@sagentlab/navarch-runtime 0.1.26 → 0.1.28

package/README.md

@@ -420,8 +420,10 @@ overrides (`codex exec --model ... -c model_reasoning_effort=...`,
 `claude -p --model ... --effort ...`, `gemini --model ...`, or `opencode run
 --model ... --variant ...`) and records the effective model, profile, and
 effort on completion. Gemini currently uses its `auto` model default and does
-not expose a reasoning-effort flag. OpenCode accepts `default` to preserve the
-authenticated account's selection, or any `provider/model` reference.
+not expose a reasoning-effort flag. OpenCode accepts any `provider/model`
+reference (the platform default is `opencode/x-preview-f-free`, Ox Alpha Free
+(Unlimited) on OpenCode Zen), or `default` to preserve the authenticated
+account's selection.
 Machine-wide extra arguments still configure other CLI behavior; dispatched
 model policy wins.
 

package/dist/session.cjs

@@ -46,6 +46,55 @@ const log = (0, logger_cjs_1.createLogger)("session");
  * isolation instead; see runtime/README.md.
  */
 async function runSession(deps, claimed, sessionId) {
+    const lifecycle = { enteredMainSessionTry: false };
+    try {
+        return await runClaimedSession(deps, claimed, sessionId, lifecycle);
+    }
+    catch (err) {
+        if (lifecycle.enteredMainSessionTry)
+            throw err;
+        // Initialization before the main worktree/adapter try (prompt creation,
+        // broker access, MCP/guard setup) must still release the lease with the
+        // explicit pre-start signal. Otherwise it waits for expiry and bypasses
+        // the repeat/backoff policy entirely.
+        const { api, config } = deps;
+        const { lease_id: leaseId, task, context_bundle: bundle } = claimed;
+        const runtime = claimed.runtime ?? config.agentType;
+        const execution = bundle.execution ?? {
+            profile: task.execution_profile ?? "standard",
+            model: runtime === "codex"
+                ? "gpt-5.6-sol"
+                : runtime === "gemini"
+                    ? "auto"
+                    : runtime === "opencode"
+                        ? "default"
+                        : "best",
+            reasoning_effort: "medium",
+        };
+        const crashDetail = describeCompletionError(err).slice(0, 1800);
+        const failureSummary = `Session crashed: ${crashDetail}`;
+        log.error(`session ${leaseId} threw before adapter start: ${crashDetail}`);
+        await api
+            .completeLease(leaseId, {
+            status: "failed",
+            report: verificationFailureReport(task.task_type, failureSummary),
+            failure_summary: failureSummary,
+            failed_before_start: true,
+            evidence_urls: [],
+            cost: { tokens_in: 0, tokens_out: 0, cost_usd: 0 },
+            exit_status: "crashed",
+            agent_type: runtime,
+            model: execution.model,
+            execution_profile: execution.profile,
+            reasoning_effort: execution.reasoning_effort,
+        })
+            .catch((completeErr) => log.warn(`complete() after initialization crash also failed: ${String(completeErr)}`));
+        await node_fs_1.promises
+            .rm(node_path_1.default.join(config.workspaceRoot, "sessions", sessionId), { recursive: true, force: true })
+            .catch(() => undefined);
+    }
+}
+async function runClaimedSession(deps, claimed, sessionId, lifecycle) {
     const { api, config } = deps;
     const { lease_id: leaseId, task, context_bundle: bundle } = claimed;
     const runtime = claimed.runtime ?? config.agentType;
@@ -111,6 +160,7 @@ async function runSession(deps, claimed, sessionId) {
             status: "failed",
             report: verificationFailureReport(task.task_type, failureSummary),
             failure_summary: failureSummary,
+            failed_before_start: true,
             evidence_urls: [],
             cost: { tokens_in: 0, tokens_out: 0, cost_usd: 0 },
             exit_status: "crashed",
@@ -175,16 +225,15 @@ async function runSession(deps, claimed, sessionId) {
         heartbeatInFlight = request;
         return request;
     };
-    const heartbeatTimer = setInterval(() => void pollLease(), config.leaseHeartbeatIntervalMs);
     const dockerAvailable = config.sandboxMode === "docker" && (await (0, sandbox_cjs_1.isDockerAvailable)());
     if (config.sandboxMode === "docker" && !dockerAvailable) {
-        clearInterval(heartbeatTimer);
         log.warn("Docker requested but not available on this machine — failing the task instead of crashing.");
         await api
             .completeLease(leaseId, {
             status: "failed",
             report: verificationFailureReport(task.task_type, "Docker sandbox unavailable on this machine."),
             failure_summary: "Docker sandbox unavailable on this machine.",
+            failed_before_start: true,
             evidence_urls: [],
             cost: { tokens_in: 0, tokens_out: 0, cost_usd: 0 },
             exit_status: "crashed",
@@ -277,6 +326,9 @@ async function runSession(deps, claimed, sessionId) {
     // Kept outside the try so the crash path can report the telemetry this
     // session actually produced instead of zeroing it.
     let lastCompletion = null;
+    let adapterStarted = false;
+    const heartbeatTimer = setInterval(() => void pollLease(), config.leaseHeartbeatIntervalMs);
+    lifecycle.enteredMainSessionTry = true;
     try {
         await gitWorktree.prepare();
         if (sandbox) {
@@ -307,6 +359,7 @@ async function runSession(deps, claimed, sessionId) {
                     : promptText);
             nextPrompt = null;
             await node_fs_1.promises.writeFile(node_path_1.default.join(workDir, "prompt.md"), runPrompt, "utf8");
+            adapterStarted = true;
             const turnResult = await adapter.run({
                 prompt: runPrompt,
                 mcpConfigPath,
@@ -498,6 +551,7 @@ async function runSession(deps, claimed, sessionId) {
                 ? `${lastCompletion.report}\n\n---\n\n${failureSummary}`
                 : failureSummary),
             failure_summary: failureSummary,
+            failed_before_start: !adapterStarted,
             evidence_urls: lastCompletion?.evidence_urls ?? [],
             cost: lastCompletion?.cost ?? { tokens_in: 0, tokens_out: 0, cost_usd: 0 },
             ...(lastCompletion?.transcript_url
@@ -571,6 +625,10 @@ const REMEDIABLE_REJECTION_CODES = new Set([
     // the file usually already exists in the worktree, so the remediation turn
     // only has to call register_artifact and PUT the upload.
     "artifact_required",
+    // A registered render whose duration misses the brief's target_duration:
+    // the source assets are still in the worktree, so a remediation turn can
+    // re-cut to length, re-register the artifact, and retry on the same lease.
+    "duration_mismatch",
     "pr_not_ready",
     "review_attestation_required",
     "review_attestation_invalid",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sagentlab/navarch-runtime",
-  "version": "0.1.26",
+  "version": "0.1.28",
   "description": "Navarch machine-side session manager: claims delivery tasks and runs them through Claude Code, Codex, Gemini, or OpenCode.",
   "type": "commonjs",
   "license": "MIT",