184 lines
6.2 KiB
TypeScript
184 lines
6.2 KiB
TypeScript
/**
|
|
* CLI-agent validator integration (CLI Agent Executor, U9).
|
|
*
|
|
* Bridges a one-shot CLI agent session into the existing validator verdict
|
|
* contract (`ValidationResult` in mission-execution-loop.ts:
|
|
* `status: "pass" | "fail" | "blocked" | "error"`).
|
|
*
|
|
* The cardinal rule: a malformed / unparseable / nonzero-exit one-shot maps to
|
|
* `error`, NEVER a silent `pass`. The verdict must be indistinguishable
|
|
* downstream from a model-executed validation run.
|
|
*/
|
|
|
|
import type {
|
|
OneShotResult,
|
|
RunOneShotOptions,
|
|
runOneShotSession as RunOneShotFn,
|
|
} from "./cli-agent/one-shot-session.js";
|
|
|
|
/** The validator verdict contract shared with model-executed runs. */
|
|
export interface ValidatorVerdict {
|
|
status: "pass" | "fail" | "blocked" | "error";
|
|
/** Per-assertion results (empty when the adapter reports a bare verdict). */
|
|
assertions: Array<{ assertionId: string; passed: boolean; message?: string }>;
|
|
summary: string;
|
|
blockedReason?: string;
|
|
}
|
|
|
|
/**
|
|
* A structured verdict an adapter may emit in its one-shot JSON result. We look
|
|
* for these fields on the parsed payload (in priority order) before falling
|
|
* back to prose inference on the `text` field.
|
|
*/
|
|
interface ParsedVerdictShape {
|
|
verdict?: unknown;
|
|
status?: unknown;
|
|
result?: unknown;
|
|
passed?: unknown;
|
|
blocked?: unknown;
|
|
is_error?: unknown;
|
|
summary?: unknown;
|
|
reason?: unknown;
|
|
assertions?: unknown;
|
|
}
|
|
|
|
/** Normalize a free-form verdict token to the contract's status set. */
|
|
export function normalizeVerdictToken(token: string): ValidatorVerdict["status"] | null {
|
|
const t = token.trim().toLowerCase();
|
|
if (["pass", "passed", "approve", "approved", "ok", "success"].includes(t)) return "pass";
|
|
if (["fail", "failed", "revise", "reject", "rejected", "failure"].includes(t)) return "fail";
|
|
if (["blocked", "block", "unavailable"].includes(t)) return "blocked";
|
|
if (["error", "errored"].includes(t)) return "error";
|
|
return null;
|
|
}
|
|
|
|
/**
|
|
* Map a parsed one-shot result payload into a validator verdict.
|
|
*
|
|
* Precedence:
|
|
* 1. explicit `verdict`/`status` string token (normalized)
|
|
* 2. boolean `passed` (true→pass, false→fail) and `blocked === true`
|
|
* 3. `is_error === true` → error (claude-shaped)
|
|
* 4. prose inference from the result text
|
|
* 5. nothing decodable → error (NEVER pass)
|
|
*/
|
|
export function mapParsedToVerdict(
|
|
parsed: Record<string, unknown>,
|
|
text: string,
|
|
): ValidatorVerdict {
|
|
const p = parsed as ParsedVerdictShape;
|
|
const summary =
|
|
(typeof p.summary === "string" && p.summary) ||
|
|
(typeof p.reason === "string" && p.reason) ||
|
|
text ||
|
|
"";
|
|
const assertions = parseAssertions(p.assertions);
|
|
|
|
// 3 (early): an adapter error flag is authoritative.
|
|
if (p.is_error === true) {
|
|
return { status: "error", assertions, summary: summary || "Adapter reported an error" };
|
|
}
|
|
|
|
// 2: explicit blocked flag.
|
|
if (p.blocked === true) {
|
|
return {
|
|
status: "blocked",
|
|
assertions,
|
|
summary: summary || "Validation blocked",
|
|
blockedReason: typeof p.reason === "string" ? p.reason : summary || "blocked",
|
|
};
|
|
}
|
|
|
|
// 1: explicit verdict / status token.
|
|
for (const candidate of [p.verdict, p.status, p.result]) {
|
|
if (typeof candidate === "string") {
|
|
const status = normalizeVerdictToken(candidate);
|
|
if (status) {
|
|
return status === "blocked"
|
|
? { status, assertions, summary: summary || "Validation blocked", blockedReason: summary }
|
|
: { status, assertions, summary };
|
|
}
|
|
}
|
|
}
|
|
|
|
// 2 (boolean): passed flag.
|
|
if (typeof p.passed === "boolean") {
|
|
return { status: p.passed ? "pass" : "fail", assertions, summary };
|
|
}
|
|
|
|
// 4: prose inference.
|
|
const inferred = inferVerdictFromProse(text);
|
|
if (inferred) return { status: inferred, assertions, summary: summary || text };
|
|
|
|
// 5: undecidable → error, never a silent pass.
|
|
return {
|
|
status: "error",
|
|
assertions,
|
|
summary: summary || "Validator produced no decodable verdict",
|
|
};
|
|
}
|
|
|
|
function parseAssertions(raw: unknown): ValidatorVerdict["assertions"] {
|
|
if (!Array.isArray(raw)) return [];
|
|
const out: ValidatorVerdict["assertions"] = [];
|
|
for (const item of raw) {
|
|
if (!item || typeof item !== "object") continue;
|
|
const a = item as Record<string, unknown>;
|
|
const assertionId =
|
|
typeof a.assertionId === "string"
|
|
? a.assertionId
|
|
: typeof a.id === "string"
|
|
? a.id
|
|
: null;
|
|
if (!assertionId) continue;
|
|
out.push({
|
|
assertionId,
|
|
passed: a.passed === true,
|
|
message: typeof a.message === "string" ? a.message : undefined,
|
|
});
|
|
}
|
|
return out;
|
|
}
|
|
|
|
/** Conservative prose inference. Only confident phrasings map to a verdict. */
|
|
export function inferVerdictFromProse(text: string): ValidatorVerdict["status"] | null {
|
|
const t = text.toLowerCase();
|
|
if (/\bblocked\b/.test(t)) return "blocked";
|
|
if (/\b(revise|revision requested|does not (pass|meet)|fail(s|ed)?\b)/.test(t)) return "fail";
|
|
if (/\b(all (assertions|checks) pass|validation pass(ed)?|approve(d)?)\b/.test(t)) return "pass";
|
|
return null;
|
|
}
|
|
|
|
/**
|
|
* Convert any one-shot result (success or failure) into a validator verdict.
|
|
* Failures (nonzero exit, unparseable, spawn-failed) → `error` with the bounded
|
|
* stderr folded into the summary.
|
|
*/
|
|
export function oneShotResultToVerdict(result: OneShotResult): ValidatorVerdict {
|
|
if (!result.ok) {
|
|
return {
|
|
status: "error",
|
|
assertions: [],
|
|
summary: `${result.message}${result.stderr ? `\n--- output tail ---\n${result.stderr}` : ""}`,
|
|
};
|
|
}
|
|
return mapParsedToVerdict(result.parsed, result.text);
|
|
}
|
|
|
|
/**
|
|
* Run a CLI-agent one-shot validation and map it to the verdict contract.
|
|
*
|
|
* This is the integration seam mission-execution-loop's `runValidation` branches
|
|
* to when the resolved validator executor is a CLI agent (adapter-backed) rather
|
|
* than a model. The `run` parameter is injected so this is unit-testable without
|
|
* a live PTY (tests pass a stubbed runner; production passes
|
|
* `runOneShotSession`).
|
|
*/
|
|
export async function runCliAgentValidation(
|
|
opts: Omit<RunOneShotOptions, "purpose">,
|
|
run: typeof RunOneShotFn,
|
|
): Promise<ValidatorVerdict> {
|
|
const result = await run({ ...opts, purpose: "validator" });
|
|
return oneShotResultToVerdict(result);
|
|
}
|