@caupulican/pi-adaptative 0.80.103 → 0.81.0

package/dist/core/reflection-controller.js

@@ -0,0 +1,412 @@
+/**
+ * Native reflection + learning-write controller.
+ *
+ * Extracted verbatim from agent-session.ts (god-file decomposition). Owns the end-of-loop reflection
+ * pass (R2), the isolated-completion primitive it runs on, and the learning-apply/rollback path that
+ * turns reflection writes into gated, audited durable memory/skill changes. It mutates NO session
+ * fields — every durable effect goes through the bundled memory tool, the session log (via deps), or
+ * the skills dir; the whole pass is best-effort and never throws into the turn loop. Reads live
+ * session state (model/agent/registry/memory/settings) through narrow deps accessors rather than the
+ * whole AgentSession.
+ */
+import { existsSync, mkdirSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+import { AUTONOMY_TELEMETRY_EVENT_TYPES } from "./autonomy/telemetry-events.js";
+import { APPLY_WRITE_REFUSED_REASON_CODE, appendLearningAuditSnapshot, contradictionsForReflectionWrite, getLearningAuditSnapshots, proposalFromReflectionWrite, rollbackPlanForReflectionWrite, } from "./learning/learning-audit.js";
+import { evaluateLearningDecision } from "./learning/learning-gate.js";
+import { ObservationStore, observationKey } from "./learning/observation-store.js";
+import { decideDemand, ReflectionEngine, } from "./learning/reflection-engine.js";
+export class ReflectionController {
+    deps;
+    constructor(deps) {
+        this.deps = deps;
+    }
+    /**
+     * Run a one-shot LLM completion fully ISOLATED from the main session — the load-bearing
+     * primitive for the native reflection engine (adaptive-agent design §6c/§7).
+     *
+     * Isolation invariants (audited by codex): builds a fresh {@link Context} (no main history), runs
+     * with `tools: []`, sets `cacheRetention: "none"`, and passes **no `sessionId`** — so it cannot
+     * mutate `agent.state.messages`, cannot append session entries, cannot touch the tool registry,
+     * and cannot churn the main session's prompt cache. Mirrors `generateSummary()`'s mechanics.
+     *
+     * Returns the result even on an error/aborted stop reason (callers — e.g. a background reflection
+     * microtask — decide whether to act); it does not throw on a model-level error.
+     */
+    async runIsolatedCompletion(opts) {
+        const model = opts.model ?? this.deps.getModel();
+        if (!model) {
+            throw new Error("runIsolatedCompletion: no model available");
+        }
+        const thinkingLevel = opts.thinkingLevel ?? "off";
+        // Fresh, isolated context: explicit messages, no tools, nothing from the main session.
+        const context = {
+            systemPrompt: opts.systemPrompt,
+            messages: opts.messages,
+            tools: [],
+        };
+        // Isolate the prompt cache and DELIBERATELY omit sessionId so no session-aware caching/routing
+        // can entangle this call with the main session.
+        const options = {
+            maxTokens: opts.maxTokens,
+            signal: opts.signal,
+            cacheRetention: opts.cacheRetention ?? "none",
+        };
+        // pi-ai's `reasoning` option does not include "off" (that's the provider default already).
+        if (thinkingLevel !== "off") {
+            options.reasoning = thinkingLevel;
+        }
+        // When streamFn is the raw streamSimple (e.g. in tests), auth must be injected explicitly.
+        // Throw only when auth genuinely fails — providers that authenticate without an API key
+        // (OAuth, local no-key) legitimately return ok with an undefined apiKey.
+        if (this.deps.isRawStreamSimple()) {
+            const auth = await this.deps.getModelRegistry().getApiKeyAndHeaders(model);
+            if (!auth.ok) {
+                throw new Error(auth.error);
+            }
+            options.apiKey = auth.apiKey;
+            options.headers = auth.headers;
+        }
+        const stream = await this.deps.getAgent().streamFn(model, context, options);
+        const result = await stream.result();
+        const text = result.content
+            .filter((c) => c.type === "text")
+            .map((c) => c.text)
+            .join("");
+        const usage = result.usage ?? {
+            input: 0,
+            output: 0,
+            cacheRead: 0,
+            cacheWrite: 0,
+            totalTokens: 0,
+            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+        };
+        return { text, usage, stopReason: result.stopReason };
+    }
+    /**
+     * Native end-of-loop reflection pass (R2). Demand-gates (zero-I/O), and when warranted runs the
+     * {@link ReflectionEngine} via an isolated completion ({@link runIsolatedCompletion}), applies the
+     * resulting memory writes through the bundled `memory` tool, and accounts the reflection's token
+     * cost via the cost-aggregation surface so it stays visible and net-negative-auditable.
+     *
+     * Returns `null` when the gate skips (or in a child session, which must not learn). The whole pass
+     * is best-effort: a model/parse error yields no writes, never throws into the caller.
+     */
+    async runReflectionPass(input) {
+        if (this.deps.isChildSession() || this.deps.isDisposed())
+            return null;
+        const plan = decideDemand(input.signals);
+        if (plan.act === "skip")
+            return null;
+        // Bug #21: tie this background pass to the session lifetime. Disposing the session aborts the
+        // in-flight completion (input.signal can add a more specific abort).
+        const signal = input.signal
+            ? AbortSignal.any([input.signal, this.deps.getReflectionSignal()])
+            : this.deps.getReflectionSignal();
+        const complete = (systemPrompt, userPrompt) => this.runIsolatedCompletion({
+            systemPrompt,
+            messages: [{ role: "user", content: [{ type: "text", text: userPrompt }], timestamp: Date.now() }],
+            model: input.model,
+            thinkingLevel: input.thinkingLevel ?? "low",
+            maxTokens: plan.tokenBudget,
+            signal,
+            // The reflection system prompt is static (#33) — let the provider cache the prefix so
+            // repeated passes only pay for the variable tail.
+            cacheRetention: "short",
+        });
+        const result = await new ReflectionEngine().reflect({
+            recentTurnText: input.recentTurnText,
+            // Read memory FRESH (not the prefix-cache-frozen system-prompt block) so confront-before-write
+            // sees writes made earlier this session.
+            existingMemory: this.deps.getMemoryManager().buildSystemPromptBlockFresh() || "",
+            plan,
+            complete,
+        });
+        // Bug #21: if the session was disposed while the completion was in flight, do NOT write memory
+        // or skills against the dead session.
+        if (this.deps.isDisposed())
+            return result;
+        // Learning apply policy: every durable write is converted to a proposal, decided by the
+        // learning gate, and audited with a rollback plan. With the policy disabled (default) the
+        // legacy direct-apply behavior is preserved — but now leaves audit records with rollback info.
+        const policy = this.deps.getSettingsManager().getLearningPolicySettings();
+        // The audit id sequence counts STORED snapshots only: it reseeds from the stored count on
+        // every pass, so advancing it for a no-op (which stores nothing) would make later passes
+        // reuse ids — and rollback keys on the id, so a collision blocks or misdirects rollback.
+        let auditSequence = getLearningAuditSnapshots(this.deps.getSessionManager().getEntries()).length;
+        // G6 evidence strength: durable proposals accumulate observation counts across passes/sessions
+        // so the gate can distinguish a one-off cue from a repeatedly-confirmed lesson. Built once per
+        // pass; every increment is best-effort (store IO must never break reflection).
+        const observationStore = ObservationStore.forAgentDir(this.deps.getAgentDir());
+        let writeIndex = 0;
+        for (const write of result.writes) {
+            writeIndex += 1;
+            const proposalId = `${input.reportId ?? "reflection"}-w${writeIndex}`;
+            const proposal = proposalFromReflectionWrite(write, proposalId);
+            const rollback = rollbackPlanForReflectionWrite(write);
+            let observations = 1;
+            if (policy.enabled) {
+                try {
+                    observations = observationStore.increment(observationKey(proposal.layer, proposal.summary));
+                }
+                catch {
+                    // A store read/write failure falls back to a fresh count of 1, which keeps the gate
+                    // proposal-first (never spuriously auto-applies) rather than crashing the pass.
+                    observations = 1;
+                }
+            }
+            const decision = policy.enabled
+                ? evaluateLearningDecision({
+                    proposal,
+                    confidence: policy.reflectionSourceConfidence,
+                    observations,
+                    // A replace/remove supersedes an existing durable fact — the reflection engine's
+                    // confront-before-write conflict signal — so it routes through approval instead of
+                    // silently overwriting prior memory. Additive writes contradict nothing.
+                    contradictions: contradictionsForReflectionWrite(write),
+                    settings: {
+                        enabled: true,
+                        autoApplyEnabled: policy.autoApplyEnabled,
+                        confidenceThreshold: policy.confidenceThreshold,
+                        minObservations: policy.minObservations,
+                        allowedAutoApplyLayers: policy.allowedAutoApplyLayers,
+                        requireRollbackPlan: policy.requireRollbackPlan,
+                        autoApplySupersessions: policy.autoApplySupersessions,
+                    },
+                })
+                : {
+                    kind: "apply",
+                    reasonCode: "learning_policy_disabled_legacy_apply",
+                    confidence: 0,
+                    summary: proposal.summary,
+                    requiresApproval: false,
+                };
+            this.deps.saveLearningDecisionSnapshot(decision);
+            // G3: learning-gate outcome. Codes/numbers only — never the proposal summary/memory text.
+            this.deps.emitAutonomyTelemetry({
+                type: AUTONOMY_TELEMETRY_EVENT_TYPES.learningDecision,
+                timestamp: new Date().toISOString(),
+                payload: {
+                    kind: decision.kind,
+                    reasonCode: decision.reasonCode,
+                    layer: proposal.layer,
+                    confidence: decision.confidence,
+                    requiresApproval: decision.requiresApproval,
+                },
+            });
+            // G8: a proposal that needs human sign-off is an approval REQUEST. Codes/layer only —
+            // never the proposal summary/memory text (those live only in the audit snapshot).
+            if (decision.requiresApproval) {
+                this.deps.emitAutonomyTelemetry({
+                    type: AUTONOMY_TELEMETRY_EVENT_TYPES.approvalRequest,
+                    timestamp: new Date().toISOString(),
+                    payload: {
+                        kind: decision.kind,
+                        reasonCode: decision.reasonCode,
+                        layer: proposal.layer,
+                    },
+                });
+            }
+            // The gate's decision and the write's actual outcome are two different questions: the memory
+            // tool can refuse a write (budget exceeded, drift, threat) via details.success:false without
+            // throwing. Capture that outcome instead of assuming "decision.kind === apply" means it landed
+            // — otherwise a refused write leaves a phantom "apply" audit whose rollback later fails
+            // not-found (or, worse, misfires against whatever now occupies that text).
+            const applied = decision.kind === "apply" ? await this._applyReflectionWrite(write, signal) : false;
+            const writeFailed = decision.kind === "apply" && !applied;
+            if (decision.kind !== "no-op") {
+                auditSequence += 1;
+                appendLearningAuditSnapshot(this.deps.getSessionManager(), {
+                    id: `audit-${auditSequence}`,
+                    proposalId,
+                    layer: proposal.layer,
+                    action: writeFailed ? "apply_failed" : decision.kind === "apply" ? "apply" : "propose",
+                    summary: proposal.summary,
+                    reasonCode: writeFailed ? APPLY_WRITE_REFUSED_REASON_CODE : decision.reasonCode,
+                    decision,
+                    // No rollback plan on a failed apply — nothing durable landed, so there is nothing to undo.
+                    rollback: writeFailed ? undefined : rollback,
+                    createdAt: new Date().toISOString(),
+                });
+            }
+        }
+        // Account the reflection's spend so it surfaces in the footer roll-up (net-token visibility).
+        // Idempotent on reportId so a retried/duplicated pass cannot double-count.
+        if (result.usage.cost.total > 0 || result.usage.totalTokens > 0) {
+            this.deps.addSpawnedUsage(result.usage, { label: "reflection", reportId: input.reportId });
+        }
+        return result;
+    }
+    getLearningAuditRecords() {
+        return getLearningAuditSnapshots(this.deps.getSessionManager().getEntries());
+    }
+    /**
+     * Roll back one applied durable learning change by executing the inverse operation recorded in
+     * its audit record (memory ops run through the same bundled memory-tool path as the original
+     * apply; promoted skills are archived). Appends a linked "rollback" audit record on success so
+     * the change history stays complete and a change cannot be rolled back twice.
+     */
+    async rollbackLearningWrite(auditId) {
+        if (this.deps.isDisposed())
+            return { ok: false, reason: "session_disposed" };
+        const audits = this.getLearningAuditRecords();
+        const audit = audits.find((record) => record.id === auditId);
+        if (!audit)
+            return { ok: false, reason: "audit_not_found" };
+        if (audit.action !== "apply")
+            return { ok: false, reason: "not_an_applied_change" };
+        if (audits.some((record) => record.action === "rollback" && record.rollbackOf === auditId)) {
+            return { ok: false, reason: "already_rolled_back" };
+        }
+        const rollback = audit.rollback;
+        if (!rollback)
+            return { ok: false, reason: "no_rollback_plan" };
+        // Every inverse must be VERIFIED-applied before the rollback audit is appended: a silently
+        // failed inverse that still recorded "rollback" would permanently self-lock the change
+        // behind already_rolled_back while the durable write is in fact still live.
+        switch (rollback.kind) {
+            case "memory_remove": {
+                if (!rollback.target)
+                    return { ok: false, reason: "missing_rollback_target" };
+                if (!(await this._applyReflectionWrite({ kind: "memory_remove", target: rollback.target }))) {
+                    return { ok: false, reason: "rollback_apply_failed" };
+                }
+                break;
+            }
+            case "memory_restore": {
+                if (!rollback.target || rollback.previous === undefined) {
+                    return { ok: false, reason: "missing_rollback_target" };
+                }
+                const applied = await this._applyReflectionWrite({
+                    kind: "memory_replace",
+                    target: rollback.target,
+                    text: rollback.previous,
+                });
+                if (!applied)
+                    return { ok: false, reason: "rollback_apply_failed" };
+                break;
+            }
+            case "memory_add": {
+                if (rollback.previous === undefined)
+                    return { ok: false, reason: "missing_rollback_target" };
+                const applied = await this._applyReflectionWrite({
+                    kind: "memory_add",
+                    section: "MEMORY",
+                    text: rollback.previous,
+                });
+                if (!applied)
+                    return { ok: false, reason: "rollback_apply_failed" };
+                break;
+            }
+            case "archive_skill": {
+                if (!rollback.target)
+                    return { ok: false, reason: "missing_rollback_target" };
+                if (!this.deps.archivePromotedSkill(rollback.target)) {
+                    return { ok: false, reason: "skill_archive_failed" };
+                }
+                break;
+            }
+        }
+        appendLearningAuditSnapshot(this.deps.getSessionManager(), {
+            id: `${audit.id}-rollback`,
+            proposalId: audit.proposalId,
+            layer: audit.layer,
+            action: "rollback",
+            summary: `Rolled back: ${audit.summary}`,
+            reasonCode: "user_requested_rollback",
+            decision: audit.decision,
+            rollbackOf: audit.id,
+            createdAt: new Date().toISOString(),
+        });
+        return { ok: true, reason: "rollback_applied" };
+    }
+    /**
+     * Apply one reflection write through the bundled `memory` tool. `memory_replace`/`memory_remove`
+     * don't carry a target file, so we try MEMORY.md first and fall back to USER.md when the substring
+     * isn't found there. Never throws (reflection must never break a turn); returns whether the write
+     * actually applied so callers that MUST know — rollback's once-only accounting — can react instead
+     * of recording a success that never happened.
+     */
+    async _applyReflectionWrite(write, signal) {
+        // R7 memory-to-behavior: a recurring procedure is compiled into an executable skill file rather
+        // than stored as a flat fact. Written under the agent skills dir so it loads like any user skill.
+        if (write.kind === "promote_skill") {
+            return this._promoteReflectionSkill(write.name, write.description, write.body);
+        }
+        const memTool = this.deps
+            .getMemoryManager()
+            .getToolDefinitions()
+            .find((t) => t.name === "memory");
+        const exec = memTool?.execute;
+        if (!exec)
+            return false;
+        const run = (params) => exec("reflection", params, signal, undefined, undefined);
+        if (write.kind === "memory_add") {
+            try {
+                const res = await run({
+                    action: "add",
+                    target: write.section === "USER" ? "user" : "memory",
+                    content: write.text,
+                });
+                return res?.details?.success === true;
+            }
+            catch {
+                // best-effort; reflection writes must never throw into the turn loop
+                return false;
+            }
+        }
+        // replace / remove carry no target file — try MEMORY.md, then USER.md. The memory tool reports
+        // outcomes via `details.success` (it catches its own errors rather than throwing). Only a
+        // genuine "not found in the file" justifies trying the other file; a real failure for a file
+        // (budget exceeded / drift) must NOT fall through and mutate the wrong target.
+        for (const target of ["memory", "user"]) {
+            try {
+                const params = write.kind === "memory_replace"
+                    ? { action: "replace", target, oldContent: write.target, content: write.text }
+                    : { action: "remove", target, oldContent: write.target };
+                const res = await run(params);
+                if (res?.details?.success === true)
+                    return true; // applied
+                if (!/not found/i.test(String(res?.details?.error ?? "")))
+                    return false; // real failure — don't misapply
+                // substring simply absent from this file — try the next target
+            }
+            catch {
+                // defensive: if the tool ever does throw, try the next target
+            }
+        }
+        return false;
+    }
+    /**
+     * R7: write a reflection-promoted skill as `<agentDir>/skills/<name>/SKILL.md` so it loads like any
+     * user skill. Best-effort; never clobbers an existing (hand-authored) skill of the same name.
+     */
+    _promoteReflectionSkill(rawName, description, body) {
+        const name = rawName
+            .trim()
+            .toLowerCase()
+            .replace(/[^a-z0-9-]+/g, "-")
+            .replace(/^-+|-+$/g, "")
+            .slice(0, 64);
+        if (!name || !body.trim())
+            return false;
+        try {
+            const dir = join(this.deps.getAgentDir(), "skills", name);
+            const file = join(dir, "SKILL.md");
+            if (existsSync(file))
+                return false; // do not overwrite an existing skill
+            mkdirSync(dir, { recursive: true });
+            const safeDescription = description.replace(/[\r\n]+/g, " ").trim();
+            // `promoted: true` marks this as reflection-generated so the curator (#32) can lifecycle-manage
+            // it (archive/consolidate) WITHOUT ever touching hand-authored user skills.
+            const content = `---\nname: ${name}\ndescription: ${safeDescription}\npromoted: true\n---\n\n<!-- Auto-generated by the reflection engine (R7 memory-to-behavior). Review and refine. -->\n\n${body.trim()}\n`;
+            writeFileSync(file, content, "utf-8");
+            return true;
+        }
+        catch {
+            // promotion must never break a turn
+            return false;
+        }
+    }
+}
+//# sourceMappingURL=reflection-controller.js.map

package/dist/core/reflection-controller.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"reflection-controller.js","sourceRoot":"","sources":["../../src/core/reflection-controller.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,aAAa,EAAE,MAAM,SAAS,CAAC;AAC/D,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAMjC,OAAO,EAAE,8BAA8B,EAA+B,MAAM,gCAAgC,CAAC;AAC7G,OAAO,EACN,+BAA+B,EAC/B,2BAA2B,EAC3B,gCAAgC,EAChC,yBAAyB,EAEzB,2BAA2B,EAC3B,8BAA8B,GAC9B,MAAM,8BAA8B,CAAC;AACtC,OAAO,EAAE,wBAAwB,EAAE,MAAM,6BAA6B,CAAC;AACvE,OAAO,EAAE,gBAAgB,EAAE,cAAc,EAAE,MAAM,iCAAiC,CAAC;AACnF,OAAO,EAEN,YAAY,EACZ,gBAAgB,GAGhB,MAAM,iCAAiC,CAAC;AAyCzC,MAAM,OAAO,oBAAoB;IACf,IAAI,CAA2B;IAEhD,YAAY,IAA8B,EAAE;QAC3C,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;IAAA,CACjB;IAED;;;;;;;;;;;OAWG;IACH,KAAK,CAAC,qBAAqB,CAAC,IAA+B,EAAqC;QAC/F,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC;QACjD,IAAI,CAAC,KAAK,EAAE,CAAC;YACZ,MAAM,IAAI,KAAK,CAAC,2CAA2C,CAAC,CAAC;QAC9D,CAAC;QACD,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa,IAAI,KAAK,CAAC;QAElD,uFAAuF;QACvF,MAAM,OAAO,GAAY;YACxB,YAAY,EAAE,IAAI,CAAC,YAAY;YAC/B,QAAQ,EAAE,IAAI,CAAC,QAAQ;YACvB,KAAK,EAAE,EAAE;SACT,CAAC;QAEF,+FAA+F;QAC/F,gDAAgD;QAChD,MAAM,OAAO,GAAwB;YACpC,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,cAAc,EAAE,IAAI,CAAC,cAAc,IAAI,MAAM;SAC7C,CAAC;QACF,2FAA2F;QAC3F,IAAI,aAAa,KAAK,KAAK,EAAE,CAAC;YAC7B,OAAO,CAAC,SAAS,GAAG,aAAa,CAAC;QACnC,CAAC;QAED,2FAA2F;QAC3F,0FAAwF;QACxF,yEAAyE;QACzE,IAAI,IAAI,CAAC,IAAI,CAAC,iBAAiB,EAAE,EAAE,CAAC;YACnC,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,gBAAgB,EAAE,CAAC,mBAAmB,CAAC,KAAK,CAAC,CAAC;YAC3E,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC;gBACd,MAAM,IAAI,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YAC7B,CAAC;YACD,OAAO,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC;YAC7B,OAAO,CAAC,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC;QAChC,CAAC;QAED,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;QAC5E,MAAM,MAAM,GAAG,MAAM,MAAM,CAAC,MAAM,EAAE,CAAC;QACrC,MAAM,IAAI,GAAG,MAAM,CAAC,OAAO;aACzB,MAAM,CAAC,CAAC,CAAC,EAAoB,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC;aAClD,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC;aAClB,IAAI,CAAC,EAAE,CAAC,CAAC;QACX,MAAM,KAAK,GAAU,MAAM,CAAC,KAAK,IAAI;YACpC,KAAK,EAAE,CAAC;YACR,MAAM,EAAE,CAAC;YACT,SAAS,EAAE,CAAC;YACZ,UAAU,EAAE,CAAC;YACb,WAAW,EAAE,CAAC;YACd,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;SACpE,CAAC;QACF,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,UAAU,EAAE,MAAM,CAAC,UAAU,EAAE,CAAC;IAAA,CACtD;IAED;;;;;;;;OAQG;IACH,KAAK,CAAC,iBAAiB,CAAC,KAQvB,EAAoC;QACpC,IAAI,IAAI,CAAC,IAAI,CAAC,cAAc,EAAE,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE;YAAE,OAAO,IAAI,CAAC;QACtE,MAAM,IAAI,GAAG,YAAY,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;QACzC,IAAI,IAAI,CAAC,GAAG,KAAK,MAAM;YAAE,OAAO,IAAI,CAAC;QAErC,8FAA8F;QAC9F,qEAAqE;QACrE,MAAM,MAAM,GAAG,KAAK,CAAC,MAAM;YAC1B,CAAC,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,EAAE,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE,CAAC,CAAC;YAClE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE,CAAC;QAEnC,MAAM,QAAQ,GAAG,CAAC,YAAoB,EAAE,UAAkB,EAAE,EAAE,CAC7D,IAAI,CAAC,qBAAqB,CAAC;YAC1B,YAAY;YACZ,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,CAAC;YAClG,KAAK,EAAE,KAAK,CAAC,KAAK;YAClB,aAAa,EAAE,KAAK,CAAC,aAAa,IAAI,KAAK;YAC3C,SAAS,EAAE,IAAI,CAAC,WAAW;YAC3B,MAAM;YACN,wFAAsF;YACtF,kDAAkD;YAClD,cAAc,EAAE,OAAO;SACvB,CAAC,CAAC;QAEJ,MAAM,MAAM,GAAG,MAAM,IAAI,gBAAgB,EAAE,CAAC,OAAO,CAAC;YACnD,cAAc,EAAE,KAAK,CAAC,cAAc;YACpC,+FAA+F;YAC/F,yCAAyC;YACzC,cAAc,EAAE,IAAI,CAAC,IAAI,CAAC,gBAAgB,EAAE,CAAC,2BAA2B,EAAE,IAAI,EAAE;YAChF,IAAI;YACJ,QAAQ;SACR,CAAC,CAAC;QAEH,+FAA+F;QAC/F,sCAAsC;QACtC,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE;YAAE,OAAO,MAAM,CAAC;QAE1C,wFAAwF;QACxF,0FAA0F;QAC1F,iGAA+F;QAC/F,MAAM,MAAM,GAAG,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,CAAC,yBAAyB,EAAE,CAAC;QAC1E,0FAA0F;QAC1F,yFAAyF;QACzF,2FAAyF;QACzF,IAAI,aAAa,GAAG,yBAAyB,CAAC,IAAI,CAAC,IAAI,CAAC,iBAAiB,EAAE,CAAC,UAAU,EAAE,CAAC,CAAC,MAAM,CAAC;QACjG,+FAA+F;QAC/F,+FAA+F;QAC/F,+EAA+E;QAC/E,MAAM,gBAAgB,GAAG,gBAAgB,CAAC,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,CAAC;QAC/E,IAAI,UAAU,GAAG,CAAC,CAAC;QACnB,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,MAAM,EAAE,CAAC;YACnC,UAAU,IAAI,CAAC,CAAC;YAChB,MAAM,UAAU,GAAG,GAAG,KAAK,CAAC,QAAQ,IAAI,YAAY,KAAK,UAAU,EAAE,CAAC;YACtE,MAAM,QAAQ,GAAG,2BAA2B,CAAC,KAAK,EAAE,UAAU,CAAC,CAAC;YAChE,MAAM,QAAQ,GAAG,8BAA8B,CAAC,KAAK,CAAC,CAAC;YACvD,IAAI,YAAY,GAAG,CAAC,CAAC;YACrB,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC;gBACpB,IAAI,CAAC;oBACJ,YAAY,GAAG,gBAAgB,CAAC,SAAS,CAAC,cAAc,CAAC,QAAQ,CAAC,KAAK,EAAE,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC;gBAC7F,CAAC;gBAAC,MAAM,CAAC;oBACR,oFAAoF;oBACpF,gFAAgF;oBAChF,YAAY,GAAG,CAAC,CAAC;gBAClB,CAAC;YACF,CAAC;YACD,MAAM,QAAQ,GAAqB,MAAM,CAAC,OAAO;gBAChD,CAAC,CAAC,wBAAwB,CAAC;oBACzB,QAAQ;oBACR,UAAU,EAAE,MAAM,CAAC,0BAA0B;oBAC7C,YAAY;oBACZ,mFAAiF;oBACjF,qFAAmF;oBACnF,yEAAyE;oBACzE,cAAc,EAAE,gCAAgC,CAAC,KAAK,CAAC;oBACvD,QAAQ,EAAE;wBACT,OAAO,EAAE,IAAI;wBACb,gBAAgB,EAAE,MAAM,CAAC,gBAAgB;wBACzC,mBAAmB,EAAE,MAAM,CAAC,mBAAmB;wBAC/C,eAAe,EAAE,MAAM,CAAC,eAAe;wBACvC,sBAAsB,EAAE,MAAM,CAAC,sBAAsB;wBACrD,mBAAmB,EAAE,MAAM,CAAC,mBAAmB;wBAC/C,sBAAsB,EAAE,MAAM,CAAC,sBAAsB;qBACrD;iBACD,CAAC;gBACH,CAAC,CAAC;oBACA,IAAI,EAAE,OAAO;oBACb,UAAU,EAAE,uCAAuC;oBACnD,UAAU,EAAE,CAAC;oBACb,OAAO,EAAE,QAAQ,CAAC,OAAO;oBACzB,gBAAgB,EAAE,KAAK;iBACvB,CAAC;YAEJ,IAAI,CAAC,IAAI,CAAC,4BAA4B,CAAC,QAAQ,CAAC,CAAC;YACjD,4FAA0F;YAC1F,IAAI,CAAC,IAAI,CAAC,qBAAqB,CAAC;gBAC/B,IAAI,EAAE,8BAA8B,CAAC,gBAAgB;gBACrD,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;gBACnC,OAAO,EAAE;oBACR,IAAI,EAAE,QAAQ,CAAC,IAAI;oBACnB,UAAU,EAAE,QAAQ,CAAC,UAAU;oBAC/B,KAAK,EAAE,QAAQ,CAAC,KAAK;oBACrB,UAAU,EAAE,QAAQ,CAAC,UAAU;oBAC/B,gBAAgB,EAAE,QAAQ,CAAC,gBAAgB;iBAC3C;aACD,CAAC,CAAC;YACH,wFAAsF;YACtF,kFAAkF;YAClF,IAAI,QAAQ,CAAC,gBAAgB,EAAE,CAAC;gBAC/B,IAAI,CAAC,IAAI,CAAC,qBAAqB,CAAC;oBAC/B,IAAI,EAAE,8BAA8B,CAAC,eAAe;oBACpD,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;oBACnC,OAAO,EAAE;wBACR,IAAI,EAAE,QAAQ,CAAC,IAAI;wBACnB,UAAU,EAAE,QAAQ,CAAC,UAAU;wBAC/B,KAAK,EAAE,QAAQ,CAAC,KAAK;qBACrB;iBACD,CAAC,CAAC;YACJ,CAAC;YACD,6FAA6F;YAC7F,6FAA6F;YAC7F,+FAA+F;YAC/F,0FAAwF;YACxF,2EAA2E;YAC3E,MAAM,OAAO,GAAG,QAAQ,CAAC,IAAI,KAAK,OAAO,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,qBAAqB,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC;YACpG,MAAM,WAAW,GAAG,QAAQ,CAAC,IAAI,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC;YAC1D,IAAI,QAAQ,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;gBAC/B,aAAa,IAAI,CAAC,CAAC;gBACnB,2BAA2B,CAAC,IAAI,CAAC,IAAI,CAAC,iBAAiB,EAAE,EAAE;oBAC1D,EAAE,EAAE,SAAS,aAAa,EAAE;oBAC5B,UAAU;oBACV,KAAK,EAAE,QAAQ,CAAC,KAAK;oBACrB,MAAM,EAAE,WAAW,CAAC,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,KAAK,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS;oBACtF,OAAO,EAAE,QAAQ,CAAC,OAAO;oBACzB,UAAU,EAAE,WAAW,CAAC,CAAC,CAAC,+BAA+B,CAAC,CAAC,CAAC,QAAQ,CAAC,UAAU;oBAC/E,QAAQ;oBACR,8FAA4F;oBAC5F,QAAQ,EAAE,WAAW,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,QAAQ;oBAC5C,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;iBACnC,CAAC,CAAC;YACJ,CAAC;QACF,CAAC;QAED,8FAA8F;QAC9F,2EAA2E;QAC3E,IAAI,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,IAAI,MAAM,CAAC,KAAK,CAAC,WAAW,GAAG,CAAC,EAAE,CAAC;YACjE,IAAI,CAAC,IAAI,CAAC,eAAe,CAAC,MAAM,CAAC,KAAK,EAAE,EAAE,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,KAAK,CAAC,QAAQ,EAAE,CAAC,CAAC;QAC5F,CAAC;QACD,OAAO,MAAM,CAAC;IAAA,CACd;IAED,uBAAuB,GAA0B;QAChD,OAAO,yBAAyB,CAAC,IAAI,CAAC,IAAI,CAAC,iBAAiB,EAAE,CAAC,UAAU,EAAE,CAAC,CAAC;IAAA,CAC7E;IAED;;;;;OAKG;IACH,KAAK,CAAC,qBAAqB,CAAC,OAAe,EAA4C;QACtF,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,kBAAkB,EAAE,CAAC;QAE7E,MAAM,MAAM,GAAG,IAAI,CAAC,uBAAuB,EAAE,CAAC;QAC9C,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,EAAE,KAAK,OAAO,CAAC,CAAC;QAC7D,IAAI,CAAC,KAAK;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,iBAAiB,EAAE,CAAC;QAC5D,IAAI,KAAK,CAAC,MAAM,KAAK,OAAO;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,uBAAuB,EAAE,CAAC;QACpF,IAAI,MAAM,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,MAAM,KAAK,UAAU,IAAI,MAAM,CAAC,UAAU,KAAK,OAAO,CAAC,EAAE,CAAC;YAC5F,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,qBAAqB,EAAE,CAAC;QACrD,CAAC;QACD,MAAM,QAAQ,GAAG,KAAK,CAAC,QAAQ,CAAC;QAChC,IAAI,CAAC,QAAQ;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,kBAAkB,EAAE,CAAC;QAEhE,2FAA2F;QAC3F,uFAAuF;QACvF,4EAA4E;QAC5E,QAAQ,QAAQ,CAAC,IAAI,EAAE,CAAC;YACvB,KAAK,eAAe,EAAE,CAAC;gBACtB,IAAI,CAAC,QAAQ,CAAC,MAAM;oBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,yBAAyB,EAAE,CAAC;gBAC9E,IAAI,CAAC,CAAC,MAAM,IAAI,CAAC,qBAAqB,CAAC,EAAE,IAAI,EAAE,eAAe,EAAE,MAAM,EAAE,QAAQ,CAAC,MAAM,EAAE,CAAC,CAAC,EAAE,CAAC;oBAC7F,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,uBAAuB,EAAE,CAAC;gBACvD,CAAC;gBACD,MAAM;YACP,CAAC;YACD,KAAK,gBAAgB,EAAE,CAAC;gBACvB,IAAI,CAAC,QAAQ,CAAC,MAAM,IAAI,QAAQ,CAAC,QAAQ,KAAK,SAAS,EAAE,CAAC;oBACzD,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,yBAAyB,EAAE,CAAC;gBACzD,CAAC;gBACD,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,qBAAqB,CAAC;oBAChD,IAAI,EAAE,gBAAgB;oBACtB,MAAM,EAAE,QAAQ,CAAC,MAAM;oBACvB,IAAI,EAAE,QAAQ,CAAC,QAAQ;iBACvB,CAAC,CAAC;gBACH,IAAI,CAAC,OAAO;oBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,uBAAuB,EAAE,CAAC;gBACpE,MAAM;YACP,CAAC;YACD,KAAK,YAAY,EAAE,CAAC;gBACnB,IAAI,QAAQ,CAAC,QAAQ,KAAK,SAAS;oBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,yBAAyB,EAAE,CAAC;gBAC7F,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,qBAAqB,CAAC;oBAChD,IAAI,EAAE,YAAY;oBAClB,OAAO,EAAE,QAAQ;oBACjB,IAAI,EAAE,QAAQ,CAAC,QAAQ;iBACvB,CAAC,CAAC;gBACH,IAAI,CAAC,OAAO;oBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,uBAAuB,EAAE,CAAC;gBACpE,MAAM;YACP,CAAC;YACD,KAAK,eAAe,EAAE,CAAC;gBACtB,IAAI,CAAC,QAAQ,CAAC,MAAM;oBAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,yBAAyB,EAAE,CAAC;gBAC9E,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,oBAAoB,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;oBACtD,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,sBAAsB,EAAE,CAAC;gBACtD,CAAC;gBACD,MAAM;YACP,CAAC;QACF,CAAC;QAED,2BAA2B,CAAC,IAAI,CAAC,IAAI,CAAC,iBAAiB,EAAE,EAAE;YAC1D,EAAE,EAAE,GAAG,KAAK,CAAC,EAAE,WAAW;YAC1B,UAAU,EAAE,KAAK,CAAC,UAAU;YAC5B,KAAK,EAAE,KAAK,CAAC,KAAK;YAClB,MAAM,EAAE,UAAU;YAClB,OAAO,EAAE,gBAAgB,KAAK,CAAC,OAAO,EAAE;YACxC,UAAU,EAAE,yBAAyB;YACrC,QAAQ,EAAE,KAAK,CAAC,QAAQ;YACxB,UAAU,EAAE,KAAK,CAAC,EAAE;YACpB,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;SACnC,CAAC,CAAC;QACH,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,kBAAkB,EAAE,CAAC;IAAA,CAChD;IAED;;;;;;OAMG;IACK,KAAK,CAAC,qBAAqB,CAAC,KAAsB,EAAE,MAAoB,EAAoB;QACnG,gGAAgG;QAChG,kGAAkG;QAClG,IAAI,KAAK,CAAC,IAAI,KAAK,eAAe,EAAE,CAAC;YACpC,OAAO,IAAI,CAAC,uBAAuB,CAAC,KAAK,CAAC,IAAI,EAAE,KAAK,CAAC,WAAW,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC;QAChF,CAAC;QAUD,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI;aACvB,gBAAgB,EAAE;aAClB,kBAAkB,EAAE;aACpB,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;QACnC,MAAM,IAAI,GAAG,OAAO,EAAE,OAAyC,CAAC;QAChE,IAAI,CAAC,IAAI;YAAE,OAAO,KAAK,CAAC;QAExB,MAAM,GAAG,GAAG,CAAC,MAA8B,EAAE,EAAE,CAAC,IAAI,CAAC,YAAY,EAAE,MAAM,EAAE,MAAM,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC;QAEzG,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY,EAAE,CAAC;YACjC,IAAI,CAAC;gBACJ,MAAM,GAAG,GAAG,MAAM,GAAG,CAAC;oBACrB,MAAM,EAAE,KAAK;oBACb,MAAM,EAAE,KAAK,CAAC,OAAO,KAAK,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,QAAQ;oBACpD,OAAO,EAAE,KAAK,CAAC,IAAI;iBACnB,CAAC,CAAC;gBACH,OAAO,GAAG,EAAE,OAAO,EAAE,OAAO,KAAK,IAAI,CAAC;YACvC,CAAC;YAAC,MAAM,CAAC;gBACR,qEAAqE;gBACrE,OAAO,KAAK,CAAC;YACd,CAAC;QACF,CAAC;QAED,iGAA+F;QAC/F,0FAA0F;QAC1F,6FAA6F;QAC7F,+EAA+E;QAC/E,KAAK,MAAM,MAAM,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAU,EAAE,CAAC;YAClD,IAAI,CAAC;gBACJ,MAAM,MAAM,GACX,KAAK,CAAC,IAAI,KAAK,gBAAgB;oBAC9B,CAAC,CAAC,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,UAAU,EAAE,KAAK,CAAC,MAAM,EAAE,OAAO,EAAE,KAAK,CAAC,IAAI,EAAE;oBAC9E,CAAC,CAAC,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,UAAU,EAAE,KAAK,CAAC,MAAM,EAAE,CAAC;gBAC3D,MAAM,GAAG,GAAG,MAAM,GAAG,CAAC,MAAM,CAAC,CAAC;gBAC9B,IAAI,GAAG,EAAE,OAAO,EAAE,OAAO,KAAK,IAAI;oBAAE,OAAO,IAAI,CAAC,CAAC,UAAU;gBAC3D,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC,MAAM,CAAC,GAAG,EAAE,OAAO,EAAE,KAAK,IAAI,EAAE,CAAC,CAAC;oBAAE,OAAO,KAAK,CAAC,CAAC,kCAAgC;gBACzG,iEAA+D;YAChE,CAAC;YAAC,MAAM,CAAC;gBACR,8DAA8D;YAC/D,CAAC;QACF,CAAC;QACD,OAAO,KAAK,CAAC;IAAA,CACb;IAED;;;OAGG;IACK,uBAAuB,CAAC,OAAe,EAAE,WAAmB,EAAE,IAAY,EAAW;QAC5F,MAAM,IAAI,GAAG,OAAO;aAClB,IAAI,EAAE;aACN,WAAW,EAAE;aACb,OAAO,CAAC,cAAc,EAAE,GAAG,CAAC;aAC5B,OAAO,CAAC,UAAU,EAAE,EAAE,CAAC;aACvB,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QACf,IAAI,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE;YAAE,OAAO,KAAK,CAAC;QACxC,IAAI,CAAC;YACJ,MAAM,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,EAAE,QAAQ,EAAE,IAAI,CAAC,CAAC;YAC1D,MAAM,IAAI,GAAG,IAAI,CAAC,GAAG,EAAE,UAAU,CAAC,CAAC;YACnC,IAAI,UAAU,CAAC,IAAI,CAAC;gBAAE,OAAO,KAAK,CAAC,CAAC,qCAAqC;YACzE,SAAS,CAAC,GAAG,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;YACpC,MAAM,eAAe,GAAG,WAAW,CAAC,OAAO,CAAC,UAAU,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;YACpE,gGAAgG;YAChG,4EAA4E;YAC5E,MAAM,OAAO,GAAG,cAAc,IAAI,kBAAkB,eAAe,4HAA4H,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC;YAC/M,aAAa,CAAC,IAAI,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;YACtC,OAAO,IAAI,CAAC;QACb,CAAC;QAAC,MAAM,CAAC;YACR,oCAAoC;YACpC,OAAO,KAAK,CAAC;QACd,CAAC;IAAA,CACD;CACD","sourcesContent":["/**\n * Native reflection + learning-write controller.\n *\n * Extracted verbatim from agent-session.ts (god-file decomposition). Owns the end-of-loop reflection\n * pass (R2), the isolated-completion primitive it runs on, and the learning-apply/rollback path that\n * turns reflection writes into gated, audited durable memory/skill changes. It mutates NO session\n * fields — every durable effect goes through the bundled memory tool, the session log (via deps), or\n * the skills dir; the whole pass is best-effort and never throws into the turn loop. Reads live\n * session state (model/agent/registry/memory/settings) through narrow deps accessors rather than the\n * whole AgentSession.\n */\n\nimport { existsSync, mkdirSync, writeFileSync } from \"node:fs\";\nimport { join } from \"node:path\";\nimport type { Agent, ThinkingLevel } from \"@caupulican/pi-agent-core\";\nimport type { SessionManager } from \"@caupulican/pi-agent-core/node\";\nimport type { Context, Model, SimpleStreamOptions, TextContent, Usage } from \"@caupulican/pi-ai\";\nimport type { IsolatedCompletionOptions, IsolatedCompletionResult } from \"./agent-session.ts\";\nimport type { LearningDecision } from \"./autonomy/contracts.ts\";\nimport { AUTONOMY_TELEMETRY_EVENT_TYPES, type AutonomyTelemetryEvent } from \"./autonomy/telemetry-events.ts\";\nimport {\n\tAPPLY_WRITE_REFUSED_REASON_CODE,\n\tappendLearningAuditSnapshot,\n\tcontradictionsForReflectionWrite,\n\tgetLearningAuditSnapshots,\n\ttype LearningAuditRecord,\n\tproposalFromReflectionWrite,\n\trollbackPlanForReflectionWrite,\n} from \"./learning/learning-audit.ts\";\nimport { evaluateLearningDecision } from \"./learning/learning-gate.ts\";\nimport { ObservationStore, observationKey } from \"./learning/observation-store.ts\";\nimport {\n\ttype DemandSignals,\n\tdecideDemand,\n\tReflectionEngine,\n\ttype ReflectionResult,\n\ttype ReflectionWrite,\n} from \"./learning/reflection-engine.ts\";\nimport type { MemoryManager } from \"./memory/memory-manager.ts\";\nimport type { ModelRegistry } from \"./model-registry.ts\";\nimport type { SettingsManager } from \"./settings-manager.ts\";\n\nexport interface ReflectionControllerDeps {\n\t/** Current session model (fallback for an isolated call that omits its own model). */\n\tgetModel(): Model<any> | undefined;\n\t/** The underlying agent — its `streamFn` runs the isolated completion. */\n\tgetAgent(): Agent;\n\t/** True when the session's stream fn is the raw `streamSimple` provider (auth must be injected). */\n\tisRawStreamSimple(): boolean;\n\t/** Model registry for API-key/header resolution on the raw-provider path. */\n\tgetModelRegistry(): ModelRegistry;\n\t/** Memory subsystem — the bundled `memory` tool applies durable writes; fresh block feeds reflection. */\n\tgetMemoryManager(): MemoryManager;\n\t/** Settings — the learning-apply policy (gate thresholds, auto-apply layers) is read here. */\n\tgetSettingsManager(): SettingsManager;\n\t/** Session log — audit snapshots and learning-audit reads go through this. */\n\tgetSessionManager(): SessionManager;\n\t/** Agent dir — reflection-promoted skills are written under `<agentDir>/skills/`. */\n\tgetAgentDir(): string;\n\t/** Child sessions must not learn — the pass returns null for them. */\n\tisChildSession(): boolean;\n\t/** Disposal short-circuits: no completion, no writes against a dead session. */\n\tisDisposed(): boolean;\n\t/** Session-lifetime abort signal — aborts an in-flight reflection completion on dispose. */\n\tgetReflectionSignal(): AbortSignal;\n\t/** Archive a promoted skill (rollback of a `promote_skill` write). */\n\tarchivePromotedSkill(name: string): boolean;\n\t/** G3/G8 autonomy telemetry sink for learning-gate outcomes and approval requests. */\n\temitAutonomyTelemetry(event: AutonomyTelemetryEvent): void;\n\t/** Account the reflection pass's token spend into the cost roll-up (idempotent on reportId). */\n\taddSpawnedUsage(\n\t\tusage: Usage,\n\t\topts?: { label?: string; sourceSessionId?: string; reportId?: string },\n\t): string | undefined;\n\t/** Persist a learning-gate decision snapshot to the session log. */\n\tsaveLearningDecisionSnapshot(decision: LearningDecision): string;\n}\n\nexport class ReflectionController {\n\tprivate readonly deps: ReflectionControllerDeps;\n\n\tconstructor(deps: ReflectionControllerDeps) {\n\t\tthis.deps = deps;\n\t}\n\n\t/**\n\t * Run a one-shot LLM completion fully ISOLATED from the main session — the load-bearing\n\t * primitive for the native reflection engine (adaptive-agent design §6c/§7).\n\t *\n\t * Isolation invariants (audited by codex): builds a fresh {@link Context} (no main history), runs\n\t * with `tools: []`, sets `cacheRetention: \"none\"`, and passes **no `sessionId`** — so it cannot\n\t * mutate `agent.state.messages`, cannot append session entries, cannot touch the tool registry,\n\t * and cannot churn the main session's prompt cache. Mirrors `generateSummary()`'s mechanics.\n\t *\n\t * Returns the result even on an error/aborted stop reason (callers — e.g. a background reflection\n\t * microtask — decide whether to act); it does not throw on a model-level error.\n\t */\n\tasync runIsolatedCompletion(opts: IsolatedCompletionOptions): Promise<IsolatedCompletionResult> {\n\t\tconst model = opts.model ?? this.deps.getModel();\n\t\tif (!model) {\n\t\t\tthrow new Error(\"runIsolatedCompletion: no model available\");\n\t\t}\n\t\tconst thinkingLevel = opts.thinkingLevel ?? \"off\";\n\n\t\t// Fresh, isolated context: explicit messages, no tools, nothing from the main session.\n\t\tconst context: Context = {\n\t\t\tsystemPrompt: opts.systemPrompt,\n\t\t\tmessages: opts.messages,\n\t\t\ttools: [],\n\t\t};\n\n\t\t// Isolate the prompt cache and DELIBERATELY omit sessionId so no session-aware caching/routing\n\t\t// can entangle this call with the main session.\n\t\tconst options: SimpleStreamOptions = {\n\t\t\tmaxTokens: opts.maxTokens,\n\t\t\tsignal: opts.signal,\n\t\t\tcacheRetention: opts.cacheRetention ?? \"none\",\n\t\t};\n\t\t// pi-ai's `reasoning` option does not include \"off\" (that's the provider default already).\n\t\tif (thinkingLevel !== \"off\") {\n\t\t\toptions.reasoning = thinkingLevel;\n\t\t}\n\n\t\t// When streamFn is the raw streamSimple (e.g. in tests), auth must be injected explicitly.\n\t\t// Throw only when auth genuinely fails — providers that authenticate without an API key\n\t\t// (OAuth, local no-key) legitimately return ok with an undefined apiKey.\n\t\tif (this.deps.isRawStreamSimple()) {\n\t\t\tconst auth = await this.deps.getModelRegistry().getApiKeyAndHeaders(model);\n\t\t\tif (!auth.ok) {\n\t\t\t\tthrow new Error(auth.error);\n\t\t\t}\n\t\t\toptions.apiKey = auth.apiKey;\n\t\t\toptions.headers = auth.headers;\n\t\t}\n\n\t\tconst stream = await this.deps.getAgent().streamFn(model, context, options);\n\t\tconst result = await stream.result();\n\t\tconst text = result.content\n\t\t\t.filter((c): c is TextContent => c.type === \"text\")\n\t\t\t.map((c) => c.text)\n\t\t\t.join(\"\");\n\t\tconst usage: Usage = result.usage ?? {\n\t\t\tinput: 0,\n\t\t\toutput: 0,\n\t\t\tcacheRead: 0,\n\t\t\tcacheWrite: 0,\n\t\t\ttotalTokens: 0,\n\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t};\n\t\treturn { text, usage, stopReason: result.stopReason };\n\t}\n\n\t/**\n\t * Native end-of-loop reflection pass (R2). Demand-gates (zero-I/O), and when warranted runs the\n\t * {@link ReflectionEngine} via an isolated completion ({@link runIsolatedCompletion}), applies the\n\t * resulting memory writes through the bundled `memory` tool, and accounts the reflection's token\n\t * cost via the cost-aggregation surface so it stays visible and net-negative-auditable.\n\t *\n\t * Returns `null` when the gate skips (or in a child session, which must not learn). The whole pass\n\t * is best-effort: a model/parse error yields no writes, never throws into the caller.\n\t */\n\tasync runReflectionPass(input: {\n\t\tsignals: DemandSignals;\n\t\trecentTurnText: string;\n\t\tmodel?: Model<any>;\n\t\tthinkingLevel?: ThinkingLevel;\n\t\tsignal?: AbortSignal;\n\t\t/** Stable id so a duplicate scheduling/retry of the same pass can't double-count its cost. */\n\t\treportId?: string;\n\t}): Promise<ReflectionResult | null> {\n\t\tif (this.deps.isChildSession() || this.deps.isDisposed()) return null;\n\t\tconst plan = decideDemand(input.signals);\n\t\tif (plan.act === \"skip\") return null;\n\n\t\t// Bug #21: tie this background pass to the session lifetime. Disposing the session aborts the\n\t\t// in-flight completion (input.signal can add a more specific abort).\n\t\tconst signal = input.signal\n\t\t\t? AbortSignal.any([input.signal, this.deps.getReflectionSignal()])\n\t\t\t: this.deps.getReflectionSignal();\n\n\t\tconst complete = (systemPrompt: string, userPrompt: string) =>\n\t\t\tthis.runIsolatedCompletion({\n\t\t\t\tsystemPrompt,\n\t\t\t\tmessages: [{ role: \"user\", content: [{ type: \"text\", text: userPrompt }], timestamp: Date.now() }],\n\t\t\t\tmodel: input.model,\n\t\t\t\tthinkingLevel: input.thinkingLevel ?? \"low\",\n\t\t\t\tmaxTokens: plan.tokenBudget,\n\t\t\t\tsignal,\n\t\t\t\t// The reflection system prompt is static (#33) — let the provider cache the prefix so\n\t\t\t\t// repeated passes only pay for the variable tail.\n\t\t\t\tcacheRetention: \"short\",\n\t\t\t});\n\n\t\tconst result = await new ReflectionEngine().reflect({\n\t\t\trecentTurnText: input.recentTurnText,\n\t\t\t// Read memory FRESH (not the prefix-cache-frozen system-prompt block) so confront-before-write\n\t\t\t// sees writes made earlier this session.\n\t\t\texistingMemory: this.deps.getMemoryManager().buildSystemPromptBlockFresh() || \"\",\n\t\t\tplan,\n\t\t\tcomplete,\n\t\t});\n\n\t\t// Bug #21: if the session was disposed while the completion was in flight, do NOT write memory\n\t\t// or skills against the dead session.\n\t\tif (this.deps.isDisposed()) return result;\n\n\t\t// Learning apply policy: every durable write is converted to a proposal, decided by the\n\t\t// learning gate, and audited with a rollback plan. With the policy disabled (default) the\n\t\t// legacy direct-apply behavior is preserved — but now leaves audit records with rollback info.\n\t\tconst policy = this.deps.getSettingsManager().getLearningPolicySettings();\n\t\t// The audit id sequence counts STORED snapshots only: it reseeds from the stored count on\n\t\t// every pass, so advancing it for a no-op (which stores nothing) would make later passes\n\t\t// reuse ids — and rollback keys on the id, so a collision blocks or misdirects rollback.\n\t\tlet auditSequence = getLearningAuditSnapshots(this.deps.getSessionManager().getEntries()).length;\n\t\t// G6 evidence strength: durable proposals accumulate observation counts across passes/sessions\n\t\t// so the gate can distinguish a one-off cue from a repeatedly-confirmed lesson. Built once per\n\t\t// pass; every increment is best-effort (store IO must never break reflection).\n\t\tconst observationStore = ObservationStore.forAgentDir(this.deps.getAgentDir());\n\t\tlet writeIndex = 0;\n\t\tfor (const write of result.writes) {\n\t\t\twriteIndex += 1;\n\t\t\tconst proposalId = `${input.reportId ?? \"reflection\"}-w${writeIndex}`;\n\t\t\tconst proposal = proposalFromReflectionWrite(write, proposalId);\n\t\t\tconst rollback = rollbackPlanForReflectionWrite(write);\n\t\t\tlet observations = 1;\n\t\t\tif (policy.enabled) {\n\t\t\t\ttry {\n\t\t\t\t\tobservations = observationStore.increment(observationKey(proposal.layer, proposal.summary));\n\t\t\t\t} catch {\n\t\t\t\t\t// A store read/write failure falls back to a fresh count of 1, which keeps the gate\n\t\t\t\t\t// proposal-first (never spuriously auto-applies) rather than crashing the pass.\n\t\t\t\t\tobservations = 1;\n\t\t\t\t}\n\t\t\t}\n\t\t\tconst decision: LearningDecision = policy.enabled\n\t\t\t\t? evaluateLearningDecision({\n\t\t\t\t\t\tproposal,\n\t\t\t\t\t\tconfidence: policy.reflectionSourceConfidence,\n\t\t\t\t\t\tobservations,\n\t\t\t\t\t\t// A replace/remove supersedes an existing durable fact — the reflection engine's\n\t\t\t\t\t\t// confront-before-write conflict signal — so it routes through approval instead of\n\t\t\t\t\t\t// silently overwriting prior memory. Additive writes contradict nothing.\n\t\t\t\t\t\tcontradictions: contradictionsForReflectionWrite(write),\n\t\t\t\t\t\tsettings: {\n\t\t\t\t\t\t\tenabled: true,\n\t\t\t\t\t\t\tautoApplyEnabled: policy.autoApplyEnabled,\n\t\t\t\t\t\t\tconfidenceThreshold: policy.confidenceThreshold,\n\t\t\t\t\t\t\tminObservations: policy.minObservations,\n\t\t\t\t\t\t\tallowedAutoApplyLayers: policy.allowedAutoApplyLayers,\n\t\t\t\t\t\t\trequireRollbackPlan: policy.requireRollbackPlan,\n\t\t\t\t\t\t\tautoApplySupersessions: policy.autoApplySupersessions,\n\t\t\t\t\t\t},\n\t\t\t\t\t})\n\t\t\t\t: {\n\t\t\t\t\t\tkind: \"apply\",\n\t\t\t\t\t\treasonCode: \"learning_policy_disabled_legacy_apply\",\n\t\t\t\t\t\tconfidence: 0,\n\t\t\t\t\t\tsummary: proposal.summary,\n\t\t\t\t\t\trequiresApproval: false,\n\t\t\t\t\t};\n\n\t\t\tthis.deps.saveLearningDecisionSnapshot(decision);\n\t\t\t// G3: learning-gate outcome. Codes/numbers only — never the proposal summary/memory text.\n\t\t\tthis.deps.emitAutonomyTelemetry({\n\t\t\t\ttype: AUTONOMY_TELEMETRY_EVENT_TYPES.learningDecision,\n\t\t\t\ttimestamp: new Date().toISOString(),\n\t\t\t\tpayload: {\n\t\t\t\t\tkind: decision.kind,\n\t\t\t\t\treasonCode: decision.reasonCode,\n\t\t\t\t\tlayer: proposal.layer,\n\t\t\t\t\tconfidence: decision.confidence,\n\t\t\t\t\trequiresApproval: decision.requiresApproval,\n\t\t\t\t},\n\t\t\t});\n\t\t\t// G8: a proposal that needs human sign-off is an approval REQUEST. Codes/layer only —\n\t\t\t// never the proposal summary/memory text (those live only in the audit snapshot).\n\t\t\tif (decision.requiresApproval) {\n\t\t\t\tthis.deps.emitAutonomyTelemetry({\n\t\t\t\t\ttype: AUTONOMY_TELEMETRY_EVENT_TYPES.approvalRequest,\n\t\t\t\t\ttimestamp: new Date().toISOString(),\n\t\t\t\t\tpayload: {\n\t\t\t\t\t\tkind: decision.kind,\n\t\t\t\t\t\treasonCode: decision.reasonCode,\n\t\t\t\t\t\tlayer: proposal.layer,\n\t\t\t\t\t},\n\t\t\t\t});\n\t\t\t}\n\t\t\t// The gate's decision and the write's actual outcome are two different questions: the memory\n\t\t\t// tool can refuse a write (budget exceeded, drift, threat) via details.success:false without\n\t\t\t// throwing. Capture that outcome instead of assuming \"decision.kind === apply\" means it landed\n\t\t\t// — otherwise a refused write leaves a phantom \"apply\" audit whose rollback later fails\n\t\t\t// not-found (or, worse, misfires against whatever now occupies that text).\n\t\t\tconst applied = decision.kind === \"apply\" ? await this._applyReflectionWrite(write, signal) : false;\n\t\t\tconst writeFailed = decision.kind === \"apply\" && !applied;\n\t\t\tif (decision.kind !== \"no-op\") {\n\t\t\t\tauditSequence += 1;\n\t\t\t\tappendLearningAuditSnapshot(this.deps.getSessionManager(), {\n\t\t\t\t\tid: `audit-${auditSequence}`,\n\t\t\t\t\tproposalId,\n\t\t\t\t\tlayer: proposal.layer,\n\t\t\t\t\taction: writeFailed ? \"apply_failed\" : decision.kind === \"apply\" ? \"apply\" : \"propose\",\n\t\t\t\t\tsummary: proposal.summary,\n\t\t\t\t\treasonCode: writeFailed ? APPLY_WRITE_REFUSED_REASON_CODE : decision.reasonCode,\n\t\t\t\t\tdecision,\n\t\t\t\t\t// No rollback plan on a failed apply — nothing durable landed, so there is nothing to undo.\n\t\t\t\t\trollback: writeFailed ? undefined : rollback,\n\t\t\t\t\tcreatedAt: new Date().toISOString(),\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\n\t\t// Account the reflection's spend so it surfaces in the footer roll-up (net-token visibility).\n\t\t// Idempotent on reportId so a retried/duplicated pass cannot double-count.\n\t\tif (result.usage.cost.total > 0 || result.usage.totalTokens > 0) {\n\t\t\tthis.deps.addSpawnedUsage(result.usage, { label: \"reflection\", reportId: input.reportId });\n\t\t}\n\t\treturn result;\n\t}\n\n\tgetLearningAuditRecords(): LearningAuditRecord[] {\n\t\treturn getLearningAuditSnapshots(this.deps.getSessionManager().getEntries());\n\t}\n\n\t/**\n\t * Roll back one applied durable learning change by executing the inverse operation recorded in\n\t * its audit record (memory ops run through the same bundled memory-tool path as the original\n\t * apply; promoted skills are archived). Appends a linked \"rollback\" audit record on success so\n\t * the change history stays complete and a change cannot be rolled back twice.\n\t */\n\tasync rollbackLearningWrite(auditId: string): Promise<{ ok: boolean; reason: string }> {\n\t\tif (this.deps.isDisposed()) return { ok: false, reason: \"session_disposed\" };\n\n\t\tconst audits = this.getLearningAuditRecords();\n\t\tconst audit = audits.find((record) => record.id === auditId);\n\t\tif (!audit) return { ok: false, reason: \"audit_not_found\" };\n\t\tif (audit.action !== \"apply\") return { ok: false, reason: \"not_an_applied_change\" };\n\t\tif (audits.some((record) => record.action === \"rollback\" && record.rollbackOf === auditId)) {\n\t\t\treturn { ok: false, reason: \"already_rolled_back\" };\n\t\t}\n\t\tconst rollback = audit.rollback;\n\t\tif (!rollback) return { ok: false, reason: \"no_rollback_plan\" };\n\n\t\t// Every inverse must be VERIFIED-applied before the rollback audit is appended: a silently\n\t\t// failed inverse that still recorded \"rollback\" would permanently self-lock the change\n\t\t// behind already_rolled_back while the durable write is in fact still live.\n\t\tswitch (rollback.kind) {\n\t\t\tcase \"memory_remove\": {\n\t\t\t\tif (!rollback.target) return { ok: false, reason: \"missing_rollback_target\" };\n\t\t\t\tif (!(await this._applyReflectionWrite({ kind: \"memory_remove\", target: rollback.target }))) {\n\t\t\t\t\treturn { ok: false, reason: \"rollback_apply_failed\" };\n\t\t\t\t}\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"memory_restore\": {\n\t\t\t\tif (!rollback.target || rollback.previous === undefined) {\n\t\t\t\t\treturn { ok: false, reason: \"missing_rollback_target\" };\n\t\t\t\t}\n\t\t\t\tconst applied = await this._applyReflectionWrite({\n\t\t\t\t\tkind: \"memory_replace\",\n\t\t\t\t\ttarget: rollback.target,\n\t\t\t\t\ttext: rollback.previous,\n\t\t\t\t});\n\t\t\t\tif (!applied) return { ok: false, reason: \"rollback_apply_failed\" };\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"memory_add\": {\n\t\t\t\tif (rollback.previous === undefined) return { ok: false, reason: \"missing_rollback_target\" };\n\t\t\t\tconst applied = await this._applyReflectionWrite({\n\t\t\t\t\tkind: \"memory_add\",\n\t\t\t\t\tsection: \"MEMORY\",\n\t\t\t\t\ttext: rollback.previous,\n\t\t\t\t});\n\t\t\t\tif (!applied) return { ok: false, reason: \"rollback_apply_failed\" };\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"archive_skill\": {\n\t\t\t\tif (!rollback.target) return { ok: false, reason: \"missing_rollback_target\" };\n\t\t\t\tif (!this.deps.archivePromotedSkill(rollback.target)) {\n\t\t\t\t\treturn { ok: false, reason: \"skill_archive_failed\" };\n\t\t\t\t}\n\t\t\t\tbreak;\n\t\t\t}\n\t\t}\n\n\t\tappendLearningAuditSnapshot(this.deps.getSessionManager(), {\n\t\t\tid: `${audit.id}-rollback`,\n\t\t\tproposalId: audit.proposalId,\n\t\t\tlayer: audit.layer,\n\t\t\taction: \"rollback\",\n\t\t\tsummary: `Rolled back: ${audit.summary}`,\n\t\t\treasonCode: \"user_requested_rollback\",\n\t\t\tdecision: audit.decision,\n\t\t\trollbackOf: audit.id,\n\t\t\tcreatedAt: new Date().toISOString(),\n\t\t});\n\t\treturn { ok: true, reason: \"rollback_applied\" };\n\t}\n\n\t/**\n\t * Apply one reflection write through the bundled `memory` tool. `memory_replace`/`memory_remove`\n\t * don't carry a target file, so we try MEMORY.md first and fall back to USER.md when the substring\n\t * isn't found there. Never throws (reflection must never break a turn); returns whether the write\n\t * actually applied so callers that MUST know — rollback's once-only accounting — can react instead\n\t * of recording a success that never happened.\n\t */\n\tprivate async _applyReflectionWrite(write: ReflectionWrite, signal?: AbortSignal): Promise<boolean> {\n\t\t// R7 memory-to-behavior: a recurring procedure is compiled into an executable skill file rather\n\t\t// than stored as a flat fact. Written under the agent skills dir so it loads like any user skill.\n\t\tif (write.kind === \"promote_skill\") {\n\t\t\treturn this._promoteReflectionSkill(write.name, write.description, write.body);\n\t\t}\n\n\t\ttype MemResult = { details?: { success?: boolean; error?: string } };\n\t\ttype MemExec = (\n\t\t\ttoolCallId: string,\n\t\t\tparams: { action: string; target: string; content?: string; oldContent?: string },\n\t\t\tsignal: AbortSignal | undefined,\n\t\t\tonUpdate: undefined,\n\t\t\tctx: undefined,\n\t\t) => Promise<MemResult>;\n\t\tconst memTool = this.deps\n\t\t\t.getMemoryManager()\n\t\t\t.getToolDefinitions()\n\t\t\t.find((t) => t.name === \"memory\");\n\t\tconst exec = memTool?.execute as unknown as MemExec | undefined;\n\t\tif (!exec) return false;\n\n\t\tconst run = (params: Parameters<MemExec>[1]) => exec(\"reflection\", params, signal, undefined, undefined);\n\n\t\tif (write.kind === \"memory_add\") {\n\t\t\ttry {\n\t\t\t\tconst res = await run({\n\t\t\t\t\taction: \"add\",\n\t\t\t\t\ttarget: write.section === \"USER\" ? \"user\" : \"memory\",\n\t\t\t\t\tcontent: write.text,\n\t\t\t\t});\n\t\t\t\treturn res?.details?.success === true;\n\t\t\t} catch {\n\t\t\t\t// best-effort; reflection writes must never throw into the turn loop\n\t\t\t\treturn false;\n\t\t\t}\n\t\t}\n\n\t\t// replace / remove carry no target file — try MEMORY.md, then USER.md. The memory tool reports\n\t\t// outcomes via `details.success` (it catches its own errors rather than throwing). Only a\n\t\t// genuine \"not found in the file\" justifies trying the other file; a real failure for a file\n\t\t// (budget exceeded / drift) must NOT fall through and mutate the wrong target.\n\t\tfor (const target of [\"memory\", \"user\"] as const) {\n\t\t\ttry {\n\t\t\t\tconst params =\n\t\t\t\t\twrite.kind === \"memory_replace\"\n\t\t\t\t\t\t? { action: \"replace\", target, oldContent: write.target, content: write.text }\n\t\t\t\t\t\t: { action: \"remove\", target, oldContent: write.target };\n\t\t\t\tconst res = await run(params);\n\t\t\t\tif (res?.details?.success === true) return true; // applied\n\t\t\t\tif (!/not found/i.test(String(res?.details?.error ?? \"\"))) return false; // real failure — don't misapply\n\t\t\t\t// substring simply absent from this file — try the next target\n\t\t\t} catch {\n\t\t\t\t// defensive: if the tool ever does throw, try the next target\n\t\t\t}\n\t\t}\n\t\treturn false;\n\t}\n\n\t/**\n\t * R7: write a reflection-promoted skill as `<agentDir>/skills/<name>/SKILL.md` so it loads like any\n\t * user skill. Best-effort; never clobbers an existing (hand-authored) skill of the same name.\n\t */\n\tprivate _promoteReflectionSkill(rawName: string, description: string, body: string): boolean {\n\t\tconst name = rawName\n\t\t\t.trim()\n\t\t\t.toLowerCase()\n\t\t\t.replace(/[^a-z0-9-]+/g, \"-\")\n\t\t\t.replace(/^-+|-+$/g, \"\")\n\t\t\t.slice(0, 64);\n\t\tif (!name || !body.trim()) return false;\n\t\ttry {\n\t\t\tconst dir = join(this.deps.getAgentDir(), \"skills\", name);\n\t\t\tconst file = join(dir, \"SKILL.md\");\n\t\t\tif (existsSync(file)) return false; // do not overwrite an existing skill\n\t\t\tmkdirSync(dir, { recursive: true });\n\t\t\tconst safeDescription = description.replace(/[\\r\\n]+/g, \" \").trim();\n\t\t\t// `promoted: true` marks this as reflection-generated so the curator (#32) can lifecycle-manage\n\t\t\t// it (archive/consolidate) WITHOUT ever touching hand-authored user skills.\n\t\t\tconst content = `---\\nname: ${name}\\ndescription: ${safeDescription}\\npromoted: true\\n---\\n\\n<!-- Auto-generated by the reflection engine (R7 memory-to-behavior). Review and refine. -->\\n\\n${body.trim()}\\n`;\n\t\t\twriteFileSync(file, content, \"utf-8\");\n\t\t\treturn true;\n\t\t} catch {\n\t\t\t// promotion must never break a turn\n\t\t\treturn false;\n\t\t}\n\t}\n}\n"]}

package/dist/core/research/model-fitness.d.ts

@@ -77,6 +77,18 @@ export declare const SEARCH_PROBE_SYSTEM_PROMPT: string;
 export declare const TOOL_CALL_PROBE_SYSTEM_PROMPT: string;
 export { CURATION_DIGEST_SYSTEM_PROMPT as DIGEST_PROBE_SYSTEM_PROMPT } from "../context/brain-curator.ts";
 export declare function runModelFitnessProbe(options: ModelFitnessOptions): Promise<ModelFitnessReport>;
+/**
+ * Pure verdict: true when the probe found ZERO successes on every LANE surface it actually graded
+ * AND the judge (if it ran) also failed. A lane/judge with total 0 (i.e. never run) carries no
+ * evidence and is excluded from the lane check — but at least one lane must actually have been
+ * graded for an all-failed verdict at all: `gradedLanes.every(...)` is vacuously true over an
+ * empty array, so a report where only the judge ran (every research/worker/search/toolCall/digest
+ * lane is ungraded) is excluded explicitly rather than misread as "all lanes failed" on zero lane
+ * evidence. An empty/degenerate report (nothing graded at all, lanes AND judge) is likewise never
+ * mistaken for a failed one. This is the gate adoption flows must consult before assigning a role —
+ * see `isProbeAllFailed` callers in interactive-mode.ts and agent-session.ts.
+ */
+export declare function isProbeAllFailed(report: ModelFitnessReport): boolean;
 /** Compact human-readable report for tool output / interactive display. Bounded, no raw dumps. */
 export declare function formatModelFitnessReport(model: string, report: ModelFitnessReport): string;
 //# sourceMappingURL=model-fitness.d.ts.map

package/dist/core/research/model-fitness.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model-fitness.d.ts","sourceRoot":"","sources":["../../../src/core/research/model-fitness.ts"],"names":[],"mappings":"AAOA;;;;;;GAMG;AAEH,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,iGAAiG;IACjG,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,kGAAkG;IAClG,MAAM,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,MAAM,eAAe,GAAG,CAAC,IAAI,EAAE;IACpC,YAAY,EAAE,MAAM,CAAC;IACrB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB,KAAK,OAAO,CAAC,iBAAiB,CAAC,CAAC;AAEjC,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,0EAA0E;IAC1E,QAAQ,EAAE,OAAO,CAAC;CAClB;AAED,qFAAqF;AACrF,eAAO,MAAM,6BAA6B,EAAE,SAAS,kBAAkB,EAOtE,CAAC;AAEF,MAAM,WAAW,mBAAmB;IACnC,QAAQ,EAAE,eAAe,CAAC;IAC1B,0CAA0C;IAC1C,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,wDAAwD;IACxD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,YAAY,CAAC,EAAE,SAAS,kBAAkB,EAAE,CAAC;IAC7C,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,gFAAgF;IAChF,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,gBAAgB;IAChC,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,yFAAyF;IACzF,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,iBAAiB;IACjC,MAAM,EAAE,MAAM,CAAC;IACf,gBAAgB,EAAE,MAAM,CAAC;IACzB,aAAa,EAAE,MAAM,CAAC;IACtB,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,qFAAqF;IACrF,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,yFAAyF;IACzF,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,MAAM,EAAE,gBAAgB,CAAC;IACzB,KAAK,EAAE,iBAAiB,CAAC;IACzB,8EAA8E;IAC9E,MAAM,EAAE,gBAAgB,CAAC;IACzB,yFAAyF;IACzF,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,qGAAqG;IACrG,MAAM,EAAE,gBAAgB,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;CACrB;AAED,yFAAyF;AACzF,eAAO,MAAM,0BAA0B,QAK3B,CAAC;AAEb,eAAO,MAAM,6BAA6B,QAK9B,CAAC;AASb,OAAO,EAAE,6BAA6B,IAAI,0BAA0B,EAAE,MAAM,6BAA6B,CAAC;AA6G1G,wBAAsB,oBAAoB,CAAC,OAAO,EAAE,mBAAmB,GAAG,OAAO,CAAC,kBAAkB,CAAC,CAkKpG;AAED,kGAAkG;AAClG,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,kBAAkB,GAAG,MAAM,CAiB1F","sourcesContent":["import { runBoundedCompletion } from \"../autonomy/bounded-completion.ts\";\nimport type { CapabilityEnvelope } from \"../autonomy/contracts.ts\";\nimport { CURATION_DIGEST_SYSTEM_PROMPT } from \"../context/brain-curator.ts\";\nimport { runWorker } from \"../delegation/worker-runner.ts\";\nimport { runRouteJudge } from \"../model-router/route-judge.ts\";\nimport { runResearch } from \"./research-runner.ts\";\n\n/**\n * Model fitness probe: measures whether a candidate model can actually drive the harness's\n * subagent contracts — the research lane, the scout-worker lane, and the routing judge — by\n * running each real runner against the model and scoring parse/success rates plus judge\n * discrimination. Provider-free: the completion executor is injected, so this works against any\n * registered model (local Ollama models included) and against faux providers in tests.\n */\n\nexport interface FitnessCompletion {\n\ttext: string;\n\tcostUsd: number;\n\tstopReason: string;\n\t/** Output tokens generated (for tok/s). Optional: providers that don't report it are skipped. */\n\toutputTokens?: number;\n\t/** Pure generation time in ms (e.g. Ollama eval_duration). Falls back to wall-clock if absent. */\n\tevalMs?: number;\n}\n\nexport type FitnessComplete = (args: {\n\tsystemPrompt: string;\n\tuserPrompt: string;\n\tsignal?: AbortSignal;\n}) => Promise<FitnessCompletion>;\n\nexport interface JudgeFitnessPrompt {\n\tprompt: string;\n\t/** True when the prompt is planning-shaped and must never route cheap. */\n\tplanning: boolean;\n}\n\n/** Default judge probe set: three planning-shaped prompts, three trivial lookups. */\nexport const DEFAULT_JUDGE_FITNESS_PROMPTS: readonly JudgeFitnessPrompt[] = [\n\t{ prompt: \"how should we plan the migration of the session storage layer?\", planning: true },\n\t{ prompt: \"design an approach for splitting the settings manager\", planning: true },\n\t{ prompt: \"draft a roadmap for the autonomy rework\", planning: true },\n\t{ prompt: \"what does the resolvePath function return?\", planning: false },\n\t{ prompt: \"list the files in the delegation module\", planning: false },\n\t{ prompt: \"why is this test flaky?\", planning: false },\n];\n\nexport interface ModelFitnessOptions {\n\tcomplete: FitnessComplete;\n\t/** Trials per lane surface. Default 3. */\n\ttrials?: number;\n\t/** Wall-clock budget per call in ms. Default 120000. */\n\tmaxWallClockMs?: number;\n\tjudgePrompts?: readonly JudgeFitnessPrompt[];\n\tsignal?: AbortSignal;\n\t/** Injected clock for latency measurement (test seam). Defaults to Date.now. */\n\tnow?: () => number;\n}\n\nexport interface LaneFitnessScore {\n\tsucceeded: number;\n\ttotal: number;\n\toutcomes: string[];\n\tmeanMs: number;\n\t/** Mean output tokens/second across the surface's calls; undefined when not reported. */\n\ttokensPerSecond?: number;\n}\n\nexport interface JudgeFitnessScore {\n\tparsed: number;\n\tplanningElevated: number;\n\tplanningTotal: number;\n\ttrivialCheap: number;\n\ttrivialTotal: number;\n\ttotal: number;\n\toutcomes: string[];\n\tmeanMs: number;\n\t/** Mean output tokens/second across the judge calls; undefined when not reported. */\n\ttokensPerSecond?: number;\n}\n\nexport interface ModelFitnessReport {\n\ttrials: number;\n\t/** Aggregate output tokens/second across ALL probe calls (the headline speed number). */\n\ttokensPerSecond?: number;\n\tresearch: LaneFitnessScore;\n\tworker: LaneFitnessScore;\n\tjudge: JudgeFitnessScore;\n\t/** Heavy-lifter surface: can the model formulate a structured search plan? */\n\tsearch: LaneFitnessScore;\n\t/** Heavy-lifter surface: can the model emit a well-formed tool call against a schema? */\n\ttoolCall: LaneFitnessScore;\n\t/** Curator surface: can the model digest a context chunk to strict JSON WITHOUT losing key facts? */\n\tdigest: LaneFitnessScore;\n\ttotalCostUsd: number;\n}\n\n/** Static prompts for the heavy-lifter surfaces (stable for provider prompt caching). */\nexport const SEARCH_PROBE_SYSTEM_PROMPT = [\n\t\"You plan code searches for a coding agent. You never answer the question yourself.\",\n\t\"Given a question about a codebase, respond with STRICT JSON only - no prose:\",\n\t'{\"queries\":[{\"pattern\":\"<regex or literal to grep>\",\"glob\":\"<file glob like **/*.ts>\"}]}',\n\t\"Return 1 to 4 queries, most specific first.\",\n].join(\"\\n\");\n\nexport const TOOL_CALL_PROBE_SYSTEM_PROMPT = [\n\t\"You operate tools for a coding agent. You have exactly one tool:\",\n\t\"grep(pattern: string, path: string) - search files under a path for a pattern.\",\n\t\"Respond to every task with STRICT JSON only - no prose:\",\n\t'{\"tool\":\"grep\",\"arguments\":{\"pattern\":\"<pattern>\",\"path\":\"<path>\"}}',\n].join(\"\\n\");\n\nconst SEARCH_PROBE_TASKS: readonly string[] = [\n\t\"Where is the retry/backoff logic for HTTP requests implemented?\",\n\t\"Which files define the settings for background research?\",\n\t\"Find where session entries of type custom are appended.\",\n];\n\n// The probe measures the REAL curation contract — same prompt the BrainCurator ships.\nexport { CURATION_DIGEST_SYSTEM_PROMPT as DIGEST_PROBE_SYSTEM_PROMPT } from \"../context/brain-curator.ts\";\n\n/**\n * Digest probe chunks each carry a NONCE identifier that cannot be guessed from the\n * instructions: acceptance requires the digest to RETAIN the nonce verbatim, so the score\n * measures extraction fidelity, not narration (a model cannot pass by paraphrasing).\n */\nconst DIGEST_PROBE_TASKS: readonly { chunk: string; nonce: string }[] = [\n\t{\n\t\tnonce: \"retryWithJitter_zx41\",\n\t\tchunk: [\n\t\t\t\"grep results for 'retry' under src/http:\",\n\t\t\t\"src/http/client.ts:88: export function retryWithJitter_zx41(fn, attempts = 3) {\",\n\t\t\t\"src/http/client.ts:112:   // exponential backoff capped at 30s\",\n\t\t\t\"src/http/pool.ts:41:   client.retry = false\",\n\t\t].join(\"\\n\"),\n\t},\n\t{\n\t\tnonce: \"ERR_QM_7734\",\n\t\tchunk: [\n\t\t\t\"$ npm run migrate\",\n\t\t\t\"migrating 14 files...\",\n\t\t\t\"error ERR_QM_7734: column 'owner_id' missing on table sessions (migration 0009)\",\n\t\t\t\"exit code 1\",\n\t\t].join(\"\\n\"),\n\t},\n\t{\n\t\tnonce: \"v3.9.2-hotfix.1\",\n\t\tchunk: [\n\t\t\t\"read package.json (34 lines):\",\n\t\t\t'  \"name\": \"acme-billing\",',\n\t\t\t'  \"version\": \"v3.9.2-hotfix.1\",',\n\t\t\t'  \"engines\": { \"node\": \">=22\" },',\n\t\t].join(\"\\n\"),\n\t},\n];\n\nfunction parseDigest(text: string, nonce: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst digest = (parsed as { digest?: unknown }).digest;\n\tif (typeof digest !== \"string\") return false;\n\tconst trimmed = digest.trim();\n\t// Bounded and faithful: short enough to be a stub annotation, still carrying the nonce fact.\n\treturn trimmed.length > 0 && trimmed.length <= 240 && trimmed.includes(nonce);\n}\n\nconst TOOL_CALL_PROBE_TASKS: readonly string[] = [\n\t\"Find usages of the function resolveCliModel under src/.\",\n\t\"Search for the string 'budget_exhausted' in the core directory.\",\n\t\"Locate where LaneTracker is instantiated under src/core.\",\n];\n\nfunction parseSearchPlan(text: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst queries = (parsed as { queries?: unknown }).queries;\n\tif (!Array.isArray(queries) || queries.length === 0 || queries.length > 8) return false;\n\treturn queries.every(\n\t\t(query) =>\n\t\t\tquery &&\n\t\t\ttypeof query === \"object\" &&\n\t\t\ttypeof (query as { pattern?: unknown }).pattern === \"string\" &&\n\t\t\t(query as { pattern: string }).pattern.trim().length > 0,\n\t);\n}\n\nfunction parseToolCall(text: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst record = parsed as { tool?: unknown; arguments?: unknown };\n\tif (record.tool !== \"grep\") return false;\n\tconst args = record.arguments;\n\tif (!args || typeof args !== \"object\" || Array.isArray(args)) return false;\n\tconst pattern = (args as { pattern?: unknown }).pattern;\n\tconst path = (args as { path?: unknown }).path;\n\treturn (\n\t\ttypeof pattern === \"string\" && pattern.trim().length > 0 && typeof path === \"string\" && path.trim().length > 0\n\t);\n}\n\nfunction extractJsonObject(text: string): unknown | undefined {\n\tconst trimmed = text.trim();\n\tconst candidates: string[] = [trimmed];\n\tconst fenced = /```(?:json)?\\s*([\\s\\S]*?)```/.exec(trimmed);\n\tif (fenced?.[1]) candidates.push(fenced[1].trim());\n\tconst start = trimmed.indexOf(\"{\");\n\tconst end = trimmed.lastIndexOf(\"}\");\n\tif (start >= 0 && end > start) candidates.push(trimmed.slice(start, end + 1));\n\tfor (const candidate of candidates) {\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(candidate);\n\t\t\tif (parsed && typeof parsed === \"object\" && !Array.isArray(parsed)) return parsed;\n\t\t} catch {\n\t\t\t// try next candidate\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction fitnessEnvelope(): CapabilityEnvelope {\n\treturn {\n\t\tid: \"model-fitness-probe\",\n\t\tcapabilities: [\"research\", \"read_files\", \"memory_read\"],\n\t\tmaxEstimatedUsd: 1,\n\t\tcreatedAt: new Date().toISOString(),\n\t};\n}\n\nexport async function runModelFitnessProbe(options: ModelFitnessOptions): Promise<ModelFitnessReport> {\n\tconst trials = Math.max(1, Math.min(options.trials ?? 3, 20));\n\tconst maxWallClockMs = options.maxWallClockMs ?? 120_000;\n\tconst judgePrompts = options.judgePrompts ?? DEFAULT_JUDGE_FITNESS_PROMPTS;\n\tconst now = options.now ?? Date.now;\n\tlet totalCostUsd = 0;\n\n\t// Token-speed instrumentation: the lane runners' own contracts carry text/cost only, so the\n\t// completer is wrapped once here and generation stats are accumulated per surface.\n\tconst overallSpeed = { tokens: 0, evalMs: 0 };\n\tlet surfaceSpeed = { tokens: 0, evalMs: 0 };\n\tconst complete: FitnessComplete = async (args) => {\n\t\tconst completion = await options.complete(args);\n\t\tconst tokens = completion.outputTokens ?? 0;\n\t\tconst evalMs = completion.evalMs ?? 0;\n\t\tif (tokens > 0 && evalMs > 0) {\n\t\t\tsurfaceSpeed.tokens += tokens;\n\t\t\tsurfaceSpeed.evalMs += evalMs;\n\t\t\toverallSpeed.tokens += tokens;\n\t\t\toverallSpeed.evalMs += evalMs;\n\t\t}\n\t\treturn completion;\n\t};\n\tconst takeSurfaceSpeed = (): number | undefined => {\n\t\tconst speed =\n\t\t\tsurfaceSpeed.evalMs > 0 ? Math.round((surfaceSpeed.tokens / surfaceSpeed.evalMs) * 1000) : undefined;\n\t\tsurfaceSpeed = { tokens: 0, evalMs: 0 };\n\t\treturn speed;\n\t};\n\n\tconst research: LaneFitnessScore = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };\n\tfor (let i = 0; i < trials; i++) {\n\t\tconst started = now();\n\t\tconst result = await runResearch({\n\t\t\tquery: `fitness:probe requirements:req-${i}`,\n\t\t\tcontext: [\n\t\t\t\t\"Goal: add a retry helper to the HTTP client module\",\n\t\t\t\t\"Open requirements:\",\n\t\t\t\t\"- Find what retry/backoff conventions the codebase already uses\",\n\t\t\t\t\"- Identify which call sites would adopt the helper\",\n\t\t\t].join(\"\\n\"),\n\t\t\tenvelope: fitnessEnvelope(),\n\t\t\tmaxUsd: 1,\n\t\t\tmaxSources: 8,\n\t\t\tmaxFindings: 5,\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tresearch.meanMs += now() - started;\n\t\ttotalCostUsd += result.costUsd;\n\t\tif (result.status === \"succeeded\") research.succeeded++;\n\t\tresearch.outcomes.push(`${result.status}/${result.reasonCode}`);\n\t}\n\tresearch.meanMs = Math.round(research.meanMs / trials);\n\tresearch.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst worker: LaneFitnessScore = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };\n\tfor (let i = 0; i < trials; i++) {\n\t\tconst started = now();\n\t\tconst outcome = await runWorker({\n\t\t\trequest: {\n\t\t\t\tid: `fitness-worker-${i}`,\n\t\t\t\tinstructions:\n\t\t\t\t\t\"Summarize in two sentences what a capability envelope is: a declared set of allowed tools, paths, and capability names that bounds what a delegated worker may do.\",\n\t\t\t\troute: { tier: \"cheap\", risk: \"read-only\", confidence: 1, reasonCode: \"fitness_probe\", reasons: [] },\n\t\t\t\tenvelope: { id: `fitness-env-${i}`, capabilities: [\"read_files\"], maxEstimatedUsd: 1 },\n\t\t\t\tmaxEstimatedUsd: 1,\n\t\t\t},\n\t\t\tmaxUsd: 1,\n\t\t\tmaxWallClockMs,\n\t\t\tusageReportId: `fitness:${i}`,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tworker.meanMs += now() - started;\n\t\ttotalCostUsd += outcome.costUsd;\n\t\tif (outcome.result.status === \"completed\" && outcome.accepted) worker.succeeded++;\n\t\tworker.outcomes.push(`${outcome.result.status}/${outcome.reasonCode}`);\n\t}\n\tworker.meanMs = Math.round(worker.meanMs / trials);\n\tworker.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst judge: JudgeFitnessScore = {\n\t\tparsed: 0,\n\t\tplanningElevated: 0,\n\t\tplanningTotal: judgePrompts.filter((entry) => entry.planning).length,\n\t\ttrivialCheap: 0,\n\t\ttrivialTotal: judgePrompts.filter((entry) => !entry.planning).length,\n\t\ttotal: judgePrompts.length,\n\t\toutcomes: [],\n\t\tmeanMs: 0,\n\t};\n\tfor (const entry of judgePrompts) {\n\t\tconst started = now();\n\t\tconst result = await runRouteJudge({\n\t\t\tprompt: entry.prompt,\n\t\t\tbaseline: { tier: \"cheap\", risk: \"read-only\", confidence: 0.5, reasonCode: \"fitness_probe\", reasons: [] },\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tjudge.meanMs += now() - started;\n\t\ttotalCostUsd += result.costUsd;\n\t\tconst tier = result.decision.tier;\n\t\tif (result.verdict) {\n\t\t\tjudge.parsed++;\n\t\t\t// A useful judge must both keep planning off the cheap tier AND actually send trivial\n\t\t\t// prompts there — all-medium verdicts are safe but save nothing.\n\t\t\tif (entry.planning && tier !== \"cheap\") judge.planningElevated++;\n\t\t\tif (!entry.planning && tier === \"cheap\") judge.trivialCheap++;\n\t\t}\n\t\tjudge.outcomes.push(\n\t\t\t`\"${entry.prompt.slice(0, 40)}\" -> ${tier}${result.fallbackReason ? ` (${result.fallbackReason})` : \"\"}`,\n\t\t);\n\t}\n\tjudge.meanMs = judgePrompts.length > 0 ? Math.round(judge.meanMs / judgePrompts.length) : 0;\n\tjudge.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst probeSurface = async (\n\t\tsystemPrompt: string,\n\t\ttasks: readonly string[],\n\t\taccepts: (text: string, taskIndex: number) => boolean,\n\t): Promise<LaneFitnessScore> => {\n\t\tconst score: LaneFitnessScore = { succeeded: 0, total: tasks.length, outcomes: [], meanMs: 0 };\n\t\tfor (const [taskIndex, task] of tasks.entries()) {\n\t\t\tconst started = now();\n\t\t\t// Same wall-clock envelope as the lane surfaces — a hung model must not hang the probe.\n\t\t\tconst bounded = await runBoundedCompletion({\n\t\t\t\tmaxWallClockMs,\n\t\t\t\tsignal: options.signal,\n\t\t\t\texecute: (signal) => complete({ systemPrompt, userPrompt: task, signal }),\n\t\t\t});\n\t\t\tif (bounded.completion) totalCostUsd += bounded.completion.costUsd;\n\t\t\tif (bounded.failure || !bounded.completion) {\n\t\t\t\tscore.outcomes.push(bounded.failure ? bounded.failure.status : \"completion_error\");\n\t\t\t} else {\n\t\t\t\tconst ok = accepts(bounded.completion.text, taskIndex);\n\t\t\t\tif (ok) score.succeeded++;\n\t\t\t\tscore.outcomes.push(ok ? \"ok\" : \"unparseable_output\");\n\t\t\t}\n\t\t\tscore.meanMs += now() - started;\n\t\t}\n\t\tscore.meanMs = tasks.length > 0 ? Math.round(score.meanMs / tasks.length) : 0;\n\t\treturn score;\n\t};\n\n\tconst search = await probeSurface(SEARCH_PROBE_SYSTEM_PROMPT, SEARCH_PROBE_TASKS, parseSearchPlan);\n\tsearch.tokensPerSecond = takeSurfaceSpeed();\n\tconst toolCall = await probeSurface(TOOL_CALL_PROBE_SYSTEM_PROMPT, TOOL_CALL_PROBE_TASKS, parseToolCall);\n\ttoolCall.tokensPerSecond = takeSurfaceSpeed();\n\tconst digest = await probeSurface(\n\t\tCURATION_DIGEST_SYSTEM_PROMPT,\n\t\tDIGEST_PROBE_TASKS.map((task) => task.chunk),\n\t\t(text, taskIndex) => parseDigest(text, DIGEST_PROBE_TASKS[taskIndex]!.nonce),\n\t);\n\tdigest.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst tokensPerSecond =\n\t\toverallSpeed.evalMs > 0 ? Math.round((overallSpeed.tokens / overallSpeed.evalMs) * 1000) : undefined;\n\n\treturn { trials, tokensPerSecond, research, worker, judge, search, toolCall, digest, totalCostUsd };\n}\n\n/** Compact human-readable report for tool output / interactive display. Bounded, no raw dumps. */\nexport function formatModelFitnessReport(model: string, report: ModelFitnessReport): string {\n\tconst speed = (tokensPerSecond: number | undefined) =>\n\t\ttokensPerSecond !== undefined ? `, ~${tokensPerSecond} tok/s` : \"\";\n\tconst lines = [\n\t\t`Model fitness: ${model} (${report.trials} trials/lane${speed(report.tokensPerSecond)})`,\n\t\t`- research lane: ${report.research.succeeded}/${report.research.total} succeeded, mean ${report.research.meanMs}ms${speed(report.research.tokensPerSecond)} [${report.research.outcomes.join(\", \")}]`,\n\t\t`- worker lane:   ${report.worker.succeeded}/${report.worker.total} completed+accepted, mean ${report.worker.meanMs}ms${speed(report.worker.tokensPerSecond)} [${report.worker.outcomes.join(\", \")}]`,\n\t\t`- search plans:  ${report.search.succeeded}/${report.search.total} well-formed, mean ${report.search.meanMs}ms${speed(report.search.tokensPerSecond)}`,\n\t\t`- tool calls:    ${report.toolCall.succeeded}/${report.toolCall.total} well-formed, mean ${report.toolCall.meanMs}ms${speed(report.toolCall.tokensPerSecond)}`,\n\t\t`- digests:       ${report.digest.succeeded}/${report.digest.total} faithful, mean ${report.digest.meanMs}ms${speed(report.digest.tokensPerSecond)}`,\n\t\t`- route judge:   parsed ${report.judge.parsed}/${report.judge.total}, planning-elevated ${report.judge.planningElevated}/${report.judge.planningTotal}, trivial-cheap ${report.judge.trivialCheap}/${report.judge.trivialTotal}, mean ${report.judge.meanMs}ms${speed(report.judge.tokensPerSecond)}`,\n\t\t...report.judge.outcomes.map((outcome) => `    ${outcome}`),\n\t];\n\tif (report.totalCostUsd > 0) {\n\t\tlines.push(`- probe cost: $${report.totalCostUsd.toFixed(4)}`);\n\t}\n\treturn lines.join(\"\\n\");\n}\n"]}
+{"version":3,"file":"model-fitness.d.ts","sourceRoot":"","sources":["../../../src/core/research/model-fitness.ts"],"names":[],"mappings":"AAOA;;;;;;GAMG;AAEH,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,iGAAiG;IACjG,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,kGAAkG;IAClG,MAAM,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,MAAM,eAAe,GAAG,CAAC,IAAI,EAAE;IACpC,YAAY,EAAE,MAAM,CAAC;IACrB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB,KAAK,OAAO,CAAC,iBAAiB,CAAC,CAAC;AAEjC,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,0EAA0E;IAC1E,QAAQ,EAAE,OAAO,CAAC;CAClB;AAED,qFAAqF;AACrF,eAAO,MAAM,6BAA6B,EAAE,SAAS,kBAAkB,EAOtE,CAAC;AAEF,MAAM,WAAW,mBAAmB;IACnC,QAAQ,EAAE,eAAe,CAAC;IAC1B,0CAA0C;IAC1C,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,wDAAwD;IACxD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,YAAY,CAAC,EAAE,SAAS,kBAAkB,EAAE,CAAC;IAC7C,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,gFAAgF;IAChF,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,gBAAgB;IAChC,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,yFAAyF;IACzF,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,iBAAiB;IACjC,MAAM,EAAE,MAAM,CAAC;IACf,gBAAgB,EAAE,MAAM,CAAC;IACzB,aAAa,EAAE,MAAM,CAAC;IACtB,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,qFAAqF;IACrF,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,kBAAkB;IAClC,MAAM,EAAE,MAAM,CAAC;IACf,yFAAyF;IACzF,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,MAAM,EAAE,gBAAgB,CAAC;IACzB,KAAK,EAAE,iBAAiB,CAAC;IACzB,8EAA8E;IAC9E,MAAM,EAAE,gBAAgB,CAAC;IACzB,yFAAyF;IACzF,QAAQ,EAAE,gBAAgB,CAAC;IAC3B,qGAAqG;IACrG,MAAM,EAAE,gBAAgB,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;CACrB;AAED,yFAAyF;AACzF,eAAO,MAAM,0BAA0B,QAK3B,CAAC;AAEb,eAAO,MAAM,6BAA6B,QAK9B,CAAC;AASb,OAAO,EAAE,6BAA6B,IAAI,0BAA0B,EAAE,MAAM,6BAA6B,CAAC;AA6G1G,wBAAsB,oBAAoB,CAAC,OAAO,EAAE,mBAAmB,GAAG,OAAO,CAAC,kBAAkB,CAAC,CAkKpG;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,kBAAkB,GAAG,OAAO,CAQpE;AAED,kGAAkG;AAClG,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,kBAAkB,GAAG,MAAM,CAiB1F","sourcesContent":["import { runBoundedCompletion } from \"../autonomy/bounded-completion.ts\";\nimport type { CapabilityEnvelope } from \"../autonomy/contracts.ts\";\nimport { CURATION_DIGEST_SYSTEM_PROMPT } from \"../context/brain-curator.ts\";\nimport { runWorker } from \"../delegation/worker-runner.ts\";\nimport { runRouteJudge } from \"../model-router/route-judge.ts\";\nimport { runResearch } from \"./research-runner.ts\";\n\n/**\n * Model fitness probe: measures whether a candidate model can actually drive the harness's\n * subagent contracts — the research lane, the scout-worker lane, and the routing judge — by\n * running each real runner against the model and scoring parse/success rates plus judge\n * discrimination. Provider-free: the completion executor is injected, so this works against any\n * registered model (local Ollama models included) and against faux providers in tests.\n */\n\nexport interface FitnessCompletion {\n\ttext: string;\n\tcostUsd: number;\n\tstopReason: string;\n\t/** Output tokens generated (for tok/s). Optional: providers that don't report it are skipped. */\n\toutputTokens?: number;\n\t/** Pure generation time in ms (e.g. Ollama eval_duration). Falls back to wall-clock if absent. */\n\tevalMs?: number;\n}\n\nexport type FitnessComplete = (args: {\n\tsystemPrompt: string;\n\tuserPrompt: string;\n\tsignal?: AbortSignal;\n}) => Promise<FitnessCompletion>;\n\nexport interface JudgeFitnessPrompt {\n\tprompt: string;\n\t/** True when the prompt is planning-shaped and must never route cheap. */\n\tplanning: boolean;\n}\n\n/** Default judge probe set: three planning-shaped prompts, three trivial lookups. */\nexport const DEFAULT_JUDGE_FITNESS_PROMPTS: readonly JudgeFitnessPrompt[] = [\n\t{ prompt: \"how should we plan the migration of the session storage layer?\", planning: true },\n\t{ prompt: \"design an approach for splitting the settings manager\", planning: true },\n\t{ prompt: \"draft a roadmap for the autonomy rework\", planning: true },\n\t{ prompt: \"what does the resolvePath function return?\", planning: false },\n\t{ prompt: \"list the files in the delegation module\", planning: false },\n\t{ prompt: \"why is this test flaky?\", planning: false },\n];\n\nexport interface ModelFitnessOptions {\n\tcomplete: FitnessComplete;\n\t/** Trials per lane surface. Default 3. */\n\ttrials?: number;\n\t/** Wall-clock budget per call in ms. Default 120000. */\n\tmaxWallClockMs?: number;\n\tjudgePrompts?: readonly JudgeFitnessPrompt[];\n\tsignal?: AbortSignal;\n\t/** Injected clock for latency measurement (test seam). Defaults to Date.now. */\n\tnow?: () => number;\n}\n\nexport interface LaneFitnessScore {\n\tsucceeded: number;\n\ttotal: number;\n\toutcomes: string[];\n\tmeanMs: number;\n\t/** Mean output tokens/second across the surface's calls; undefined when not reported. */\n\ttokensPerSecond?: number;\n}\n\nexport interface JudgeFitnessScore {\n\tparsed: number;\n\tplanningElevated: number;\n\tplanningTotal: number;\n\ttrivialCheap: number;\n\ttrivialTotal: number;\n\ttotal: number;\n\toutcomes: string[];\n\tmeanMs: number;\n\t/** Mean output tokens/second across the judge calls; undefined when not reported. */\n\ttokensPerSecond?: number;\n}\n\nexport interface ModelFitnessReport {\n\ttrials: number;\n\t/** Aggregate output tokens/second across ALL probe calls (the headline speed number). */\n\ttokensPerSecond?: number;\n\tresearch: LaneFitnessScore;\n\tworker: LaneFitnessScore;\n\tjudge: JudgeFitnessScore;\n\t/** Heavy-lifter surface: can the model formulate a structured search plan? */\n\tsearch: LaneFitnessScore;\n\t/** Heavy-lifter surface: can the model emit a well-formed tool call against a schema? */\n\ttoolCall: LaneFitnessScore;\n\t/** Curator surface: can the model digest a context chunk to strict JSON WITHOUT losing key facts? */\n\tdigest: LaneFitnessScore;\n\ttotalCostUsd: number;\n}\n\n/** Static prompts for the heavy-lifter surfaces (stable for provider prompt caching). */\nexport const SEARCH_PROBE_SYSTEM_PROMPT = [\n\t\"You plan code searches for a coding agent. You never answer the question yourself.\",\n\t\"Given a question about a codebase, respond with STRICT JSON only - no prose:\",\n\t'{\"queries\":[{\"pattern\":\"<regex or literal to grep>\",\"glob\":\"<file glob like **/*.ts>\"}]}',\n\t\"Return 1 to 4 queries, most specific first.\",\n].join(\"\\n\");\n\nexport const TOOL_CALL_PROBE_SYSTEM_PROMPT = [\n\t\"You operate tools for a coding agent. You have exactly one tool:\",\n\t\"grep(pattern: string, path: string) - search files under a path for a pattern.\",\n\t\"Respond to every task with STRICT JSON only - no prose:\",\n\t'{\"tool\":\"grep\",\"arguments\":{\"pattern\":\"<pattern>\",\"path\":\"<path>\"}}',\n].join(\"\\n\");\n\nconst SEARCH_PROBE_TASKS: readonly string[] = [\n\t\"Where is the retry/backoff logic for HTTP requests implemented?\",\n\t\"Which files define the settings for background research?\",\n\t\"Find where session entries of type custom are appended.\",\n];\n\n// The probe measures the REAL curation contract — same prompt the BrainCurator ships.\nexport { CURATION_DIGEST_SYSTEM_PROMPT as DIGEST_PROBE_SYSTEM_PROMPT } from \"../context/brain-curator.ts\";\n\n/**\n * Digest probe chunks each carry a NONCE identifier that cannot be guessed from the\n * instructions: acceptance requires the digest to RETAIN the nonce verbatim, so the score\n * measures extraction fidelity, not narration (a model cannot pass by paraphrasing).\n */\nconst DIGEST_PROBE_TASKS: readonly { chunk: string; nonce: string }[] = [\n\t{\n\t\tnonce: \"retryWithJitter_zx41\",\n\t\tchunk: [\n\t\t\t\"grep results for 'retry' under src/http:\",\n\t\t\t\"src/http/client.ts:88: export function retryWithJitter_zx41(fn, attempts = 3) {\",\n\t\t\t\"src/http/client.ts:112:   // exponential backoff capped at 30s\",\n\t\t\t\"src/http/pool.ts:41:   client.retry = false\",\n\t\t].join(\"\\n\"),\n\t},\n\t{\n\t\tnonce: \"ERR_QM_7734\",\n\t\tchunk: [\n\t\t\t\"$ npm run migrate\",\n\t\t\t\"migrating 14 files...\",\n\t\t\t\"error ERR_QM_7734: column 'owner_id' missing on table sessions (migration 0009)\",\n\t\t\t\"exit code 1\",\n\t\t].join(\"\\n\"),\n\t},\n\t{\n\t\tnonce: \"v3.9.2-hotfix.1\",\n\t\tchunk: [\n\t\t\t\"read package.json (34 lines):\",\n\t\t\t'  \"name\": \"acme-billing\",',\n\t\t\t'  \"version\": \"v3.9.2-hotfix.1\",',\n\t\t\t'  \"engines\": { \"node\": \">=22\" },',\n\t\t].join(\"\\n\"),\n\t},\n];\n\nfunction parseDigest(text: string, nonce: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst digest = (parsed as { digest?: unknown }).digest;\n\tif (typeof digest !== \"string\") return false;\n\tconst trimmed = digest.trim();\n\t// Bounded and faithful: short enough to be a stub annotation, still carrying the nonce fact.\n\treturn trimmed.length > 0 && trimmed.length <= 240 && trimmed.includes(nonce);\n}\n\nconst TOOL_CALL_PROBE_TASKS: readonly string[] = [\n\t\"Find usages of the function resolveCliModel under src/.\",\n\t\"Search for the string 'budget_exhausted' in the core directory.\",\n\t\"Locate where LaneTracker is instantiated under src/core.\",\n];\n\nfunction parseSearchPlan(text: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst queries = (parsed as { queries?: unknown }).queries;\n\tif (!Array.isArray(queries) || queries.length === 0 || queries.length > 8) return false;\n\treturn queries.every(\n\t\t(query) =>\n\t\t\tquery &&\n\t\t\ttypeof query === \"object\" &&\n\t\t\ttypeof (query as { pattern?: unknown }).pattern === \"string\" &&\n\t\t\t(query as { pattern: string }).pattern.trim().length > 0,\n\t);\n}\n\nfunction parseToolCall(text: string): boolean {\n\tconst parsed = extractJsonObject(text);\n\tif (!parsed) return false;\n\tconst record = parsed as { tool?: unknown; arguments?: unknown };\n\tif (record.tool !== \"grep\") return false;\n\tconst args = record.arguments;\n\tif (!args || typeof args !== \"object\" || Array.isArray(args)) return false;\n\tconst pattern = (args as { pattern?: unknown }).pattern;\n\tconst path = (args as { path?: unknown }).path;\n\treturn (\n\t\ttypeof pattern === \"string\" && pattern.trim().length > 0 && typeof path === \"string\" && path.trim().length > 0\n\t);\n}\n\nfunction extractJsonObject(text: string): unknown | undefined {\n\tconst trimmed = text.trim();\n\tconst candidates: string[] = [trimmed];\n\tconst fenced = /```(?:json)?\\s*([\\s\\S]*?)```/.exec(trimmed);\n\tif (fenced?.[1]) candidates.push(fenced[1].trim());\n\tconst start = trimmed.indexOf(\"{\");\n\tconst end = trimmed.lastIndexOf(\"}\");\n\tif (start >= 0 && end > start) candidates.push(trimmed.slice(start, end + 1));\n\tfor (const candidate of candidates) {\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(candidate);\n\t\t\tif (parsed && typeof parsed === \"object\" && !Array.isArray(parsed)) return parsed;\n\t\t} catch {\n\t\t\t// try next candidate\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction fitnessEnvelope(): CapabilityEnvelope {\n\treturn {\n\t\tid: \"model-fitness-probe\",\n\t\tcapabilities: [\"research\", \"read_files\", \"memory_read\"],\n\t\tmaxEstimatedUsd: 1,\n\t\tcreatedAt: new Date().toISOString(),\n\t};\n}\n\nexport async function runModelFitnessProbe(options: ModelFitnessOptions): Promise<ModelFitnessReport> {\n\tconst trials = Math.max(1, Math.min(options.trials ?? 3, 20));\n\tconst maxWallClockMs = options.maxWallClockMs ?? 120_000;\n\tconst judgePrompts = options.judgePrompts ?? DEFAULT_JUDGE_FITNESS_PROMPTS;\n\tconst now = options.now ?? Date.now;\n\tlet totalCostUsd = 0;\n\n\t// Token-speed instrumentation: the lane runners' own contracts carry text/cost only, so the\n\t// completer is wrapped once here and generation stats are accumulated per surface.\n\tconst overallSpeed = { tokens: 0, evalMs: 0 };\n\tlet surfaceSpeed = { tokens: 0, evalMs: 0 };\n\tconst complete: FitnessComplete = async (args) => {\n\t\tconst completion = await options.complete(args);\n\t\tconst tokens = completion.outputTokens ?? 0;\n\t\tconst evalMs = completion.evalMs ?? 0;\n\t\tif (tokens > 0 && evalMs > 0) {\n\t\t\tsurfaceSpeed.tokens += tokens;\n\t\t\tsurfaceSpeed.evalMs += evalMs;\n\t\t\toverallSpeed.tokens += tokens;\n\t\t\toverallSpeed.evalMs += evalMs;\n\t\t}\n\t\treturn completion;\n\t};\n\tconst takeSurfaceSpeed = (): number | undefined => {\n\t\tconst speed =\n\t\t\tsurfaceSpeed.evalMs > 0 ? Math.round((surfaceSpeed.tokens / surfaceSpeed.evalMs) * 1000) : undefined;\n\t\tsurfaceSpeed = { tokens: 0, evalMs: 0 };\n\t\treturn speed;\n\t};\n\n\tconst research: LaneFitnessScore = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };\n\tfor (let i = 0; i < trials; i++) {\n\t\tconst started = now();\n\t\tconst result = await runResearch({\n\t\t\tquery: `fitness:probe requirements:req-${i}`,\n\t\t\tcontext: [\n\t\t\t\t\"Goal: add a retry helper to the HTTP client module\",\n\t\t\t\t\"Open requirements:\",\n\t\t\t\t\"- Find what retry/backoff conventions the codebase already uses\",\n\t\t\t\t\"- Identify which call sites would adopt the helper\",\n\t\t\t].join(\"\\n\"),\n\t\t\tenvelope: fitnessEnvelope(),\n\t\t\tmaxUsd: 1,\n\t\t\tmaxSources: 8,\n\t\t\tmaxFindings: 5,\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tresearch.meanMs += now() - started;\n\t\ttotalCostUsd += result.costUsd;\n\t\tif (result.status === \"succeeded\") research.succeeded++;\n\t\tresearch.outcomes.push(`${result.status}/${result.reasonCode}`);\n\t}\n\tresearch.meanMs = Math.round(research.meanMs / trials);\n\tresearch.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst worker: LaneFitnessScore = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };\n\tfor (let i = 0; i < trials; i++) {\n\t\tconst started = now();\n\t\tconst outcome = await runWorker({\n\t\t\trequest: {\n\t\t\t\tid: `fitness-worker-${i}`,\n\t\t\t\tinstructions:\n\t\t\t\t\t\"Summarize in two sentences what a capability envelope is: a declared set of allowed tools, paths, and capability names that bounds what a delegated worker may do.\",\n\t\t\t\troute: { tier: \"cheap\", risk: \"read-only\", confidence: 1, reasonCode: \"fitness_probe\", reasons: [] },\n\t\t\t\tenvelope: { id: `fitness-env-${i}`, capabilities: [\"read_files\"], maxEstimatedUsd: 1 },\n\t\t\t\tmaxEstimatedUsd: 1,\n\t\t\t},\n\t\t\tmaxUsd: 1,\n\t\t\tmaxWallClockMs,\n\t\t\tusageReportId: `fitness:${i}`,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tworker.meanMs += now() - started;\n\t\ttotalCostUsd += outcome.costUsd;\n\t\tif (outcome.result.status === \"completed\" && outcome.accepted) worker.succeeded++;\n\t\tworker.outcomes.push(`${outcome.result.status}/${outcome.reasonCode}`);\n\t}\n\tworker.meanMs = Math.round(worker.meanMs / trials);\n\tworker.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst judge: JudgeFitnessScore = {\n\t\tparsed: 0,\n\t\tplanningElevated: 0,\n\t\tplanningTotal: judgePrompts.filter((entry) => entry.planning).length,\n\t\ttrivialCheap: 0,\n\t\ttrivialTotal: judgePrompts.filter((entry) => !entry.planning).length,\n\t\ttotal: judgePrompts.length,\n\t\toutcomes: [],\n\t\tmeanMs: 0,\n\t};\n\tfor (const entry of judgePrompts) {\n\t\tconst started = now();\n\t\tconst result = await runRouteJudge({\n\t\t\tprompt: entry.prompt,\n\t\t\tbaseline: { tier: \"cheap\", risk: \"read-only\", confidence: 0.5, reasonCode: \"fitness_probe\", reasons: [] },\n\t\t\tmaxWallClockMs,\n\t\t\tsignal: options.signal,\n\t\t\tcomplete,\n\t\t});\n\t\tjudge.meanMs += now() - started;\n\t\ttotalCostUsd += result.costUsd;\n\t\tconst tier = result.decision.tier;\n\t\tif (result.verdict) {\n\t\t\tjudge.parsed++;\n\t\t\t// A useful judge must both keep planning off the cheap tier AND actually send trivial\n\t\t\t// prompts there — all-medium verdicts are safe but save nothing.\n\t\t\tif (entry.planning && tier !== \"cheap\") judge.planningElevated++;\n\t\t\tif (!entry.planning && tier === \"cheap\") judge.trivialCheap++;\n\t\t}\n\t\tjudge.outcomes.push(\n\t\t\t`\"${entry.prompt.slice(0, 40)}\" -> ${tier}${result.fallbackReason ? ` (${result.fallbackReason})` : \"\"}`,\n\t\t);\n\t}\n\tjudge.meanMs = judgePrompts.length > 0 ? Math.round(judge.meanMs / judgePrompts.length) : 0;\n\tjudge.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst probeSurface = async (\n\t\tsystemPrompt: string,\n\t\ttasks: readonly string[],\n\t\taccepts: (text: string, taskIndex: number) => boolean,\n\t): Promise<LaneFitnessScore> => {\n\t\tconst score: LaneFitnessScore = { succeeded: 0, total: tasks.length, outcomes: [], meanMs: 0 };\n\t\tfor (const [taskIndex, task] of tasks.entries()) {\n\t\t\tconst started = now();\n\t\t\t// Same wall-clock envelope as the lane surfaces — a hung model must not hang the probe.\n\t\t\tconst bounded = await runBoundedCompletion({\n\t\t\t\tmaxWallClockMs,\n\t\t\t\tsignal: options.signal,\n\t\t\t\texecute: (signal) => complete({ systemPrompt, userPrompt: task, signal }),\n\t\t\t});\n\t\t\tif (bounded.completion) totalCostUsd += bounded.completion.costUsd;\n\t\t\tif (bounded.failure || !bounded.completion) {\n\t\t\t\tscore.outcomes.push(bounded.failure ? bounded.failure.status : \"completion_error\");\n\t\t\t} else {\n\t\t\t\tconst ok = accepts(bounded.completion.text, taskIndex);\n\t\t\t\tif (ok) score.succeeded++;\n\t\t\t\tscore.outcomes.push(ok ? \"ok\" : \"unparseable_output\");\n\t\t\t}\n\t\t\tscore.meanMs += now() - started;\n\t\t}\n\t\tscore.meanMs = tasks.length > 0 ? Math.round(score.meanMs / tasks.length) : 0;\n\t\treturn score;\n\t};\n\n\tconst search = await probeSurface(SEARCH_PROBE_SYSTEM_PROMPT, SEARCH_PROBE_TASKS, parseSearchPlan);\n\tsearch.tokensPerSecond = takeSurfaceSpeed();\n\tconst toolCall = await probeSurface(TOOL_CALL_PROBE_SYSTEM_PROMPT, TOOL_CALL_PROBE_TASKS, parseToolCall);\n\ttoolCall.tokensPerSecond = takeSurfaceSpeed();\n\tconst digest = await probeSurface(\n\t\tCURATION_DIGEST_SYSTEM_PROMPT,\n\t\tDIGEST_PROBE_TASKS.map((task) => task.chunk),\n\t\t(text, taskIndex) => parseDigest(text, DIGEST_PROBE_TASKS[taskIndex]!.nonce),\n\t);\n\tdigest.tokensPerSecond = takeSurfaceSpeed();\n\n\tconst tokensPerSecond =\n\t\toverallSpeed.evalMs > 0 ? Math.round((overallSpeed.tokens / overallSpeed.evalMs) * 1000) : undefined;\n\n\treturn { trials, tokensPerSecond, research, worker, judge, search, toolCall, digest, totalCostUsd };\n}\n\n/**\n * Pure verdict: true when the probe found ZERO successes on every LANE surface it actually graded\n * AND the judge (if it ran) also failed. A lane/judge with total 0 (i.e. never run) carries no\n * evidence and is excluded from the lane check — but at least one lane must actually have been\n * graded for an all-failed verdict at all: `gradedLanes.every(...)` is vacuously true over an\n * empty array, so a report where only the judge ran (every research/worker/search/toolCall/digest\n * lane is ungraded) is excluded explicitly rather than misread as \"all lanes failed\" on zero lane\n * evidence. An empty/degenerate report (nothing graded at all, lanes AND judge) is likewise never\n * mistaken for a failed one. This is the gate adoption flows must consult before assigning a role —\n * see `isProbeAllFailed` callers in interactive-mode.ts and agent-session.ts.\n */\nexport function isProbeAllFailed(report: ModelFitnessReport): boolean {\n\tconst lanes = [report.research, report.worker, report.search, report.toolCall, report.digest];\n\tconst gradedLanes = lanes.filter((lane) => lane.total > 0);\n\tconst judgeGraded = report.judge.total > 0;\n\tif (gradedLanes.length === 0 && !judgeGraded) return false;\n\tconst lanesAllFailed = gradedLanes.length > 0 && gradedLanes.every((lane) => lane.succeeded === 0);\n\tconst judgeFailed = !judgeGraded || report.judge.parsed === 0;\n\treturn lanesAllFailed && judgeFailed;\n}\n\n/** Compact human-readable report for tool output / interactive display. Bounded, no raw dumps. */\nexport function formatModelFitnessReport(model: string, report: ModelFitnessReport): string {\n\tconst speed = (tokensPerSecond: number | undefined) =>\n\t\ttokensPerSecond !== undefined ? `, ~${tokensPerSecond} tok/s` : \"\";\n\tconst lines = [\n\t\t`Model fitness: ${model} (${report.trials} trials/lane${speed(report.tokensPerSecond)})`,\n\t\t`- research lane: ${report.research.succeeded}/${report.research.total} succeeded, mean ${report.research.meanMs}ms${speed(report.research.tokensPerSecond)} [${report.research.outcomes.join(\", \")}]`,\n\t\t`- worker lane:   ${report.worker.succeeded}/${report.worker.total} completed+accepted, mean ${report.worker.meanMs}ms${speed(report.worker.tokensPerSecond)} [${report.worker.outcomes.join(\", \")}]`,\n\t\t`- search plans:  ${report.search.succeeded}/${report.search.total} well-formed, mean ${report.search.meanMs}ms${speed(report.search.tokensPerSecond)}`,\n\t\t`- tool calls:    ${report.toolCall.succeeded}/${report.toolCall.total} well-formed, mean ${report.toolCall.meanMs}ms${speed(report.toolCall.tokensPerSecond)}`,\n\t\t`- digests:       ${report.digest.succeeded}/${report.digest.total} faithful, mean ${report.digest.meanMs}ms${speed(report.digest.tokensPerSecond)}`,\n\t\t`- route judge:   parsed ${report.judge.parsed}/${report.judge.total}, planning-elevated ${report.judge.planningElevated}/${report.judge.planningTotal}, trivial-cheap ${report.judge.trivialCheap}/${report.judge.trivialTotal}, mean ${report.judge.meanMs}ms${speed(report.judge.tokensPerSecond)}`,\n\t\t...report.judge.outcomes.map((outcome) => `    ${outcome}`),\n\t];\n\tif (report.totalCostUsd > 0) {\n\t\tlines.push(`- probe cost: $${report.totalCostUsd.toFixed(4)}`);\n\t}\n\treturn lines.join(\"\\n\");\n}\n"]}