gentle-pi 4.2.0 → 4.3.0

package/lib/review-risk-assessment.ts

@@ -267,8 +267,9 @@ export type VerificationTier = (typeof VERIFICATION_TIER)[keyof typeof VERIFICAT
 // ---------------------------------------------------------------------------
 // Agent-raised risk (gentle-shell#1494). The agent that made the change may
 // raise its risk to high by citing one item of the high-risk list in the
-// always-on Task Size section; it can never lower a tier. Lowering stays with
-// the deterministic native assessment.
+// always-on Task Size section; it can never lower a tier. That escalation is
+// the only way a plan gets an independent verifier: a native high or a failed
+// assessment informs but never adds one by itself.
 // ---------------------------------------------------------------------------
 
 export const HIGH_RISK_ITEMS = Object.freeze({
@@ -305,10 +306,17 @@ export function decodeAgentRiskEscalation(value: unknown): AgentRiskEscalation {
 	return Object.freeze({ item: item as HighRiskItem, reason });
 }
 
-/** Raises passive or medium to high when the agent escalated; never lowers. */
-export function escalatedRisk(risk: VerificationTier, escalation: AgentRiskEscalation | undefined): VerificationTier {
-	if (escalation === undefined) return risk;
-	return risk === VERIFICATION_TIER.PASSIVE || risk === VERIFICATION_TIER.MEDIUM ? VERIFICATION_TIER.HIGH : risk;
+/**
+ * The risk tier the verification plan follows. The independent verifier is
+ * for concrete occasions only: the plan is high only when the agent escalated
+ * a named Task Size high-risk item. Without an escalation a native passive
+ * stays passive, and every other native tier -- medium, high, or unassessable
+ * (a failed or unavailable assessment) -- is medium: it informs, but adds no
+ * verifier by itself. The native tier stays visible to callers as `nativeRisk`.
+ */
+export function verificationRisk(nativeRisk: VerificationTier, escalation: AgentRiskEscalation | undefined): VerificationTier {
+	if (escalation !== undefined) return VERIFICATION_TIER.HIGH;
+	return nativeRisk === VERIFICATION_TIER.PASSIVE ? VERIFICATION_TIER.PASSIVE : VERIFICATION_TIER.MEDIUM;
 }
 
 export const RDD_LINE = {
@@ -342,7 +350,9 @@ export type NativeReviewOutcome = (typeof NATIVE_REVIEW_OUTCOME)[keyof typeof NA
 
 export interface VerificationPlanInput {
 	readonly rddLine: RddLine;
+	/** The plan risk from `verificationRisk`, not the raw native tier. */
 	readonly risk: VerificationTier;
+	/** Reported context only: the writer's model or effort never changes the plan. */
 	readonly writerProfile: WriterProfile;
 	/**
 	 * Native review outcome for the current candidate. Only consulted when
@@ -366,7 +376,7 @@ export interface VerificationPlan {
  * naming which branch fired, up to and including the word "and"; the
  * tier-specific clause is appended after it.
  */
-function riskGatedPlan(gatePrefix: string, risk: VerificationTier, writerProfile: WriterProfile): VerificationPlan {
+function riskGatedPlan(gatePrefix: string, risk: VerificationTier): VerificationPlan {
 	if (risk === VERIFICATION_TIER.PASSIVE) {
 		return Object.freeze({
 			writerSelfVerification: false,
@@ -376,26 +386,22 @@ function riskGatedPlan(gatePrefix: string, risk: VerificationTier, writerProfile
 		});
 	}
 
-	if (risk === VERIFICATION_TIER.MEDIUM) {
-		const smallModelBias = writerProfile === WRITER_PROFILE.SMALL;
+	if (risk === VERIFICATION_TIER.HIGH) {
 		return Object.freeze({
 			writerSelfVerification: true,
 			structuralReadbackOnly: false,
-			independentVerifier: smallModelBias,
-			reason: smallModelBias
-				? `${gatePrefix} the native assessment reports medium risk, but the small-model bias raises it one tier to high for verification purposes: the writer self-verifies and a separate independent verifier also runs.`
-				: `${gatePrefix} the native assessment reports medium risk with a large writer profile: the writer's self-verification stands and no separate verifier is required.`,
+			independentVerifier: true,
+			reason: `${gatePrefix} an agent escalation names a Task Size high-risk item: the writer self-verifies and a separate independent verifier runs.`,
 		});
 	}
 
-	// high or unassessable.
+	// medium, or an unassessable tier that did not go through verificationRisk:
+	// both inform only.
 	return Object.freeze({
 		writerSelfVerification: true,
 		structuralReadbackOnly: false,
-		independentVerifier: true,
-		reason: risk === VERIFICATION_TIER.UNASSESSABLE
-			? `${gatePrefix} the native risk assessment could not be produced, so the candidate is treated as high risk: the writer self-verifies and a separate independent verifier always runs.`
-			: `${gatePrefix} the native assessment reports high risk: the writer self-verifies and a separate independent verifier always runs.`,
+		independentVerifier: false,
+		reason: `${gatePrefix} no Task Size high-risk item was escalated: the writer's self-verification stands and no separate verifier runs, whatever the native tier or the model that wrote the change.`,
 	});
 }
 
@@ -416,24 +422,23 @@ function nonClosedOutcomeClause(outcome: NativeReviewOutcome): string {
 
 /**
  * Computes who verifies a delegated writer's change. Verification follows the
- * native risk tier whatever the RDD line and native review outcome
+ * plan risk tier whatever the RDD line and native review outcome
  * (verify-always-rdd-high S1, superseding the gentle-pi#662/#668 on-path): the
  * native review is an additional outside view and never replaces or skips the
  * tier's verification. `nativeReviewOutcome` only shapes the reason text.
  *
+ * `risk` is the plan risk from `verificationRisk`; `writerProfile` is never
+ * consulted, so the model or effort that wrote the change never adds a
+ * verifier.
+ *
  * - `passive`: structural readback by the parent only; no verifier, no tests.
- * - `medium`: the writer self-verifies; a separate independent verifier runs
- *   only when `writerProfile` is `"small"` -- the small-model bias raises the
- *   tier by one for verification purposes (medium -> high). An unknown or
- *   omitted writer profile resolves to `"small"` before it reaches this
- *   function (see `resolveWriterProfile`), so an unrecognized profile never
- *   lowers a tier below what a known small model would get.
- * - `high` or `unassessable` (a failed or unrecognized native assessment is
- *   always treated as high): the writer self-verifies and an independent
- *   verifier always runs.
+ * - `medium` (also an `unassessable` tier passed directly): the writer
+ *   self-verifies; no separate verifier.
+ * - `high` (only through an agent escalation naming a Task Size high-risk
+ *   item): the writer self-verifies and an independent verifier runs.
  */
 export function verificationPlan(input: VerificationPlanInput): VerificationPlan {
-	const { rddLine, risk, writerProfile } = input;
+	const { rddLine, risk } = input;
 
 	if (rddLine === RDD_LINE.ON) {
 		const outcome = input.nativeReviewOutcome ?? NATIVE_REVIEW_OUTCOME.UNKNOWN;
@@ -443,14 +448,13 @@ export function verificationPlan(input: VerificationPlanInput): VerificationPlan
 		return riskGatedPlan(
 			`receipt-driven development is on and ${clause}; the native review is an additional outside view that never replaces verification, so the risk-gated path applies exactly as receipt-driven development off, and`,
 			risk,
-			writerProfile,
 		);
 	}
 
 	// rddLine is "off" or "unknown" here -- both gate on the native risk
 	// assessment identically, so an unknown RDD line never lowers a tier, and
 	// neither line ever consults nativeReviewOutcome.
-	return riskGatedPlan(`receipt-driven development is ${rddLine} and`, risk, writerProfile);
+	return riskGatedPlan(`receipt-driven development is ${rddLine} and`, risk);
 }
 
 // ---------------------------------------------------------------------------
@@ -475,15 +479,16 @@ export interface WriterModelProfileLike {
 const SMALL_MODEL_ID_TOKEN = /(?:^|[-_./:\s])mini(?:$|[-_./:\s])/i;
 
 /**
- * A writer profile counts as a "small model" for the medium-risk bias when:
+ * A writer profile counts as a "small model" (reported by ASSESS as
+ * `writerProfile`; it never changes the verification plan) when:
  * - its resolved effort is `low`, or
  * - its resolved model id carries `mini` as a whole token (see
  *   `SMALL_MODEL_ID_TOKEN`), or
  * - it carries NO signal at all -- no profile object, or a profile with
  *   neither a model id nor a resolved effort. An unknown or omitted writer
- *   profile fails closed to small rather than defaulting to large, so a
- *   caller that cannot resolve the writer's profile never gets a weaker
- *   verification plan than a known small model would (gentle-pi#662).
+ *   profile fails closed to small rather than defaulting to large. The
+ *   result is reported in the assessment only: the verification plan never
+ *   reads the writer profile, so it cannot add or remove a verifier.
  *
  * A profile that carries a model id or an effort, just not a matching one
  * (for example `{model: {id: "claude-sonnet-5"}, thinking: "high"}`), is
@@ -501,9 +506,9 @@ export function isSmallWriterProfile(profile: WriterModelProfileLike | undefined
 }
 
 /**
- * Maps a resolved writer profile to the `writerProfile` input `verificationPlan`
- * expects. An unknown or omitted profile never lowers a tier: it resolves to
- * `"small"`, exactly like `isSmallWriterProfile` (gentle-pi#662).
+ * Maps a resolved writer profile to the reported `writerProfile`. An unknown
+ * or omitted profile resolves to `"small"`, exactly like
+ * `isSmallWriterProfile` (gentle-pi#662).
  */
 export function resolveWriterProfile(profile: WriterModelProfileLike | undefined): WriterProfile {
 	return isSmallWriterProfile(profile) ? WRITER_PROFILE.SMALL : WRITER_PROFILE.LARGE;

package/lib/review-session-standing-permission.ts

@@ -15,6 +15,7 @@ export interface GitCommandOptions {
 	timeout: number;
 	maxBuffer: number;
 	windowsHide?: boolean;
+	stdio?: ["ignore", "pipe", "pipe"];
 }
 
 export type GitAsyncRunner = (command: string, args: readonly string[], options: GitCommandOptions) => Promise<{ stdout: string }>;
@@ -133,7 +134,8 @@ export function resolveCanonicalGitRepositoryIdentitySync(
 	run: GitSyncRunner = execFileSync as unknown as GitSyncRunner,
 ): string | undefined {
 	try {
-		const output = run("git", ["-C", cwd, "rev-parse", "--git-common-dir"], { encoding: "utf8", timeout: 5_000, maxBuffer: 64 * 1024, windowsHide: true }).trim();
+		// execFileSync otherwise echoes stderr on failure, even when caught below.
+		const output = run("git", ["-C", cwd, "rev-parse", "--git-common-dir"], { encoding: "utf8", timeout: 5_000, maxBuffer: 64 * 1024, windowsHide: true, stdio: ["ignore", "pipe", "pipe"] }).trim();
 		if (output.length === 0 || output.includes("\n") || output.includes("\r")) return undefined;
 		return canonicalRepositoryIdentity(realpathSync(resolve(cwd, output)));
 	} catch {

package/lib/session-profile-freeze.ts

@@ -27,9 +27,9 @@
 // `gentle-pi.session-profile/v1` entry with origin `local | repo | global` is
 // the integration point with slice 3b-i.
 import { createHash } from "node:crypto";
+import { normalizeModelConfig, type AgentModelConfig } from "./model-routing-authority.ts";
 import { profileRoleEntries, profilesFilePath, readProfilesFileResult } from "./agent-profiles.ts";
 import { readProfilePinStatus, resolveProfilePin, resolveUnversionedProjectProfile } from "./agent-profile-pin.ts";
-import type { AgentModelConfig } from "./model-routing-authority.ts";
 import { readSessionProfileBinding } from "./session-profile-binding.ts";
 import type { WorktreeIdentity, WorktreeResolver } from "./session-worktree-registry.ts";
 
@@ -242,6 +242,7 @@ function describe(profile: InheritedProfile | undefined): string {
  */
 export function inheritedProfileDriftNotice(options: SessionProfileOptions): string | undefined {
 	const resolution = resolveSessionProfile(options);
+	if (resolution.kind === "explicit") return bindingDefinitionDriftNotice(options, resolution);
 	if (resolution.kind !== "inherited") return undefined;
 	const current = resolveInheritedProfile(options);
 	const key = fingerprint(current);
@@ -254,6 +255,34 @@ export function inheritedProfileDriftNotice(options: SessionProfileOptions): str
 	return `el Gentleman: the default profile changed to ${describe(current)}, this session keeps ${describe(resolution.profile)}. Press Enter on a profile in /gentle:profiles to adopt it.`;
 }
 
+// gentle-shell#2228: an explicit binding re-copies the profile definition at
+// selection, so later edits to that entry in profiles.json no-op silently until
+// the operator re-Enters. The same once-per-change fingerprint machine as the
+// inherited notice names the drift; the comparison never adopts the edit.
+function bindingDefinitionDriftNotice(
+	options: SessionProfileOptions,
+	resolution: { kind: "explicit"; name: string; modelProfiles: AgentModelConfig },
+): string | undefined {
+	const file = readProfilesFileResult(profilesFilePath(options.configHome));
+	if (file.status !== "valid") return undefined;
+	const definition = file.file.profiles[resolution.name];
+	if (definition === undefined) return undefined;
+	// Both sides normalize through the store's own route normalizer: a bound
+	// snapshot taken from the panel holds normalized routes, while raw string
+	// forms from other callers must hash equal to the same definition.
+	const normalized = (config: AgentModelConfig) =>
+		Object.fromEntries(Object.entries(config).map(([agent, route]) => [agent, normalizeModelConfig(route)]));
+	const key = createHash("sha256").update(stableJson({ modelProfiles: normalized(definition) })).digest("hex");
+	const current = createHash("sha256").update(stableJson({ modelProfiles: normalized(resolution.modelProfiles) })).digest("hex");
+	if (key === current) return undefined;
+	const notified = registry().notified;
+	const seen = notified.get(options.sessionId!) ?? new Set<string>();
+	if (seen.has(key)) return undefined;
+	seen.add(key);
+	notified.set(options.sessionId!, seen);
+	return `el Gentleman: the profile definition changed to "${resolution.name}" (profiles.json), this session keeps the binding from selection. Press Enter on a profile in /gentle:profiles to adopt it.`;
+}
+
 /** Testing-only reset of every frozen snapshot and drift memory. */
 export function resetFrozenInheritedProfilesForTesting(): void {
 	registry().frozen.clear();