instar 1.2.57 → 1.2.59

package/.claude/skills/autonomous/hooks/autonomous-stop-hook.sh

@@ -146,6 +146,7 @@ DURATION_SECONDS=$(fm_get duration_seconds)
 STARTED_AT=$(fm_get started_at)
 COMPLETION_PROMISE=$(fm_get completion_promise)
 COMPLETION_CONDITION=$(fm_get completion_condition)
+GOAL_MODE=$(fm_get goal_mode)   # "native" = the framework's own /goal loop drives completion
 
 # Validate recorded session_id is a real UUID. Claude sometimes writes a custom
 # string instead of $CLAUDE_CODE_SESSION_ID; non-UUID values are treated as
@@ -228,6 +229,36 @@ if [[ "$OWNER" != "true" ]]; then
   exit 0
 fi
 
+# ── Native /goal delegation: the framework's own /goal loop owns completion. ──
+# instar injected "/goal <condition>" at start (goal_mode=native), so we DEFER the
+# continue/stop decision to native /goal's own Stop hook (we approve/exit; its hook
+# blocks until the condition is met — block wins over approve, so it stays in control).
+# instar still owns its terminal STOP concerns (emergency-stop, duration) and enforces
+# them by CLEARING native /goal first (inject "/goal clear" via the server).
+if [[ "$GOAL_MODE" == "native" ]]; then
+  native_goal_clear() {
+    local port auth
+    port=$(python3 -c "import json;print(json.load(open('.instar/config.json')).get('port',4040))" 2>/dev/null || echo 4040)
+    auth=$(python3 -c "import json;print(json.load(open('.instar/config.json')).get('authToken',''))" 2>/dev/null || echo "")
+    jq -nc --arg t "$REPORT_TOPIC" '{topicId:$t}' \
+      | curl -s -m 10 -H "Authorization: Bearer $auth" -H 'Content-Type: application/json' \
+        --data-binary @- "http://localhost:${port}/autonomous/native-goal/clear" >/dev/null 2>&1 || true
+  }
+  if [[ -f ".instar/autonomous-emergency-stop" ]]; then
+    native_goal_clear; rm -f "$STATE_FILE"
+    echo "[autonomous] emergency stop — native /goal cleared" >&2; exit 0
+  fi
+  if [[ "$DURATION_SECONDS" =~ ^[0-9]+$ ]] && [[ $DURATION_SECONDS -gt 0 ]]; then
+    NG_START=$(date -u -j -f "%Y-%m-%dT%H:%M:%SZ" "$STARTED_AT" +%s 2>/dev/null || date -d "$STARTED_AT" +%s 2>/dev/null || echo 0)
+    if [[ "$NG_START" =~ ^[0-9]+$ ]] && [[ $NG_START -gt 0 ]] && [[ $(( $(date +%s) - NG_START )) -ge $DURATION_SECONDS ]]; then
+      native_goal_clear; rm -f "$STATE_FILE"
+      echo "[autonomous] duration expired — native /goal cleared" >&2; exit 0
+    fi
+  fi
+  # Not terminal → let native /goal decide completion. Approve (its hook keeps control).
+  exit 0
+fi
+
 # ── This IS the autonomous session. Terminal checks first. ────────────
 
 # Validate iteration

package/.claude/skills/autonomous/scripts/setup-autonomous.sh

@@ -197,6 +197,31 @@ To complete, ALL of these must be true:
 - Then output: <promise>$COMPLETION_PROMISE</promise>
 EOF
 
+# ── Native /goal delegation: where the framework has a native /goal loop, hand the
+# condition to it (instar injects "/goal <condition>" via the server → SessionManager
+# send-keys) and mark goal_mode:native so the stop-hook defers completion to it. Only
+# with a condition + a per-topic job + Claude Code >= 2.1.139. Best-effort: on any miss
+# we leave goal_mode empty and instar's own independent evaluator drives (Phase 1).
+if [[ -n "$COMPLETION_CONDITION" ]] && [[ -n "$REPORT_TOPIC" ]]; then
+  CLAUDE_VER=$(claude --version 2>/dev/null | grep -oE '[0-9]+\.[0-9]+\.[0-9]+' | head -1 || echo "")
+  NATIVE_GOAL_OK="false"
+  if [[ -n "$CLAUDE_VER" ]]; then
+    # /goal requires Claude Code >= 2.1.139.
+    if [[ "$(printf '%s\n2.1.139\n' "$CLAUDE_VER" | sort -V | head -1)" == "2.1.139" ]]; then
+      NATIVE_GOAL_OK="true"
+    fi
+  fi
+  if [[ "$NATIVE_GOAL_OK" == "true" ]]; then
+    NG_PORT=$(python3 -c "import json;print(json.load(open('.instar/config.json')).get('port',4040))" 2>/dev/null || echo 4040)
+    NG_AUTH=$(python3 -c "import json;print(json.load(open('.instar/config.json')).get('authToken',''))" 2>/dev/null || echo "")
+    jq -nc --arg t "$REPORT_TOPIC" --arg c "$COMPLETION_CONDITION" '{topicId:$t,condition:$c}' \
+      | curl -s -m 8 -H "Authorization: Bearer $NG_AUTH" -H 'Content-Type: application/json' \
+        --data-binary @- "http://localhost:${NG_PORT}/autonomous/native-goal/set" >/dev/null 2>&1 \
+      && echo "  Native /goal: handed condition to Claude Code's /goal loop" \
+      || echo "  Native /goal: unavailable — using instar's own completion evaluator"
+  fi
+fi
+
 echo "🔄 Autonomous mode activated!"
 echo ""
 echo "Goal: $GOAL"

package/dist/commands/server.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"server.d.ts","sourceRoot":"","sources":["../../src/commands/server.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAqQH,UAAU,YAAY;IACpB,UAAU,CAAC,EAAE,OAAO,CAAC;IACrB,GAAG,CAAC,EAAE,MAAM,CAAC;IACb;2DACuD;IACvD,QAAQ,CAAC,EAAE,OAAO,CAAC;CACpB;AAiqDD,wBAAsB,WAAW,CAAC,OAAO,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAoyLtE;AAED,wBAAsB,UAAU,CAAC,OAAO,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,OAAO,CAAC,IAAI,CAAC,CAsDzE;AAED;;;;;;;;;;;GAWG;AACH,wBAAsB,aAAa,CAAC,OAAO,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,OAAO,CAAC,IAAI,CAAC,CAuD5E"}
+{"version":3,"file":"server.d.ts","sourceRoot":"","sources":["../../src/commands/server.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAsQH,UAAU,YAAY;IACpB,UAAU,CAAC,EAAE,OAAO,CAAC;IACrB,GAAG,CAAC,EAAE,MAAM,CAAC;IACb;2DACuD;IACvD,QAAQ,CAAC,EAAE,OAAO,CAAC;CACpB;AAiqDD,wBAAsB,WAAW,CAAC,OAAO,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CA0zLtE;AAED,wBAAsB,UAAU,CAAC,OAAO,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,OAAO,CAAC,IAAI,CAAC,CAsDzE;AAED;;;;;;;;;;;GAWG;AACH,wBAAsB,aAAa,CAAC,OAAO,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,OAAO,CAAC,IAAI,CAAC,CAuD5E"}

package/dist/commands/server.js

@@ -91,6 +91,7 @@ import { pickupGitSyncMessages } from '../messaging/GitSyncTransport.js';
 import { DeliveryRetryManager } from '../messaging/DeliveryRetryManager.js';
 import { SpawnRequestManager } from '../messaging/SpawnRequestManager.js';
 import { ThreadlineRouter } from '../threadline/ThreadlineRouter.js';
+import { resolveThreadlineMcpEntry } from '../threadline/mcpEntry.js';
 import { ThreadResumeMap } from '../threadline/ThreadResumeMap.js';
 import { ListenerSessionManager } from '../threadline/ListenerSessionManager.js';
 import { SystemReviewer } from '../monitoring/SystemReviewer.js';
@@ -2116,7 +2117,17 @@ export async function startServer(options) {
                 console.log(pc.yellow(`  Git sync setup: ${err instanceof Error ? err.message : String(err)}`));
             }
         }
-        const sessionManager = new SessionManager(config.sessions, state);
+        // Per-agent Codex threadline MCP override. Codex reads a SHARED
+        // ~/.codex/config.toml whose [mcp_servers."threadline"] is last-writer-wins
+        // across every codex agent on the machine — so a codex worker could load a
+        // DIFFERENT agent's threadline identity and its threadline_send would be
+        // misaddressed. Pin this agent's own entry per-spawn (the launch builders
+        // emit `-c mcp_servers.threadline.*`). Only when threadline is configured;
+        // ignored by non-codex launches. See CODEX-MULTIAGENT-THREADLINE-SPEC.
+        const codexThreadlineMcp = config.threadline
+            ? resolveThreadlineMcpEntry(config.sessions.projectDir, config.stateDir, config.projectName)
+            : undefined;
+        const sessionManager = new SessionManager(codexThreadlineMcp ? { ...config.sessions, codexThreadlineMcp } : config.sessions, state);
         // Input Guard is constructed later (after sharedIntelligence is available)
         // so the topic coherence reviewer can route through the IntelligenceProvider
         // abstraction instead of calling Anthropic directly.
@@ -5970,6 +5981,12 @@ export async function startServer(options) {
                     // §4.5: honor SpawnRequestManager's provenance tag so drain-spawned
                     // sessions are distinguishable from inline-spawned ones in logs/stream.
                     triggeredBy: opts?.triggeredBy ?? 'spawn-request',
+                    // This is the Threadline inbound-reply spawn: the worker must call
+                    // the threadline_send MCP tool to reply, which a codex worker can only
+                    // do under full bypass (codex cancels MCP calls in any sandbox). Jobs
+                    // do NOT set this and stay sandboxed. Bounded: Threadline only accepts
+                    // messages from trusted agents.
+                    codexAllowMcpTools: true,
                 });
                 return session.id;
             },
@@ -6160,6 +6177,9 @@ export async function startServer(options) {
                     minIqsBand: pipeConfig?.minIqsBand ?? 70,
                     framework: pipeFramework,
                     binaryPath: pipeBinaryPath,
+                    // Same per-agent codex threadline MCP override as SessionManager, so a
+                    // codex pipe-reply worker uses THIS agent's threadline MCP.
+                    ...(codexThreadlineMcp ? { codexThreadlineMcp } : {}),
                 });
                 console.log(pc.dim(`  Pipe sessions: enabled (model: ${pipeConfig?.model ?? 'sonnet'}, max: ${pipeConfig?.maxConcurrent ?? 5})`));
             }