fix: fail closed on unparsed review output

CHANGELOG.md

@@ -6,6 +6,46 @@
 
 - **Human `/plan` respects `maxChildrenPerTask` but not `dailyLimit` (AGT-4123).** `POST /api/plan/dispatch` now refuses with `decomposition_child_cap` when an approved plan would leave more children than `autonomous.decomposition.maxChildrenPerTask` on the new parent (sharing `refuseForChildCap` with the autonomous path). The automation-pacing `dailyLimit` / `reserveDailyCreations` budget remains runner-only — an explicitly confirmed plan is not refused because the daemon already spent today's slots.
 
+## 0.24.1 — 2026-09-20
+
+### Added
+
+- **Cost-capped GitHub Action reviews.** `openswarm review --model <id>` is now
+  exposed by the CLI and the composite action forwards its optional `model`
+  input through the same reviewed call chain. The action refuses as
+  gate-not-run when an older installed package lacks the option, rather than
+  silently falling back to an uncontrolled default model. The dogfood workflow
+  pins `deepseek/deepseek-v4-flash`.
+- **Backfill for pre-publication parked branches (AGT-4124).** `openswarm
+  drain-parks --path <repo>` finds unowned `NEEDS_HUMAN` rows, creates or
+  reuses their draft PRs, and records zero-ahead or missing-branch reasons with
+  CAS-protected ledger updates. `--dry-run` performs no GitHub or ledger write.
+- **Resource-aware test fan-out.** Test parallelism now scales with available
+  CPU and memory instead of unconditionally consuming the host.
+
+### Changed
+
+- **OpenSwarm is documented as an agentic CI/CD gate first.** The README now
+  leads with review → fix → re-review → deterministic verification and presents
+  the Linear daemon as an optional orchestration layer.
+- **Fleet supervision exposes thinking effort.** The supervisor can configure
+  reasoning effort for every role instead of treating it as an inaccessible
+  model default.
+
+### Fixed
+
+- **Reproducible package installation on every supported platform.** LanceDB is
+  pinned to the last stable release whose declared macOS, Linux, and Windows
+  optional native packages all exist in the registry, so `npm ci` no longer
+  rejects the committed lockfile.
+- **OpenRouter flash routing fails over to GLM.** A failing or unavailable
+  `deepseek/deepseek-v4-flash` review can promote to `z-ai/glm-5.3-flash` under
+  the existing fallback policy.
+- **Durable run state has stronger recovery evidence.** Session transcripts,
+  watchdog causes, staged memory promotion, write scopes, and parked
+  publication records now survive the retries and handoffs that previously
+  obscured them.
+
 
 ## 0.24.0 — 2026-09-10
 

action.yml

@@ -219,18 +219,41 @@ runs:
         if [ -n "${REVIEW_MODEL:-}" ]; then ARGS+=(--model "$REVIEW_MODEL"); fi
         if [ "${READ_ONLY:-true}" = "true" ]; then ARGS+=(--read-only); fi
 
-        # stdout is the JSON document; diagnostics go to stderr and stay on the
-        # job log.
+        # The CLI's JSON contract puts diagnostics on stderr. A dependency can
+        # still emit a late stdout diagnostic from a background task, so parse
+        # the one review document into a clean runner-temp file before trusting
+        # it for outputs or the exit contract below.
         set +e
         openswarm "${ARGS[@]}" > "$REVIEW_JSON"
         STATUS=$?
         set -e
 
         DECISION=""
         GATE_RAN="false"
-        if [ -s "$REVIEW_JSON" ]; then
-          DECISION=$(REVIEW_JSON="$REVIEW_JSON" node -e 'try{process.stdout.write(String(JSON.parse(require("fs").readFileSync(process.env.REVIEW_JSON,"utf8")).decision??""))}catch{}')
-          GATE_RAN=$(REVIEW_JSON="$REVIEW_JSON" node -e 'try{process.stdout.write(String(JSON.parse(require("fs").readFileSync(process.env.REVIEW_JSON,"utf8")).gateRan??false))}catch{process.stdout.write("false")}')
+        PARSED_REVIEW_JSON="$RUNNER_TEMP/openswarm-review-parsed.json"
+        if [ -s "$REVIEW_JSON" ] && REVIEW_JSON="$REVIEW_JSON" PARSED_REVIEW_JSON="$PARSED_REVIEW_JSON" node -e '
+          const fs = require("fs");
+          const raw = fs.readFileSync(process.env.REVIEW_JSON, "utf8");
+          let review;
+          try { review = JSON.parse(raw); } catch {
+            // A review document is emitted at the start of a line. Keep the
+            // first complete JSON document with the required gate fields,
+            // rather than treating a diagnostic prefix as a successful pass.
+            for (const match of raw.matchAll(/^[ \t]*\{/gm)) {
+              try {
+                const candidate = JSON.parse(raw.slice(match.index));
+                if (typeof candidate?.gateRan === "boolean") { review = candidate; break; }
+              } catch {}
+            }
+          }
+          if (!review || typeof review.gateRan !== "boolean") process.exit(1);
+          fs.writeFileSync(process.env.PARSED_REVIEW_JSON, JSON.stringify(review));
+        '; then
+          REVIEW_JSON="$PARSED_REVIEW_JSON"
+          DECISION=$(REVIEW_JSON="$REVIEW_JSON" node -e 'process.stdout.write(String(JSON.parse(require("fs").readFileSync(process.env.REVIEW_JSON,"utf8")).decision??""))')
+          GATE_RAN=$(REVIEW_JSON="$REVIEW_JSON" node -e 'process.stdout.write(String(JSON.parse(require("fs").readFileSync(process.env.REVIEW_JSON,"utf8")).gateRan))')
+        elif [ -s "$REVIEW_JSON" ]; then
+          echo "::warning::OpenSwarm produced output but no parseable review document"
         fi
 
         echo "decision=$DECISION" >> "$GITHUB_OUTPUT"
@@ -278,7 +301,7 @@ runs:
         # an unknown option, a crash, output that is not the JSON document — is
         # a gate that did not run, and calling that a reject both misreports the
         # change and hides the real failure.
-        if [ "$STATUS" != "0" ] && [ "$GATE_RAN" != "true" ]; then
+        if [ "$GATE_RAN" != "true" ]; then
           echo "::error::OpenSwarm exited $STATUS without producing a verdict — the gate did not run"
           if [ "$FAIL_ON_GATE_NOT_RUN" = "true" ]; then exit 2; fi
           exit 0

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@intrect/openswarm",
-  "version": "0.24.0",
+  "version": "0.24.1",
   "description": "Autonomous AI agent orchestrator — Claude, GPT, Codex, and local models (Ollama/LMStudio/llama.cpp)",
   "license": "MIT",
   "type": "module",
@@ -50,7 +50,7 @@
     "@huggingface/transformers": "^4.2.0",
     "@inquirer/prompts": "^8.5.2",
     "@intrect/cxt": "^0.3.0",
-    "@lancedb/lancedb": "^0.23.0",
+    "@lancedb/lancedb": "0.22.3",
     "@linear/sdk": "^19.0.0",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@types/better-sqlite3": "^7.6.13",