@caupulican/pi-adaptative 0.80.85 → 0.80.88

package/dist/core/research/model-fitness.js

@@ -0,0 +1,257 @@
+import { runBoundedCompletion } from "../autonomy/bounded-completion.js";
+import { runWorker } from "../delegation/worker-runner.js";
+import { runRouteJudge } from "../model-router/route-judge.js";
+import { runResearch } from "./research-runner.js";
+/** Default judge probe set: three planning-shaped prompts, three trivial lookups. */
+export const DEFAULT_JUDGE_FITNESS_PROMPTS = [
+    { prompt: "how should we plan the migration of the session storage layer?", planning: true },
+    { prompt: "design an approach for splitting the settings manager", planning: true },
+    { prompt: "draft a roadmap for the autonomy rework", planning: true },
+    { prompt: "what does the resolvePath function return?", planning: false },
+    { prompt: "list the files in the delegation module", planning: false },
+    { prompt: "why is this test flaky?", planning: false },
+];
+/** Static prompts for the heavy-lifter surfaces (stable for provider prompt caching). */
+export const SEARCH_PROBE_SYSTEM_PROMPT = [
+    "You plan code searches for a coding agent. You never answer the question yourself.",
+    "Given a question about a codebase, respond with STRICT JSON only - no prose:",
+    '{"queries":[{"pattern":"<regex or literal to grep>","glob":"<file glob like **/*.ts>"}]}',
+    "Return 1 to 4 queries, most specific first.",
+].join("\n");
+export const TOOL_CALL_PROBE_SYSTEM_PROMPT = [
+    "You operate tools for a coding agent. You have exactly one tool:",
+    "grep(pattern: string, path: string) - search files under a path for a pattern.",
+    "Respond to every task with STRICT JSON only - no prose:",
+    '{"tool":"grep","arguments":{"pattern":"<pattern>","path":"<path>"}}',
+].join("\n");
+const SEARCH_PROBE_TASKS = [
+    "Where is the retry/backoff logic for HTTP requests implemented?",
+    "Which files define the settings for background research?",
+    "Find where session entries of type custom are appended.",
+];
+const TOOL_CALL_PROBE_TASKS = [
+    "Find usages of the function resolveCliModel under src/.",
+    "Search for the string 'budget_exhausted' in the core directory.",
+    "Locate where LaneTracker is instantiated under src/core.",
+];
+function parseSearchPlan(text) {
+    const parsed = extractJsonObject(text);
+    if (!parsed)
+        return false;
+    const queries = parsed.queries;
+    if (!Array.isArray(queries) || queries.length === 0 || queries.length > 8)
+        return false;
+    return queries.every((query) => query &&
+        typeof query === "object" &&
+        typeof query.pattern === "string" &&
+        query.pattern.trim().length > 0);
+}
+function parseToolCall(text) {
+    const parsed = extractJsonObject(text);
+    if (!parsed)
+        return false;
+    const record = parsed;
+    if (record.tool !== "grep")
+        return false;
+    const args = record.arguments;
+    if (!args || typeof args !== "object" || Array.isArray(args))
+        return false;
+    const pattern = args.pattern;
+    const path = args.path;
+    return (typeof pattern === "string" && pattern.trim().length > 0 && typeof path === "string" && path.trim().length > 0);
+}
+function extractJsonObject(text) {
+    const trimmed = text.trim();
+    const candidates = [trimmed];
+    const fenced = /```(?:json)?\s*([\s\S]*?)```/.exec(trimmed);
+    if (fenced?.[1])
+        candidates.push(fenced[1].trim());
+    const start = trimmed.indexOf("{");
+    const end = trimmed.lastIndexOf("}");
+    if (start >= 0 && end > start)
+        candidates.push(trimmed.slice(start, end + 1));
+    for (const candidate of candidates) {
+        try {
+            const parsed = JSON.parse(candidate);
+            if (parsed && typeof parsed === "object" && !Array.isArray(parsed))
+                return parsed;
+        }
+        catch {
+            // try next candidate
+        }
+    }
+    return undefined;
+}
+function fitnessEnvelope() {
+    return {
+        id: "model-fitness-probe",
+        capabilities: ["research", "read_files", "memory_read"],
+        maxEstimatedUsd: 1,
+        createdAt: new Date().toISOString(),
+    };
+}
+export async function runModelFitnessProbe(options) {
+    const trials = Math.max(1, Math.min(options.trials ?? 3, 20));
+    const maxWallClockMs = options.maxWallClockMs ?? 120_000;
+    const judgePrompts = options.judgePrompts ?? DEFAULT_JUDGE_FITNESS_PROMPTS;
+    const now = options.now ?? Date.now;
+    let totalCostUsd = 0;
+    // Token-speed instrumentation: the lane runners' own contracts carry text/cost only, so the
+    // completer is wrapped once here and generation stats are accumulated per surface.
+    const overallSpeed = { tokens: 0, evalMs: 0 };
+    let surfaceSpeed = { tokens: 0, evalMs: 0 };
+    const complete = async (args) => {
+        const completion = await options.complete(args);
+        const tokens = completion.outputTokens ?? 0;
+        const evalMs = completion.evalMs ?? 0;
+        if (tokens > 0 && evalMs > 0) {
+            surfaceSpeed.tokens += tokens;
+            surfaceSpeed.evalMs += evalMs;
+            overallSpeed.tokens += tokens;
+            overallSpeed.evalMs += evalMs;
+        }
+        return completion;
+    };
+    const takeSurfaceSpeed = () => {
+        const speed = surfaceSpeed.evalMs > 0 ? Math.round((surfaceSpeed.tokens / surfaceSpeed.evalMs) * 1000) : undefined;
+        surfaceSpeed = { tokens: 0, evalMs: 0 };
+        return speed;
+    };
+    const research = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };
+    for (let i = 0; i < trials; i++) {
+        const started = now();
+        const result = await runResearch({
+            query: `fitness:probe requirements:req-${i}`,
+            context: [
+                "Goal: add a retry helper to the HTTP client module",
+                "Open requirements:",
+                "- Find what retry/backoff conventions the codebase already uses",
+                "- Identify which call sites would adopt the helper",
+            ].join("\n"),
+            envelope: fitnessEnvelope(),
+            maxUsd: 1,
+            maxSources: 8,
+            maxFindings: 5,
+            maxWallClockMs,
+            signal: options.signal,
+            complete,
+        });
+        research.meanMs += now() - started;
+        totalCostUsd += result.costUsd;
+        if (result.status === "succeeded")
+            research.succeeded++;
+        research.outcomes.push(`${result.status}/${result.reasonCode}`);
+    }
+    research.meanMs = Math.round(research.meanMs / trials);
+    research.tokensPerSecond = takeSurfaceSpeed();
+    const worker = { succeeded: 0, total: trials, outcomes: [], meanMs: 0 };
+    for (let i = 0; i < trials; i++) {
+        const started = now();
+        const outcome = await runWorker({
+            request: {
+                id: `fitness-worker-${i}`,
+                instructions: "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.",
+                route: { tier: "cheap", risk: "read-only", confidence: 1, reasonCode: "fitness_probe", reasons: [] },
+                envelope: { id: `fitness-env-${i}`, capabilities: ["read_files"], maxEstimatedUsd: 1 },
+                maxEstimatedUsd: 1,
+            },
+            maxUsd: 1,
+            maxWallClockMs,
+            usageReportId: `fitness:${i}`,
+            signal: options.signal,
+            complete,
+        });
+        worker.meanMs += now() - started;
+        totalCostUsd += outcome.costUsd;
+        if (outcome.result.status === "completed" && outcome.accepted)
+            worker.succeeded++;
+        worker.outcomes.push(`${outcome.result.status}/${outcome.reasonCode}`);
+    }
+    worker.meanMs = Math.round(worker.meanMs / trials);
+    worker.tokensPerSecond = takeSurfaceSpeed();
+    const judge = {
+        parsed: 0,
+        planningElevated: 0,
+        planningTotal: judgePrompts.filter((entry) => entry.planning).length,
+        trivialCheap: 0,
+        trivialTotal: judgePrompts.filter((entry) => !entry.planning).length,
+        total: judgePrompts.length,
+        outcomes: [],
+        meanMs: 0,
+    };
+    for (const entry of judgePrompts) {
+        const started = now();
+        const result = await runRouteJudge({
+            prompt: entry.prompt,
+            baseline: { tier: "cheap", risk: "read-only", confidence: 0.5, reasonCode: "fitness_probe", reasons: [] },
+            maxWallClockMs,
+            signal: options.signal,
+            complete,
+        });
+        judge.meanMs += now() - started;
+        totalCostUsd += result.costUsd;
+        const tier = result.decision.tier;
+        if (result.verdict) {
+            judge.parsed++;
+            // A useful judge must both keep planning off the cheap tier AND actually send trivial
+            // prompts there — all-medium verdicts are safe but save nothing.
+            if (entry.planning && tier !== "cheap")
+                judge.planningElevated++;
+            if (!entry.planning && tier === "cheap")
+                judge.trivialCheap++;
+        }
+        judge.outcomes.push(`"${entry.prompt.slice(0, 40)}" -> ${tier}${result.fallbackReason ? ` (${result.fallbackReason})` : ""}`);
+    }
+    judge.meanMs = judgePrompts.length > 0 ? Math.round(judge.meanMs / judgePrompts.length) : 0;
+    judge.tokensPerSecond = takeSurfaceSpeed();
+    const probeSurface = async (systemPrompt, tasks, accepts) => {
+        const score = { succeeded: 0, total: tasks.length, outcomes: [], meanMs: 0 };
+        for (const task of tasks) {
+            const started = now();
+            // Same wall-clock envelope as the lane surfaces — a hung model must not hang the probe.
+            const bounded = await runBoundedCompletion({
+                maxWallClockMs,
+                signal: options.signal,
+                execute: (signal) => complete({ systemPrompt, userPrompt: task, signal }),
+            });
+            if (bounded.completion)
+                totalCostUsd += bounded.completion.costUsd;
+            if (bounded.failure || !bounded.completion) {
+                score.outcomes.push(bounded.failure ? bounded.failure.status : "completion_error");
+            }
+            else {
+                const ok = accepts(bounded.completion.text);
+                if (ok)
+                    score.succeeded++;
+                score.outcomes.push(ok ? "ok" : "unparseable_output");
+            }
+            score.meanMs += now() - started;
+        }
+        score.meanMs = tasks.length > 0 ? Math.round(score.meanMs / tasks.length) : 0;
+        return score;
+    };
+    const search = await probeSurface(SEARCH_PROBE_SYSTEM_PROMPT, SEARCH_PROBE_TASKS, parseSearchPlan);
+    search.tokensPerSecond = takeSurfaceSpeed();
+    const toolCall = await probeSurface(TOOL_CALL_PROBE_SYSTEM_PROMPT, TOOL_CALL_PROBE_TASKS, parseToolCall);
+    toolCall.tokensPerSecond = takeSurfaceSpeed();
+    const tokensPerSecond = overallSpeed.evalMs > 0 ? Math.round((overallSpeed.tokens / overallSpeed.evalMs) * 1000) : undefined;
+    return { trials, tokensPerSecond, research, worker, judge, search, toolCall, totalCostUsd };
+}
+/** Compact human-readable report for tool output / interactive display. Bounded, no raw dumps. */
+export function formatModelFitnessReport(model, report) {
+    const speed = (tokensPerSecond) => tokensPerSecond !== undefined ? `, ~${tokensPerSecond} tok/s` : "";
+    const lines = [
+        `Model fitness: ${model} (${report.trials} trials/lane${speed(report.tokensPerSecond)})`,
+        `- research lane: ${report.research.succeeded}/${report.research.total} succeeded, mean ${report.research.meanMs}ms${speed(report.research.tokensPerSecond)} [${report.research.outcomes.join(", ")}]`,
+        `- worker lane:   ${report.worker.succeeded}/${report.worker.total} completed+accepted, mean ${report.worker.meanMs}ms${speed(report.worker.tokensPerSecond)} [${report.worker.outcomes.join(", ")}]`,
+        `- search plans:  ${report.search.succeeded}/${report.search.total} well-formed, mean ${report.search.meanMs}ms${speed(report.search.tokensPerSecond)}`,
+        `- tool calls:    ${report.toolCall.succeeded}/${report.toolCall.total} well-formed, mean ${report.toolCall.meanMs}ms${speed(report.toolCall.tokensPerSecond)}`,
+        `- 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)}`,
+        ...report.judge.outcomes.map((outcome) => `    ${outcome}`),
+    ];
+    if (report.totalCostUsd > 0) {
+        lines.push(`- probe cost: $${report.totalCostUsd.toFixed(4)}`);
+    }
+    return lines.join("\n");
+}
+//# sourceMappingURL=model-fitness.js.map

package/dist/core/research/model-fitness.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-fitness.js","sourceRoot":"","sources":["../../../src/core/research/model-fitness.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,oBAAoB,EAAE,MAAM,mCAAmC,CAAC;AAEzE,OAAO,EAAE,SAAS,EAAE,MAAM,gCAAgC,CAAC;AAC3D,OAAO,EAAE,aAAa,EAAE,MAAM,gCAAgC,CAAC;AAC/D,OAAO,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AAgCnD,qFAAqF;AACrF,MAAM,CAAC,MAAM,6BAA6B,GAAkC;IAC3E,EAAE,MAAM,EAAE,gEAAgE,EAAE,QAAQ,EAAE,IAAI,EAAE;IAC5F,EAAE,MAAM,EAAE,uDAAuD,EAAE,QAAQ,EAAE,IAAI,EAAE;IACnF,EAAE,MAAM,EAAE,yCAAyC,EAAE,QAAQ,EAAE,IAAI,EAAE;IACrE,EAAE,MAAM,EAAE,4CAA4C,EAAE,QAAQ,EAAE,KAAK,EAAE;IACzE,EAAE,MAAM,EAAE,yCAAyC,EAAE,QAAQ,EAAE,KAAK,EAAE;IACtE,EAAE,MAAM,EAAE,yBAAyB,EAAE,QAAQ,EAAE,KAAK,EAAE;CACtD,CAAC;AAkDF,yFAAyF;AACzF,MAAM,CAAC,MAAM,0BAA0B,GAAG;IACzC,oFAAoF;IACpF,8EAA8E;IAC9E,0FAA0F;IAC1F,6CAA6C;CAC7C,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAEb,MAAM,CAAC,MAAM,6BAA6B,GAAG;IAC5C,kEAAkE;IAClE,gFAAgF;IAChF,yDAAyD;IACzD,qEAAqE;CACrE,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAEb,MAAM,kBAAkB,GAAsB;IAC7C,iEAAiE;IACjE,0DAA0D;IAC1D,yDAAyD;CACzD,CAAC;AAEF,MAAM,qBAAqB,GAAsB;IAChD,yDAAyD;IACzD,iEAAiE;IACjE,0DAA0D;CAC1D,CAAC;AAEF,SAAS,eAAe,CAAC,IAAY,EAAW;IAC/C,MAAM,MAAM,GAAG,iBAAiB,CAAC,IAAI,CAAC,CAAC;IACvC,IAAI,CAAC,MAAM;QAAE,OAAO,KAAK,CAAC;IAC1B,MAAM,OAAO,GAAI,MAAgC,CAAC,OAAO,CAAC;IAC1D,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,KAAK,CAAC;IACxF,OAAO,OAAO,CAAC,KAAK,CACnB,CAAC,KAAK,EAAE,EAAE,CACT,KAAK;QACL,OAAO,KAAK,KAAK,QAAQ;QACzB,OAAQ,KAA+B,CAAC,OAAO,KAAK,QAAQ;QAC3D,KAA6B,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CACzD,CAAC;AAAA,CACF;AAED,SAAS,aAAa,CAAC,IAAY,EAAW;IAC7C,MAAM,MAAM,GAAG,iBAAiB,CAAC,IAAI,CAAC,CAAC;IACvC,IAAI,CAAC,MAAM;QAAE,OAAO,KAAK,CAAC;IAC1B,MAAM,MAAM,GAAG,MAAiD,CAAC;IACjE,IAAI,MAAM,CAAC,IAAI,KAAK,MAAM;QAAE,OAAO,KAAK,CAAC;IACzC,MAAM,IAAI,GAAG,MAAM,CAAC,SAAS,CAAC;IAC9B,IAAI,CAAC,IAAI,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC;QAAE,OAAO,KAAK,CAAC;IAC3E,MAAM,OAAO,GAAI,IAA8B,CAAC,OAAO,CAAC;IACxD,MAAM,IAAI,GAAI,IAA2B,CAAC,IAAI,CAAC;IAC/C,OAAO,CACN,OAAO,OAAO,KAAK,QAAQ,IAAI,OAAO,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAC9G,CAAC;AAAA,CACF;AAED,SAAS,iBAAiB,CAAC,IAAY,EAAuB;IAC7D,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;IAC5B,MAAM,UAAU,GAAa,CAAC,OAAO,CAAC,CAAC;IACvC,MAAM,MAAM,GAAG,8BAA8B,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;IAC5D,IAAI,MAAM,EAAE,CAAC,CAAC,CAAC;QAAE,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;IACnD,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IACnC,MAAM,GAAG,GAAG,OAAO,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;IACrC,IAAI,KAAK,IAAI,CAAC,IAAI,GAAG,GAAG,KAAK;QAAE,UAAU,CAAC,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC;IAC9E,KAAK,MAAM,SAAS,IAAI,UAAU,EAAE,CAAC;QACpC,IAAI,CAAC;YACJ,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC;YACrC,IAAI,MAAM,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC;gBAAE,OAAO,MAAM,CAAC;QACnF,CAAC;QAAC,MAAM,CAAC;YACR,qBAAqB;QACtB,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAAA,CACjB;AAED,SAAS,eAAe,GAAuB;IAC9C,OAAO;QACN,EAAE,EAAE,qBAAqB;QACzB,YAAY,EAAE,CAAC,UAAU,EAAE,YAAY,EAAE,aAAa,CAAC;QACvD,eAAe,EAAE,CAAC;QAClB,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;KACnC,CAAC;AAAA,CACF;AAED,MAAM,CAAC,KAAK,UAAU,oBAAoB,CAAC,OAA4B,EAA+B;IACrG,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;IAC9D,MAAM,cAAc,GAAG,OAAO,CAAC,cAAc,IAAI,OAAO,CAAC;IACzD,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,IAAI,6BAA6B,CAAC;IAC3E,MAAM,GAAG,GAAG,OAAO,CAAC,GAAG,IAAI,IAAI,CAAC,GAAG,CAAC;IACpC,IAAI,YAAY,GAAG,CAAC,CAAC;IAErB,4FAA4F;IAC5F,mFAAmF;IACnF,MAAM,YAAY,GAAG,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;IAC9C,IAAI,YAAY,GAAG,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;IAC5C,MAAM,QAAQ,GAAoB,KAAK,EAAE,IAAI,EAAE,EAAE,CAAC;QACjD,MAAM,UAAU,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;QAChD,MAAM,MAAM,GAAG,UAAU,CAAC,YAAY,IAAI,CAAC,CAAC;QAC5C,MAAM,MAAM,GAAG,UAAU,CAAC,MAAM,IAAI,CAAC,CAAC;QACtC,IAAI,MAAM,GAAG,CAAC,IAAI,MAAM,GAAG,CAAC,EAAE,CAAC;YAC9B,YAAY,CAAC,MAAM,IAAI,MAAM,CAAC;YAC9B,YAAY,CAAC,MAAM,IAAI,MAAM,CAAC;YAC9B,YAAY,CAAC,MAAM,IAAI,MAAM,CAAC;YAC9B,YAAY,CAAC,MAAM,IAAI,MAAM,CAAC;QAC/B,CAAC;QACD,OAAO,UAAU,CAAC;IAAA,CAClB,CAAC;IACF,MAAM,gBAAgB,GAAG,GAAuB,EAAE,CAAC;QAClD,MAAM,KAAK,GACV,YAAY,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,YAAY,CAAC,MAAM,GAAG,YAAY,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QACtG,YAAY,GAAG,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;QACxC,OAAO,KAAK,CAAC;IAAA,CACb,CAAC;IAEF,MAAM,QAAQ,GAAqB,EAAE,SAAS,EAAE,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;IAC5F,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACjC,MAAM,OAAO,GAAG,GAAG,EAAE,CAAC;QACtB,MAAM,MAAM,GAAG,MAAM,WAAW,CAAC;YAChC,KAAK,EAAE,kCAAkC,CAAC,EAAE;YAC5C,OAAO,EAAE;gBACR,oDAAoD;gBACpD,oBAAoB;gBACpB,iEAAiE;gBACjE,oDAAoD;aACpD,CAAC,IAAI,CAAC,IAAI,CAAC;YACZ,QAAQ,EAAE,eAAe,EAAE;YAC3B,MAAM,EAAE,CAAC;YACT,UAAU,EAAE,CAAC;YACb,WAAW,EAAE,CAAC;YACd,cAAc;YACd,MAAM,EAAE,OAAO,CAAC,MAAM;YACtB,QAAQ;SACR,CAAC,CAAC;QACH,QAAQ,CAAC,MAAM,IAAI,GAAG,EAAE,GAAG,OAAO,CAAC;QACnC,YAAY,IAAI,MAAM,CAAC,OAAO,CAAC;QAC/B,IAAI,MAAM,CAAC,MAAM,KAAK,WAAW;YAAE,QAAQ,CAAC,SAAS,EAAE,CAAC;QACxD,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,MAAM,CAAC,MAAM,IAAI,MAAM,CAAC,UAAU,EAAE,CAAC,CAAC;IACjE,CAAC;IACD,QAAQ,CAAC,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC;IACvD,QAAQ,CAAC,eAAe,GAAG,gBAAgB,EAAE,CAAC;IAE9C,MAAM,MAAM,GAAqB,EAAE,SAAS,EAAE,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;IAC1F,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACjC,MAAM,OAAO,GAAG,GAAG,EAAE,CAAC;QACtB,MAAM,OAAO,GAAG,MAAM,SAAS,CAAC;YAC/B,OAAO,EAAE;gBACR,EAAE,EAAE,kBAAkB,CAAC,EAAE;gBACzB,YAAY,EACX,oKAAoK;gBACrK,KAAK,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,CAAC,EAAE,UAAU,EAAE,eAAe,EAAE,OAAO,EAAE,EAAE,EAAE;gBACpG,QAAQ,EAAE,EAAE,EAAE,EAAE,eAAe,CAAC,EAAE,EAAE,YAAY,EAAE,CAAC,YAAY,CAAC,EAAE,eAAe,EAAE,CAAC,EAAE;gBACtF,eAAe,EAAE,CAAC;aAClB;YACD,MAAM,EAAE,CAAC;YACT,cAAc;YACd,aAAa,EAAE,WAAW,CAAC,EAAE;YAC7B,MAAM,EAAE,OAAO,CAAC,MAAM;YACtB,QAAQ;SACR,CAAC,CAAC;QACH,MAAM,CAAC,MAAM,IAAI,GAAG,EAAE,GAAG,OAAO,CAAC;QACjC,YAAY,IAAI,OAAO,CAAC,OAAO,CAAC;QAChC,IAAI,OAAO,CAAC,MAAM,CAAC,MAAM,KAAK,WAAW,IAAI,OAAO,CAAC,QAAQ;YAAE,MAAM,CAAC,SAAS,EAAE,CAAC;QAClF,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,MAAM,IAAI,OAAO,CAAC,UAAU,EAAE,CAAC,CAAC;IACxE,CAAC;IACD,MAAM,CAAC,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC;IACnD,MAAM,CAAC,eAAe,GAAG,gBAAgB,EAAE,CAAC;IAE5C,MAAM,KAAK,GAAsB;QAChC,MAAM,EAAE,CAAC;QACT,gBAAgB,EAAE,CAAC;QACnB,aAAa,EAAE,YAAY,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,MAAM;QACpE,YAAY,EAAE,CAAC;QACf,YAAY,EAAE,YAAY,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,MAAM;QACpE,KAAK,EAAE,YAAY,CAAC,MAAM;QAC1B,QAAQ,EAAE,EAAE;QACZ,MAAM,EAAE,CAAC;KACT,CAAC;IACF,KAAK,MAAM,KAAK,IAAI,YAAY,EAAE,CAAC;QAClC,MAAM,OAAO,GAAG,GAAG,EAAE,CAAC;QACtB,MAAM,MAAM,GAAG,MAAM,aAAa,CAAC;YAClC,MAAM,EAAE,KAAK,CAAC,MAAM;YACpB,QAAQ,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,GAAG,EAAE,UAAU,EAAE,eAAe,EAAE,OAAO,EAAE,EAAE,EAAE;YACzG,cAAc;YACd,MAAM,EAAE,OAAO,CAAC,MAAM;YACtB,QAAQ;SACR,CAAC,CAAC;QACH,KAAK,CAAC,MAAM,IAAI,GAAG,EAAE,GAAG,OAAO,CAAC;QAChC,YAAY,IAAI,MAAM,CAAC,OAAO,CAAC;QAC/B,MAAM,IAAI,GAAG,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC;QAClC,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC;YACpB,KAAK,CAAC,MAAM,EAAE,CAAC;YACf,sFAAsF;YACtF,mEAAiE;YACjE,IAAI,KAAK,CAAC,QAAQ,IAAI,IAAI,KAAK,OAAO;gBAAE,KAAK,CAAC,gBAAgB,EAAE,CAAC;YACjE,IAAI,CAAC,KAAK,CAAC,QAAQ,IAAI,IAAI,KAAK,OAAO;gBAAE,KAAK,CAAC,YAAY,EAAE,CAAC;QAC/D,CAAC;QACD,KAAK,CAAC,QAAQ,CAAC,IAAI,CAClB,IAAI,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,QAAQ,IAAI,GAAG,MAAM,CAAC,cAAc,CAAC,CAAC,CAAC,KAAK,MAAM,CAAC,cAAc,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CACxG,CAAC;IACH,CAAC;IACD,KAAK,CAAC,MAAM,GAAG,YAAY,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,YAAY,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAC5F,KAAK,CAAC,eAAe,GAAG,gBAAgB,EAAE,CAAC;IAE3C,MAAM,YAAY,GAAG,KAAK,EACzB,YAAoB,EACpB,KAAwB,EACxB,OAAkC,EACN,EAAE,CAAC;QAC/B,MAAM,KAAK,GAAqB,EAAE,SAAS,EAAE,CAAC,EAAE,KAAK,EAAE,KAAK,CAAC,MAAM,EAAE,QAAQ,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;QAC/F,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;YAC1B,MAAM,OAAO,GAAG,GAAG,EAAE,CAAC;YACtB,0FAAwF;YACxF,MAAM,OAAO,GAAG,MAAM,oBAAoB,CAAC;gBAC1C,cAAc;gBACd,MAAM,EAAE,OAAO,CAAC,MAAM;gBACtB,OAAO,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,QAAQ,CAAC,EAAE,YAAY,EAAE,UAAU,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;aACzE,CAAC,CAAC;YACH,IAAI,OAAO,CAAC,UAAU;gBAAE,YAAY,IAAI,OAAO,CAAC,UAAU,CAAC,OAAO,CAAC;YACnE,IAAI,OAAO,CAAC,OAAO,IAAI,CAAC,OAAO,CAAC,UAAU,EAAE,CAAC;gBAC5C,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,kBAAkB,CAAC,CAAC;YACpF,CAAC;iBAAM,CAAC;gBACP,MAAM,EAAE,GAAG,OAAO,CAAC,OAAO,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC;gBAC5C,IAAI,EAAE;oBAAE,KAAK,CAAC,SAAS,EAAE,CAAC;gBAC1B,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,oBAAoB,CAAC,CAAC;YACvD,CAAC;YACD,KAAK,CAAC,MAAM,IAAI,GAAG,EAAE,GAAG,OAAO,CAAC;QACjC,CAAC;QACD,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAC9E,OAAO,KAAK,CAAC;IAAA,CACb,CAAC;IAEF,MAAM,MAAM,GAAG,MAAM,YAAY,CAAC,0BAA0B,EAAE,kBAAkB,EAAE,eAAe,CAAC,CAAC;IACnG,MAAM,CAAC,eAAe,GAAG,gBAAgB,EAAE,CAAC;IAC5C,MAAM,QAAQ,GAAG,MAAM,YAAY,CAAC,6BAA6B,EAAE,qBAAqB,EAAE,aAAa,CAAC,CAAC;IACzG,QAAQ,CAAC,eAAe,GAAG,gBAAgB,EAAE,CAAC;IAE9C,MAAM,eAAe,GACpB,YAAY,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,YAAY,CAAC,MAAM,GAAG,YAAY,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;IAEtG,OAAO,EAAE,MAAM,EAAE,eAAe,EAAE,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,YAAY,EAAE,CAAC;AAAA,CAC5F;AAED,kGAAkG;AAClG,MAAM,UAAU,wBAAwB,CAAC,KAAa,EAAE,MAA0B,EAAU;IAC3F,MAAM,KAAK,GAAG,CAAC,eAAmC,EAAE,EAAE,CACrD,eAAe,KAAK,SAAS,CAAC,CAAC,CAAC,MAAM,eAAe,QAAQ,CAAC,CAAC,CAAC,EAAE,CAAC;IACpE,MAAM,KAAK,GAAG;QACb,kBAAkB,KAAK,KAAK,MAAM,CAAC,MAAM,eAAe,KAAK,CAAC,MAAM,CAAC,eAAe,CAAC,GAAG;QACxF,oBAAoB,MAAM,CAAC,QAAQ,CAAC,SAAS,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,oBAAoB,MAAM,CAAC,QAAQ,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,QAAQ,CAAC,eAAe,CAAC,KAAK,MAAM,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG;QACtM,oBAAoB,MAAM,CAAC,MAAM,CAAC,SAAS,IAAI,MAAM,CAAC,MAAM,CAAC,KAAK,6BAA6B,MAAM,CAAC,MAAM,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC,eAAe,CAAC,KAAK,MAAM,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG;QACrM,oBAAoB,MAAM,CAAC,MAAM,CAAC,SAAS,IAAI,MAAM,CAAC,MAAM,CAAC,KAAK,sBAAsB,MAAM,CAAC,MAAM,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC,eAAe,CAAC,EAAE;QACvJ,oBAAoB,MAAM,CAAC,QAAQ,CAAC,SAAS,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,sBAAsB,MAAM,CAAC,QAAQ,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,QAAQ,CAAC,eAAe,CAAC,EAAE;QAC/J,2BAA2B,MAAM,CAAC,KAAK,CAAC,MAAM,IAAI,MAAM,CAAC,KAAK,CAAC,KAAK,uBAAuB,MAAM,CAAC,KAAK,CAAC,gBAAgB,IAAI,MAAM,CAAC,KAAK,CAAC,aAAa,mBAAmB,MAAM,CAAC,KAAK,CAAC,YAAY,IAAI,MAAM,CAAC,KAAK,CAAC,YAAY,UAAU,MAAM,CAAC,KAAK,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,eAAe,CAAC,EAAE;QACtS,GAAG,MAAM,CAAC,KAAK,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,OAAO,EAAE,CAAC;KAC3D,CAAC;IACF,IAAI,MAAM,CAAC,YAAY,GAAG,CAAC,EAAE,CAAC;QAC7B,KAAK,CAAC,IAAI,CAAC,kBAAkB,MAAM,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IAChE,CAAC;IACD,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAAA,CACxB","sourcesContent":["import { runBoundedCompletion } from \"../autonomy/bounded-completion.ts\";\nimport type { CapabilityEnvelope } from \"../autonomy/contracts.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\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\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) => boolean,\n\t): Promise<LaneFitnessScore> => {\n\t\tconst score: LaneFitnessScore = { succeeded: 0, total: tasks.length, outcomes: [], meanMs: 0 };\n\t\tfor (const task of tasks) {\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);\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\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, 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`- 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"]}

package/dist/core/research/research-gate.d.ts

@@ -0,0 +1,11 @@
+import type { CapabilityEnvelope, EvidenceSourceKind, GateOutcome } from "../autonomy/contracts.ts";
+interface ResearchRequestArgs {
+    envelope?: CapabilityEnvelope | null;
+    sourceKind: EvidenceSourceKind | string;
+    estimatedUsd: number;
+    maxEstimatedUsd?: number;
+    privateHistoryAllowed?: boolean;
+}
+export declare function evaluateResearchRequest(args: ResearchRequestArgs): GateOutcome;
+export {};
+//# sourceMappingURL=research-gate.d.ts.map

package/dist/core/research/research-gate.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"research-gate.d.ts","sourceRoot":"","sources":["../../../src/core/research/research-gate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAkB,kBAAkB,EAAE,WAAW,EAAE,MAAM,0BAA0B,CAAC;AAEpH,UAAU,mBAAmB;IAC5B,QAAQ,CAAC,EAAE,kBAAkB,GAAG,IAAI,CAAC;IACrC,UAAU,EAAE,kBAAkB,GAAG,MAAM,CAAC;IACxC,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,qBAAqB,CAAC,EAAE,OAAO,CAAC;CAChC;AAmBD,wBAAgB,uBAAuB,CAAC,IAAI,EAAE,mBAAmB,GAAG,WAAW,CAyE9E","sourcesContent":["import type { CapabilityEnvelope, CapabilityName, EvidenceSourceKind, GateOutcome } from \"../autonomy/contracts.ts\";\n\ninterface ResearchRequestArgs {\n\tenvelope?: CapabilityEnvelope | null;\n\tsourceKind: EvidenceSourceKind | string;\n\testimatedUsd: number;\n\tmaxEstimatedUsd?: number;\n\tprivateHistoryAllowed?: boolean;\n}\n\nfunction missingCapabilityOutcome(sourceKind: string, capabilities: readonly CapabilityName[]): GateOutcome {\n\treturn {\n\t\toutcome: \"block\",\n\t\tgate: \"research_gate\",\n\t\treasonCode: \"missing_capability\",\n\t\tmessage: `Source kind '${sourceKind}' requires capability ${capabilities.join(\" or \")}.`,\n\t};\n}\n\nfunction hasAnyCapability(envelope: CapabilityEnvelope, capabilities: readonly CapabilityName[]): boolean {\n\treturn capabilities.some((capability) => envelope.capabilities.includes(capability));\n}\n\nfunction isWellFormedEnvelope(value: CapabilityEnvelope | null | undefined): value is CapabilityEnvelope {\n\treturn Boolean(value) && typeof value?.id === \"string\" && Array.isArray(value.capabilities);\n}\n\nexport function evaluateResearchRequest(args: ResearchRequestArgs): GateOutcome {\n\tif (!isWellFormedEnvelope(args.envelope)) {\n\t\treturn {\n\t\t\toutcome: \"block\",\n\t\t\tgate: \"research_gate\",\n\t\t\treasonCode: \"missing_envelope\",\n\t\t\tmessage: \"Missing or malformed capability envelope.\",\n\t\t};\n\t}\n\n\tconst { envelope, sourceKind, estimatedUsd, maxEstimatedUsd, privateHistoryAllowed } = args;\n\tconst limit = maxEstimatedUsd ?? envelope.maxEstimatedUsd;\n\tif (limit !== undefined && estimatedUsd > limit) {\n\t\treturn {\n\t\t\toutcome: \"ask-user\",\n\t\t\tgate: \"research_gate\",\n\t\t\treasonCode: \"over_budget\",\n\t\t\tmessage: `Estimated cost (${estimatedUsd}) exceeds maximum allowed (${limit}).`,\n\t\t};\n\t}\n\n\tswitch (sourceKind) {\n\t\tcase \"workspace\":\n\t\tcase \"tool\":\n\t\tcase \"user\": {\n\t\t\tconst requiredCapabilities: readonly CapabilityName[] = [\"read_files\", \"research\"];\n\t\t\tif (!hasAnyCapability(envelope, requiredCapabilities)) {\n\t\t\t\treturn missingCapabilityOutcome(sourceKind, requiredCapabilities);\n\t\t\t}\n\t\t\tbreak;\n\t\t}\n\n\t\tcase \"transcript\":\n\t\t\tif (!hasAnyCapability(envelope, [\"memory_read\"])) {\n\t\t\t\treturn missingCapabilityOutcome(sourceKind, [\"memory_read\"]);\n\t\t\t}\n\t\t\tbreak;\n\n\t\tcase \"automata\":\n\t\t\tif (!hasAnyCapability(envelope, [\"memory_read\"])) {\n\t\t\t\treturn missingCapabilityOutcome(sourceKind, [\"memory_read\"]);\n\t\t\t}\n\t\t\tif (!privateHistoryAllowed) {\n\t\t\t\treturn {\n\t\t\t\t\toutcome: \"ask-user\",\n\t\t\t\t\tgate: \"research_gate\",\n\t\t\t\t\treasonCode: \"private_history_denied\",\n\t\t\t\t\tmessage: \"Automata source requires privateHistoryAllowed=true.\",\n\t\t\t\t};\n\t\t\t}\n\t\t\tbreak;\n\n\t\tcase \"web\":\n\t\t\tif (!hasAnyCapability(envelope, [\"network\"])) {\n\t\t\t\treturn missingCapabilityOutcome(sourceKind, [\"network\"]);\n\t\t\t}\n\t\t\tbreak;\n\n\t\tdefault:\n\t\t\treturn {\n\t\t\t\toutcome: \"block\",\n\t\t\t\tgate: \"research_gate\",\n\t\t\t\treasonCode: \"unknown_source_kind\",\n\t\t\t\tmessage: `Unknown source kind '${sourceKind}'.`,\n\t\t\t};\n\t}\n\n\treturn {\n\t\toutcome: \"allow\",\n\t\tgate: \"research_gate\",\n\t\treasonCode: \"allowed\",\n\t\tmessage: \"Research request allowed.\",\n\t};\n}\n"]}

package/dist/core/research/research-gate.js

@@ -0,0 +1,82 @@
+function missingCapabilityOutcome(sourceKind, capabilities) {
+    return {
+        outcome: "block",
+        gate: "research_gate",
+        reasonCode: "missing_capability",
+        message: `Source kind '${sourceKind}' requires capability ${capabilities.join(" or ")}.`,
+    };
+}
+function hasAnyCapability(envelope, capabilities) {
+    return capabilities.some((capability) => envelope.capabilities.includes(capability));
+}
+function isWellFormedEnvelope(value) {
+    return Boolean(value) && typeof value?.id === "string" && Array.isArray(value.capabilities);
+}
+export function evaluateResearchRequest(args) {
+    if (!isWellFormedEnvelope(args.envelope)) {
+        return {
+            outcome: "block",
+            gate: "research_gate",
+            reasonCode: "missing_envelope",
+            message: "Missing or malformed capability envelope.",
+        };
+    }
+    const { envelope, sourceKind, estimatedUsd, maxEstimatedUsd, privateHistoryAllowed } = args;
+    const limit = maxEstimatedUsd ?? envelope.maxEstimatedUsd;
+    if (limit !== undefined && estimatedUsd > limit) {
+        return {
+            outcome: "ask-user",
+            gate: "research_gate",
+            reasonCode: "over_budget",
+            message: `Estimated cost (${estimatedUsd}) exceeds maximum allowed (${limit}).`,
+        };
+    }
+    switch (sourceKind) {
+        case "workspace":
+        case "tool":
+        case "user": {
+            const requiredCapabilities = ["read_files", "research"];
+            if (!hasAnyCapability(envelope, requiredCapabilities)) {
+                return missingCapabilityOutcome(sourceKind, requiredCapabilities);
+            }
+            break;
+        }
+        case "transcript":
+            if (!hasAnyCapability(envelope, ["memory_read"])) {
+                return missingCapabilityOutcome(sourceKind, ["memory_read"]);
+            }
+            break;
+        case "automata":
+            if (!hasAnyCapability(envelope, ["memory_read"])) {
+                return missingCapabilityOutcome(sourceKind, ["memory_read"]);
+            }
+            if (!privateHistoryAllowed) {
+                return {
+                    outcome: "ask-user",
+                    gate: "research_gate",
+                    reasonCode: "private_history_denied",
+                    message: "Automata source requires privateHistoryAllowed=true.",
+                };
+            }
+            break;
+        case "web":
+            if (!hasAnyCapability(envelope, ["network"])) {
+                return missingCapabilityOutcome(sourceKind, ["network"]);
+            }
+            break;
+        default:
+            return {
+                outcome: "block",
+                gate: "research_gate",
+                reasonCode: "unknown_source_kind",
+                message: `Unknown source kind '${sourceKind}'.`,
+            };
+    }
+    return {
+        outcome: "allow",
+        gate: "research_gate",
+        reasonCode: "allowed",
+        message: "Research request allowed.",
+    };
+}
+//# sourceMappingURL=research-gate.js.map

package/dist/core/research/research-gate.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"research-gate.js","sourceRoot":"","sources":["../../../src/core/research/research-gate.ts"],"names":[],"mappings":"AAUA,SAAS,wBAAwB,CAAC,UAAkB,EAAE,YAAuC,EAAe;IAC3G,OAAO;QACN,OAAO,EAAE,OAAO;QAChB,IAAI,EAAE,eAAe;QACrB,UAAU,EAAE,oBAAoB;QAChC,OAAO,EAAE,gBAAgB,UAAU,yBAAyB,YAAY,CAAC,IAAI,CAAC,MAAM,CAAC,GAAG;KACxF,CAAC;AAAA,CACF;AAED,SAAS,gBAAgB,CAAC,QAA4B,EAAE,YAAuC,EAAW;IACzG,OAAO,YAAY,CAAC,IAAI,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,QAAQ,CAAC,YAAY,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC,CAAC;AAAA,CACrF;AAED,SAAS,oBAAoB,CAAC,KAA4C,EAA+B;IACxG,OAAO,OAAO,CAAC,KAAK,CAAC,IAAI,OAAO,KAAK,EAAE,EAAE,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,YAAY,CAAC,CAAC;AAAA,CAC5F;AAED,MAAM,UAAU,uBAAuB,CAAC,IAAyB,EAAe;IAC/E,IAAI,CAAC,oBAAoB,CAAC,IAAI,CAAC,QAAQ,CAAC,EAAE,CAAC;QAC1C,OAAO;YACN,OAAO,EAAE,OAAO;YAChB,IAAI,EAAE,eAAe;YACrB,UAAU,EAAE,kBAAkB;YAC9B,OAAO,EAAE,2CAA2C;SACpD,CAAC;IACH,CAAC;IAED,MAAM,EAAE,QAAQ,EAAE,UAAU,EAAE,YAAY,EAAE,eAAe,EAAE,qBAAqB,EAAE,GAAG,IAAI,CAAC;IAC5F,MAAM,KAAK,GAAG,eAAe,IAAI,QAAQ,CAAC,eAAe,CAAC;IAC1D,IAAI,KAAK,KAAK,SAAS,IAAI,YAAY,GAAG,KAAK,EAAE,CAAC;QACjD,OAAO;YACN,OAAO,EAAE,UAAU;YACnB,IAAI,EAAE,eAAe;YACrB,UAAU,EAAE,aAAa;YACzB,OAAO,EAAE,mBAAmB,YAAY,8BAA8B,KAAK,IAAI;SAC/E,CAAC;IACH,CAAC;IAED,QAAQ,UAAU,EAAE,CAAC;QACpB,KAAK,WAAW,CAAC;QACjB,KAAK,MAAM,CAAC;QACZ,KAAK,MAAM,EAAE,CAAC;YACb,MAAM,oBAAoB,GAA8B,CAAC,YAAY,EAAE,UAAU,CAAC,CAAC;YACnF,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE,oBAAoB,CAAC,EAAE,CAAC;gBACvD,OAAO,wBAAwB,CAAC,UAAU,EAAE,oBAAoB,CAAC,CAAC;YACnE,CAAC;YACD,MAAM;QACP,CAAC;QAED,KAAK,YAAY;YAChB,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE,CAAC,aAAa,CAAC,CAAC,EAAE,CAAC;gBAClD,OAAO,wBAAwB,CAAC,UAAU,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC;YAC9D,CAAC;YACD,MAAM;QAEP,KAAK,UAAU;YACd,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE,CAAC,aAAa,CAAC,CAAC,EAAE,CAAC;gBAClD,OAAO,wBAAwB,CAAC,UAAU,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC;YAC9D,CAAC;YACD,IAAI,CAAC,qBAAqB,EAAE,CAAC;gBAC5B,OAAO;oBACN,OAAO,EAAE,UAAU;oBACnB,IAAI,EAAE,eAAe;oBACrB,UAAU,EAAE,wBAAwB;oBACpC,OAAO,EAAE,sDAAsD;iBAC/D,CAAC;YACH,CAAC;YACD,MAAM;QAEP,KAAK,KAAK;YACT,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE,CAAC,SAAS,CAAC,CAAC,EAAE,CAAC;gBAC9C,OAAO,wBAAwB,CAAC,UAAU,EAAE,CAAC,SAAS,CAAC,CAAC,CAAC;YAC1D,CAAC;YACD,MAAM;QAEP;YACC,OAAO;gBACN,OAAO,EAAE,OAAO;gBAChB,IAAI,EAAE,eAAe;gBACrB,UAAU,EAAE,qBAAqB;gBACjC,OAAO,EAAE,wBAAwB,UAAU,IAAI;aAC/C,CAAC;IACJ,CAAC;IAED,OAAO;QACN,OAAO,EAAE,OAAO;QAChB,IAAI,EAAE,eAAe;QACrB,UAAU,EAAE,SAAS;QACrB,OAAO,EAAE,2BAA2B;KACpC,CAAC;AAAA,CACF","sourcesContent":["import type { CapabilityEnvelope, CapabilityName, EvidenceSourceKind, GateOutcome } from \"../autonomy/contracts.ts\";\n\ninterface ResearchRequestArgs {\n\tenvelope?: CapabilityEnvelope | null;\n\tsourceKind: EvidenceSourceKind | string;\n\testimatedUsd: number;\n\tmaxEstimatedUsd?: number;\n\tprivateHistoryAllowed?: boolean;\n}\n\nfunction missingCapabilityOutcome(sourceKind: string, capabilities: readonly CapabilityName[]): GateOutcome {\n\treturn {\n\t\toutcome: \"block\",\n\t\tgate: \"research_gate\",\n\t\treasonCode: \"missing_capability\",\n\t\tmessage: `Source kind '${sourceKind}' requires capability ${capabilities.join(\" or \")}.`,\n\t};\n}\n\nfunction hasAnyCapability(envelope: CapabilityEnvelope, capabilities: readonly CapabilityName[]): boolean {\n\treturn capabilities.some((capability) => envelope.capabilities.includes(capability));\n}\n\nfunction isWellFormedEnvelope(value: CapabilityEnvelope | null | undefined): value is CapabilityEnvelope {\n\treturn Boolean(value) && typeof value?.id === \"string\" && Array.isArray(value.capabilities);\n}\n\nexport function evaluateResearchRequest(args: ResearchRequestArgs): GateOutcome {\n\tif (!isWellFormedEnvelope(args.envelope)) {\n\t\treturn {\n\t\t\toutcome: \"block\",\n\t\t\tgate: \"research_gate\",\n\t\t\treasonCode: \"missing_envelope\",\n\t\t\tmessage: \"Missing or malformed capability envelope.\",\n\t\t};\n\t}\n\n\tconst { envelope, sourceKind, estimatedUsd, maxEstimatedUsd, privateHistoryAllowed } = args;\n\tconst limit = maxEstimatedUsd ?? envelope.maxEstimatedUsd;\n\tif (limit !== undefined && estimatedUsd > limit) {\n\t\treturn {\n\t\t\toutcome: \"ask-user\",\n\t\t\tgate: \"research_gate\",\n\t\t\treasonCode: \"over_budget\",\n\t\t\tmessage: `Estimated cost (${estimatedUsd}) exceeds maximum allowed (${limit}).`,\n\t\t};\n\t}\n\n\tswitch (sourceKind) {\n\t\tcase \"workspace\":\n\t\tcase \"tool\":\n\t\tcase \"user\": {\n\t\t\tconst requiredCapabilities: readonly CapabilityName[] = [\"read_files\", \"research\"];\n\t\t\tif (!hasAnyCapability(envelope, requiredCapabilities)) {\n\t\t\t\treturn missingCapabilityOutcome(sourceKind, requiredCapabilities);\n\t\t\t}\n\t\t\tbreak;\n\t\t}\n\n\t\tcase \"transcript\":\n\t\t\tif (!hasAnyCapability(envelope, [\"memory_read\"])) {\n\t\t\t\treturn missingCapabilityOutcome(sourceKind, [\"memory_read\"]);\n\t\t\t}\n\t\t\tbreak;\n\n\t\tcase \"automata\":\n\t\t\tif (!hasAnyCapability(envelope, [\"memory_read\"])) {\n\t\t\t\treturn missingCapabilityOutcome(sourceKind, [\"memory_read\"]);\n\t\t\t}\n\t\t\tif (!privateHistoryAllowed) {\n\t\t\t\treturn {\n\t\t\t\t\toutcome: \"ask-user\",\n\t\t\t\t\tgate: \"research_gate\",\n\t\t\t\t\treasonCode: \"private_history_denied\",\n\t\t\t\t\tmessage: \"Automata source requires privateHistoryAllowed=true.\",\n\t\t\t\t};\n\t\t\t}\n\t\t\tbreak;\n\n\t\tcase \"web\":\n\t\t\tif (!hasAnyCapability(envelope, [\"network\"])) {\n\t\t\t\treturn missingCapabilityOutcome(sourceKind, [\"network\"]);\n\t\t\t}\n\t\t\tbreak;\n\n\t\tdefault:\n\t\t\treturn {\n\t\t\t\toutcome: \"block\",\n\t\t\t\tgate: \"research_gate\",\n\t\t\t\treasonCode: \"unknown_source_kind\",\n\t\t\t\tmessage: `Unknown source kind '${sourceKind}'.`,\n\t\t\t};\n\t}\n\n\treturn {\n\t\toutcome: \"allow\",\n\t\tgate: \"research_gate\",\n\t\treasonCode: \"allowed\",\n\t\tmessage: \"Research request allowed.\",\n\t};\n}\n"]}

package/dist/core/research/research-runner.d.ts

@@ -0,0 +1,59 @@
+import type { CapabilityEnvelope, EvidenceBundle, GateOutcome } from "../autonomy/contracts.ts";
+/**
+ * Pure orchestration for one autonomous research pass: gate -> bounded isolated completion ->
+ * parse -> evidence bundle. The model executor is injected so this stays provider-free and
+ * session-free; production wires `AgentSession.runIsolatedCompletion` in.
+ *
+ * The lane is read-only by construction: the executor receives text prompts only, and the output
+ * is an `EvidenceBundle` whose model-synthesized findings are marked untrusted.
+ */
+/** Static across calls so callers can use `cacheRetention: "short"` and only pay for the variable tail. */
+export declare const RESEARCH_LANE_SYSTEM_PROMPT: string;
+export interface ResearchCompletion {
+    text: string;
+    costUsd: number;
+    stopReason: string;
+}
+export interface ResearchRunnerOptions {
+    query: string;
+    /** Bounded, pre-redacted context handed to the research model (goal text, open requirements). */
+    context?: string;
+    /** Stripped research envelope - never the foreground/architect envelope. */
+    envelope: CapabilityEnvelope;
+    /** Budget for this pass; a post-hoc breach marks the run budget_exhausted (spend stays visible). */
+    maxUsd: number;
+    maxSources: number;
+    maxFindings: number;
+    /** Wall-clock budget in milliseconds; 0 disables. */
+    maxWallClockMs: number;
+    /** Executes one isolated completion. Production: AgentSession.runIsolatedCompletion. */
+    complete: (args: {
+        systemPrompt: string;
+        userPrompt: string;
+        signal?: AbortSignal;
+    }) => Promise<ResearchCompletion>;
+    /** External cancellation (e.g. session disposal). */
+    signal?: AbortSignal;
+}
+export type ResearchRunStatus = "succeeded" | "failed" | "canceled" | "timeout" | "budget_exhausted";
+export interface ResearchRunResult {
+    status: ResearchRunStatus;
+    reasonCode: string;
+    gateOutcome: GateOutcome;
+    bundle?: EvidenceBundle;
+    costUsd: number;
+}
+export declare function buildResearchUserPrompt(args: {
+    query: string;
+    context?: string;
+    maxFindings: number;
+}): string;
+export interface ParsedResearchFindings {
+    findings: Array<{
+        summary: string;
+        confidence?: number;
+    }>;
+}
+export declare function parseResearchFindings(text: string, maxFindings: number): ParsedResearchFindings | undefined;
+export declare function runResearch(options: ResearchRunnerOptions): Promise<ResearchRunResult>;
+//# sourceMappingURL=research-runner.d.ts.map

package/dist/core/research/research-runner.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"research-runner.d.ts","sourceRoot":"","sources":["../../../src/core/research/research-runner.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,kBAAkB,EAAE,cAAc,EAAwB,WAAW,EAAE,MAAM,0BAA0B,CAAC;AAItH;;;;;;;GAOG;AAEH,2GAA2G;AAC3G,eAAO,MAAM,2BAA2B,QAM5B,CAAC;AAEb,MAAM,WAAW,kBAAkB;IAClC,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,qBAAqB;IACrC,KAAK,EAAE,MAAM,CAAC;IACd,iGAAiG;IACjG,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,4EAA4E;IAC5E,QAAQ,EAAE,kBAAkB,CAAC;IAC7B,oGAAoG;IACpG,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,qDAAqD;IACrD,cAAc,EAAE,MAAM,CAAC;IACvB,wFAAwF;IACxF,QAAQ,EAAE,CAAC,IAAI,EAAE;QAAE,YAAY,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,WAAW,CAAA;KAAE,KAAK,OAAO,CAAC,kBAAkB,CAAC,CAAC;IACpH,qDAAqD;IACrD,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAED,MAAM,MAAM,iBAAiB,GAAG,WAAW,GAAG,QAAQ,GAAG,UAAU,GAAG,SAAS,GAAG,kBAAkB,CAAC;AAErG,MAAM,WAAW,iBAAiB;IACjC,MAAM,EAAE,iBAAiB,CAAC;IAC1B,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,WAAW,CAAC;IACzB,MAAM,CAAC,EAAE,cAAc,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;CAChB;AAED,wBAAgB,uBAAuB,CAAC,IAAI,EAAE;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,MAAM,CAAC;IAAC,WAAW,EAAE,MAAM,CAAA;CAAE,GAAG,MAAM,CAO9G;AAED,MAAM,WAAW,sBAAsB;IACtC,QAAQ,EAAE,KAAK,CAAC;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;CAC1D;AAED,wBAAgB,qBAAqB,CAAC,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,GAAG,sBAAsB,GAAG,SAAS,CAwC3G;AA+BD,wBAAsB,WAAW,CAAC,OAAO,EAAE,qBAAqB,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAwD5F","sourcesContent":["import { runBoundedCompletion } from \"../autonomy/bounded-completion.ts\";\nimport type { CapabilityEnvelope, EvidenceBundle, EvidenceRef, Finding, GateOutcome } from \"../autonomy/contracts.ts\";\nimport { createEvidenceBundle } from \"./evidence-bundle.ts\";\nimport { evaluateResearchRequest } from \"./research-gate.ts\";\n\n/**\n * Pure orchestration for one autonomous research pass: gate -> bounded isolated completion ->\n * parse -> evidence bundle. The model executor is injected so this stays provider-free and\n * session-free; production wires `AgentSession.runIsolatedCompletion` in.\n *\n * The lane is read-only by construction: the executor receives text prompts only, and the output\n * is an `EvidenceBundle` whose model-synthesized findings are marked untrusted.\n */\n\n/** Static across calls so callers can use `cacheRetention: \"short\"` and only pay for the variable tail. */\nexport const RESEARCH_LANE_SYSTEM_PROMPT = [\n\t\"You are a read-only research lane for a coding agent.\",\n\t\"You receive a research query plus bounded context and produce findings that help satisfy open goal requirements.\",\n\t\"Respond with STRICT JSON only - no prose, no markdown fences:\",\n\t'{\"findings\":[{\"summary\":\"<one concrete, actionable finding>\",\"confidence\":<0..1>}]}',\n\t\"Base findings only on the provided context. Never invent file paths, APIs, or facts.\",\n].join(\"\\n\");\n\nexport interface ResearchCompletion {\n\ttext: string;\n\tcostUsd: number;\n\tstopReason: string;\n}\n\nexport interface ResearchRunnerOptions {\n\tquery: string;\n\t/** Bounded, pre-redacted context handed to the research model (goal text, open requirements). */\n\tcontext?: string;\n\t/** Stripped research envelope - never the foreground/architect envelope. */\n\tenvelope: CapabilityEnvelope;\n\t/** Budget for this pass; a post-hoc breach marks the run budget_exhausted (spend stays visible). */\n\tmaxUsd: number;\n\tmaxSources: number;\n\tmaxFindings: number;\n\t/** Wall-clock budget in milliseconds; 0 disables. */\n\tmaxWallClockMs: number;\n\t/** Executes one isolated completion. Production: AgentSession.runIsolatedCompletion. */\n\tcomplete: (args: { systemPrompt: string; userPrompt: string; signal?: AbortSignal }) => Promise<ResearchCompletion>;\n\t/** External cancellation (e.g. session disposal). */\n\tsignal?: AbortSignal;\n}\n\nexport type ResearchRunStatus = \"succeeded\" | \"failed\" | \"canceled\" | \"timeout\" | \"budget_exhausted\";\n\nexport interface ResearchRunResult {\n\tstatus: ResearchRunStatus;\n\treasonCode: string;\n\tgateOutcome: GateOutcome;\n\tbundle?: EvidenceBundle;\n\tcostUsd: number;\n}\n\nexport function buildResearchUserPrompt(args: { query: string; context?: string; maxFindings: number }): string {\n\tconst parts = [`Research query: ${args.query}`];\n\tif (args.context && args.context.length > 0) {\n\t\tparts.push(\"\", \"Context:\", args.context);\n\t}\n\tparts.push(\"\", `Return at most ${args.maxFindings} findings.`);\n\treturn parts.join(\"\\n\");\n}\n\nexport interface ParsedResearchFindings {\n\tfindings: Array<{ summary: string; confidence?: number }>;\n}\n\nexport function parseResearchFindings(text: string, maxFindings: number): ParsedResearchFindings | 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\n\tfor (const candidate of candidates) {\n\t\tlet parsed: unknown;\n\t\ttry {\n\t\t\tparsed = JSON.parse(candidate);\n\t\t} catch {\n\t\t\tcontinue;\n\t\t}\n\t\tif (!parsed || typeof parsed !== \"object\" || Array.isArray(parsed)) continue;\n\t\tconst findingsRaw = (parsed as { findings?: unknown }).findings;\n\t\tif (!Array.isArray(findingsRaw)) continue;\n\n\t\tconst findings: Array<{ summary: string; confidence?: number }> = [];\n\t\tfor (const item of findingsRaw) {\n\t\t\tif (!item || typeof item !== \"object\" || Array.isArray(item)) continue;\n\t\t\tconst summary = (item as { summary?: unknown }).summary;\n\t\t\tif (typeof summary !== \"string\" || summary.trim().length === 0) continue;\n\t\t\tconst confidenceRaw = (item as { confidence?: unknown }).confidence;\n\t\t\tconst confidence =\n\t\t\t\ttypeof confidenceRaw === \"number\" && Number.isFinite(confidenceRaw)\n\t\t\t\t\t? Math.min(Math.max(confidenceRaw, 0), 1)\n\t\t\t\t\t: undefined;\n\t\t\tfindings.push({ summary: summary.trim(), confidence });\n\t\t\tif (findings.length >= maxFindings) break;\n\t\t}\n\t\t// A well-formed-but-empty findings array is a valid \"nothing found\"; a findings array whose\n\t\t// every item is malformed is not.\n\t\tif (findings.length > 0 || findingsRaw.length === 0) {\n\t\t\treturn { findings };\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction truncateExcerpt(text: string, maxChars: number): string {\n\tif (text.length <= maxChars) return text;\n\treturn `${text.slice(0, Math.max(0, maxChars - 1))}…`;\n}\n\nfunction buildBundle(options: ResearchRunnerOptions, parsed: ParsedResearchFindings): EvidenceBundle {\n\tconst contextRef: EvidenceRef = {\n\t\tid: \"src-context\",\n\t\tkind: \"user\",\n\t\ttitle: \"Goal/context provided to the research lane\",\n\t\ttrusted: true,\n\t\texcerpt: truncateExcerpt(options.context && options.context.length > 0 ? options.context : options.query, 2000),\n\t};\n\tconst synthesisRef: EvidenceRef = {\n\t\tid: \"src-synthesis\",\n\t\tkind: \"tool\",\n\t\ttitle: \"Research-model synthesis\",\n\t\ttrusted: false,\n\t};\n\tconst sources = [contextRef, synthesisRef].slice(0, Math.max(1, options.maxSources));\n\tconst findings: Finding[] = parsed.findings.slice(0, options.maxFindings).map((finding, index) => ({\n\t\tid: `finding-${index + 1}`,\n\t\tsummary: finding.summary,\n\t\tevidenceIds: [synthesisRef.id],\n\t\t...(finding.confidence !== undefined ? { confidence: finding.confidence } : {}),\n\t}));\n\treturn createEvidenceBundle({ query: options.query, sources, findings });\n}\n\nexport async function runResearch(options: ResearchRunnerOptions): Promise<ResearchRunResult> {\n\tconst gateOutcome = evaluateResearchRequest({\n\t\tenvelope: options.envelope,\n\t\tsourceKind: \"tool\",\n\t\testimatedUsd: options.maxUsd,\n\t});\n\tif (gateOutcome.outcome !== \"allow\") {\n\t\t// Skip-and-record, never prompt: gate denials inform diagnostics instead of blocking anything.\n\t\tconst status: ResearchRunStatus = gateOutcome.reasonCode === \"over_budget\" ? \"budget_exhausted\" : \"failed\";\n\t\treturn { status, reasonCode: gateOutcome.reasonCode, gateOutcome, costUsd: 0 };\n\t}\n\n\tconst bounded = await runBoundedCompletion({\n\t\tmaxWallClockMs: options.maxWallClockMs,\n\t\tsignal: options.signal,\n\t\texecute: (signal) =>\n\t\t\toptions.complete({\n\t\t\t\tsystemPrompt: RESEARCH_LANE_SYSTEM_PROMPT,\n\t\t\t\tuserPrompt: buildResearchUserPrompt(options),\n\t\t\t\tsignal,\n\t\t\t}),\n\t});\n\tif (bounded.failure) {\n\t\treturn {\n\t\t\tstatus: bounded.failure.status,\n\t\t\treasonCode: bounded.failure.reasonCode,\n\t\t\tgateOutcome,\n\t\t\tcostUsd: bounded.completion?.costUsd ?? 0,\n\t\t};\n\t}\n\tconst completion = bounded.completion;\n\tif (!completion) {\n\t\treturn { status: \"failed\", reasonCode: \"completion_error\", gateOutcome, costUsd: 0 };\n\t}\n\tif (completion.stopReason === \"error\" || completion.stopReason === \"aborted\") {\n\t\treturn { status: \"failed\", reasonCode: \"model_error\", gateOutcome, costUsd: completion.costUsd };\n\t}\n\n\tconst parsed = parseResearchFindings(completion.text, options.maxFindings);\n\tif (!parsed) {\n\t\treturn { status: \"failed\", reasonCode: \"unparseable_output\", gateOutcome, costUsd: completion.costUsd };\n\t}\n\n\tconst bundle = buildBundle(options, parsed);\n\tconst overBudget = options.maxUsd > 0 && completion.costUsd > options.maxUsd;\n\treturn {\n\t\tstatus: overBudget ? \"budget_exhausted\" : \"succeeded\",\n\t\treasonCode: overBudget\n\t\t\t? \"cost_budget_exceeded\"\n\t\t\t: parsed.findings.length === 0\n\t\t\t\t? \"no_findings\"\n\t\t\t\t: \"research_completed\",\n\t\tgateOutcome,\n\t\tbundle,\n\t\tcostUsd: completion.costUsd,\n\t};\n}\n"]}