Bump claude-agent-sdk ^0.1.0 → ^0.3.x (unlocks native agentic harness; fixes cross-turn ScheduleWakeup/background)

[edspencer]

Summary

Bump @anthropic-ai/claude-agent-sdk in packages/core from ^0.1.0 (resolves 0.1.77) to ^0.3.x (latest 0.3.205). The pinned ^0.1.0 range cannot cross 0.x minors, so herdctl is frozen on a stale SDK line whose bundled harness lacks the current agentic toolset (ScheduleWakeup, ToolSearch, Cron*, Monitor, …). This is the root cause of a real cross-turn-autonomy failure for consumers running long-lived chat agents (e.g. Paddock keepers): an agent that schedules a wakeup or a background task to "continue later" is never re-invoked.

packages/core/package.json:21:

"@anthropic-ai/claude-agent-sdk": "^0.1.0",

Why it matters (evidence)

Root-caused why a keeper's "schedule a wakeup / continue later" never fires:

• 0.1.x bundles its own JS cli.js harness — grep of the installed 0.1.77 cli.js: 0 ScheduleWakeup refs. So the SDK runtime's openSession() hands the agent a harness that doesn't even contain the tool.
• System claude 2.1.199 (what the CLI runtime spawns): 9 ScheduleWakeup refs — the tool exists, but the CLI runtime runs it one-shot (claude -p, print-and-exit), so a wakeup timer has no live process to fire in.
• 0.3.x changed architecture: it drops the bundled JS harness and instead extracts/runs the native Claude Code 2.1.205 binary (per its manifest.json, linux-x64 ~257 MB) — the same 2.1.x line that has the agentic tools.

Empirical confirmation (reproducible): installed claude-agent-sdk@0.3.205, drove a persistent streaming query() (the openSession() path), had the agent call ScheduleWakeup(delaySeconds: 60), then held the session open and idle. The harness re-invoked the same session at the scheduled time and did real work:

tools=34  (ScheduleWakeup loaded via ToolSearch)
RESULT #1  → turn ends, session idle
   … ~83s idle, one live streaming process, no input …
system/init (SAME session id) → TOOL_USE Bash → file written → ASSISTANT: WOKE → RESULT #2
VERDICT: wake fired during idle ✅

The identical test against 0.1.77: ScheduleWakeup / ToolSearch are not available to the agent at all.

Scope / risk

Not a trivial version nudge — the SDK changed substantially between 0.1 and 0.3:

• Packaging changed: 0.1.x bundled cli.js; 0.3.x ships bridge.mjs + extractFromBunfs.js + a native-binary manifest.json (extracted at runtime). Executable-resolution / startup differs.
• Verify the surface used by runner/sdk-adapter.ts, runner/runtime/sdk-runtime.ts, and runner/runtime/message-queue.ts still typechecks and behaves: query() signature, Options (systemPrompt preset, permissionMode, allowedTools, canUseTool, cwd, includePartialMessages), and the Query control methods (interrupt, supportedCommands, setModel) + streaming-input AsyncIterable<SDKUserMessage>.
• Confirm Max-plan auth still resolves via CLAUDE_CODE_OAUTH_TOKEN (verified working with 0.3.205 in the repro).
• Observed SDK warning on 0.3.205: bare allowedTools entries auto-approve and shadow the canUseTool callback — relevant to per-call permission gating; use a PreToolUse hook if callback gating is required.

Downstream/consumer impact

The bump unlocks the fix but does not deliver it alone. runtime.execute() (the per-turn job path behind trigger()) still runs one query() per turn and ends when the turn's result arrives — so wakeups only survive when a consumer uses the persistent openSession() path, which is SDK-runtime-only (CLI/Docker leave openSession undefined). Consumers wanting cross-turn autonomy must (a) run the agent on runtime: "sdk" and (b) drive turns through openChatSession / openSession, not one-shot trigger(). (Tracked separately on the Paddock side.)

Acceptance criteria

• packages/core depends on @anthropic-ai/claude-agent-sdk@^0.3.x; lockfile updated.
• sdk-adapter / sdk-runtime / message-queue build and pass tests against 0.3.x.
• A test (or documented manual repro) shows a persistent openSession() agent scheduling ScheduleWakeup and being re-invoked at fire time.
• Changeset added per RELEASING.md.