Context
Upgraded @anthropic-ai/claude-agent-sdk from ^0.2.44 to ^0.2.70 on 2026-03-06. Full analysis in docs/pending-updates/0.2.47-0.2.70-update.md.
What Changed (key features)
- v0.2.51: Memory leak fix (unbounded UUID tracking in long sessions)
- v0.2.51:
task_progress events for background agent progress
- v0.2.53:
listSessions() API
- v0.2.59:
getSessionMessages() with pagination
- v0.2.63:
supportedAgents() method; MCP unprefixed tool name fix
- v0.2.69:
agent_id/agent_type guaranteed on all hook events; MCP disconnect fix
- v0.2.70:
AgentToolInput.subagent_type optional; HTTP MCP 406 fix
What to Monitor
Potential Regressions
Known Upstream Issues (open on anthropics/claude-agent-sdk-typescript)
- #192 — Random UUID in Bash tool description busts input cache
- #188 — 1-hour cache TTL default instead of 5-minute
- #25 —
maxThinkingTokens still breaks hooks (unchanged)
- #14 —
defer_loading still blocked on upstream
Breaking Change Warning
system:init and result events currently emit 'Task' as the Agent tool name (reverted in v0.2.69). The next minor release will change this to 'Agent' — must audit string matching before upgrading past v0.2.x.
Files Modified
package.json — ^0.2.44 → ^0.2.70
src/hooks/sdkHooks.js — added v0.2.69+ comments on agent_id/agent_type guarantees
src/utils/hookSSEBridge.js — updated NOTE block re: agent field guarantees, agentRegistry fallback context
Validation Plan
- Run existing test suite (
npm test)
- Run a small memo generation (1-2 documents) and verify:
- All subagents spawn and complete
- Gate checks pass
- SSE events reach frontend
- No memory growth anomalies
- Run a full 29-subagent memo and compare token usage against pre-upgrade baseline
Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
Context
Upgraded
@anthropic-ai/claude-agent-sdkfrom^0.2.44to^0.2.70on 2026-03-06. Full analysis indocs/pending-updates/0.2.47-0.2.70-update.md.What Changed (key features)
task_progressevents for background agent progresslistSessions()APIgetSessionMessages()with paginationsupportedAgents()method; MCP unprefixed tool name fixagent_id/agent_typeguaranteed on all hook events; MCP disconnect fixAgentToolInput.subagent_typeoptional; HTTP MCP 406 fixWhat to Monitor
Potential Regressions
agent_typevalues: Confirm string values match existingclassifyAgent()patterns inhookSSEBridge.js(25 includes-checks)GATE_CHECK_CONFIGSkey matching viaagent_id?.includes(type) || agent_type?.includes(type)still works (sdkHooks.js:1203)subagent_start,subagent_stop,document_ready,code_execution_*eventsKnown Upstream Issues (open on anthropics/claude-agent-sdk-typescript)
maxThinkingTokensstill breaks hooks (unchanged)defer_loadingstill blocked on upstreamBreaking Change Warning
system:initandresultevents currently emit'Task'as the Agent tool name (reverted in v0.2.69). The next minor release will change this to'Agent'— must audit string matching before upgrading past v0.2.x.Files Modified
package.json—^0.2.44→^0.2.70src/hooks/sdkHooks.js— added v0.2.69+ comments onagent_id/agent_typeguaranteessrc/utils/hookSSEBridge.js— updated NOTE block re: agent field guarantees, agentRegistry fallback contextValidation Plan
npm test)Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com