prizmkit 1.1.74 → 1.1.77

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (29) hide show
  1. package/bundled/VERSION.json +3 -3
  2. package/bundled/dev-pipeline/.env.example +1 -0
  3. package/bundled/dev-pipeline/lib/common.sh +53 -0
  4. package/bundled/dev-pipeline/lib/heartbeat.sh +29 -6
  5. package/bundled/dev-pipeline/run-bugfix.sh +5 -0
  6. package/bundled/dev-pipeline/run-feature.sh +5 -0
  7. package/bundled/dev-pipeline/run-recovery.sh +6 -0
  8. package/bundled/dev-pipeline/run-refactor.sh +5 -0
  9. package/bundled/dev-pipeline/scripts/parse-stream-progress.py +229 -7
  10. package/bundled/dev-pipeline-windows/lib/common.ps1 +23 -0
  11. package/bundled/dev-pipeline-windows/lib/heartbeat.sh +439 -0
  12. package/bundled/dev-pipeline-windows/lib/pipeline.ps1 +7 -0
  13. package/bundled/dev-pipeline-windows/run-recovery.ps1 +11 -0
  14. package/bundled/dev-pipeline-windows/scripts/parse-stream-progress.py +229 -7
  15. package/bundled/skills/_metadata.json +1 -1
  16. package/bundled/skills/app-planner/SKILL.md +13 -10
  17. package/bundled/skills/app-planner/references/rules/backend/question-manifest.json +46 -0
  18. package/bundled/skills/app-planner/references/rules/database/question-manifest.json +39 -0
  19. package/bundled/skills/app-planner/references/rules/frontend/question-manifest.json +51 -0
  20. package/bundled/skills/app-planner/references/rules/mobile/question-manifest.json +47 -0
  21. package/bundled/skills/bugfix-pipeline-launcher/SKILL.md +1 -0
  22. package/bundled/skills/feature-pipeline-launcher/SKILL.md +1 -0
  23. package/bundled/skills/refactor-pipeline-launcher/SKILL.md +1 -0
  24. package/bundled/skills-windows/app-planner/SKILL.md +13 -10
  25. package/bundled/skills-windows/app-planner/references/rules/backend/question-manifest.json +46 -0
  26. package/bundled/skills-windows/app-planner/references/rules/database/question-manifest.json +39 -0
  27. package/bundled/skills-windows/app-planner/references/rules/frontend/question-manifest.json +51 -0
  28. package/bundled/skills-windows/app-planner/references/rules/mobile/question-manifest.json +47 -0
  29. package/package.json +1 -1
@@ -1,5 +1,5 @@
1
1
  {
2
- "frameworkVersion": "1.1.74",
3
- "bundledAt": "2026-06-12T13:22:51.065Z",
4
- "bundledFrom": "8817522"
2
+ "frameworkVersion": "1.1.77",
3
+ "bundledAt": "2026-06-14T15:13:42.592Z",
4
+ "bundledFrom": "2327b8b"
5
5
  }
@@ -44,6 +44,7 @@
44
44
  # STALE_KILL_THRESHOLD=900 # Auto-kill after N seconds without parent log progress (0 = disabled)
45
45
  # CODEX_WAIT_STALE_KILL_THRESHOLD=3600 # Longer no-log window while Codex waits on subagents
46
46
  # CODEX_SUBAGENT_TIMEOUT_SECONDS=3300 # Codex subagent max runtime; defaults to wait threshold - 300
47
+ # CB_SUBAGENT_STALE_KILL_THRESHOLD=3600 # Longer no-log window while CodeBuddy sub-agents are running
47
48
  # LOG_CLEANUP_ENABLED=1 # Periodic log cleanup (1=on, 0=off)
48
49
  # LOG_RETENTION_DAYS=14 # Delete logs older than N days
49
50
  # LOG_MAX_TOTAL_MB=1024 # Keep total logs under N MB via oldest-first cleanup
@@ -338,6 +338,41 @@ prizm_detect_cli_and_platform() {
338
338
  export PRIZMKIT_PLATFORM="$PLATFORM"
339
339
  }
340
340
 
341
+ # Validate PRIZMKIT_EFFORT against the detected platform's supported values.
342
+ # Exits with a friendly error if the value is unsupported.
343
+ prizm_validate_effort() {
344
+ [[ -n "${PRIZMKIT_EFFORT:-}" ]] || return 0
345
+
346
+ local session_platform
347
+ session_platform="$(_prizm_session_platform)"
348
+
349
+ local allowed_values
350
+ case "$session_platform" in
351
+ claude) allowed_values="low, medium, high, xhigh, max" ;;
352
+ codex) allowed_values="low, medium, high, xhigh" ;;
353
+ *) allowed_values="low, medium, high, xhigh" ;;
354
+ esac
355
+
356
+ local valid=0
357
+ local val
358
+ for val in $allowed_values; do
359
+ val="${val%,}"
360
+ [[ "$val" == "$PRIZMKIT_EFFORT" ]] && valid=1 && break
361
+ done
362
+
363
+ if [[ "$valid" -ne 1 ]]; then
364
+ local platform_label="$session_platform"
365
+ [[ "$session_platform" == "claude" ]] && platform_label="Claude Code"
366
+ [[ "$session_platform" == "codebuddy" ]] && platform_label="CodeBuddy"
367
+ [[ "$session_platform" == "codex" ]] && platform_label="Codex"
368
+
369
+ log_error "PRIZMKIT_EFFORT='$PRIZMKIT_EFFORT' is not supported by the detected CLI ($platform_label)."
370
+ log_error " Supported values for $platform_label: $allowed_values"
371
+ log_error " Set PRIZMKIT_EFFORT to one of the above, or unset it to use the CLI default."
372
+ exit 1
373
+ fi
374
+ }
375
+
341
376
  # Start an AI CLI session in the background.
342
377
  # Usage: prizm_start_ai_session <prompt_path> <log_path> <model>
343
378
  # Sets PRIZM_AI_PID to the spawned process PID. Do not call this function via
@@ -380,6 +415,9 @@ prizm_start_ai_session() {
380
415
  if [[ "$USE_STREAM_JSON" == "true" ]]; then
381
416
  claude_args+=(--output-format stream-json)
382
417
  fi
418
+ if [[ -n "${PRIZMKIT_EFFORT:-}" ]]; then
419
+ claude_args+=(--effort "$PRIZMKIT_EFFORT")
420
+ fi
383
421
  if [[ -n "$model" ]]; then
384
422
  claude_args+=(--model "$model")
385
423
  fi
@@ -399,6 +437,9 @@ prizm_start_ai_session() {
399
437
  if [[ -n "$model" ]]; then
400
438
  codex_args+=(--model "$model")
401
439
  fi
440
+ if [[ -n "${PRIZMKIT_EFFORT:-}" ]]; then
441
+ codex_args+=(--config "model_reasoning_effort=$PRIZMKIT_EFFORT")
442
+ fi
402
443
  "$CLI_CMD" "${codex_args[@]}" - < "$prompt_path" > "$log_path" 2>&1 &
403
444
  ;;
404
445
  *)
@@ -409,6 +450,9 @@ prizm_start_ai_session() {
409
450
  if [[ "$USE_STREAM_JSON" == "true" ]]; then
410
451
  cb_args+=(--output-format stream-json)
411
452
  fi
453
+ if [[ -n "${PRIZMKIT_EFFORT:-}" ]]; then
454
+ cb_args+=(--effort "$PRIZMKIT_EFFORT")
455
+ fi
412
456
  if [[ -n "$model" ]]; then
413
457
  cb_args+=(--model "$model")
414
458
  fi
@@ -900,6 +944,9 @@ prizm_run_ai_session() {
900
944
  case "$session_platform" in
901
945
  claude)
902
946
  local claude_args=(-p "$(cat "$prompt_path")" --dangerously-skip-permissions)
947
+ if [[ -n "${PRIZMKIT_EFFORT:-}" ]]; then
948
+ claude_args+=(--effort "$PRIZMKIT_EFFORT")
949
+ fi
903
950
  if [[ -n "$model" ]]; then
904
951
  claude_args+=(--model "$model")
905
952
  fi
@@ -919,10 +966,16 @@ prizm_run_ai_session() {
919
966
  if [[ -n "$model" ]]; then
920
967
  codex_args+=(--model "$model")
921
968
  fi
969
+ if [[ -n "${PRIZMKIT_EFFORT:-}" ]]; then
970
+ codex_args+=(--config "model_reasoning_effort=$PRIZMKIT_EFFORT")
971
+ fi
922
972
  "$CLI_CMD" "${codex_args[@]}" - < "$prompt_path" > "$log_path" 2>&1
923
973
  ;;
924
974
  *)
925
975
  local cb_args=(--print -y)
976
+ if [[ -n "${PRIZMKIT_EFFORT:-}" ]]; then
977
+ cb_args+=(--effort "$PRIZMKIT_EFFORT")
978
+ fi
926
979
  if [[ -n "$model" ]]; then
927
980
  cb_args+=(--model "$model")
928
981
  fi
@@ -92,8 +92,8 @@ PY
92
92
 
93
93
  local effective_stale_kill_threshold="$stale_kill_threshold"
94
94
  if [[ $stale_kill_threshold -gt 0 && -f "$progress_json" ]]; then
95
- local codex_wait_threshold
96
- codex_wait_threshold=$(python3 - "$progress_json" "$stale_kill_threshold" <<'PY' 2>/dev/null || true
95
+ local extended_threshold
96
+ extended_threshold=$(python3 - "$progress_json" "$stale_kill_threshold" <<'PY' 2>/dev/null || true
97
97
  import json
98
98
  import os
99
99
  import sys
@@ -106,14 +106,21 @@ with open(progress_path, "r", encoding="utf-8") as fh:
106
106
 
107
107
  spawn_count = 0
108
108
  for tool in progress.get("tool_calls", []):
109
- if isinstance(tool, dict) and tool.get("name") == "spawn_agent":
109
+ if isinstance(tool, dict) and tool.get("name") in ("spawn_agent", "Agent", "TaskCreate"):
110
110
  try:
111
111
  spawn_count += int(tool.get("count", 0))
112
112
  except (TypeError, ValueError):
113
113
  pass
114
114
 
115
+ # Also check the subagent_spawn_count field (set by _record_cb_agent_tool_call)
116
+ if not spawn_count:
117
+ spawn_count = int(progress.get("subagent_spawn_count", 0))
118
+
119
+ fmt = progress.get("event_format", "")
120
+
121
+ # Codex: current_tool == "wait" means parent is blocked on spawn_agent completion
115
122
  if (
116
- progress.get("event_format") == "codex-json"
123
+ fmt == "codex-json"
117
124
  and progress.get("current_tool") == "wait"
118
125
  and spawn_count > 0
119
126
  ):
@@ -124,10 +131,26 @@ if (
124
131
  wait_threshold = max(base_threshold * 4, 3600)
125
132
  if wait_threshold > base_threshold:
126
133
  print(wait_threshold)
134
+
135
+ # CodeBuddy: Agent tool blocks synchronously; Task* tools imply bg agents.
136
+ # Extend the stale window when sub-agents have been spawned so the heartbeat
137
+ # doesn't kill the parent while children are still running.
138
+ if (
139
+ fmt == "stream-json"
140
+ and spawn_count > 0
141
+ and progress.get("cb_session_id", "")
142
+ ):
143
+ configured = os.environ.get("CB_SUBAGENT_STALE_KILL_THRESHOLD", "")
144
+ try:
145
+ cb_threshold = int(configured)
146
+ except ValueError:
147
+ cb_threshold = max(base_threshold * 4, 3600)
148
+ if cb_threshold > base_threshold:
149
+ print(cb_threshold)
127
150
  PY
128
151
  )
129
- if [[ "$codex_wait_threshold" =~ ^[0-9]+$ && "$codex_wait_threshold" -gt "$stale_kill_threshold" ]]; then
130
- effective_stale_kill_threshold="$codex_wait_threshold"
152
+ if [[ "$extended_threshold" =~ ^[0-9]+$ && "$extended_threshold" -gt "$stale_kill_threshold" ]]; then
153
+ effective_stale_kill_threshold="$extended_threshold"
131
154
  fi
132
155
  fi
133
156
 
@@ -51,6 +51,7 @@ LOG_RETENTION_DAYS=${LOG_RETENTION_DAYS:-14}
51
51
  LOG_MAX_TOTAL_MB=${LOG_MAX_TOTAL_MB:-1024}
52
52
  VERBOSE=${VERBOSE:-0}
53
53
  MODEL=${MODEL:-""}
54
+ PRIZMKIT_EFFORT=${PRIZMKIT_EFFORT:-""}
54
55
  DEV_BRANCH=${DEV_BRANCH:-""}
55
56
  AUTO_PUSH=${AUTO_PUSH:-0}
56
57
  STOP_ON_FAILURE=${STOP_ON_FAILURE:-0}
@@ -59,6 +60,9 @@ ENABLE_DEPLOY=${ENABLE_DEPLOY:-0}
59
60
  # Source shared common helpers (CLI/platform detection + logs + deps)
60
61
  prizm_detect_cli_and_platform
61
62
 
63
+ # Validate PRIZMKIT_EFFORT early (fail-fast before any sessions are spawned)
64
+ prizm_validate_effort
65
+
62
66
  # Source shared heartbeat library
63
67
  source "$SCRIPT_DIR/lib/heartbeat.sh"
64
68
 
@@ -1197,6 +1201,7 @@ show_help() {
1197
1201
  echo " MAX_RETRIES Max retries per bug (default: 3)"
1198
1202
  echo " SESSION_TIMEOUT Session timeout in seconds (default: 0 = no limit)"
1199
1203
  echo " MODEL Default AI model (overridden by per-bug model in bug list)"
1204
+ echo " PRIZMKIT_EFFORT AI reasoning effort (low|medium|high|xhigh|max; max is Claude Code only)"
1200
1205
  echo " PIPELINE_MODE Default pipeline mode: lite|standard|full (overridden by --mode)"
1201
1206
  echo " ENABLE_CRITIC Enable/disable critic: true|false|1|0 (overridden by --critic/--no-critic)"
1202
1207
  echo " AI_CLI AI CLI command name (auto-detected: cbc, claude, or codex)"
@@ -54,6 +54,7 @@ LOG_RETENTION_DAYS=${LOG_RETENTION_DAYS:-14}
54
54
  LOG_MAX_TOTAL_MB=${LOG_MAX_TOTAL_MB:-1024}
55
55
  VERBOSE=${VERBOSE:-0}
56
56
  MODEL=${MODEL:-""}
57
+ PRIZMKIT_EFFORT=${PRIZMKIT_EFFORT:-""}
57
58
  DEV_BRANCH=${DEV_BRANCH:-""}
58
59
  AUTO_PUSH=${AUTO_PUSH:-0}
59
60
  STOP_ON_FAILURE=${STOP_ON_FAILURE:-0}
@@ -62,6 +63,9 @@ ENABLE_DEPLOY=${ENABLE_DEPLOY:-0}
62
63
  # Source shared common helpers (CLI/platform detection + logs + deps)
63
64
  prizm_detect_cli_and_platform
64
65
 
66
+ # Validate PRIZMKIT_EFFORT early (fail-fast before any sessions are spawned)
67
+ prizm_validate_effort
68
+
65
69
  # Source shared heartbeat library
66
70
  source "$SCRIPT_DIR/lib/heartbeat.sh"
67
71
 
@@ -1483,6 +1487,7 @@ show_help() {
1483
1487
  echo " SESSION_TIMEOUT Session timeout in seconds (default: 0 = no limit)"
1484
1488
  echo " AI_CLI AI CLI command name (auto-detected: cbc, claude, or codex)"
1485
1489
  echo " MODEL AI model ID (e.g. claude-opus-4.6, claude-sonnet-4.6, claude-haiku-4.5)"
1490
+ echo " PRIZMKIT_EFFORT AI reasoning effort (low|medium|high|xhigh|max; max is Claude Code only)"
1486
1491
  echo " HEARTBEAT_INTERVAL Heartbeat log interval in seconds (default: 30)"
1487
1492
  echo " STALE_KILL_THRESHOLD Auto-kill session after N seconds of no progress (default: 900)"
1488
1493
  echo " HEARTBEAT_STALE_THRESHOLD Heartbeat stale threshold in seconds (default: 600)"
@@ -67,11 +67,15 @@ HEARTBEAT_INTERVAL=${HEARTBEAT_INTERVAL:-30}
67
67
  STALE_KILL_THRESHOLD=${STALE_KILL_THRESHOLD:-900}
68
68
  VERBOSE=${VERBOSE:-0}
69
69
  MODEL=${MODEL:-""}
70
+ PRIZMKIT_EFFORT=${PRIZMKIT_EFFORT:-""}
70
71
  AUTO_PUSH=${AUTO_PUSH:-0}
71
72
 
72
73
  # Source shared common helpers (CLI/platform detection + logs + deps)
73
74
  prizm_detect_cli_and_platform
74
75
 
76
+ # Validate PRIZMKIT_EFFORT early (fail-fast before any sessions are spawned)
77
+ prizm_validate_effort
78
+
75
79
  # Source shared heartbeat library
76
80
  source "$SCRIPT_DIR/lib/heartbeat.sh"
77
81
 
@@ -106,6 +110,7 @@ Options (for 'run' command):
106
110
  Environment Variables:
107
111
  SESSION_TIMEOUT Session timeout in seconds (default: 0 = no limit)
108
112
  MODEL AI model to use
113
+ PRIZMKIT_EFFORT AI reasoning effort (low|medium|high|xhigh|max; max is Claude Code only)
109
114
  VERBOSE Set to 1 for verbose AI CLI output
110
115
  HEARTBEAT_INTERVAL Heartbeat log interval in seconds (default: 30)
111
116
  STALE_KILL_THRESHOLD Auto-kill after N seconds no progress (default: 900)
@@ -117,6 +122,7 @@ Examples:
117
122
  ./run-recovery.sh run --dry-run # Generate prompt, don't execute
118
123
  ./run-recovery.sh run --yes # Skip confirmation prompt
119
124
  ./run-recovery.sh run --model claude-opus-4.6
125
+ PRIZMKIT_EFFORT=high ./run-recovery.sh # Use high reasoning effort
120
126
  HELP
121
127
  }
122
128
 
@@ -52,6 +52,7 @@ LOG_RETENTION_DAYS=${LOG_RETENTION_DAYS:-14}
52
52
  LOG_MAX_TOTAL_MB=${LOG_MAX_TOTAL_MB:-1024}
53
53
  VERBOSE=${VERBOSE:-0}
54
54
  MODEL=${MODEL:-""}
55
+ PRIZMKIT_EFFORT=${PRIZMKIT_EFFORT:-""}
55
56
  DEV_BRANCH=${DEV_BRANCH:-""}
56
57
  AUTO_PUSH=${AUTO_PUSH:-0}
57
58
  STOP_ON_FAILURE=${STOP_ON_FAILURE:-0}
@@ -61,6 +62,9 @@ ENABLE_DEPLOY=${ENABLE_DEPLOY:-0}
61
62
  # Source shared common helpers (CLI/platform detection + logs + deps)
62
63
  prizm_detect_cli_and_platform
63
64
 
65
+ # Validate PRIZMKIT_EFFORT early (fail-fast before any sessions are spawned)
66
+ prizm_validate_effort
67
+
64
68
  # Source shared heartbeat library
65
69
  source "$SCRIPT_DIR/lib/heartbeat.sh"
66
70
 
@@ -1232,6 +1236,7 @@ show_help() {
1232
1236
  echo " MAX_RETRIES Max retries per refactor (default: 3)"
1233
1237
  echo " SESSION_TIMEOUT Session timeout in seconds (default: 0 = no limit)"
1234
1238
  echo " MODEL Default AI model (overridden by per-refactor model in refactor list)"
1239
+ echo " PRIZMKIT_EFFORT AI reasoning effort (low|medium|high|xhigh|max; max is Claude Code only)"
1235
1240
  echo " PIPELINE_MODE Default pipeline mode: lite|standard|full (overridden by --mode)"
1236
1241
  echo " ENABLE_CRITIC Enable/disable critic: true|false|1|0 (overridden by --critic/--no-critic)"
1237
1242
  echo " AI_CLI AI CLI command name (auto-detected: cbc, claude, or codex)"
@@ -137,7 +137,10 @@ class ProgressTracker:
137
137
  self.event_format = ""
138
138
  self.active_subagent_count = 0
139
139
  self.subagent_status_counts = Counter()
140
+ self._subagent_spawn_count = 0
140
141
  self.codex_child_thread_ids = set()
142
+ self.cb_session_id = ""
143
+ self.cb_cwd = ""
141
144
  self.claude_session_id = ""
142
145
  self.claude_cwd = ""
143
146
  self.claude_task_states = {}
@@ -147,6 +150,7 @@ class ProgressTracker:
147
150
  self.last_child_activity_at = ""
148
151
  self._codex_child_session_paths = {}
149
152
  self._claude_child_session_files = []
153
+ self._cb_child_session_files = []
150
154
  self._last_child_scan_at = 0.0
151
155
  self._last_claude_fallback_scan_at = 0.0
152
156
  self._last_claude_fallback_scan_key = ""
@@ -158,10 +162,13 @@ class ProgressTracker:
158
162
  def process_event(self, event):
159
163
  """Process a single stream-json event and update state.
160
164
 
161
- Supports two formats:
165
+ Supports three formats:
162
166
  1. Claude API raw stream: message_start, content_block_start, content_block_delta, etc.
163
167
  2. Claude Code verbose stream-json: assistant, user, tool_result, system, etc.
164
168
  (produced by claude/claude-internal --verbose --output-format stream-json)
169
+ 3. CodeBuddy stream-json: message/function_call/function_call_result events
170
+ (produced by cbc --print -y --output-format stream-json — message type
171
+ with sessionId/cwd metadata, function_call for tool invocations).
165
172
  """
166
173
  event_type = event.get("type", "")
167
174
 
@@ -233,6 +240,93 @@ class ProgressTracker:
233
240
 
234
241
  return
235
242
 
243
+ # ── CodeBuddy stream-json format ──────────────────────────────
244
+ # cbc --print -y --output-format stream-json emits:
245
+ # message (role=user/assistant), function_call, function_call_result,
246
+ # file-history-snapshot, error. The first user message carries
247
+ # sessionId and cwd metadata; later function_call items carry the
248
+ # same fields.
249
+ # NOTE: cbc emits tool invocations as EITHER tool_use blocks inside a
250
+ # message (role=assistant) event OR as standalone function_call events,
251
+ # but not both for the same invocation. The two handlers are
252
+ # complementary rather than redundant, so no deduplication is needed.
253
+ # Detect via event_type=="message" with a "sessionId" field.
254
+ # Must run before the Claude Code assistant / tool_result branches
255
+ # because those also accept those types but use different internals.
256
+ if event_type == "message" and isinstance(event.get("sessionId"), str):
257
+ self.event_format = self.event_format or "stream-json"
258
+ role = event.get("role", "")
259
+ sid = event.get("sessionId", "")
260
+ cwd = event.get("cwd", "")
261
+ if sid and not self.cb_session_id:
262
+ self.cb_session_id = sid.strip()
263
+ if cwd and not self.cb_cwd:
264
+ self.cb_cwd = cwd.strip()
265
+ if role == "assistant":
266
+ self.message_count += 1
267
+ self.is_active = True
268
+ content = event.get("content", [])
269
+ if isinstance(content, list):
270
+ for block in content:
271
+ block_type = block.get("type", "")
272
+ if block_type == "tool_use":
273
+ tool_name = block.get("name", "unknown")
274
+ self.current_tool = tool_name
275
+ self.tool_call_counts[tool_name] += 1
276
+ self.total_tool_calls += 1
277
+ tool_input = block.get("input", {})
278
+ if isinstance(tool_input, dict):
279
+ self._extract_tool_summary_from_dict(tool_input)
280
+ self._detect_phase(
281
+ json.dumps(tool_input, ensure_ascii=False)[:500]
282
+ )
283
+ # Track CodeBuddy Agent/Task tool invocations as
284
+ # potential sub-agent spawns (Gap 3 fix).
285
+ self._record_cb_agent_tool_call(tool_name, tool_input)
286
+ elif block_type == "text":
287
+ text = block.get("text", "")
288
+ if text.strip():
289
+ self.last_text_snippet = text.strip()[:120]
290
+ self._detect_phase(text)
291
+ self._detect_terminal_error(
292
+ text, require_error_context=True
293
+ )
294
+ return
295
+
296
+ # CodeBuddy function_call / function_call_result in stream-json
297
+ if event_type in ("function_call", "function_call_result"):
298
+ self.event_format = self.event_format or "stream-json"
299
+ sid = event.get("sessionId", "")
300
+ cwd = event.get("cwd", "")
301
+ if sid and not self.cb_session_id:
302
+ self.cb_session_id = sid.strip()
303
+ if cwd and not self.cb_cwd:
304
+ self.cb_cwd = cwd.strip()
305
+ if event_type == "function_call":
306
+ tool_name = event.get("name", "unknown")
307
+ self.current_tool = tool_name
308
+ self.tool_call_counts[tool_name] += 1
309
+ self.total_tool_calls += 1
310
+ self.is_active = True
311
+ # Extract summary from arguments
312
+ raw_args = event.get("arguments", "")
313
+ if isinstance(raw_args, str) and raw_args.strip():
314
+ self._extract_tool_summary(raw_args)
315
+ self._detect_phase(raw_args[:500])
316
+ # Track Agent/Task tool invocations as sub-agent spawns
317
+ self._record_cb_agent_tool_call(tool_name, raw_args)
318
+ elif event_type == "function_call_result":
319
+ self.current_tool = None
320
+ # Check result text for phase / error hints
321
+ result_text = event.get("output") or ""
322
+ if result_text and isinstance(result_text, str):
323
+ stripped = result_text.strip()
324
+ if stripped:
325
+ self.last_text_snippet = stripped[:120]
326
+ self._detect_phase(stripped)
327
+ self._detect_terminal_error(stripped, require_error_context=True)
328
+ return
329
+
236
330
  # ── Claude Code verbose format ──────────────────────────────
237
331
  if event_type == "assistant":
238
332
  self.event_format = self.event_format or "stream-json"
@@ -259,7 +353,10 @@ class ProgressTracker:
259
353
  self._detect_phase(text)
260
354
  self._detect_terminal_error(text, require_error_context=True)
261
355
 
262
- elif event_type == "tool_result" or event_type == "user":
356
+ elif event_type == "tool_result" or (
357
+ event_type == "user"
358
+ and not isinstance(event.get("sessionId"), str)
359
+ ):
263
360
  # tool_result contains output from tool execution
264
361
  self.event_format = self.event_format or "stream-json"
265
362
  self.is_active = True
@@ -269,11 +366,11 @@ class ProgressTracker:
269
366
  # B) Nested user events: event["message"]["content"][] has type=="tool_result" items
270
367
  content_candidates = []
271
368
 
272
- # Format A: top-level tool_result
369
+ # Format A: top-level tool_result (Claude Code)
273
370
  if event_type == "tool_result":
274
371
  content_candidates.append(str(event.get("content", "")))
275
372
 
276
- # Format B: nested inside user event
373
+ # Format B: nested inside user event (Claude Code, NOT CodeBuddy)
277
374
  if event_type == "user":
278
375
  message = event.get("message", {})
279
376
  content_list = message.get("content", [])
@@ -465,15 +562,60 @@ class ProgressTracker:
465
562
  self._current_tool_input_parts = []
466
563
 
467
564
  elif event_type == "error":
565
+ self.event_format = self.event_format or "stream-json"
468
566
  error_msg = event.get("error", {}).get("message", "Unknown error")
567
+ errors = event.get("errors") or []
568
+ if isinstance(errors, list) and errors:
569
+ error_msg = "; ".join(str(e) for e in errors[:3])
469
570
  self.errors.append(error_msg)
470
571
  self._detect_terminal_error(str(error_msg))
471
572
 
573
+ # ── CodeBuddy file-history-snapshot (ignore) ─────────────────
574
+ elif event_type == "file-history-snapshot":
575
+ return
576
+
472
577
  # Check for subagent indicator
473
578
  if event.get("parent_tool_use_id"):
474
579
  # This is a sub-agent event; tool name is still tracked normally
475
580
  pass
476
581
 
582
+ def _record_cb_agent_tool_call(self, tool_name, raw_args):
583
+ """Record a CodeBuddy Agent/Task* tool invocation for sub-agent tracking.
584
+
585
+ CodeBuddy spawns sub-agents via:
586
+ - Agent(prompt=..., run_in_background=True/False) — synchronous or bg fork
587
+ - TaskCreate(subject=..., description=...) — scheduled task
588
+
589
+ (TaskUpdate/TaskOutput exist but are lifecycle-only — they track
590
+ existing tasks rather than creating new sub-agents, so we don't
591
+ count them as spawns.)
592
+
593
+ We don't get child session ids from these tool calls in the stream,
594
+ but recording the count lets the heartbeat code in heartbeat.sh apply
595
+ an extended stale-kill window just as it does for Codex spawn_agent.
596
+ """
597
+ if tool_name not in ("Agent", "TaskCreate"):
598
+ return
599
+ # Both dict (from message events) and str/JSON (from function_call events)
600
+ # are supported.
601
+ if isinstance(raw_args, dict):
602
+ args = raw_args
603
+ elif isinstance(raw_args, str):
604
+ try:
605
+ args = json.loads(raw_args)
606
+ except (json.JSONDecodeError, TypeError):
607
+ return
608
+ else:
609
+ return
610
+ # For Agent tool, subagent_type or run_in_background hints at delegation
611
+ if tool_name == "Agent":
612
+ subagent_type = args.get("subagent_type", "")
613
+ prompt = args.get("prompt", "")
614
+ if subagent_type or prompt:
615
+ self._subagent_spawn_count += 1
616
+ elif tool_name == "TaskCreate":
617
+ self._subagent_spawn_count += 1
618
+
477
619
  def _record_terminal_result(self, text="", is_error=False, api_error_status=None, api_error_code=""):
478
620
  """Record a Claude Code terminal result event."""
479
621
  terminal_text = str(text or "")
@@ -757,17 +899,85 @@ class ProgressTracker:
757
899
  return []
758
900
  return matches
759
901
 
902
+ def _cb_projects_dir(self):
903
+ """Return the CodeBuddy projects directory for transcript lookup.
904
+
905
+ CodeBuddy stores per-project session transcripts and sub-agent data
906
+ under ~/.codebuddy/projects/{projectDir}/{sessionId}/.
907
+ """
908
+ cb_config = os.environ.get("CODEBUDDY_CONFIG_DIR")
909
+ if cb_config:
910
+ return Path(cb_config).expanduser() / "projects"
911
+ cb_home = os.environ.get("CODEBUDDY_HOME")
912
+ if cb_home:
913
+ return Path(cb_home).expanduser() / "projects"
914
+ return Path.home() / ".codebuddy" / "projects"
915
+
916
+ def _cb_project_key(self):
917
+ """Encode cwd the same way CodeBuddy stores project subdirs.
918
+
919
+ CodeBuddy uses the same sanitisation as Claude Code (\\, /, : → -)
920
+ but strips the leading '-' so "/Users/test/MyProject" becomes
921
+ "Users-test-MyProject".
922
+ """
923
+ cwd = self.cb_cwd
924
+ if not cwd:
925
+ return ""
926
+ return cwd.replace("\\", "-").replace("/", "-").replace(":", "").lstrip("-")
927
+
928
+ def _find_cb_child_session_files(self):
929
+ """Find CodeBuddy subagent transcript data for this parent session.
930
+
931
+ CodeBuddy sub-agents store tool-results/{callId}.txt files; conversation
932
+ transcripts are not (as of 2026) written as agent-*.jsonl files in the
933
+ subagents/ directory. We track the tool-results .txt files as a proxy
934
+ for child activity so the heartbeat monitor can extend the stale-kill
935
+ window while sub-agents are running.
936
+ """
937
+ if not self.cb_session_id:
938
+ return []
939
+
940
+ projects_dir = self._cb_projects_dir()
941
+ if not projects_dir.exists():
942
+ return []
943
+
944
+ project_key = self._cb_project_key()
945
+ if not project_key:
946
+ return []
947
+
948
+ subagents_dir = (
949
+ projects_dir / project_key / self.cb_session_id / "subagents"
950
+ )
951
+ if not subagents_dir.exists():
952
+ return []
953
+
954
+ # Collect all files under each agent-{id} subdirectory. These are
955
+ # currently tool-results/{callId}.txt files, but the glob also picks
956
+ # up future agent-*.jsonl transcripts should CodeBuddy add them.
957
+ try:
958
+ return sorted(
959
+ p for p in subagents_dir.glob("**/*")
960
+ if p.is_file()
961
+ )
962
+ except OSError:
963
+ return []
964
+
760
965
  def refresh_child_session_activity(self, force=False):
761
966
  """Refresh child transcript file stats.
762
967
 
763
968
  The heartbeat monitor uses this activity signature to treat subagent
764
969
  transcript growth as real progress while the parent session is blocked
765
- waiting for a child agent/tool result. Supports Codex child threads and
766
- Claude Code in-process teammate transcripts.
970
+ waiting for a child agent/tool result. Supports Codex child threads,
971
+ Claude Code in-process teammate transcripts, and CodeBuddy sub-agent
972
+ file activity under subagents/.
767
973
  """
768
974
  previous_signature = self.child_activity_signature
769
975
 
770
- if not self.codex_child_thread_ids and not self.claude_session_id:
976
+ if (
977
+ not self.codex_child_thread_ids
978
+ and not self.claude_session_id
979
+ and not self.cb_session_id
980
+ ):
771
981
  self.child_session_files = []
772
982
  self.child_total_bytes = 0
773
983
  self.child_activity_signature = ""
@@ -788,6 +998,7 @@ class ProgressTracker:
788
998
  if found:
789
999
  self._codex_child_session_paths[thread_id] = found
790
1000
  self._claude_child_session_files = self._find_claude_child_session_files()
1001
+ self._cb_child_session_files = self._find_cb_child_session_files()
791
1002
  self._last_child_scan_at = now
792
1003
 
793
1004
  files = []
@@ -826,6 +1037,14 @@ class ProgressTracker:
826
1037
  for path in self._claude_child_session_files:
827
1038
  add_file("claude", path.stem, path)
828
1039
 
1040
+ for path in self._cb_child_session_files:
1041
+ # Identifier for CodeBuddy sub-agent files: use the parent
1042
+ # directory name (agent-{id}) so heartbeat can distinguish
1043
+ # activity across different sub-agent instances.
1044
+ parent_name = path.parent.name if hasattr(path, "parent") else ""
1045
+ identifier = parent_name if parent_name.startswith("agent-") else path.name
1046
+ add_file("codebuddy", identifier, path)
1047
+
829
1048
  self.child_session_files = files
830
1049
  self.child_total_bytes = total_bytes
831
1050
  self.child_activity_signature = "|".join(signature_parts)
@@ -861,11 +1080,14 @@ class ProgressTracker:
861
1080
  "total_tool_calls": self.total_tool_calls,
862
1081
  "active_subagent_count": self.active_subagent_count,
863
1082
  "subagent_states": subagent_states,
1083
+ "subagent_spawn_count": self._subagent_spawn_count,
864
1084
  "child_thread_ids": sorted(self.codex_child_thread_ids),
865
1085
  "child_session_files": self.child_session_files,
866
1086
  "child_total_bytes": self.child_total_bytes,
867
1087
  "child_activity_signature": self.child_activity_signature,
868
1088
  "last_child_activity_at": self.last_child_activity_at,
1089
+ "cb_session_id": self.cb_session_id,
1090
+ "cb_cwd": self.cb_cwd,
869
1091
  "last_text_snippet": self.last_text_snippet,
870
1092
  "last_result_is_error": self.last_result_is_error,
871
1093
  "api_error_status": self.api_error_status,