0.16.7

# Letta Server 0.16.7 Release Notes

**173 commits since 0.16.6** | Released March 31, 2026

## Highlights

**Self-hosted users: this is a big upgrade.** The default global context window is raised from 32k to 128k, the context window reset bug (LET-7991) is fixed, and compaction has been overhauled. If you've been running curl commands to patch your config after every ADE load, most of that pain should be gone.

## Breaking Changes

- **Block limits are no longer enforced** -- block limit validation has been deprecated and removed from the git memory sync path (#9977, #9983). Blocks can now grow freely. If you were relying on limits to cap per-turn cost, you'll need to manage block size via other means.

## Context Window & Compaction (21 fixes)

The biggest category of fixes. Self-hosted users were hit hardest by these.

- **Global context window default raised from 32k to 128k** (#9993) -- self-hosted servers no longer default to 32k for unknown models
- **Context window preserved on conversation model override** (LET-7991, #9986) -- the bug where non-default conversations fell back to 32k is fixed
- **Compaction overflow fixes** (#9897) -- addresses the double-compaction and runaway compaction loops
- **Compaction model resets on agent model change** (#10031) -- switching your agent's model no longer leaves the old summarizer model behind
- **Summarizer prompt improved** (#10314) -- now remembers plan files, GitHub PRs, and other structured content during summarization
- **BYOK summarization fixed** (#10152) -- summarizer provider fallback no longer fires for BYOK requests
- **Better error surfacing** -- context window exceeded errors now have descriptive messages (#10135, #10171), and system prompt size warnings during compaction (#10058)

## Gemini (2 fixes)

- **thought_signature preserved on function calls without reasoning** (LET-8166, #10237) -- the bug blocking all Gemini 2.5+/3.x multi-turn tool calling is fixed
- **Streaming interface crash fixed** (#10306) -- `self.model` now initialized in `SimpleGeminiStreamingInterface` constructor (LET-8129)

## Memory & memfs (10 fixes, 4 features)

- **available_skills block no longer duplicates in system prompt** (#10006, #10011, #10021) -- three separate fixes for the skills block multiplying and inflating context (LET-8013)
- **Git memory sync deferred until stream close** (#9951) -- reduces mid-stream sync failures
- **System prompt recompiles on agent creation with git memory** (#9950) -- new git-enabled agents no longer start with empty compiled context
- **Projection-style git memory rendering** (#10211) -- new rendering approach for memfs content in system prompts
- **Manual block edits via API trigger recompile** (#9775) -- no more stale context after API block updates
- **Conversation recompile endpoint** (#9848) -- `POST /v1/conversations/{id}/recompile` is now available

## Conversations (7 features)

- **Conversation forking** (#10234, #10263) -- fork conversations with shared message history, including the default conversation
- **Sort conversations by last_message_at** (#10190)
- **Idempotent conversation streaming** (#10147) -- OTID-based retry safety
- **Request-scoped system overrides** (#10227) -- per-request system prompt modifications

## Streaming & Reliability (13 fixes)

- **OTID retry hardening** (#10229, #10209) -- stream resume with backoff, race condition fixes
- **Conversation lock released earlier** (#10203) -- reduces contention on concurrent requests
- **Better error messages** (#10207) -- known LLM errors now surface descriptive messages instead of generic failures
- **BYOK error tagging** (#10204, #10311) -- errors now include `is_byok` flag for debugging
- **stream_incomplete diagnosis** (#10033) -- BaseException catching to identify root causes

## Model Support (14 features, 20 fixes)

- **GPT-5.4** -- full support including mini, nano, and fast variants (#9798, #10043)
- **GLM-5** -- GLM-5, GLM-5.1, GLM-5 Turbo, GLM-4.7 (#10317, #10285, #9994)
- **MiniMax M2.7** (#10093)
- **Baseten** -- added as provider with full frontend integration, serverless auto mode, reasoning support (#10250, #9846, #9998)
- **Fireworks** (#9780) and **zAI coding provider** (#10064)
- **Opus 4.6 / Sonnet 4.6** -- adaptive thinking tokens no longer incorrectly capped (#9795)
- **OpenAI proxy cleanup** -- extra fields removed (#9949), parallel tool calling supported (#9879)

## Security

- **Local filesystem access blocked via ImageContent bypass** (#3256, #10329) -- `file:///` URLs in images are now rejected
- **Internal MCP server targets blocked** (#10009)
- **SECURITY.md added** (#3228)

## Infrastructure

- Readiness enforcement scaffold (M1-M3 metrics pipeline) -- request pressure, DB pool, SSE lifecycle, event loop lag monitoring
- Multi-agent tools moved to less privileged execution environment (#9779)
- Subagent agents auto-hidden on create (#10096)
- WebSocket transport for OpenAI Responses API (#9841)

---

**For self-hosted users upgrading from 0.16.6:** This release addresses the majority of issues reported in the community over the past month. The context window default change alone (#9993) eliminates the most common source of "everything breaks when I open ADE" complaints.

**Full Changelog**: https://github.com/letta-ai/letta/compare/0.16.6...0.16.7