@yusukeshib/pi-babysit 0.3.11 → 0.3.13

package/README.md

@@ -33,7 +33,7 @@ reachable from anywhere (`~/.pi-babysit/<pi-session-id>/`). Two kinds:
 | kind | started by | completion | on completion |
 | ---- | ---------- | ---------- | ------------- |
 | **process** | `babysit_run { command }` | process **exit** | automatic notification message (`triggerTurn`), batched for all exits observed in the same poll — the agent may end its turn after starting and is resumed on exit, same contract as the old `process` tool |
-| **subagent** | `babysit_run { profile: "subagent", task }` | `agent_end` in the RPC event stream (process stays alive) | none — the agent polls `babysit_check` or blocks on `babysit_wait`; the idle session accepts follow-up tasks |
+| **subagent** | `babysit_run { profile: "subagent", task }` | `agent_settled` in the RPC event stream (process stays alive) | none — the agent polls `babysit_check` or blocks on `babysit_wait`; the idle session accepts follow-up tasks |
 
 The **profile is a tool parameter, not a separate tool set**: domain knowledge
 (RPC bookkeeping, per-task byte offsets, parked-turn detection, PTY-safe
@@ -62,8 +62,11 @@ programs** (installers, wizards, REPLs): type with `babysit_send`
 | `babysit_wait` | Block until done: process exit (or `expect: "regex"` readiness marker), subagent task completion. Multi-wait: `ids` + `mode: "any"\|"all"` |
 | `babysit_kill` | Terminate a session, verify terminal state, then suppress the exit notification |
 
-A `tool_call` hook blocks shell backgrounding (`… &`, `nohup`, `setsid`,
-`disown`) and redirects all direct `bash` commands to `babysit_run`.
+The built-in `bash` tool is removed from the active tool set so the model does
+not waste a failed tool turn before choosing `babysit_run`. A fallback
+`tool_call` hook still blocks direct shell calls if another extension or preset
+re-enables `bash`, including shell backgrounding (`… &`, `nohup`, `setsid`,
+`disown`). Set `PI_BABYSIT_ALLOW_BASH=1` to retain direct `bash` explicitly.
 
 ## Commands (human)
 
@@ -91,9 +94,10 @@ babysit_check { id: "cargo-test", lines: 50 }
 babysit_check { id: "cargo-test", pattern: "FAIL|ERROR", lines: 50 }
 ```
 
-Tail and search results are capped at 200 lines and clipped to 8 KB. Pattern
-search returns the latest matching lines with line numbers. Do not read a
-potentially large log file in full.
+Tail and search results are capped at 200 lines, and the complete returned tool
+result (including lifecycle headers) is clipped to 8 KB. Pattern search returns
+the latest matching lines with line numbers. Prefer a targeted pattern over a
+broad tail, and do not read a potentially large log file in full.
 
 Subagent logs also compact Pi's streaming `message_update` events before they
 are recorded. Pi repeats the complete growing assistant message and partial
@@ -103,7 +107,8 @@ events untouched. This keeps long RPC sessions approximately linear in emitted
 content without changing final answers, completion detection, or follow-up
 behavior. Live `/babysit` and attach views render the retained deltas.
 
-All shell commands, including `pwd` and Git, are redirected to `babysit_run`.
+All shell commands, including `pwd` and Git, go through `babysit_run`. Bundle
+closely related tiny observations when doing so safely reduces tool turns.
 Set `PI_BABYSIT_ALLOW_BASH=1` only as an explicit emergency escape hatch.
 
 ## Unexpected worker loss
@@ -117,18 +122,24 @@ because blindly rerunning an arbitrary command can duplicate side effects.
 
 ## How completion detection works
 
-- **Process**: a 2.5s poller watches for running→exited transitions and injects
-  one `pi.sendMessage(…, { triggerTurn: true, deliverAs: "steer" })` containing
-  every deliverable exit observed in that poll (deduped via `meta/<id>.json`).
+- **Process**: a 2.5s poller watches for running→exited transitions and, once the
+  parent agent is idle, injects one
+  `pi.sendMessage(…, { triggerTurn: true, deliverAs: "steer" })` containing every
+  deliverable exit observed in that poll (deduped via `meta/<id>.json`). Waiting
+  for idleness prevents an immediately-following `babysit_wait` from racing the
+  poller and receiving a duplicate completion.
   `babysit_kill` and an exit already reported by `babysit_wait` suppress the
   notification.
-- **Subagent**: `babysit_wait` blocks on `babysit expect '"type":"agent_end"'`.
-  An `agent_end` whose last message is a **parked** toolResult — a
-  `babysit_run { command }` result carrying the `[notify-on-exit]` marker (or
-  the legacy `process` tool) — only means "turn parked awaiting a process-exit
-  notification; pi resumes on its own", so the wait continues. Any other
-  `agent_end` is real completion. Per-task byte offsets scope check/wait to the
-  CURRENT task, which is what makes follow-up tasks work.
+- **Subagent**: `babysit_wait` blocks on `babysit expect '"type":"agent_settled"'`.
+  Unlike `agent_end`, `agent_settled` cannot precede an automatic retry,
+  compaction retry, or queued continuation. A settled run containing a
+  **parked** toolResult — a `babysit_run { command }` result carrying the
+  `[notify-on-exit]` marker (or the legacy `process` tool) — only means "turn
+  parked awaiting a process-exit notification; pi resumes on its own", so the
+  wait continues. The marker-bearing result may be followed by a short
+  assistant note without defeating parked detection. Per-task byte offsets
+  scope check/wait to the CURRENT task, and appended log bytes are parsed
+  incrementally, which keeps follow-up tasks efficient.
 
 Subagents load `self-reap.ts`, which exits an idle finished subagent after a
 grace window (`PI_BABYSIT_REAP_AFTER`, default 120s) using the same parked-turn

package/agents.ts

@@ -56,20 +56,34 @@ function loadAgentsFromDir(
 			continue;
 		}
 
-		const { frontmatter, body } =
-			parseFrontmatter<Record<string, string>>(content);
-		if (!frontmatter.name || !frontmatter.description) continue;
+		let parsed: ReturnType<typeof parseFrontmatter<Record<string, unknown>>>;
+		try {
+			parsed = parseFrontmatter<Record<string, unknown>>(content);
+		} catch {
+			// One malformed definition must not prevent every other named agent
+			// from being discovered.
+			continue;
+		}
+		const { frontmatter, body } = parsed;
+		if (
+			typeof frontmatter.name !== "string" ||
+			typeof frontmatter.description !== "string"
+		) {
+			continue;
+		}
 
-		const tools = frontmatter.tools
-			?.split(",")
-			.map((t) => t.trim())
-			.filter(Boolean);
+		const tools = Array.isArray(frontmatter.tools)
+			? frontmatter.tools.filter((tool): tool is string => typeof tool === "string")
+			: typeof frontmatter.tools === "string"
+				? frontmatter.tools.split(",")
+				: [];
+		const normalizedTools = tools.map((tool) => tool.trim()).filter(Boolean);
 
 		agents.push({
 			name: frontmatter.name,
 			description: frontmatter.description,
-			tools: tools && tools.length > 0 ? tools : undefined,
-			model: frontmatter.model,
+			tools: normalizedTools.length > 0 ? normalizedTools : undefined,
+			model: typeof frontmatter.model === "string" ? frontmatter.model : undefined,
 			systemPrompt: body,
 			source,
 			filePath,