@mutagent/cli 0.1.146 → 0.1.147

package/dist/bin/cli.js

@@ -863,12 +863,24 @@ class SDKClientWrapper {
       this.handleError(error);
     }
   }
+  async getOptimizationScorecard(jobId) {
+    try {
+      const res = await this.request(`/api/optimization/${jobId}/results`);
+      if (res.scorecard?.rendered) {
+        return { rendered: res.scorecard.rendered, data: res.scorecard.data ?? [] };
+      }
+      return null;
+    } catch {
+      return null;
+    }
+  }
   async getOptimizationResults(jobId) {
     try {
       const job = await this.request(`/api/optimization/${jobId}`);
       const progress = await this.request(`/api/optimization/${jobId}/progress`);
       const prompt = await this.getPrompt(String(job.promptId ?? ""));
       const statesRes = await this.request(`/api/optimization/${jobId}/states`).catch(() => ({ states: [] }));
+      const scorecardDigest = await this.getOptimizationScorecard(jobId);
       const latestState = statesRes.states[statesRes.states.length - 1];
       const rawState = latestState?.state ?? {};
       const iterCtx = rawState.iterationContext ?? rawState.current?.context;
@@ -895,7 +907,8 @@ class SDKClientWrapper {
         datasetResults: extracted.datasetResults,
         failureModes: extracted.failureModes,
         mutations: extracted.mutations,
-        evaluationDetails: extracted.evaluationDetails
+        evaluationDetails: extracted.evaluationDetails,
+        ...scorecardDigest ? { scorecard: scorecardDigest } : {}
       };
     } catch (error) {
       this.handleError(error);
@@ -5253,7 +5266,7 @@ function createWatchClient(opts) {
     }
     if (msg.type !== "event" || !msg.event)
       return;
-    const { eventType, iteration, stage, data } = msg.event;
+    const { eventType, iteration, stage, data, scorecardPayload } = msg.event;
     switch (eventType) {
       case "stage:completed":
         opts.onStageComplete(stage ?? "unknown", iteration ?? 0, data?.stageResult ?? data ?? { type: "unknown" });
@@ -5271,6 +5284,11 @@ function createWatchClient(opts) {
         opts.onError(new Error(data?.error ?? "Optimization job failed"));
         close();
         break;
+      case "scorecard:update":
+        if (scorecardPayload && opts.onScorecardUpdate) {
+          opts.onScorecardUpdate(scorecardPayload);
+        }
+        break;
       default:
         break;
     }
@@ -5824,6 +5842,30 @@ async function startWatchStream(jobId, isJson, maxIterations, baselineScore) {
           console.error(chalk14.red(`✗ Watch error: ${error.message}`));
         }
         reject(error);
+      },
+      onScorecardUpdate: (payload) => {
+        if (payload.error) {
+          if (isJson) {
+            console.log(JSON.stringify({
+              type: "scorecard.update",
+              iteration: payload.iteration,
+              error: payload.error
+            }));
+          } else {
+            console.warn(chalk14.yellow(`⚠ Scorecard unavailable for iteration ${String(payload.iteration)} (${payload.error.type})`));
+          }
+          return;
+        }
+        if (isJson) {
+          console.log(JSON.stringify({
+            type: "scorecard.update",
+            iteration: payload.iteration,
+            scorecard: payload.scorecard
+          }));
+        } else {
+          console.log(payload.renderedScorecard);
+          console.log("");
+        }
       }
     });
     client.connect().catch(reject);
@@ -6370,7 +6412,12 @@ After viewing results:
         output.output({ ...resultData, _links: { optimizer: optimizerLink(resultsPromptId, jobId) }, _directive: directive });
         echoDirectiveToStderr(directive);
       } else {
-        renderScorecard(resultData);
+        const serverScorecard = resultData.scorecard;
+        if (serverScorecard?.rendered) {
+          console.log(serverScorecard.rendered);
+        } else {
+          renderScorecard(resultData);
+        }
         const jobData = resultData.job;
         const isCompleted = jobData?.status === "completed";
         if (options.diff) {
@@ -9978,6 +10025,227 @@ Use the delimiter field to:
 - [concepts/eval-criteria.md](./eval-criteria.md) → MVC (Minimum Viable Context) — uses delimiter to enumerate input params
 - Source: \`mutagent-cli/src/lib/explorer.ts\` → \`inferPromptVariables()\` and \`DiscoveredPrompt.delimiter\`
 - Tests: \`mutagent-cli/src/__tests__/lib/explorer.test.ts\`
+`,
+  "concepts/scorecard-output.md": `---
+name: mutagent-cli-concepts-scorecard-output
+description: |
+  Per-iteration structured scorecard emitted by the optimizer.
+  Covers the ScorecardData shape, where agents see it (watch vs results),
+  how to consume NDJSON streams, cost-gate patterns, and criterion drill-down.
+triggers:
+  - "scorecard"
+  - "scorecard output"
+  - "optimize scorecard"
+  - "scorecard data"
+  - "iteration scorecard"
+  - "nextAction"
+  - "stop-stagnation"
+  - "cumulativeUsd"
+  - "optimizer results json"
+  - "watch optimizer"
+---
+
+# Concept — Scorecard Output
+
+> Per-iteration structured scorecard emitted by the optimizer, available
+> progressively via \`--watch\` and as a post-hoc digest via \`optimize results\`.
+
+## What it is
+
+Every completed optimizer iteration produces a \`ScorecardData\` record: a
+structured snapshot of scores, criterion pass-rates, cumulative cost, and the
+optimizer's own next-action decision. The scorecard is the primary signal
+surface for agents monitoring or reacting to a running optimization job.
+
+Two delivery modes:
+
+- **Progressive (streaming)** — emitted once per completed iteration over the
+  WebSocket event bus while the job runs.
+- **Post-hoc (batch)** — the full collection of per-iteration scorecards
+  returned in the \`/api/optimization/:id/results\` response after the job
+  finishes (or to inspect a paused job).
+
+---
+
+## Where an agent sees it
+
+### 1. \`optimize start --watch --json\`
+
+Streams NDJSON to stdout while the job runs. Each iteration emits one line:
+
+\`\`\`
+{ "type": "scorecard.update", "iteration": 1, "scorecard": { ...ScorecardData } }
+\`\`\`
+
+Other line types (e.g. \`job.started\`, \`stage.completed\`) may appear in the
+stream — branch on \`type === "scorecard.update"\` to isolate scorecard events.
+
+### 2. \`optimize watch <id> --json\`
+
+Attaches to an already-running job and streams the same NDJSON shape. Useful
+when the agent started the job in a previous session or on behalf of a
+background process.
+
+### 3. \`optimize results <id> --json\`
+
+Returns the complete result digest after job completion:
+
+\`\`\`json
+{
+  "job": { "id": "...", "status": "completed", ... },
+  "prompt": { ... },
+  "scorecard": {
+    "rendered": "<ASCII scorecard string>",
+    "data": [ ...ScorecardData[] ]
+  }
+}
+\`\`\`
+
+The \`scorecard.data\` array contains one entry per completed iteration in
+chronological order. \`scorecard.rendered\` is the ASCII terminal render of the
+final iteration (same bytes as the terminal output), included as a convenience
+for agents that want to surface a human-readable summary without re-rendering.
+
+---
+
+## \`ScorecardData\` shape
+
+Reproduced from
+\`mutagent/src/framework/metatuner/output/scorecard.ts\`:
+
+\`\`\`typescript
+interface ScorecardData {
+  jobId: string;
+  iteration: number;
+  totalIterations: number | null;  // null when maxIterations not configured
+  timestamp: string;               // ISO-8601
+  stage: string;                   // always "result-analysis" for completed iterations
+
+  scores: {
+    bestScore: number;             // highest score seen across all iterations so far
+    currentScore: number;          // score for this iteration
+    targetScore: number | null;    // convergence threshold; null if not set
+  };
+
+  criteria: Array<{
+    name: string;                  // criterion name from the evaluation rubric
+    scoreAvg: number;              // mean LLM score for this criterion this iteration
+    passRate: number;              // fraction of dataset items passing (0.0–1.0)
+  }>;
+
+  costs: {
+    iterationUsd: number;          // LLM cost for this iteration only
+    cumulativeUsd: number;         // total cost from job start through this iteration
+  };
+
+  durations: {
+    iterationMs: number;           // wall-clock duration of this iteration
+    stageBreakdown: Record<string, number>; // token counts keyed by stage name
+  };
+
+  nextAction:
+    | 'continue'          // optimizer will run another iteration
+    | 'stop-converged'    // currentScore >= targetScore
+    | 'stop-budget'       // cost or token budget exhausted
+    | 'stop-stagnation'   // stagnationCount >= patience threshold
+    | 'stop-max-iter';    // iteration count reached maxIterations
+}
+\`\`\`
+
+Key fields at a glance:
+
+| Field | Why it matters |
+|---|---|
+| \`scores.bestScore\` | The best the optimizer has achieved — compare to \`targetScore\` to estimate progress |
+| \`scores.currentScore\` | This iteration's score — use to detect regressions |
+| \`scores.targetScore\` | Convergence threshold; \`null\` means run until \`maxIterations\` |
+| \`criteria[].passRate\` | Per-criterion health — values below \`0.5\` flag a specific rubric as blocking |
+| \`costs.cumulativeUsd\` | Running cost — gate against any user-approved budget |
+| \`nextAction\` | The optimizer's own routing decision — surface \`stop-stagnation\` proactively |
+
+---
+
+## How an agent should use it
+
+### Per-iteration monitoring (streaming)
+
+While consuming \`optimize start --watch --json\` or \`optimize watch <id> --json\`:
+
+1. **Score trend** — compare \`scores.currentScore\` against the previous
+   iteration's \`currentScore\`. A sustained downward trend (3+ iterations)
+   warrants surfacing to the user before the budget is spent.
+
+2. **\`nextAction\` gate** — if \`nextAction === "stop-stagnation"\`, the optimizer
+   has detected a plateau and is about to halt. Surface this to the user with
+   the current \`scores.bestScore\` and \`costs.cumulativeUsd\` so they can decide
+   whether to continue with a higher patience budget or accept the result.
+
+3. **Cost gate** — if \`costs.cumulativeUsd\` crosses a threshold the user
+   approved for, alert immediately. The optimizer does not know about
+   user-defined soft budgets; the agent is the enforcement layer.
+
+### Post-hoc analysis (\`optimize results\`)
+
+1. **Criterion drill-down** — iterate \`scorecard.data\` (all iterations) and
+   collect \`criteria[].passRate\` per criterion. Any criterion with a mean
+   \`passRate < 0.5\` across iterations did not benefit from optimization and
+   should be flagged back to the user as potentially mis-specified or
+   conflicting with another criterion.
+
+2. **Best-iteration identification** — find the entry where
+   \`scores.currentScore === Math.max(...data.map(d => d.scores.currentScore))\`.
+   This is the iteration the optimizer treats as \`bestIteration\`. Compare its
+   \`criteria\` snapshot to the final iteration to detect regressions in
+   individual rubrics even when the composite score improved.
+
+3. **Cost/iteration trade-off** — divide \`costs.cumulativeUsd\` by
+   \`iteration\` to get average cost per iteration. Present this when asking the
+   user whether to re-run with more iterations.
+
+---
+
+## Parsing example
+
+Minimal Node.js / Bun snippet — consume NDJSON from \`optimize start --watch --json\`
+and branch on \`nextAction\`:
+
+\`\`\`js
+import { spawn } from 'node:child_process';
+import * as readline from 'node:readline';
+
+const proc = spawn('mutagent', ['optimize', 'start', JOB_ID, '--watch', '--json']);
+const rl = readline.createInterface({ input: proc.stdout });
+
+rl.on('line', (line) => {
+  let event;
+  try { event = JSON.parse(line); } catch { return; }
+  if (event.type !== 'scorecard.update') return;
+
+  const { scorecard } = event;
+  const { scores, costs, criteria, nextAction } = scorecard;
+
+  console.log(\`Iter \${scorecard.iteration}: score=\${scores.currentScore.toFixed(3)} cost=$\${costs.cumulativeUsd.toFixed(4)}\`);
+
+  if (nextAction === 'stop-stagnation') {
+    // Surface to user — optimizer is about to halt without converging
+    console.warn(\`[ALERT] Optimizer stagnated at score \${scores.bestScore}. Consider raising patience.\`);
+  }
+
+  const weakCriteria = criteria.filter(c => c.passRate < 0.5).map(c => c.name);
+  if (weakCriteria.length > 0) {
+    console.warn(\`[ALERT] Low-pass criteria: \${weakCriteria.join(', ')}\`);
+  }
+});
+\`\`\`
+
+---
+
+## Cross-references
+
+- [concepts/eval-criteria.md](./eval-criteria.md) — how evaluation criteria are
+  defined; \`criteria[].name\` in \`ScorecardData\` maps to \`name\` in the rubric.
+- [workflows/optimization.md](../workflows/optimization.md) — full optimization
+  loop; the scorecard is produced at Step 8 (watch) and Step 9 (results).
 `,
   "workflows/agents.md": `---
 name: mutagent-cli-workflows-agents
@@ -12018,5 +12286,5 @@ if (isInteractive && !isSkillCommand) {
 }
 program.parse();
 
-//# debugId=7B63A61BDC73994664756E2164756E21
+//# debugId=D625C9A316600B1764756E2164756E21
 //# sourceMappingURL=cli.js.map