23e9fb4 · 316c9c6 · 4bdab48 · 4703c17 · 08767eb
@@ -11,6 +11,7 @@
import type { Measure } from "../core/measure.ts";
import type { LimitWindow } from "../core/standing.ts";
import type { RunLimits } from "./run-cell.ts";
import type { SandboxPolicy } from "./sandbox.ts";
import type { ToolNames } from "./brief.ts";
import { detectClaude, runClaude } from "./claude.ts";
export interface AgentRun {
@@ -116,6 +117,12 @@
export interface AgentRunner {
  run(input: AgentRun): Promise<AgentOutcome>;
  /** Whether `binary` is an agent this runner can drive. */
  detect(binary: string): Promise<AgentDetection>;
  /**
   * What this runner's harness calls the forge's tools, where it names them
   * differently from the forge (a prefix, another convention). The brief
   * says these names; absent, it says the neutral ones (`DEFAULT_TOOLS`).
   */
  readonly tools?: Partial<ToolNames> | undefined;
}
// Wrapped, not bound, so the two modules may load in either order.
@@ -129,6 +136,11 @@
export function currentRunner(): AgentRunner {
  return claude;
}
/** The forge tool names a brief should say for `runner` (#12128); the current runner by default. */
export function briefTools(runner: AgentRunner = currentRunner()): Partial<ToolNames> | undefined {
  return runner.tools;
}
export function runAgent(input: AgentRun): Promise<AgentOutcome> {
  return currentRunner().run(input);
}
@@ -1,6 +1,6 @@
/**
 * What an agent is told. Two texts: a system prompt appended to Claude
 * Code's own — who the agent is, where it is, how to finish, what will be
 * What an agent is told. Two texts: a system prompt appended to the runner
 * agent's own — who the agent is, where it is, how to finish, what will be
 * refused — and the task itself, which differs by why the run exists: an
 * assignment, feedback on a pull, or a check that failed. The golden path
 * is written down here so it is prompted, not hoped for; the mechanical
@@ -60,8 +60,21 @@
export interface BriefInput {
   * it writes obey the cells and the cells it does not touch stay decided.
   */
  readonly policy?: PolicyBrief | undefined;
  /**
   * What the runner calls the forge's tools, when its harness names them
   * differently (a prefix, another convention). The brief carries only
   * these names, never a harness's own tool vocabulary.
   */
  readonly tools?: Partial<ToolNames> | undefined;
      
        sentinelcrew
        1d ago
        
      
      tools is never set by any caller outside tests. Wire it from the runner where the brief input is built (with a test), or remove the field and file a follow-up; an unused seam does not satisfy box 2.
    
}
/** The forge tools a brief names; neutral names unless the runner says otherwise. */
export interface ToolNames {
  readonly keepPlan: string;
  readonly estimate: string;
}
export const DEFAULT_TOOLS: ToolNames = { keepPlan: "keep_plan", estimate: "estimate" };
export interface PolicyBrief {
  /** The identities the forge signs, floor first. */
  readonly identities: readonly string[];
@@ -87,7 +100,7 @@
export const SANDBOX_LINES = [
 * listed would run everything, so the line guards on its exit code.
 */
export const PUSH_STEPS_LINE =
  "- Before you push, run `bun install` (only when node_modules is missing), then, when the clone has `scripts/gate.ts`, `bun scripts/gate.ts` (the deterministic invariants: exclusions, migration pins), then `bun x tsc --noEmit`, and the tests that cover your change, plus any the repository's CLAUDE.md or AGENTS.md says always to run. Do not push while any of them is red: fix it first, so the pull never exists red. Where your sandbox cannot run one of them, say which and why in your last message. When the clone has `scripts/fast.ts`, name them with `files=$(bun scripts/fast.ts) && bun test $files`: exit 3 means the change reaches no test, so run none; exit 2 means the diff is too broad to name, so pick and run the test files that cover it yourself. Never a bare `bun test` with no files listed: it runs everything. Not the whole suite: the forge runs the full check once on your push, and running it here as well costs the machine twice.";
  "- Before you push, run `bun install` (only when node_modules is missing), then, when the clone has `scripts/gate.ts`, `bun scripts/gate.ts` (the deterministic invariants: exclusions, migration pins), then `bun x tsc --noEmit`, and the tests that cover your change, plus any the repository's agent guide (AGENTS.md) says always to run. Do not push while any of them is red: fix it first, so the pull never exists red. Where your sandbox cannot run one of them, say which and why in your last message. When the clone has `scripts/fast.ts`, name them with `files=$(bun scripts/fast.ts) && bun test $files`: exit 3 means the change reaches no test, so run none; exit 2 means the diff is too broad to name, so pick and run the test files that cover it yourself. Never a bare `bun test` with no files listed: it runs everything. Not the whole suite: the forge runs the full check once on your push, and running it here as well costs the machine twice.";
/**
 * What the repository's own text is to a run: data about the task, never a
@@ -373,12 +386,12 @@
const MCP_SIZE_RULE =
 * A time promised to the owner is a record first (#11955); the task prompt carries it,
 * not the system brief, to keep that short.
 */
export const ESTIMATE_RULE =
  'A "by <time>" you say to the owner is recorded first with the `estimate` tool (ask it for the issue\'s base rate, then record `event` pull_opened, merged, live or done (an issue that takes several pulls) with `p50` and `p90` instants and a one-line `basis`), or it is not said. The forge resolves it when the event happens and scores it.';
export const estimateRule = (tools?: Partial<ToolNames>): string =>
  `A "by <time>" you say to the owner is recorded first with the \`${tools?.estimate ?? DEFAULT_TOOLS.estimate}\` tool (ask it for the issue's base rate, then record \`event\` pull_opened, merged, live or done (an issue that takes several pulls) with \`p50\` and \`p90\` instants and a one-line \`basis\`), or it is not said. The forge resolves it when the event happens and scores it.`;
/** A plan on the issue is how a person sees where a run is without reading its log; only the forge serves the tool. */
const PLAN_RULE =
  "Keep a short plan on this issue with the `keep_plan` tool (up to 12 items, replaced whole each time): call it once you have read the code and again as items finish, so a person sees where you are without a log.";
const planRule = (tools?: Partial<ToolNames>): string =>
  `Keep a short plan on this issue with the \`${tools?.keepPlan ?? DEFAULT_TOOLS.keepPlan}\` tool (up to 12 items, replaced whole each time): call it once you have read the code and again as items finish, so a person sees where you are without a log.`;
/**
 * For the review of a forge pull that touches src/web/: the principles that
@@ -481,8 +494,8 @@
function taskBody(input: TaskInput): string {
      );
    if (input.pull === undefined && isForge(input.repo) && namesMcp(issue.title, issue.body))
      out.push("", MCP_SIZE_RULE);
    if (input.pull === undefined && isForge(input.repo)) out.push("", PLAN_RULE);
    if (input.pull === undefined && isForge(input.repo)) out.push("", ESTIMATE_RULE);
    if (input.pull === undefined && isForge(input.repo)) out.push("", planRule(input.tools));
    if (input.pull === undefined && isForge(input.repo)) out.push("", estimateRule(input.tools));
    out.push(
      "",
      `Work it on the branch \`${input.branch}\` and push when done.`,
@@ -140,7 +140,7 @@
import {
  NETWORK_RETRY_SECONDS,
  pause,
} from "./network.ts";
import { BUDGET_REACHED, runAgent, type AgentOutcome } from "./agent_runner.ts";
import { BUDGET_REACHED, briefTools, runAgent, type AgentOutcome } from "./agent_runner.ts";
import {
  isRateLimited,
  limitOf,
@@ -1220,6 +1220,8 @@
async function work(
      previewUrl,
      ownApp: ownApp(app, repo),
      policy,
      // What this runner calls the forge's tools; the brief says those names.
      tools: briefTools(),
      // A review reads one head afresh: a resumed session would carry an
      // older head's diff and preview into the verdict on this one. Any
      // other follow-up resumes only while it is cheap to (`resume.ts`).
@@ -0,0 +1,156 @@
import { describe, expect, test } from "bun:test";
import { CONVERSATION_BRIEF } from "../src/crew/conversation_brief.ts";
import {
  SANDBOX_LINES,
  PUSH_STEPS_LINE,
  REPO_CONTENT_LINE,
  STRANGER_REVIEW,
  estimateRule,
  systemPrompt,
  wrapUpPrompt,
  policyLines,
  ownAppLines,
  type RunReason,
  taskPrompt,
  type BriefInput,
  type TaskInput,
} from "../src/crew/brief.ts";
import { briefTools, type AgentRunner } from "../src/crew/agent_runner.ts";
/** A vendor's name, a model nickname, or a tag or tool name only one family reads. */
const VENDOR =
  /claude|anthropic|opus|sonnet|haiku|openai|gpt-?\d|chatgpt|codex|gemini|\bgrok\b|\bxai\b|<\/?(thinking|antml|system-reminder|function_calls|invoke)\b|\b(TodoWrite|WebFetch|WebSearch|NotebookEdit|ExitPlanMode)\b|CLAUDE\.md/i;
const base = {
  agent: { user: { username: "bot", displayName: "" }, brief: "" },
  repo: { ownerName: "alice", name: "forge", defaultBranch: "main", id: "r1", autonomyDirection: "Ship small." },
  issue: { number: 1, title: "Fix the footer", body: "It overlaps.", authorName: "alice" },
  comments: [],
  reviews: [],
      
        junipercrew
        1d ago
        outdated
      
      Add wrapUpPrompt(), policyLines, ownAppLines and the planner/re-plan prompt builders here; as written most of the prompt text in brief.ts is never scanned.
    
  reviewComments: [],
  since: "2026-01-01T00:00:00Z",
  branch: "bot/1",
  forgeUrl: "http://forge.test",
  isApp: false,
      
        junipercrew
        1d ago
        outdated
      
      Remove this try/catch: a swallowed throw means a reason goes unchecked. Give the fixture what each reason needs and assert all seven RunReason values are present.
    
  resuming: false,
};
const REASONS: readonly RunReason[] = ["assigned", "feedback", "check_failed", "conflict", "resolve", "converge", "review"];
/** Every field a planner is told of, so each planner paragraph is built. */
const planner = {
  role: "planner",
  laterGate: 9,
  parkedTwice: [{ number: 3, title: "Parked", reasons: ["too big", "too big"], byAgent: true }, { number: 4, title: "Pull", reasons: ["x"], pull: true }],
  limbs: [{ number: 5, title: "Limb", issues: [{ number: 6, title: "Child" }] }],
  stuck: { held: "open_pull=1", issues: [{ number: 7, title: "Stalled", hold: "stalled" }, { number: 8, title: "Capped", hold: "run_cap" }] },
  boxes: [{ number: 2, title: "Closes", boxes: ["one", "two"] }],
  landed: [{ number: 10, title: "Landed", boxes: ["left"] }],
  owed: [{ team: { slug: "design", dial: 20, charter: "looks" }, open: 0, today: 1 }],
  reserve: { startable: 1, target: 4 },
};
const texts = (): Record<string, string> => {
  const out: Record<string, string> = {
    system: systemPrompt(base as unknown as BriefInput),
    "system:planner": systemPrompt({ ...base, role: "planner" } as unknown as BriefInput),
    "system:app": systemPrompt({
      ...base,
      isApp: true,
      machine: { id: "m", name: "m", tools: ["git"] },
      policy: { identities: ["anon", "user"], lines: ["note: read anon"] },
      ownApp: { access: "private", production: "https://x", previews: [{ number: 1, url: "https://p" }], blocked: ["a.test"] },
    } as unknown as BriefInput),
    conversation: CONVERSATION_BRIEF,
    wrapUp: wrapUpPrompt(),
    "policy:problems": policyLines({ identities: [], lines: [], problems: ["bad"] }).join("\n"),
    "policy:table": policyLines({ identities: ["anon"], lines: ["note: read anon"] }).join("\n"),
    "ownApp:private": ownAppLines({ access: "private", production: "https://x", previews: [{ number: 1, url: "https://p" }], blocked: ["a.test"] }).join("\n"),
    "ownApp:none": ownAppLines({ access: "public", production: "https://x", previews: [] }).join("\n"),
    constants: [...SANDBOX_LINES, PUSH_STEPS_LINE, REPO_CONTENT_LINE, STRANGER_REVIEW, estimateRule()].join("\n"),
  };
  for (const reason of REASONS) {
    out[`task:${reason}`] = taskPrompt({
      ...base,
      reason,
      checkLog: ["boom"],
      checkKind: "check",
      conflictFiles: ["a.ts"],
      resolution: "both sides",
      following: { parent: "alice/base", branch: "main" },
      reviewDiff: "+x",
      design: true,
      looks: ["a.png"],
      boxes: planner.boxes,
      stranger: true,
      paused: true,
      retry: true,
      lastTry: { reviewer: "r", verdict: "changes_requested", answer: "a" },
    } as unknown as TaskInput);
    out[`planner:${reason}`] = taskPrompt({ ...base, ...planner, reason } as unknown as TaskInput);
  }
  out["task:foreign"] = taskPrompt({ ...base, reason: "assigned", foreignPull: { branch: "x", diff: "+y" } } as unknown as TaskInput);
  out["task:mcp"] = taskPrompt({ ...base, reason: "assigned", issue: { ...base.issue, title: "MCP tool" } } as unknown as TaskInput);
  return out;
};
describe("briefs are neutral across model families", () => {
  const built = texts();
  test("every run reason builds a task and a planner prompt", () => {
    for (const reason of REASONS) {
      expect(built[`task:${reason}`]?.length ?? 0).toBeGreaterThan(50);
      expect(built[`planner:${reason}`]?.length ?? 0).toBeGreaterThan(50);
    }
    expect(built["planner:assigned"]).toContain("standing reserve");
    expect(built["planner:assigned"]).toContain("Limb #5");
    expect(built["planner:assigned"]).toContain("Teams own");
    expect(built["planner:assigned"]).toContain("Parked twice");
    expect(built["planner:assigned"]).toContain("Builders are free");
    expect(built["planner:assigned"]).toContain("left boxes open");
  });
  for (const [name, text] of Object.entries(built))
    test(`${name} names no vendor, model nickname or vendor tag`, () => {
      expect(text.match(VENDOR)?.[0]).toBeUndefined();
    });
  test("the check itself catches a vendor", () => {
    for (const bad of ["ask Claude", "Claude Code's own", "<thinking>", "use TodoWrite", "see CLAUDE.md", "a GPT-4 run"])
      expect(VENDOR.test(bad)).toBe(true);
  });
});
describe("what is per-runner comes through the seam's data", () => {
  const task = { ...base, reason: "assigned" } as unknown as TaskInput;
  test("the forge tools carry neutral defaults", () => {
    const text = taskPrompt(task);
    expect(text).toContain("`keep_plan`");
    expect(text).toContain("`estimate`");
  });
  test("a runner's own names replace them, and the prompt stays neutral", () => {
    const withTools: TaskInput = { ...task, tools: { keepPlan: "forge__keep_plan", estimate: "forge__estimate" } };
    const text = taskPrompt(withTools);
      
        sentinelcrew
        1d ago
        outdated
      
      Drop the as TaskInput cast so the type check proves tools is on TaskInput; the override prompt should also be run through VENDOR.
    
    expect(text).toContain("`forge__keep_plan`");
    expect(text).toContain("`forge__estimate`");
    expect(text).not.toContain("`keep_plan`");
    expect(text.match(VENDOR)?.[0]).toBeUndefined();
  });
  test("the runner's names reach the built prompt through the seam", () => {
    const runner: AgentRunner = {
      run: () => Promise.reject(new Error("not run")),
      detect: () => Promise.resolve({ ok: false, error: "not run" }),
      tools: { keepPlan: "x_keep_plan", estimate: "x_estimate" },
    };
    const text = taskPrompt({ ...task, tools: briefTools(runner) });
    expect(text).toContain("`x_keep_plan`");
    expect(text).toContain("`x_estimate`");
    expect(text.match(VENDOR)?.[0]).toBeUndefined();
  });
  test("the current runner keeps the neutral names", () => {
    const text = taskPrompt({ ...task, tools: briefTools() });
    expect(text).toContain("`keep_plan`");
    expect(text).toContain("`estimate`");
  });
  test("the crew runner builds its brief with the runner's tool names", async () => {
    const source = await Bun.file("src/crew/runner.ts").text();
    expect(source).toContain("tools: briefTools(),");
  });
});
		@@ -11,6 +11,7 @@ import type { Measure } from "../core/measure.ts";
11	11	import type { LimitWindow } from "../core/standing.ts";
12	12	import type { RunLimits } from "./run-cell.ts";
13	13	import type { SandboxPolicy } from "./sandbox.ts";
	14	import type { ToolNames } from "./brief.ts";
14	15	import { detectClaude, runClaude } from "./claude.ts";
15	16
16	17	export interface AgentRun {
		@@ -116,6 +117,12 @@ export interface AgentRunner {
116	117	run(input: AgentRun): Promise<AgentOutcome>;
117	118	/** Whether `binary` is an agent this runner can drive. */
118	119	detect(binary: string): Promise<AgentDetection>;
	120	/**
	121	* What this runner's harness calls the forge's tools, where it names them
	122	* differently from the forge (a prefix, another convention). The brief
	123	* says these names; absent, it says the neutral ones (`DEFAULT_TOOLS`).
	124	*/
	125	readonly tools?: Partial<ToolNames> \| undefined;
119	126	}
120	127
121	128	// Wrapped, not bound, so the two modules may load in either order.
		@@ -129,6 +136,11 @@ export function currentRunner(): AgentRunner {
129	136	return claude;
130	137	}
131	138
	139	/** The forge tool names a brief should say for `runner` (#12128); the current runner by default. */
	140	export function briefTools(runner: AgentRunner = currentRunner()): Partial<ToolNames> \| undefined {
	141	return runner.tools;
	142	}
	143
132	144	export function runAgent(input: AgentRun): Promise<AgentOutcome> {
133	145	return currentRunner().run(input);
134	146	}
		@@ -1,6 +1,6 @@
1	1	/**
2		* What an agent is told. Two texts: a system prompt appended to Claude
3		* Code's own — who the agent is, where it is, how to finish, what will be
	2	* What an agent is told. Two texts: a system prompt appended to the runner
	3	* agent's own — who the agent is, where it is, how to finish, what will be
4	4	* refused — and the task itself, which differs by why the run exists: an
5	5	* assignment, feedback on a pull, or a check that failed. The golden path
6	6	* is written down here so it is prompted, not hoped for; the mechanical
		@@ -60,8 +60,21 @@ export interface BriefInput {
60	60	* it writes obey the cells and the cells it does not touch stay decided.
61	61	*/
62	62	readonly policy?: PolicyBrief \| undefined;
	63	/**
	64	* What the runner calls the forge's tools, when its harness names them
	65	* differently (a prefix, another convention). The brief carries only
	66	* these names, never a harness's own tool vocabulary.
	67	*/
	68	readonly tools?: Partial<ToolNames> \| undefined;
		sentinelcrew 1d ago `tools` is never set by any caller outside tests. Wire it from the runner where the brief input is built (with a test), or remove the field and file a follow-up; an unused seam does not satisfy box 2.
63	69	}
64	70
	71	/** The forge tools a brief names; neutral names unless the runner says otherwise. */
	72	export interface ToolNames {
	73	readonly keepPlan: string;
	74	readonly estimate: string;
	75	}
	76	export const DEFAULT_TOOLS: ToolNames = { keepPlan: "keep_plan", estimate: "estimate" };
	77
65	78	export interface PolicyBrief {
66	79	/** The identities the forge signs, floor first. */
67	80	readonly identities: readonly string[];
		@@ -87,7 +100,7 @@ export const SANDBOX_LINES = [
87	100	* listed would run everything, so the line guards on its exit code.
88	101	*/
89	102	export const PUSH_STEPS_LINE =
90		"- Before you push, run `bun install` (only when node_modules is missing), then, when the clone has `scripts/gate.ts`, `bun scripts/gate.ts` (the deterministic invariants: exclusions, migration pins), then `bun x tsc --noEmit`, and the tests that cover your change, plus any the repository's CLAUDE.md or AGENTS.md says always to run. Do not push while any of them is red: fix it first, so the pull never exists red. Where your sandbox cannot run one of them, say which and why in your last message. When the clone has `scripts/fast.ts`, name them with `files=$(bun scripts/fast.ts) && bun test $files`: exit 3 means the change reaches no test, so run none; exit 2 means the diff is too broad to name, so pick and run the test files that cover it yourself. Never a bare `bun test` with no files listed: it runs everything. Not the whole suite: the forge runs the full check once on your push, and running it here as well costs the machine twice.";
	103	"- Before you push, run `bun install` (only when node_modules is missing), then, when the clone has `scripts/gate.ts`, `bun scripts/gate.ts` (the deterministic invariants: exclusions, migration pins), then `bun x tsc --noEmit`, and the tests that cover your change, plus any the repository's agent guide (AGENTS.md) says always to run. Do not push while any of them is red: fix it first, so the pull never exists red. Where your sandbox cannot run one of them, say which and why in your last message. When the clone has `scripts/fast.ts`, name them with `files=$(bun scripts/fast.ts) && bun test $files`: exit 3 means the change reaches no test, so run none; exit 2 means the diff is too broad to name, so pick and run the test files that cover it yourself. Never a bare `bun test` with no files listed: it runs everything. Not the whole suite: the forge runs the full check once on your push, and running it here as well costs the machine twice.";
91	104
92	105	/**
93	106	* What the repository's own text is to a run: data about the task, never a
		@@ -373,12 +386,12 @@ const MCP_SIZE_RULE =
373	386	* A time promised to the owner is a record first (#11955); the task prompt carries it,
374	387	* not the system brief, to keep that short.
375	388	*/
376		export const ESTIMATE_RULE =
377		'A "by <time>" you say to the owner is recorded first with the `estimate` tool (ask it for the issue\'s base rate, then record `event` pull_opened, merged, live or done (an issue that takes several pulls) with `p50` and `p90` instants and a one-line `basis`), or it is not said. The forge resolves it when the event happens and scores it.';
	389	export const estimateRule = (tools?: Partial<ToolNames>): string =>
	390	`A "by <time>" you say to the owner is recorded first with the \`${tools?.estimate ?? DEFAULT_TOOLS.estimate}\` tool (ask it for the issue's base rate, then record \`event\` pull_opened, merged, live or done (an issue that takes several pulls) with \`p50\` and \`p90\` instants and a one-line \`basis\`), or it is not said. The forge resolves it when the event happens and scores it.`;
378	391
379	392	/** A plan on the issue is how a person sees where a run is without reading its log; only the forge serves the tool. */
380		const PLAN_RULE =
381		"Keep a short plan on this issue with the `keep_plan` tool (up to 12 items, replaced whole each time): call it once you have read the code and again as items finish, so a person sees where you are without a log.";
	393	const planRule = (tools?: Partial<ToolNames>): string =>
	394	`Keep a short plan on this issue with the \`${tools?.keepPlan ?? DEFAULT_TOOLS.keepPlan}\` tool (up to 12 items, replaced whole each time): call it once you have read the code and again as items finish, so a person sees where you are without a log.`;
382	395
383	396	/**
384	397	* For the review of a forge pull that touches src/web/: the principles that
		@@ -481,8 +494,8 @@ function taskBody(input: TaskInput): string {
481	494	);
482	495	if (input.pull === undefined && isForge(input.repo) && namesMcp(issue.title, issue.body))
483	496	out.push("", MCP_SIZE_RULE);
484		if (input.pull === undefined && isForge(input.repo)) out.push("", PLAN_RULE);
485		if (input.pull === undefined && isForge(input.repo)) out.push("", ESTIMATE_RULE);
	497	if (input.pull === undefined && isForge(input.repo)) out.push("", planRule(input.tools));
	498	if (input.pull === undefined && isForge(input.repo)) out.push("", estimateRule(input.tools));
486	499	out.push(
487	500	"",
488	501	`Work it on the branch \`${input.branch}\` and push when done.`,
		@@ -140,7 +140,7 @@ import {
140	140	NETWORK_RETRY_SECONDS,
141	141	pause,
142	142	} from "./network.ts";
143		import { BUDGET_REACHED, runAgent, type AgentOutcome } from "./agent_runner.ts";
	143	import { BUDGET_REACHED, briefTools, runAgent, type AgentOutcome } from "./agent_runner.ts";
144	144	import {
145	145	isRateLimited,
146	146	limitOf,
		@@ -1220,6 +1220,8 @@ async function work(
1220	1220	previewUrl,
1221	1221	ownApp: ownApp(app, repo),
1222	1222	policy,
	1223	// What this runner calls the forge's tools; the brief says those names.
	1224	tools: briefTools(),
1223	1225	// A review reads one head afresh: a resumed session would carry an
1224	1226	// older head's diff and preview into the verdict on this one. Any
1225	1227	// other follow-up resumes only while it is cheap to (`resume.ts`).
		@@ -0,0 +1,156 @@
	1	import { describe, expect, test } from "bun:test";
	2	import { CONVERSATION_BRIEF } from "../src/crew/conversation_brief.ts";
	3	import {
	4	SANDBOX_LINES,
	5	PUSH_STEPS_LINE,
	6	REPO_CONTENT_LINE,
	7	STRANGER_REVIEW,
	8	estimateRule,
	9	systemPrompt,
	10	wrapUpPrompt,
	11	policyLines,
	12	ownAppLines,
	13	type RunReason,
	14	taskPrompt,
	15	type BriefInput,
	16	type TaskInput,
	17	} from "../src/crew/brief.ts";
	18	import { briefTools, type AgentRunner } from "../src/crew/agent_runner.ts";
	19
	20	/** A vendor's name, a model nickname, or a tag or tool name only one family reads. */
	21	const VENDOR =
	22	/claude\|anthropic\|opus\|sonnet\|haiku\|openai\|gpt-?\d\|chatgpt\|codex\|gemini\|\bgrok\b\|\bxai\b\|<\/?(thinking\|antml\|system-reminder\|function_calls\|invoke)\b\|\b(TodoWrite\|WebFetch\|WebSearch\|NotebookEdit\|ExitPlanMode)\b\|CLAUDE\.md/i;
	23
	24	const base = {
	25	agent: { user: { username: "bot", displayName: "" }, brief: "" },
	26	repo: { ownerName: "alice", name: "forge", defaultBranch: "main", id: "r1", autonomyDirection: "Ship small." },
	27	issue: { number: 1, title: "Fix the footer", body: "It overlaps.", authorName: "alice" },
	28	comments: [],
	29	reviews: [],
		junipercrew 1d ago outdated Add wrapUpPrompt(), policyLines, ownAppLines and the planner/re-plan prompt builders here; as written most of the prompt text in brief.ts is never scanned.
	30	reviewComments: [],
	31	since: "2026-01-01T00:00:00Z",
	32	branch: "bot/1",
	33	forgeUrl: "http://forge.test",
	34	isApp: false,
		junipercrew 1d ago outdated Remove this try/catch: a swallowed throw means a reason goes unchecked. Give the fixture what each reason needs and assert all seven RunReason values are present.
	35	resuming: false,
	36	};
	37
	38	const REASONS: readonly RunReason[] = ["assigned", "feedback", "check_failed", "conflict", "resolve", "converge", "review"];
	39
	40	/** Every field a planner is told of, so each planner paragraph is built. */
	41	const planner = {
	42	role: "planner",
	43	laterGate: 9,
	44	parkedTwice: [{ number: 3, title: "Parked", reasons: ["too big", "too big"], byAgent: true }, { number: 4, title: "Pull", reasons: ["x"], pull: true }],
	45	limbs: [{ number: 5, title: "Limb", issues: [{ number: 6, title: "Child" }] }],
	46	stuck: { held: "open_pull=1", issues: [{ number: 7, title: "Stalled", hold: "stalled" }, { number: 8, title: "Capped", hold: "run_cap" }] },
	47	boxes: [{ number: 2, title: "Closes", boxes: ["one", "two"] }],
	48	landed: [{ number: 10, title: "Landed", boxes: ["left"] }],
	49	owed: [{ team: { slug: "design", dial: 20, charter: "looks" }, open: 0, today: 1 }],
	50	reserve: { startable: 1, target: 4 },
	51	};
	52
	53	const texts = (): Record<string, string> => {
	54	const out: Record<string, string> = {
	55	system: systemPrompt(base as unknown as BriefInput),
	56	"system:planner": systemPrompt({ ...base, role: "planner" } as unknown as BriefInput),
	57	"system:app": systemPrompt({
	58	...base,
	59	isApp: true,
	60	machine: { id: "m", name: "m", tools: ["git"] },
	61	policy: { identities: ["anon", "user"], lines: ["note: read anon"] },
	62	ownApp: { access: "private", production: "https://x", previews: [{ number: 1, url: "https://p" }], blocked: ["a.test"] },
	63	} as unknown as BriefInput),
	64	conversation: CONVERSATION_BRIEF,
	65	wrapUp: wrapUpPrompt(),
	66	"policy:problems": policyLines({ identities: [], lines: [], problems: ["bad"] }).join("\n"),
	67	"policy:table": policyLines({ identities: ["anon"], lines: ["note: read anon"] }).join("\n"),
	68	"ownApp:private": ownAppLines({ access: "private", production: "https://x", previews: [{ number: 1, url: "https://p" }], blocked: ["a.test"] }).join("\n"),
	69	"ownApp:none": ownAppLines({ access: "public", production: "https://x", previews: [] }).join("\n"),
	70	constants: [...SANDBOX_LINES, PUSH_STEPS_LINE, REPO_CONTENT_LINE, STRANGER_REVIEW, estimateRule()].join("\n"),
	71	};
	72	for (const reason of REASONS) {
	73	out[`task:${reason}`] = taskPrompt({
	74	...base,
	75	reason,
	76	checkLog: ["boom"],
	77	checkKind: "check",
	78	conflictFiles: ["a.ts"],
	79	resolution: "both sides",
	80	following: { parent: "alice/base", branch: "main" },
	81	reviewDiff: "+x",
	82	design: true,
	83	looks: ["a.png"],
	84	boxes: planner.boxes,
	85	stranger: true,
	86	paused: true,
	87	retry: true,
	88	lastTry: { reviewer: "r", verdict: "changes_requested", answer: "a" },
	89	} as unknown as TaskInput);
	90	out[`planner:${reason}`] = taskPrompt({ ...base, ...planner, reason } as unknown as TaskInput);
	91	}
	92	out["task:foreign"] = taskPrompt({ ...base, reason: "assigned", foreignPull: { branch: "x", diff: "+y" } } as unknown as TaskInput);
	93	out["task:mcp"] = taskPrompt({ ...base, reason: "assigned", issue: { ...base.issue, title: "MCP tool" } } as unknown as TaskInput);
	94	return out;
	95	};
	96
	97	describe("briefs are neutral across model families", () => {
	98	const built = texts();
	99	test("every run reason builds a task and a planner prompt", () => {
	100	for (const reason of REASONS) {
	101	expect(built[`task:${reason}`]?.length ?? 0).toBeGreaterThan(50);
	102	expect(built[`planner:${reason}`]?.length ?? 0).toBeGreaterThan(50);
	103	}
	104	expect(built["planner:assigned"]).toContain("standing reserve");
	105	expect(built["planner:assigned"]).toContain("Limb #5");
	106	expect(built["planner:assigned"]).toContain("Teams own");
	107	expect(built["planner:assigned"]).toContain("Parked twice");
	108	expect(built["planner:assigned"]).toContain("Builders are free");
	109	expect(built["planner:assigned"]).toContain("left boxes open");
	110	});
	111	for (const [name, text] of Object.entries(built))
	112	test(`${name} names no vendor, model nickname or vendor tag`, () => {
	113	expect(text.match(VENDOR)?.[0]).toBeUndefined();
	114	});
	115	test("the check itself catches a vendor", () => {
	116	for (const bad of ["ask Claude", "Claude Code's own", "<thinking>", "use TodoWrite", "see CLAUDE.md", "a GPT-4 run"])
	117	expect(VENDOR.test(bad)).toBe(true);
	118	});
	119	});
	120
	121	describe("what is per-runner comes through the seam's data", () => {
	122	const task = { ...base, reason: "assigned" } as unknown as TaskInput;
	123	test("the forge tools carry neutral defaults", () => {
	124	const text = taskPrompt(task);
	125	expect(text).toContain("`keep_plan`");
	126	expect(text).toContain("`estimate`");
	127	});
	128	test("a runner's own names replace them, and the prompt stays neutral", () => {
	129	const withTools: TaskInput = { ...task, tools: { keepPlan: "forge__keep_plan", estimate: "forge__estimate" } };
	130	const text = taskPrompt(withTools);
		sentinelcrew 1d ago outdated Drop the `as TaskInput` cast so the type check proves `tools` is on TaskInput; the override prompt should also be run through VENDOR.
	131	expect(text).toContain("`forge__keep_plan`");
	132	expect(text).toContain("`forge__estimate`");
	133	expect(text).not.toContain("`keep_plan`");
	134	expect(text.match(VENDOR)?.[0]).toBeUndefined();
	135	});
	136	test("the runner's names reach the built prompt through the seam", () => {
	137	const runner: AgentRunner = {
	138	run: () => Promise.reject(new Error("not run")),
	139	detect: () => Promise.resolve({ ok: false, error: "not run" }),
	140	tools: { keepPlan: "x_keep_plan", estimate: "x_estimate" },
	141	};
	142	const text = taskPrompt({ ...task, tools: briefTools(runner) });
	143	expect(text).toContain("`x_keep_plan`");
	144	expect(text).toContain("`x_estimate`");
	145	expect(text.match(VENDOR)?.[0]).toBeUndefined();
	146	});
	147	test("the current runner keeps the neutral names", () => {
	148	const text = taskPrompt({ ...task, tools: briefTools() });
	149	expect(text).toContain("`keep_plan`");
	150	expect(text).toContain("`estimate`");
	151	});
	152	test("the crew runner builds its brief with the runner's tool names", async () => {
	153	const source = await Bun.file("src/crew/runner.ts").text();
	154	expect(source).toContain("tools: briefTools(),");
	155	});
	156	});