opencode-usage-coach 0.13.2 → 0.13.4

package/dist/cli.js

@@ -350,7 +350,8 @@ var DEFAULT_HARNESS_CONFIG = {
   generator: "opencode/deepseek-v4-flash-free",
   grader: "opencode/mimo-v2.5-free",
   provider: "",
-  lighterModel: ""
+  lighterModel: "",
+  maxSteps: 30
 };
 function resolveAgentSourceFile() {
   const scriptDir = dirname(fileURLToPath(import.meta.url));

package/dist/index.js

@@ -434,6 +434,9 @@ async function searchContext(query, frameworks, keyDeps, timeoutMs) {
 var PLUGIN_NAME = "opencode-usage-coach";
 var TTL_MS = Number(process.env.UC_TTL_MS ?? 6e4);
 var DEFAULT_MAX_STEPS = Number(process.env.UC_MAX_STEPS ?? 30) || 30;
+function resolveMaxSteps(cfg, explicit) {
+  return explicit ?? cfg.maxSteps ?? DEFAULT_MAX_STEPS;
+}
 var WATCHDOG_POLL_MS = Math.max(1e3, Number(process.env.UC_WATCHDOG_POLL_MS ?? 3e3) || 3e3);
 var WALL_TIMEOUT_MS = Math.max(1, Number(process.env.UC_WALL_TIMEOUT_MIN ?? 30) || 30) * 60 * 1e3;
 var DEFAULT_MAX_QUESTIONS = Math.max(1, Math.round(Number(process.env.UC_MAX_QUESTIONS ?? 7)) || 7);
@@ -1821,7 +1824,7 @@ async function UsageCoachPlugin(input) {
         const agent = await resolveAgent(input.client, _input.sessionID);
         currentAgent = agent;
         refreshBackground();
-        const harnessTools = ["unknown_scan", "question", "generate", "generate_batch", "grade", "investigate", "verify_diagnosis", "generalize", "harness_start", "task_update", "harness_done", "record_failure", "reverse_interview"];
+        const harnessTools = ["unknown_scan", "question", "generate", "generate_batch", "grade", "grade_batch", "investigate", "verify_diagnosis", "generalize", "harness_start", "task_update", "harness_done", "record_failure", "reverse_interview"];
         if (!harnessTools.includes(_input.tool)) return;
         if (!isHarnessAgent(agent)) {
           throw new Error(`[${PLUGIN_NAME}] '${_input.tool}' is restricted to agent mode ${JSON.stringify(HARNESS_AGENTS)} (current: ${JSON.stringify(agent || "unknown")}). Switch to that agent mode to use it.`);
@@ -1885,13 +1888,16 @@ DETERMINISTIC LOOP \u2014 first classify the tasks:
   INDEPENDENT = task B does NOT need task A's output  ->  use PATH A (parallel, faster)
   DEPENDENT   = task B needs task A's output          ->  use PATH B (sequential)
 
-PATH A \u2014 INDEPENDENT (parallel via generate_batch):
+PATH A \u2014 INDEPENDENT (parallel via generate_batch + grade_batch):
   1. task_update(1..${args.total}, title, "generating")
   2. generate_batch({tasks: [{id:1, prompt:"Task: <title1>. Perform it."}, ...]})  -> all results + NEXT
-  3. for each i: task_update(i, title, "grading") + grade({prompt:"Evaluate... PASS/FAIL first line. Task: <title>"})  -> verdict + NEXT
-  4. for each i: PASS -> task_update(i, title, "completed", "PASS"); FAIL -> revise (up to 2x) or task_update(i, title, "failed", "FAIL")
+  3. task_update(1..${args.total}, title, "grading")
+  4. grade_batch({tasks: [{id:1, prompt:"Evaluate... PASS/FAIL first line. Task: <title1>"}, ...]})  -> all verdicts + per-task NEXT
+  5. for each i: PASS -> task_update(i, title, "completed", "PASS"); FAIL -> revise (up to 2x) or task_update(i, title, "failed", "FAIL")
+  Note: grade_batch runs all grades IN PARALLEL \u2014 much faster than grading one-by-one.
+  After revisions, you can grade_batch the revised tasks together too.
 
-PATH B \u2014 DEPENDENT (sequential):
+PATH B \u2014 DEPENDENT (sequential \u2014 one task at a time):
   Optional: task_update(1..N, title, "pending")  \u2190 pre-register all tasks first
   for i in 1..${args.total}:
     1. task_update(i, title, "generating")
@@ -1899,6 +1905,7 @@ PATH B \u2014 DEPENDENT (sequential):
     3. task_update(i, title, "grading")
     4. grade(...)  -> verdict + NEXT
     5. PASS -> task_update(i, title, "completed", "PASS"); FAIL -> revise (up to 2x) or failed
+  Note: use single grade() here \u2014 tasks are dependent, grading is one at a time.
 
 Then: harness_done(). Follow the [usage-coach NEXT] directive each tool returns. Do NOT improvise the sequence.`;
           }
@@ -2336,7 +2343,7 @@ ${gate.summary}
 ` + prefix;
             }
             const genTaskId = findActiveTaskId(ctx.sessionID, "generating");
-            const maxSteps = args.max_steps ?? DEFAULT_MAX_STEPS;
+            const maxSteps = resolveMaxSteps(cfg, args.max_steps);
             const out = await runModel(
               input.client,
               model,
@@ -2397,7 +2404,7 @@ ${gate.summary}
             const limit = decision === "THROTTLE" ? 2 : args.tasks.length;
             const rules = readRules();
             const priorNotes = readImplNotes(5);
-            const maxSteps = args.max_steps ?? DEFAULT_MAX_STEPS;
+            const maxSteps = resolveMaxSteps(cfg, args.max_steps);
             const gate = checkScanGate(ctx.sessionID);
             const gatePrefix = gate.warning ? `${gate.warning}
 
@@ -2521,6 +2528,89 @@ The next generate call will automatically include the new rule.`;
             return out + "\n" + next;
           }
         }),
+        grade_batch: tool({
+          description: "Run the GRADER model on MULTIPLE tasks IN PARALLEL. Quota-aware: GO = full parallel; THROTTLE = concurrency capped at 2; STOP = refused. Use after generate_batch returns results for all INDEPENDENT tasks \u2014 grade them all at once instead of one-by-one. Each result includes a per-task PASS/FAIL verdict + NEXT directive. Resilient: failed grades are retried sequentially (once).",
+          args: {
+            tasks: tool.schema.array(tool.schema.object({
+              id: tool.schema.number().describe("Task ID (must match the task_update ID)."),
+              prompt: tool.schema.string().describe("The grading prompt for this task (same format as grade()).")
+            })).describe("One entry per task to grade. All run in parallel on GO.")
+          },
+          async execute(args, ctx) {
+            const cfg = readHarnessCfg(ctx.directory);
+            const model = cfg.grader ?? cfg.generator;
+            if (!model) {
+              return 'ERROR: no grader/generator model configured. Set "grader" or "generator" in harness.config.json.';
+            }
+            let decision = "GO";
+            try {
+              decision = current().decision;
+            } catch {
+            }
+            if (decision === "STOP") {
+              return 'ERROR: quota STOP \u2014 halt the harness loop now. Call task_update(current, "halted_quota") and stop.';
+            }
+            const limit = decision === "THROTTLE" ? 2 : args.tasks.length;
+            const gradeOne = async (t) => {
+              const gradeTaskId = findActiveTaskId(ctx.sessionID, "grading");
+              const out = await runModel(
+                input.client,
+                model,
+                t.prompt,
+                ctx.directory,
+                gradeTaskId ? { sessionID: ctx.sessionID, taskId: gradeTaskId } : void 0
+              );
+              let verdict = "FAIL";
+              if (!out.startsWith("ERROR:")) {
+                const f = (out.split("\n").find((l) => l.trim()) ?? "").trim();
+                if (/^pass\b/i.test(f)) verdict = "PASS";
+                else if (/^fail\b/i.test(f)) verdict = "FAIL";
+              }
+              const next = verdict === "PASS" ? `[usage-coach NEXT] task ${t.id}: PASS -> task_update(${t.id}, title, "completed", "PASS"), then proceed.` : `[usage-coach NEXT] task ${t.id}: FAIL -> if revisions < 2: task_update(${t.id}, title, "revising") + generate({prompt:"Apply feedback:\\n${out.slice(0, 500)}\\nTask: {title}"}); else: record_failure + investigate + verify + generalize, then task_update(${t.id}, title, "failed", "FAIL").`;
+              return { id: t.id, result: `[task ${t.id}] ${verdict}
+${out}
+${next}` };
+            };
+            const results = [];
+            const failed = [];
+            for (let i = 0; i < args.tasks.length; i += limit) {
+              const batch = args.tasks.slice(i, i + limit);
+              const settled = await Promise.allSettled(batch.map((t) => gradeOne(t)));
+              for (let j = 0; j < settled.length; j++) {
+                const s = settled[j];
+                const t = batch[j];
+                if (s.status === "fulfilled") {
+                  results.push(s.value.result);
+                } else {
+                  const err = String(s.reason ?? "unknown rejection");
+                  log(`grade_batch task ${t.id} REJECTED: ${err}`);
+                  failed.push({ id: t.id, prompt: t.prompt, error: err });
+                }
+              }
+            }
+            if (failed.length > 0) {
+              log(`grade_batch: retrying ${failed.length} failed grade(s) sequentially`);
+              for (const f of failed) {
+                try {
+                  const r = await gradeOne(f);
+                  results.push(r.result.replace(`[task ${r.id}]`, `[task ${r.id}] (retry)`));
+                } catch (e2) {
+                  const err2 = String(e2 ?? "unknown rejection on retry");
+                  log(`grade_batch task ${f.id} retry ALSO FAILED: ${err2}`);
+                  results.push(`[task ${f.id}] FAIL
+(ERROR: grade failed on both batch and retry: ${err2})
+[usage-coach NEXT] task ${f.id}: grade error \u2014 re-run grade() individually for this task.`);
+                }
+              }
+            }
+            const throttleNote = decision === "THROTTLE" ? `
+[usage-coach] quota THROTTLE \u2014 grader concurrency capped at ${limit}.` : "";
+            const passCount = results.filter((r) => /\[task \d+\] PASS/.test(r)).length;
+            const summary = `
+[usage-coach] Graded ${results.length} task(s): ${passCount} PASS, ${results.length - passCount} FAIL.`;
+            return results.join("\n\n") + summary + throttleNote;
+          }
+        }),
         reverse_interview: tool({
           description: "Reverse interview: identify ambiguities in the task and ask the user one question at a time, highest design-impact first. Call WITHOUT answer to start or get the next question. Call WITH answer to record the user's response and advance. Returns the next question or a completion summary. ALWAYS present the question to the user verbatim \u2014 do NOT answer it yourself.",
           args: {
@@ -2656,21 +2746,23 @@ If the task is already well-specified with no significant ambiguities, return {"
           }
         }),
         coach_config: tool({
-          description: 'View or update harness model configuration (generator, grader, lighterModel, provider). Call with no args to view current config. Pass any combination of generator/grader/lighterModel/provider to update. Example: coach_config({ generator: "anthropic/claude-sonnet-4-20250514", grader: "opencode/mimo-v2.5-free" })',
+          description: 'View or update harness model configuration (generator, grader, lighterModel, provider, maxSteps). Call with no args to view current config. Pass any combination of fields to update. Example: coach_config({ generator: "anthropic/claude-sonnet-4-20250514", grader: "opencode/mimo-v2.5-free", maxSteps: 15 })',
           args: {
             generator: tool.schema.string().optional(),
             grader: tool.schema.string().optional(),
             lighterModel: tool.schema.string().optional(),
-            provider: tool.schema.string().optional()
+            provider: tool.schema.string().optional(),
+            maxSteps: tool.schema.number().optional().describe("Max steps per generate/grade sub-session (default 30). Lower = faster but may timeout on complex tasks.")
           },
           async execute(args, ctx) {
             const dir = ctx?.directory ?? input.directory;
             const current2 = readHarnessCfg(dir);
-            const hasUpdates = args.generator !== void 0 || args.grader !== void 0 || args.lighterModel !== void 0 || args.provider !== void 0;
+            const hasUpdates = args.generator !== void 0 || args.grader !== void 0 || args.lighterModel !== void 0 || args.provider !== void 0 || args.maxSteps !== void 0;
             if (!hasUpdates) {
               const envOverrides = [];
               if (process.env.UC_PROVIDER) envOverrides.push(`UC_PROVIDER=${process.env.UC_PROVIDER}`);
               if (process.env.UC_LIGHTER_MODEL) envOverrides.push(`UC_LIGHTER_MODEL=${process.env.UC_LIGHTER_MODEL}`);
+              if (process.env.UC_MAX_STEPS) envOverrides.push(`UC_MAX_STEPS=${process.env.UC_MAX_STEPS}`);
               return [
                 `Harness Configuration`,
                 `\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550`,
@@ -2678,6 +2770,7 @@ If the task is already well-specified with no significant ambiguities, return {"
                 `  grader:        ${current2.grader ?? "(defaults to generator)"}`,
                 `  lighterModel:  ${current2.lighterModel ?? "(not set \u2014 no THROTTLE fallback)"}`,
                 `  provider:      ${current2.provider ?? "(auto-detected from model)"}`,
+                `  maxSteps:      ${current2.maxSteps ?? 30}  (env: UC_MAX_STEPS)`,
                 `\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550`,
                 envOverrides.length > 0 ? `
 Environment overrides (take precedence):
@@ -2685,7 +2778,7 @@ Environment overrides (take precedence):
                 `
 Config file: ~/.config/opencode-usage-coach/harness.config.json`,
                 `
-To update, call: coach_config({ generator: "provider/model-id", ... })`
+To update, call: coach_config({ generator: "provider/model-id", maxSteps: 15, ... })`
               ].filter(Boolean).join("\n");
             }
             const updates = {};
@@ -2693,6 +2786,7 @@ To update, call: coach_config({ generator: "provider/model-id", ... })`
             if (args.grader !== void 0) updates.grader = args.grader.trim();
             if (args.lighterModel !== void 0) updates.lighterModel = args.lighterModel.trim();
             if (args.provider !== void 0) updates.provider = args.provider.trim();
+            if (args.maxSteps !== void 0) updates.maxSteps = args.maxSteps;
             const writtenPath = writeHarnessCfg(updates);
             const updated = readHarnessCfg(dir);
             return [
@@ -2702,6 +2796,7 @@ To update, call: coach_config({ generator: "provider/model-id", ... })`
               `  grader:        ${updated.grader ?? "(defaults to generator)"}`,
               `  lighterModel:  ${updated.lighterModel ?? "(not set)"}`,
               `  provider:      ${updated.provider ?? "(auto-detected)"}`,
+              `  maxSteps:      ${updated.maxSteps ?? 30}`,
               `\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550`,
               `
 Saved to: ${writtenPath}`,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-usage-coach",
-  "version": "0.13.2",
+  "version": "0.13.4",
   "description": "opencode closed-loop usage coach — quota SENSE -> coaching DECIDE -> loop ACT + TUI integration",
   "type": "module",
   "main": "./dist/index.js",