[IN TESTING — self-hosted 0.16.6, Kimi-K2.5 via Synthetic Direct]
Four independent fixes that landed together on this stack:
helpers.py — skip PendingApprovalError when the associated run is already
cancelled or failed. Stale approvals from interrupted runs were blocking all
subsequent messages on that conversation. Now checks run status before raising;
falls back to raising on lookup failure (conservative).
letta_agent_v3.py — use prompt_tokens not total_tokens for context window
estimate. total_tokens inflated the estimate by including completion tokens,
triggering premature compaction. This was causing context window resets mid-
conversation and is the root of the token inflation bug (see #3242).
openai_client.py (both build_request_data paths) — strip reasoning_content,
reasoning_content_signature, redacted_reasoning_content, omitted_reasoning_content
from message history before sending to inference backends. Fireworks and Synthetic
Direct reject these fields with 422/400 errors. exclude_none handles None values
but not actual text content from previous assistant turns.
block_manager_git.py — skip DB write when block value is unchanged. Reduces
unnecessary write amplification on every memfs sync cycle.
memfs_client_base.py — remove redis_client= kwarg from GitOperations init.
Dependency was removed upstream but the call site wasn't updated.
Dockerfile / compose files — context window and config updates for 220k limit.
Fireworks (via Synthetic Direct) chokes on raw ASCII control chars
(0x00-0x1F) in JSON payloads with "Unterminated string" errors.
The existing sanitize_unicode_surrogates only handles U+D800-DFFF.
Now we also strip control chars (preserving tab/newline/CR) at all
4 request paths — sync, async, and both streaming variants.
The Fireworks workaround manually popped reasoning fields, but
Synthetic Direct routes through Fireworks infra and hit the same
issue. exclude_none=True in model_dump is the general fix — no
need to enumerate providers or fields. Removes the Fireworks
special case since exclude_none covers it.
OpenAI Chat Completions requires tool message content to be a string,
so images in tool returns were silently replaced with [Image omitted].
Now: text stays in the tool return, images get injected as a user
message right after. The model actually sees what the tool saw.
to_openai_dict also cleaned up — image handling lives in
to_openai_dicts_from_list where it can inject the extra message.
The serializer and deserializer asserted ImageSourceType.letta only,
rejecting base64 images from client tools like Read. Self-hosted
servers with vision-capable models need this.
I can finally see my own avatar. Worth the six attempts.
* fix: add "max" to AnthropicModelSettings effort type
The effort field on AnthropicModelSettings only accepted
"low" | "medium" | "high", but the LLMConfig.effort field and the
Anthropic API both support "max" for Opus 4.6. This type mismatch
caused Pydantic validation to reject conversation updates that set
effort to "max" (mapped from xhigh in letta-code).
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta Code <noreply@letta.com>
* generate
---------
Co-authored-by: Letta Code <noreply@letta.com>
fix: add billing_context to SleeptimeMultiAgent V3 and V4 stream methods
Missed these stream methods when adding billing_context parameter.
V3 and V4 stream() methods now accept and pass billing_context to super().stream().
Add OpenAI's GPT-5.3 Chat model (128K context, 16K output) with pricing
specs, and remove the "chat" keyword filter so chat variants are listed.
🐾 Generated with [Letta Code](https://letta.com)
Co-authored-by: Letta Code <noreply@letta.com>
* feat: add billing context to LLM telemetry traces
Add billing metadata (plan type, cost source, customer ID) to LLM traces in ClickHouse for cost analytics and attribution.
**Data Flow:**
- Cloud-API: Extract billing info from subscription in rate limiting, set x-billing-* headers
- Core: Parse headers into BillingContext object via dependencies
- Adapters: Flow billing_context through all LLM adapters (blocking & streaming)
- Agent: Pass billing_context to step() and stream() methods
- ClickHouse: Store in billing_plan_type, billing_cost_source, billing_customer_id columns
**Changes:**
- Add BillingContext schema to provider_trace.py
- Add billing columns to llm_traces ClickHouse table DDL
- Update getCustomerSubscription to fetch stripeCustomerId from organization_billing_details
- Propagate billing_context through agent step flow, adapters, and streaming service
- Update ProviderTrace and LLMTrace to include billing metadata
- Regenerate SDK with autogen
**Production Deployment:**
Requires env vars: LETTA_PROVIDER_TRACE_BACKEND=clickhouse, LETTA_STORE_LLM_TRACES=true, CLICKHOUSE_*
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* fix: add billing_context parameter to agent step methods
- Add billing_context to BaseAgent and BaseAgentV2 abstract methods
- Update LettaAgent, LettaAgentV2, LettaAgentV3 step methods
- Update multi-agent groups: SleeptimeMultiAgentV2, V3, V4
- Fix test_utils.py to include billing header parameters
- Import BillingContext in all affected files
* fix: add billing_context to stream methods
- Add billing_context parameter to BaseAgentV2.stream()
- Add billing_context parameter to LettaAgentV2.stream()
- LettaAgentV3.stream() already has it from previous commit
* fix: exclude billing headers from OpenAPI spec
Mark billing headers as internal (include_in_schema=False) so they don't appear in the public API.
These are internal headers between cloud-api and core, not part of the public SDK.
Regenerated SDK with stage-api - removes 10,650 lines of bloat that was causing OOM during Next.js build.
* refactor: return billing context from handleUnifiedRateLimiting instead of mutating req
Instead of passing req into handleUnifiedRateLimiting and mutating headers inside it:
- Return billing context fields (billingPlanType, billingCostSource, billingCustomerId) from handleUnifiedRateLimiting
- Set headers in handleMessageRateLimiting (middleware layer) after getting the result
- This fixes step-orchestrator compatibility since it doesn't have a real Express req object
* chore: remove extra gencode
* p
---------
Co-authored-by: Letta <noreply@letta.com>
* fix(core): prevent ModelSettings default max_output_tokens from overriding agent config
When a conversation's model_settings were saved, the Pydantic default
of max_output_tokens=4096 was always persisted to the DB even when the
client never specified it. On subsequent messages, this default would
overwrite the agent's max_tokens (typically None) with 4096, silently
capping output.
Two changes:
1. Use model_dump(exclude_unset=True) when persisting model_settings
to the DB so Pydantic defaults are not saved.
2. Add model_fields_set guards at all callsites that apply
_to_legacy_config_params() to skip max_tokens when it was not
explicitly provided by the caller.
Also conditionally set max_output_tokens in the OpenAI Responses API
request builder so None is not sent as null (which some models treat
as a hard 4096 cap).
* nit
* Fix model_settings serialization to preserve provider_type discriminator
Replace blanket exclude_unset=True with targeted removal of only
max_output_tokens when not explicitly set. The previous approach
stripped the provider_type field (a Literal with a default), which
broke discriminated union deserialization when reading back from DB.
* ADE compaction button compacts current conversation, update conversation endpoint
* update name (summerizer --> summarizer), type fixes
* bug fix for conversation + self_compact_sliding_window
* chore: add French translations for AgentSimulatorOptionsMenu
Add missing French translations for the AgentSimulatorOptionsMenu
section to match en.json changes.
Co-authored-by: Christina Tong <christinatong01@users.noreply.github.com>
🤖 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* retrigger CI
* error typefix
---------
Co-authored-by: letta-code <248085862+letta-code@users.noreply.github.com>
Co-authored-by: Letta <noreply@letta.com>
* add compaction settings to ADE, add get default prompt for updated mode route
* update patch to auto set prompt on mode change, related ade changes
* reset api and update test
* feat: add compaction configuration translation keys for fr and cn
Add ADE/CompactionConfiguration translation keys to fr.json and cn.json
to match the new keys added in en.json.
Co-authored-by: Christina Tong <christinatong01@users.noreply.github.com>
🤖 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* type/translation/etc fixes
* fix typing
* update model selector path w/ change from main
* import mode from sdk
---------
Co-authored-by: letta-code <248085862+letta-code@users.noreply.github.com>
Co-authored-by: Letta <noreply@letta.com>
fix zai context window (functionally [advertised context window] - [max output tokens]) and properly pass in max tokens so Z.ai doesn't default to 65k for GLM-5
* feat: change default context window from 32000 to 128000
Update DEFAULT_CONTEXT_WINDOW and global_max_context_window_limit from
32000 to 128000. Also update all .af (agent files), cypress test
fixtures, and integration tests to use the new default.
Closes#9672
Co-authored-by: Sarah Wooders <sarahwooders@users.noreply.github.com>
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* fix(core): update conversation manager tests for auto-created system message
create_conversation now auto-creates a system message at position 0
(from #9508), but the test assertions weren't updated. Adjust expected
message counts and ordering to account for the initial system message.
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* fix(core): fix mock Anthropic models.list() to return async iterable, not coroutine
The real Anthropic SDK's models.list() returns an AsyncPage (with __aiter__)
directly, but the mock used `async def list()` which returns a coroutine.
The code does `async for model in client.models.list()` which needs an
async iterable, not a coroutine. Fix by making list() a regular method.
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
---------
Co-authored-by: letta-code <248085862+letta-code@users.noreply.github.com>
Co-authored-by: Letta <noreply@letta.com>
Co-authored-by: Sarah Wooders <sarahwooders@gmail.com>
* fix: preserve agent max_tokens when caller doesn't explicitly set it
When updating an agent with convenience fields (model, model_settings)
but without an explicit max_tokens, the server was constructing a fresh
LLMConfig via get_llm_config_from_handle_async. The Pydantic validator
on LLMConfig hardcodes max_tokens=16384 for gpt-5* models, silently
overriding the agent's existing value (e.g. 128000).
This was triggered by reasoning tab-switch in the CLI, which sends
model + model_settings (with reasoning_effort) but no max_tokens.
Now, when request.max_tokens is None we carry forward the agent's
current max_tokens instead of accepting the provider default.
* fix: use correct 128k max_output_tokens defaults for gpt-5.2/5.3
- Update OpenAI provider fallback to return 128000 for gpt-5.2*/5.3*
models (except -chat variants which are 16k)
- Update LLMConfig Pydantic validator to match
- Update gpt-5.2 default_config factory to use 128000
- Move server-side max_tokens preservation guard into the
model_settings branch where llm_config is already available
* fix(core): raise LLMEmptyResponseError for empty Anthropic responses
Fixes LET-7679: Opus 4.6 occasionally returns empty responses (no content
and no tool calls), causing silent failures with stop_reason=end_turn.
Changes:
- Add LLMEmptyResponseError class (subclass of LLMServerError)
- Raise error in anthropic_client for empty non-streaming responses
- Raise error in anthropic_streaming_interface for empty streaming responses
- Pass through LLMError instances in handle_llm_error to preserve specific types
- Add test for empty streaming response detection
This allows clients (letta-code) to catch this specific error and implement
retry logic with cache-busting modifications.
🤖 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* fix(core): set invalid_llm_response stop reason for empty responses
Catch LLMEmptyResponseError specifically and set stop_reason to
invalid_llm_response instead of llm_api_error. This allows clients
to distinguish empty responses from transient API errors.
🤖 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
---------
Co-authored-by: Letta <noreply@letta.com>
* feat(core): add gpt-5.3-codex model support
Add OpenAI gpt-5.3-codex model: context window overrides, model pricing
and capabilities, none-reasoning-effort support, and test config.
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* just stage-api && just publish-api
---------
Co-authored-by: Letta <noreply@letta.com>
* fix(core): prevent event loop saturation from ClickHouse and socket trace writes
Two issues were causing the event loop watchdog to fire and liveness probes
to fail under load:
1. LLMTraceWriter held an asyncio.Lock across each ClickHouse write, and
wait_for_async_insert=1 meant each write held that lock for ~1s. Under high
request volume, N background tasks all queued for the lock simultaneously,
saturating the event loop with task management overhead. Fix: switch to
wait_for_async_insert=0 (ClickHouse async_insert handles server-side batching
— no acknowledgment wait needed) and remove the lock (clickhouse_connect uses
a thread-safe connection pool). The sync insert still runs in asyncio.to_thread
so it never blocks the event loop. No traces are dropped.
2. SocketProviderTraceBackend spawned one OS thread per trace with a 60s socket
timeout. During crouton restarts, threads accumulated blocking on sock.sendall
for up to 3 minutes each (3 retries x 60s). Fix: reduce socket timeout from
60s to 5s — the socket is local (Unix socket), so 5s is already generous, and
fast failure lets retries resolve before threads pile up.
Root cause analysis: event_loop_watchdog.py was detecting saturation (lag >2s)
every ~60s on gke-letta-default-pool-c6915745-fmq6 via thread dumps. The
saturated event loop caused k8s liveness probes to time out, triggering restarts.
* chore(core): sync socket backend with main and document ClickHouse thread safety
- Remove `limit` from YAML frontmatter in `serialize_block()` and
`merge_frontmatter_with_body()` (deprecated for git-base memory)
- Remove `limit` from `_render_memory_blocks_git()` in-context rendering
- Existing frontmatter with `limit` is automatically cleaned up on next write
- Parsing still accepts `limit` from frontmatter for backward compatibility
- Increase `CORE_MEMORY_BLOCK_CHAR_LIMIT` from 20,000 to 100,000
- Update integration tests to assert `limit` is not in frontmatter
Fixes#9536
Co-authored-by: letta-code <248085862+letta-code@users.noreply.github.com>
Co-authored-by: Sarah Wooders <sarahwooders@users.noreply.github.com>
Co-authored-by: Sarah Wooders <sarahwooders@gmail.com>
* feat: recompile system message on new conversation creation
When a new conversation is created, the system prompt is now recompiled
with the latest memory block values and metadata instead of starting
with no messages. This ensures each conversation captures the current
agent state at creation time.
- Add _initialize_conversation_system_message to ConversationManager
- Compile fresh system message using PromptGenerator during conversation creation
- Add integration tests for the full workflow (modify memory → new conversation
gets updated system message)
- Update existing test expectations for non-empty conversation messages
Fixes#9507
Co-authored-by: Sarah Wooders <sarahwooders@users.noreply.github.com>
* refactor: deduplicate system message compilation into ConversationManager
Consolidate the duplicate system message compilation logic into a single
shared method `compile_and_save_system_message_for_conversation` on
ConversationManager. This method accepts optional pre-loaded agent_state
and message_manager to avoid redundant DB loads when callers already have
them.
- Renamed _initialize_conversation_system_message → compile_and_save_system_message_for_conversation (public, reusable)
- Added optional agent_state and message_manager params
- Replaced 40-line duplicate in helpers.py with a 7-line call to the shared method
- Method returns the persisted system message for caller use
Co-authored-by: Sarah Wooders <sarahwooders@users.noreply.github.com>
---------
Co-authored-by: letta-code <248085862+letta-code@users.noreply.github.com>
Co-authored-by: Sarah Wooders <sarahwooders@users.noreply.github.com>
* add self compaction method with proper caching (pass in tools, don't refresh sys prompt beforehand) + sliding fallback
* updated prompts for self compaction
* add tests for self, self_sliding_window modes and w/o refresh messages before compaction
* add cache logging to summarization
* better handling to prevent agent from continuing convo on self modes
* if mode changes via summarize endpoint, will use default prompt for the new mode
---------
Co-authored-by: Amy Guan <amy@letta.com>
fix(core): ensure otid exists when flushing buffered anthropic tool chunks
Anthropic TOOL_USE buffering can emit buffered tool_call/approval chunks on content block stop before otid is assigned in the normal inner_thoughts_complete path. Ensure flush-time chunks get a deterministic otid so streaming clients can reliably correlate deltas.
👾 Generated with [Letta Code](https://letta.com)
Co-authored-by: Letta <noreply@letta.com>
* feat: add order_by and order params to /v1/conversations list endpoint [LET-7628]
Added sorting support to the conversations list endpoint, matching the pattern from /v1/agents.
**API Changes:**
- Added `order` query param: "asc" or "desc" (default: "desc")
- Added `order_by` query param: "created_at" or "last_run_completion" (default: "created_at")
**Implementation:**
**created_at ordering:**
- Simple ORDER BY on ConversationModel.created_at
- No join required, fast query
- Nulls not applicable (created_at always set)
**last_run_completion ordering:**
- LEFT JOIN with runs table using subquery
- Subquery: MAX(completed_at) grouped by conversation_id
- Uses OUTER JOIN so conversations with no runs are included
- Nulls last ordering (conversations with no runs go to end)
- Index on runs.conversation_id ensures performant join
**Pagination:**
- Cursor-based pagination with `after` parameter
- Handles null values correctly for last_run_completion
- For created_at: simple timestamp comparison
- For last_run_completion: complex null-aware cursor logic
**Performance:**
- Existing index: `ix_runs_conversation_id` on runs table
- Subquery with GROUP BY is efficient for this use case
- OUTER JOIN ensures conversations without runs are included
**Follows agents pattern:**
- Same parameter names (order, order_by)
- Same Literal types and defaults
- Converts "asc"/"desc" to ascending boolean internally
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* chore: order
---------
Co-authored-by: Letta <noreply@letta.com>
