,6 +787,8 @@ export class MissionRun {
   private readonly askTimers = new Map<string, { cancel(): void }>();
   /** The director's streaming prompt; steering pushes into it. */
   private directorInput?: MessageStream;
+  /** The project's notes as a worker-prompt section, read once per start; empty when there are none. */
+  private memoryForWorkers = '';
   /** Director cost is cumulative per query; track the last figure for deltas. */
   private directorCostSeen = 0;
   private wasInterrupted = false;
@@ -834,6 +867,8 @@ export class MissionRun {
      */
     private readonly host: {
       exposeService?: (runId: string, port: number, label: string) => Promise<{ ok: true; url: string; path: string } | { ok: false; reason: string }>;
+      /** The browser channel detected at dispatch (src/browser.ts); Chrome when the host says nothing. */
+      browserChannel?: string;
     } = {},
   ) {
     this.meta = meta;
@@ -1108,9 +1143,12 @@ export class MissionRun {
         'actually complete. Keep completed work; do not rewrite files that already ' +
         'satisfy their milestone. Update the doc to match reality, then continue ' +
         'the mission to DONE WHEN. ' +
-        this.budgetNote()
+        this.gitLine() +
+        this.budgetNote() +
+        memorySection((await readMemory(this.meta.folder)).text, 'director')
       : `MISSION: ${this.meta.mission}\n\n${this.budgetLine()} ` +
-        `Working directory: ${this.meta.folder}. Begin by writing .foreman/MISSION.md, then execute the plan.`;
+        `Working directory: ${this.meta.folder}. ${this.gitLine()}Begin by writing .foreman/MISSION.md, then execute the plan.` +
+        memorySection((await readMemory(this.meta.folder)).text, 'director');
 
     try {
       // The mission doc directory ignores itself wholesale (`*`, which also
@@ -1135,6 +1173,7 @@ export class MissionRun {
         void captureBaseline(this.meta.folder, this.meta.id).catch(() => {});
       }
       await mkdir(path.join(this.meta.folder, WORK_DIR), { recursive: true }).catch(() => {});
+      this.memoryForWorkers = memorySection((await readMemory(this.meta.folder)).text, 'worker');
       // Covers a .claude/ left by an earlier run; the one this run creates is
       // handled again on the way out.
       await this.ignoreLocalSettings();
@@ -1360,9 +1399,9 @@ export class MissionRun {
         // Absolute path: the server runs with the mission folder as cwd,
         // where npx cannot resolve Foreman's own dependency.
         command: process.execPath,
-        // Chrome channel by default (the machine's own Chrome, no download);
-        // FOREMAN_BROWSER picks another channel or Playwright's Chromium.
-        args: [PLAYWRIGHT_MCP_CLI, '--headless', '--isolated', '--browser', process.env.FOREMAN_BROWSER || 'chrome'],
+        // The channel the server detected for this machine: FOREMAN_BROWSER,
+        // else its Chrome, else Playwright's own Chromium (src/browser.ts).
+        args: [PLAYWRIGHT_MCP_CLI, '--headless', '--isolated', '--browser', this.host.browserChannel ?? process.env.FOREMAN_BROWSER ?? 'chrome'],
       },
     };
   }
@@ -1441,14 +1480,6 @@ export class MissionRun {
   }
 
 
-  /**
-   * Fold in the SDK's own dollar figure for a role.
-   *
-   * Ignored outright where Foreman holds that role's real rates: the SDK
-   * prices every response with Anthropic's table, so on a gateway role its
-   * number is fiction — and adding fiction to a figure computed from the
-   * endpoint's own published rates would corrupt the one honest total.
-   */
   /**
    * `costUsd` from its parts. The upstream's own figure, once it has given
    * one, replaces the rated figure for gateway tokens rather than adding to
@@ -1475,12 +1506,6 @@ export class MissionRun {
     this.enforceBudget();
   }
 
-  /**
-   * Folds one result message's token usage into the run total and persists
-   * it. Called alongside addCost() from the same two call sites (director
-   * loop, runWorker) so usage and cost are always in step — the honest
-   * counterpart to a dollar figure that is not honest on every provider.
-   */
   /**
    * The one event that carries a run's economics. Emitted whenever either half
    * changes — dollars OR tokens — because through a gateway the SDK often
@@ -1621,6 +1646,15 @@ export class MissionRun {
     }
   }
 
+  /** The mission's branch, when it has one: stay on it, and leave merging and pushing alone. */
+  private gitLine(): string {
+    const g = this.meta.git;
+    if (!g) return '';
+    return `This mission runs on git branch ${g.branch}, created for it from ${g.base}. Stay on it: do not switch branches, ` +
+      'do not merge, do not push, do not rebase or reset. You may commit as you go; Foreman commits whatever ' +
+      'is left uncommitted when the mission ends. ';
+  }
+
   /**
    * What the director is told about its budget, in the units that are true.
    *
@@ -1633,18 +1667,19 @@ export class MissionRun {
    */
   private budgetLine(): string {
     const turns = this.meta.maxTurns ?? DEFAULT_MAX_TURNS;
+    const tokens = tokenCapLabel(this.meta.maxTokens ?? DEFAULT_MAX_TOKENS);
     switch (costBasisOf(this.meta)) {
       case 'free':
         return `This run costs nothing per token — it is served by hardware the ` +
           `operator already owns — so there is no spend cap. It is bounded by ` +
-          `${turns} director turns.`;
+          `${turns} director turns and ${tokens}.`;
       case 'unpriced':
         // Deliberately still "no spend cap", and deliberately not silent about
         // the spend. A director told only that money is being spent, with no
         // figure and no cap, invents a limit and winds itself down early.
         return `This run does draw on a paid account, but Foreman cannot price it, ` +
           `so there is no dollar cap and no figure to reason about — do not ration ` +
-          `yourself against one. It is bounded by ${turns} director turns.`;
+          `yourself against one. It is bounded by ${turns} director turns and ${tokens}.`;
       case 'priced':
         return `Budget: $${this.meta.budgetUsd.toFixed(2)} total for this run.`;
     }
@@ -1714,10 +1749,21 @@ export class MissionRun {
       return `TIME CAP REACHED: ${Math.round(elapsed / 60)} minutes.`;
     }
     // Money only binds where the figure is real. Enforcing it through a
-    // gateway ends working runs over spend that never happened.
-    if (!isPriced(this.meta)) return null;
-    if (this.meta.costUsd < this.meta.budgetUsd) return null;
-    return `BUDGET CAP REACHED: $${this.meta.costUsd.toFixed(2)} of $${this.meta.budgetUsd.toFixed(2)}.`;
+    // gateway ends working runs over spend that never happened. Where it is
+    // real it is the cap, full stop: a priced run is never ended by tokens,
+    // which would cut a mission the human funded to its dollar figure.
+    if (isPriced(this.meta)) {
+      if (this.meta.costUsd < this.meta.budgetUsd) return null;
+      return `BUDGET CAP REACHED: $${this.meta.costUsd.toFixed(2)} of $${this.meta.budgetUsd.toFixed(2)}.`;
+    }
+    // Unpriced or free: tokens are the bound dollars cannot be. Live usage
+    // rather than the persisted total, so a gateway's interim tokens count too.
+    const u = this.liveUsage();
+    const total = u.inputTokens + u.outputTokens + u.cacheReadTokens + u.cacheWriteTokens;
+    if (total >= (this.meta.maxTokens ?? DEFAULT_MAX_TOKENS)) {
+      return `TOKEN CAP REACHED: ${(total / 1e6).toFixed(1)}M tokens.`;
+    }
+    return null;
   }
 
   private overBudget(): string | null {
@@ -1768,22 +1814,6 @@ export class MissionRun {
     );
   }
 
-  /**
-   * Starts a worker and returns at once; the session runs in the background.
-   *
-   * The split between this and {@link runWorker} is the asynchronous design in
-   * one place: everything the director can observe about a worker — the record
-   * in the map, `worker_started`, the `done` promise wait_for_worker races, the
-   * stored report and `worker_finished` at the end — is settled here, around a
-   * runWorker that only drives the SDK session. The promise is kept on the
-   * record rather than dropped, so a worker is never a floating promise, and
-   * the outcome is written onto the record rather than returned once, because
-   * the director now reads it back whenever it asks.
-   *
-   * Synchronous up to the point runWorker takes over: by the time this returns
-   * the record exists and `worker_started` has been emitted, which is exactly
-   * the guarantee spawn_worker's immediate reply relies on.
-   */
   /**
    * Item 6, decided as "ask": a worker on a gateway provider has stalled or
    * looped. Rather than letting the director cope alone or retrying somewhere
@@ -1860,6 +1890,22 @@ export class MissionRun {
     };
   }
 
+  /**
+   * Starts a worker and returns at once; the session runs in the background.
+   *
+   * The split between this and {@link runWorker} is the asynchronous design in
+   * one place: everything the director can observe about a worker — the record
+   * in the map, `worker_started`, the `done` promise wait_for_worker races, the
+   * stored report and `worker_finished` at the end — is settled here, around a
+   * runWorker that only drives the SDK session. The promise is kept on the
+   * record rather than dropped, so a worker is never a floating promise, and
+   * the outcome is written onto the record rather than returned once, because
+   * the director now reads it back whenever it asks.
+   *
+   * Synchronous up to the point runWorker takes over: by the time this returns
+   * the record exists and `worker_started` has been emitted, which is exactly
+   * the guarantee spawn_worker's immediate reply relies on.
+   */
   private launchWorker(workerId: string, prompt: string, resumeSessionId?: string, overrides?: WorkerOverrides): WorkerRuntime {
     const existing = this.workers.get(workerId);
     const w: WorkerRuntime = existing ?? {
@@ -2137,7 +2183,9 @@ export class MissionRun {
     const stop = this.overBudget();
     if (stop) return stop;
     const id = `worker-${++this.workerSeq}`;
-    this.launchWorker(id, task);
+    // The worker gets the project's notes with its brief: what earlier crews
+    // learned is exactly what a fresh session lacks.
+    this.launchWorker(id, this.memoryForWorkers ? `${task}\n${this.memoryForWorkers}` : task);
     return `[${id} started] status: running. It works in the background — use check_workers ` +
       `to watch it, and wait_for_worker when you need its result.`;
   }
@@ -2279,7 +2327,6 @@ export class MissionRun {
           this.pendingQuestions.set(id, resolve);
           this.armAsk(id, (afterMs) => {
             if (!this.pendingQuestions.delete(id)) return;
-        this.askMeta.delete(id);
             this.askMeta.delete(id);
             timedOut = true;
             this.emit('question_timeout', { id, afterMs });
@@ -2317,9 +2364,24 @@ export class MissionRun {
           'running while they may want to look, and put the URL in your report.' }] };
       },
     );
+    const remember = tool(
+      'remember',
+      'Rewrite the project memory (.foreman/MEMORY.md): the whole page as it should read now, ' +
+      'for the next crew. Facts about this project — how to run and test it, ports and paths, ' +
+      'conventions and why, traps — one line each, a page at most. Replace stale lines; do not ' +
+      'append forever. Never a secret; anything that looks like one is stripped.',
+      { text: z.string().max(20_000).describe('The complete memory file content, Markdown') },
+      async ({ text: body }) => {
+        const r = await writeMemory(this.meta.folder, body);
+        this.emit('memory_updated', { bytes: r.bytes, redacted: r.redacted, trimmed: r.trimmed,
+          text: `Project memory rewritten (${r.bytes} bytes${r.redacted ? `, ${r.redacted} secret-looking value${r.redacted === 1 ? '' : 's'} stripped` : ''}${r.trimmed ? ', trimmed to the cap' : ''}).` });
+        return { content: [{ type: 'text' as const, text:
+          `Memory written (${r.bytes} bytes).${r.redacted ? ` ${r.redacted} value(s) that looked like secrets were replaced with [redacted]; do not put credentials in memory.` : ''}${r.trimmed ? ' It was longer than the cap and has been cut at the end — prune it to a page.' : ''}` }] };
+      },
+    );
     return createSdkMcpServer({
       name: 'foreman',
-      tools: [spawnWorker, checkWorkers, waitForWorker, messageWorker, askHuman, exposeService],
+      tools: [spawnWorker, checkWorkers, waitForWorker, messageWorker, askHuman, exposeService, remember],
     });
   }
 }

package/src/planner.ts

@@ -29,6 +29,7 @@ import {
 } from '@anthropic-ai/claude-agent-sdk';
 import type { AgentEnv } from './provider.js';
 import type { MissionProposal } from './types.js';
+import { memorySection, readMemory } from './memory.js';
 import {
   armAskTimeout, formatAnswers, normaliseQuestions,
   type AskAnswers, type AskQuestion, type PendingAsk,
@@ -172,6 +173,8 @@ export interface PlannerModel {
   providerLabel: string;
   costBasis: 'priced' | 'free' | 'unpriced';
   note?: string;
+  /** Its track record here, from the run ledger: "here: director 4/5 done (~$0.78, ~18 min)". */
+  record?: string;
 }
 
 /**
@@ -205,7 +208,7 @@ export function pickKnownModel(
 export function modelsSection(models: PlannerModel[] | undefined): string {
   if (!models?.length) return '';
   const lines = models.map((m) =>
-    `  - ${m.id} — ${m.providerLabel} · ${m.costBasis}${m.note ? ` · ${m.note}` : ''}`);
+    `  - ${m.id} — ${m.providerLabel} · ${m.costBasis}${m.note ? ` · ${m.note}` : ''}${m.record ? ` · ${m.record}` : ''}`);
   return `\nMODELS AVAILABLE ON THIS MACHINE (use these exact ids in propose_mission):\n${lines.join('\n')}\n`;
 }
 
@@ -214,6 +217,8 @@ export interface PlanningTurn {
   projectId: string;
   /** What the machine can run, so recommendations are real ids, not guesses. */
   models?: PlannerModel[];
+  /** What missions have cost here and across the fleet, from the ledger; '' when too little history. */
+  anchor?: string;
   /** Session to resume; absent starts a fresh conversation. */
   sessionId?: string;
   folder: string;
@@ -407,7 +412,7 @@ export async function runPlanningTurn(turn: PlanningTurn): Promise<PlanningResul
         // thinking about the proposal, not discover it by asking.
         systemPrompt: {
           type: 'preset', preset: 'claude_code',
-          append: PLANNER_CHARTER + modelsSection(turn.models),
+          append: PLANNER_CHARTER + modelsSection(turn.models) + (turn.anchor ?? '') + memorySection((await readMemory(turn.folder)).text, 'planner'),
         },
         mcpServers: { foreman: createSdkMcpServer({ name: 'foreman', tools: [proposeMission, askUser] }) },
         canUseTool,

package/src/preflight.ts

@@ -21,6 +21,7 @@ import { defaultInstance, describeInstance, effectiveConfigDir } from './instanc
 import { discoverOllama, ollamaHost } from './ollama.js';
 import { serveHint, tailnetUrl, type Tailnet } from './tailscale.js';
 import { codexHome, codexModels, readCodexAuth } from './codex.js';
+import { browserCheck } from './browser.js';
 import { createRequire } from 'node:module';
 import { fileURLToPath } from 'node:url';
 
@@ -116,12 +117,17 @@ async function checkAuth(): Promise<Check> {
   const name = 'Credentials';
   const { mode, source, account } = await detectAuth();
 
+  // Missing credentials are a warning, not a refusal to start. The dashboard
+  // is where a first-time user reads what to do and where a project gets an
+  // API key or an endpoint of its own; a server that will not come up until
+  // something is logged in leaves them with a terminal line and no next step.
+  // Missions cannot start meanwhile — dispatch checks the provider itself.
   if (mode === 'none') {
     return {
       name,
-      status: 'error',
-      detail: 'none found — no missions can run',
-      fix: 'Log in with Claude Code, or: export ANTHROPIC_API_KEY=sk-ant-...',
+      status: 'warn',
+      detail: 'none found — missions cannot start until a provider is set up',
+      fix: 'Sign in with Claude Code on this machine (claude, then /login) and restart · or export ANTHROPIC_API_KEY · or give a project an API key or endpoint in Settings → Provider',
     };
   }
 
@@ -142,8 +148,12 @@ async function checkAuth(): Promise<Check> {
   return { name, status: 'ok', detail: `${MODE_LABEL[mode]} — ${who}` };
 }
 
-function checkInstance(): Check {
-  return { name: 'Claude Code', status: 'ok', detail: describeInstance(defaultInstance()) };
+async function checkInstance(): Promise<Check> {
+  const detail = describeInstance(defaultInstance());
+  // A tick beside an install nobody is signed into contradicts the line
+  // above it; say so here too, and leave the fix to the Credentials row.
+  if ((await detectAuth()).mode === 'none') return { name: 'Claude Code', status: 'warn', detail: `${detail} · not signed in` };
+  return { name: 'Claude Code', status: 'ok', detail };
 }
 
 /**
@@ -204,7 +214,7 @@ async function checkCodex(): Promise<Check | null> {
   };
 }
 
-async function checkPort(port: number): Promise<Check> {
+async function checkPort(port: number, envVar: 'PORT' | 'FOREMAN_SERVICES_PORT' = 'PORT'): Promise<Check> {
   const name = `Port ${port}`;
   const inUse = await new Promise<boolean>((resolve) => {
     const probe = net.createServer();
@@ -213,12 +223,17 @@ async function checkPort(port: number): Promise<Check> {
     probe.listen(port, '127.0.0.1');
   });
 
+  // The main port taken is fatal: nothing can be served. The services port
+  // taken is a fact to report — the dashboard still works, and only the
+  // crew's exposed previews are off until the port is freed or moved.
+  // PORT + 1 is a guess, and a second Foreman (a dev server beside the
+  // installed one) is the likeliest thing sitting on it.
   return inUse
     ? {
         name,
-        status: 'error',
-        detail: 'already in use',
-        fix: `Another Foreman may be running. Stop it, or: PORT=${port + 1} npm start`,
+        status: envVar === 'PORT' ? 'error' : 'warn',
+        detail: envVar === 'PORT' ? 'already in use' : 'already in use — exposed dev servers will be unavailable',
+        fix: `Another Foreman may be running. Stop it, or: ${envVar}=${port + 1} npm start`,
       }
     : { name, status: 'ok', detail: 'free' };
 }
@@ -247,40 +262,8 @@ async function checkHome(root: string): Promise<Check> {
  * `npx playwright install chromium`).
  */
 async function checkBrowser(): Promise<Check> {
-  const name = 'Browser';
-  const want = (process.env.FOREMAN_BROWSER || 'chrome').toLowerCase();
-  const candidates: Record<string, string[]> = {
-    chrome: process.platform === 'darwin'
-      ? ['/Applications/Google Chrome.app/Contents/MacOS/Google Chrome', path.join(os.homedir(), 'Applications/Google Chrome.app/Contents/MacOS/Google Chrome')]
-      : process.platform === 'win32'
-        ? [
-            path.join(process.env.ProgramFiles ?? 'C:\\Program Files', 'Google', 'Chrome', 'Application', 'chrome.exe'),
-            path.join(process.env['ProgramFiles(x86)'] ?? 'C:\\Program Files (x86)', 'Google', 'Chrome', 'Application', 'chrome.exe'),
-            path.join(process.env.LOCALAPPDATA ?? '', 'Google', 'Chrome', 'Application', 'chrome.exe'),
-          ]
-        : ['/usr/bin/google-chrome', '/usr/bin/google-chrome-stable', '/opt/google/chrome/chrome', '/usr/bin/chromium', '/usr/bin/chromium-browser', '/snap/bin/chromium'],
-    msedge: process.platform === 'darwin' ? ['/Applications/Microsoft Edge.app/Contents/MacOS/Microsoft Edge']
-      : process.platform === 'win32' ? [path.join(process.env['ProgramFiles(x86)'] ?? 'C:\\Program Files (x86)', 'Microsoft', 'Edge', 'Application', 'msedge.exe')]
-      : ['/usr/bin/microsoft-edge', '/usr/bin/microsoft-edge-stable'],
-    firefox: process.platform === 'darwin' ? ['/Applications/Firefox.app/Contents/MacOS/firefox'] : ['/usr/bin/firefox'],
-  };
-  let found: string | null = null;
-  if (want === 'chromium') {
-    // Playwright's own build, wherever the MCP's playwright-core says it lives.
-    try {
-      const req = createRequire(fileURLToPath(new URL('../node_modules/@playwright/mcp/cli.js', import.meta.url)));
-      const { chromium } = req('playwright-core') as { chromium: { executablePath(): string } };
-      const p = chromium.executablePath();
-      if (await exists(p)) found = p;
-    } catch { /* no playwright-core to ask */ }
-    return found
-      ? { name, status: 'ok', detail: `Playwright Chromium — ${found}` }
-      : { name, status: 'warn', detail: 'FOREMAN_BROWSER=chromium but Playwright Chromium is not installed; browser missions will fail', fix: 'npx playwright install chromium' };
-  }
-  for (const p of candidates[want] ?? []) if (await exists(p)) { found = p; break; }
-  return found
-    ? { name, status: 'ok', detail: `${want === 'chrome' ? 'Google Chrome' : want} — ${found}` }
-    : { name, status: 'warn', detail: `no ${want === 'chrome' ? 'Google Chrome' : want} found; missions with browser on will fail`, fix: 'Install Google Chrome, or: npx playwright install chromium && FOREMAN_BROWSER=chromium foreman' };
+  // One detection for the row and for the crew: see src/browser.ts.
+  return { name: 'Browser', ...(await browserCheck()) };
 }
 
 /** Where the phone can reach this. Says so plainly either way — the answer decides which links work. */
@@ -309,6 +292,8 @@ async function checkUi(distDir: string): Promise<Check> {
  */
 export async function preflight(opts: {
   port: number;
+  /** The second listener, for the crew's exposed dev servers — see services.ts. */
+  servicesPort: number;
   foremanHome: string;
   distDir: string;
   /** Detected by the server before preflight; null when not on a tailnet. */
@@ -321,6 +306,7 @@ export async function preflight(opts: {
     checkOllama(),
     checkCodex(),
     checkPort(opts.port),
+    checkPort(opts.servicesPort, 'FOREMAN_SERVICES_PORT'),
     checkBrowser(),
     checkHome(opts.foremanHome),
     checkUi(opts.distDir),

package/src/role-provider.test.ts

@@ -0,0 +1,55 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { reconcileRole } from './role-provider.js';
+
+const models = [
+  { id: 'fable' }, { id: 'opus' }, { id: 'sonnet' }, { id: 'haiku' },
+  { id: 'gemma4:12b', providerId: 'ollama-local' },
+  { id: 'gpt-5.5', providerId: 'codex-local' },
+  { id: 'kimi-k3', providerId: 'prov-abc' },
+];
+
+test('reconcileRole: a consistent pairing passes through untouched', () => {
+  assert.deepEqual(reconcileRole('director', 'fable', undefined, models), { providerId: undefined });
+  assert.deepEqual(reconcileRole('workers', 'gpt-5.5', 'codex-local', models), { providerId: 'codex-local' });
+  assert.deepEqual(reconcileRole('workers', 'gemma4:12b', 'ollama-local', models), { providerId: 'ollama-local' });
+});
+
+test('reconcileRole: an Anthropic alias on the Codex provider is corrected — the bug this exists for', () => {
+  const r = reconcileRole('director', 'fable', 'codex-local', models);
+  assert.equal(r.providerId, undefined);
+  assert.match(r.note ?? '', /fable is served by the project's provider, not codex-local/);
+});
+
+test('reconcileRole: a listed model moves to whichever provider lists it', () => {
+  assert.equal(reconcileRole('workers', 'gemma4:12b', undefined, models).providerId, 'ollama-local');
+  assert.equal(reconcileRole('workers', 'gpt-5.5', 'ollama-local', models).providerId, 'codex-local');
+  assert.equal(reconcileRole('director', 'kimi-k3', 'codex-local', models).providerId, 'prov-abc');
+  assert.equal(reconcileRole('director', 'FABLE', 'codex-local', models).providerId, undefined);
+});
+
+test('reconcileRole: a provider pin without a model is stale and dropped', () => {
+  const r = reconcileRole('director', undefined, 'codex-local', models);
+  assert.equal(r.providerId, undefined);
+  assert.match(r.note ?? '', /no model was chosen/);
+  assert.deepEqual(reconcileRole('director', undefined, undefined, models), { providerId: undefined });
+  assert.deepEqual(reconcileRole('director', '', '  ', models), { providerId: undefined });
+});
+
+test('reconcileRole: an unlisted model keeps its provider, unless it is a Claude id on a closed provider', () => {
+  // A custom endpoint's model the discovery call missed: trust the pairing.
+  assert.deepEqual(reconcileRole('workers', 'mystery-9b', 'prov-abc', models), { providerId: 'prov-abc' });
+  assert.deepEqual(reconcileRole('workers', 'mystery-9b', undefined, models), { providerId: undefined });
+  // A full claude-* id nobody listed still cannot go through Codex or Ollama.
+  const r = reconcileRole('director', 'claude-opus-4-1', 'codex-local', models);
+  assert.equal(r.providerId, undefined);
+  assert.match(r.note ?? '', /Anthropic model/);
+  assert.equal(reconcileRole('director', 'claude-opus-4-1', 'ollama-local', models).providerId, undefined);
+  // On a custom endpoint a claude-* id may be a real proxy route; left alone.
+  assert.equal(reconcileRole('director', 'claude-opus-4-1', 'prov-abc', models).providerId, 'prov-abc');
+});
+
+test('reconcileRole: works with an empty model list (discovery failed) by the closed-provider rule alone', () => {
+  assert.equal(reconcileRole('director', 'fable', 'codex-local', []).providerId, undefined);
+  assert.equal(reconcileRole('director', 'gpt-5.5', 'codex-local', []).providerId, 'codex-local');
+});

package/src/role-provider.ts

@@ -0,0 +1,65 @@
+/**
+ * A role's provider, checked against its model at dispatch.
+ *
+ * A run carries, per role, a model and the provider that serves it. The two
+ * are picked together in every picker, but they travel separately — through
+ * settings files, drafts, older builds and resumes — and once they drift a
+ * mission fails on its first turn with an upstream 400 nobody chose: the
+ * Claude alias `fable` sent through the Codex gateway, say, which then
+ * answers that the model "is not supported when using Codex". Nothing was
+ * checking that the provider actually serves the model.
+ *
+ * This is that check. The model list is the authority: a model the machine
+ * offers is served by whichever provider lists it, and a role is corrected
+ * to that provider whatever it arrived with. A model the list does not know
+ * is left alone, except for the one pairing that can never work: an
+ * Anthropic model on a provider that only serves its own catalogue.
+ */
+import { ANTHROPIC_ALIASES } from './anthropic-models.js';
+
+export interface KnownModel {
+  id: string;
+  /** Absent means the project's own provider (Anthropic by default). */
+  providerId?: string;
+}
+
+/** Providers that serve exactly the models they list and nothing else. */
+const CLOSED_PROVIDERS = new Set(['codex-local', 'ollama-local']);
+
+export interface Reconciled {
+  providerId?: string;
+  /** Set when the pairing had to change; one sentence, for the transcript. */
+  note?: string;
+}
+
+const isAnthropic = (model: string) => ANTHROPIC_ALIASES.has(model) || /^claude-/i.test(model);
+
+/**
+ * The provider a role should run on, given what it was handed and what the
+ * machine offers. Returns the same provider id and no note when nothing is
+ * wrong, which is the common case.
+ */
+export function reconcileRole(
+  role: 'director' | 'workers', model: string | undefined, providerId: string | undefined, models: KnownModel[],
+): Reconciled {
+  const given = providerId?.trim() || undefined;
+  // No model means "inherit the project's default", which lives on the
+  // project's provider. A provider id left behind without a model is stale.
+  if (!model) {
+    return given ? { providerId: undefined, note: `${role}: no model was chosen, so the ${given} provider pin was dropped.` } : { providerId: undefined };
+  }
+  const want = model.trim().toLowerCase();
+  const known = models.find((m) => m.id.toLowerCase() === want);
+  if (known) {
+    const served = known.providerId || undefined;
+    if (served === given) return { providerId: given };
+    return {
+      providerId: served,
+      note: `${role}: ${model} is served by ${served ?? 'the project\'s provider'}, not ${given ?? 'the project\'s provider'}; corrected.`,
+    };
+  }
+  if (given && CLOSED_PROVIDERS.has(given) && isAnthropic(model)) {
+    return { providerId: undefined, note: `${role}: ${model} is an Anthropic model and ${given} cannot serve it; running it on the project's provider instead.` };
+  }
+  return { providerId: given };
+}