docs(examples): update supervised workflows

Document --supervised usage and update demo workflows/scripts to run review via OPENCODE and completion_check via CODEX_CLI, plus fix todo-loop and todo-check grep pattern.

Author: reoring

docs/guide/daemon.md

@@ -642,6 +642,7 @@ bun run src/daemon/cli.ts <daemon.yaml> [options]
 |-----------|------|
 | `--workspace`, `-w` | override `workspace` via CLI |
 | `--verbose`, `-v` | enable verbose logging |
+| `--supervised` | run workflows via Core IPC (Supervisor -> Core -> Worker) |
 | `--help`, `-h` | show help |
 
 Notes:
@@ -661,10 +662,12 @@ bun run src/daemon/cli.ts my-daemon.yaml --verbose
 On `Ctrl+C` (SIGINT) or SIGTERM, Daemon performs a graceful shutdown:
 
 1. stop receiving new events
-2. stop all event sources
-3. wait for running workflows to finish (30s timeout)
-4. stop webhook server
-5. flush state and exit
+2. abort running workflows (best-effort)
+3. stop all event sources
+4. wait for running workflows to finish (30s timeout)
+5. stop webhook server
+6. (supervised mode) shut down Core
+7. flush state and exit
 
 ---
 

docs/guide/workflow.md

@@ -478,6 +478,7 @@ bun run src/workflow/run.ts <workflow.yaml>
 |-----------|--------|------|----------|
 | `--workspace <dir>` | `-w` | working directory | temp directory |
 | `--verbose` | `-v` | show step output | off |
+| `--supervised` | - | delegate steps via Core IPC (Supervisor -> Core -> Worker) | off |
 | `--help` | `-h` | show help | - |
 
 ### Examples
@@ -491,6 +492,9 @@ bun run src/workflow/run.ts examples/build-test-report.yaml --workspace /tmp/my-
 
 # verbose output
 bun run src/workflow/run.ts examples/todo-loop.yaml --verbose
+
+# supervised mode (Core spawns worker processes)
+bun run src/workflow/run.ts examples/agent-pr-loop.yaml --supervised --verbose
 ```
 
 ### Reading results

examples/agent-pr-loop-demo/run-in-tmp.sh

@@ -88,6 +88,7 @@ fi
 RUNNER_ARGS=(
   "${AGENTCORE_ROOT}/src/workflow/run.ts"
   "${AGENTCORE_ROOT}/examples/agent-pr-loop.yaml"
+  --supervised
   --workspace "${TARGET}"
 )
 

examples/agent-pr-loop.yaml

@@ -8,7 +8,7 @@
 #   AGENTCORE_ROOT=/path/to/agentcore
 #   TARGET=/path/to/your/repo
 #   AGENTCORE_ROOT="$AGENTCORE_ROOT" bun run --cwd "$AGENTCORE_ROOT" ./src/workflow/run.ts \
-#     "$AGENTCORE_ROOT/examples/agent-pr-loop.yaml" --workspace "$TARGET" --verbose
+#     "$AGENTCORE_ROOT/examples/agent-pr-loop.yaml" --workspace "$TARGET" --supervised --verbose
 
 name: agent-pr-loop
 version: "1"
@@ -140,21 +140,67 @@ steps:
 
   review_fix_loop:
     description: "Review by OpenCode; if fixes needed, apply via Codex CLI; repeat until PASS"
-    worker: CUSTOM
+    worker: OPENCODE
+    model: "openai/gpt-5.2"
     depends_on: [implement]
     instructions: |
-      bun run "${AGENTCORE_ROOT:-.}/scripts/agent-pr-loop/review-fix-once.ts"
-    capabilities: [READ, EDIT, RUN_TESTS, RUN_COMMANDS]
-    timeout: "30m"
+      You are OpenCode.
+
+      Task:
+      - Review the work against the request.
+      - Write/update:
+        - .agentcore-loop/review.md
+        - .agentcore-loop/review.verdict (PASS or FAIL only)
+      - If verdict is FAIL, also write/update:
+        - .agentcore-loop/fix.md (non-empty; concrete, actionable steps; include file paths)
+
+      Before reviewing, generate these review inputs (workspace is a git repo):
+      - mkdir -p .agentcore-loop
+      - BASE_BRANCH=$(cat .agentcore-loop/base-branch.txt 2>/dev/null || echo main)
+      - git fetch origin "$BASE_BRANCH" >/dev/null 2>&1 || true
+      - if git show-ref --verify --quiet "refs/remotes/origin/$BASE_BRANCH"; then BASE_REF="origin/$BASE_BRANCH"; elif git show-ref --verify --quiet "refs/heads/$BASE_BRANCH"; then BASE_REF="$BASE_BRANCH"; else BASE_REF="HEAD"; fi
+      - git diff --no-color "$BASE_REF" > .agentcore-loop/review.diff || true
+      - git status --porcelain=v1 > .agentcore-loop/review.status || true
+      - git ls-files --others --exclude-standard > .agentcore-loop/review.untracked || true
+
+      You must read at least:
+      - .agentcore-loop/request.md
+      - .agentcore-loop/design.md (if present)
+      - .agentcore-loop/todo.md (if present)
+      - .agentcore-loop/review.diff
+      - .agentcore-loop/review.status
+      - .agentcore-loop/review.untracked
+
+      Format for .agentcore-loop/review.md:
+      - Sections: Summary, Strengths, Issues, Verification
+    capabilities: [READ, EDIT, RUN_COMMANDS]
+    timeout: "15m"
     on_failure: retry
     max_retries: 2
 
     completion_check:
-      worker: CUSTOM
+      worker: CODEX_CLI
+      model: "gpt-5.3-codex"
       instructions: |
-        bash "${AGENTCORE_ROOT:-.}/scripts/agent-pr-loop/verdict-check.sh"
-      capabilities: [READ, RUN_COMMANDS]
-      timeout: "30s"
+        You are Codex CLI.
+
+        Read:
+        - .agentcore-loop/review.verdict
+        - .agentcore-loop/fix.md (if present)
+        - .agentcore-loop/todo.md (if present)
+
+        Task:
+        - If .agentcore-loop/review.verdict is PASS:
+          - Print EXACTLY "COMPLETE" on its own line as the final line and exit successfully.
+        - If verdict is FAIL:
+          - Apply the fixes described in .agentcore-loop/fix.md (minimal changes).
+          - Run tests (bun test if present; otherwise choose the most appropriate command).
+          - Print EXACTLY "INCOMPLETE" on its own line as the final line and exit successfully.
+
+        Output rules:
+        - Always ensure the last non-empty line is either COMPLETE or INCOMPLETE.
+      capabilities: [READ, EDIT, RUN_TESTS, RUN_COMMANDS]
+      timeout: "20m"
 
     max_iterations: 5
     on_iterations_exhausted: abort

examples/todo-loop.yaml

@@ -29,11 +29,13 @@ steps:
     depends_on: [setup]
     instructions: |
       # Find and process the first incomplete task
-      TASK=$(grep -m1 '^\- \[ \]' todo.txt || true)
-      if [ -z "$TASK" ]; then
+      MATCH=$(grep -n -m1 '^\- \[ \]' todo.txt || true)
+      if [ -z "$MATCH" ]; then
         echo "All tasks complete!"
         exit 0
       fi
+      LINE_NO=${MATCH%%:*}
+      TASK=${MATCH#*:}
       echo "Processing: $TASK"
 
       # Perform work based on the task
@@ -53,9 +55,7 @@ steps:
       esac
 
       # Mark the task as completed
-      ESCAPED=$(echo "$TASK" | sed 's/[[\]/\\&]/\\&/g')
-      DONE=$(echo "$TASK" | sed 's/\- \[ \]/- [x]/')
-      sed -i "s|${TASK}|${DONE}|" todo.txt
+      sed -i "${LINE_NO}s/^- \\[ \\] /- [x] /" todo.txt
       echo "Updated todo.txt:"
       cat todo.txt
     capabilities: [READ, EDIT, RUN_COMMANDS]

scripts/agent-pr-loop/todo-check.sh

@@ -4,7 +4,7 @@ set -euo pipefail
 TODO=.agentcore-loop/todo.md
 test -s "${TODO}" || { echo "CHECK: missing todo"; exit 1; }
 
-COUNT=$(grep -c '^\- \[ \] ' "${TODO}" || true)
+COUNT=$(grep -c '^- \[ \] ' "${TODO}" || true)
 if [ "${COUNT}" -lt 6 ]; then
   echo "CHECK: too few tasks: ${COUNT}"
   exit 1