Repository navigation
feat: Store Provider Output Items and Pass the Stable Context as the SDK Conversation Block - #17149
Draft
Shabonasar wants to merge 21 commits into
Draft
Shabonasar wants to merge 21 commits into
Shabonasar wants to merge 21 commits into
Conversation
The per-request web search context (current date/time with milliseconds) was joined before additional_instructions, which carry mostly static text such as the artifacts prompt. Provider-side prefix caches stopped at the timestamp, so the static text after it was recomputed on every request. Move the dynamic context to the end via buildAdditionalInstructions and cover the ordering with unit tests.
…che-order # Conflicts: # packages/api/src/agents/initialize.ts
…e date changes in history The web search runtime context and temporal special variables were anchored to the turn start with millisecond precision. The system message precedes the whole history, so every turn invalidated the cached history and new conversations could not reuse the system prompt (LibreChat-AI#16877). - Web search runtime context states the conversation start in the user's timezone with minute precision; temporal variables resolve against the same persisted anchor. - A user message sent on a later local date than the model last saw gets a <context_update> block in its prompt copy only, derived from stored createdAt values so it renders identically on later turns; compaction boundaries follow the agents SDK formatter. - Optional endpoints.agents.promptMessageTimestamps adds the send time to every user message; the turn's user message is stored with createdAt = turnStartedAt on every save path. - Provider payload stability tests through createRun for OpenAI, an OpenAI-compatible endpoint, Anthropic, OpenRouter, Bedrock and Google.
… routes The browser resolved promptPrefix with its own clock on every send, so turn 1 of a spec-backed conversation carried the raw template while turn 2 carried resolved text in a different system-prompt position. Only Assistants routes resolve on the client now; the server resolves against the conversation start.
… of a conversation
Adds `endpoints.agents.promptCache.stableContext` (default off). Off, the prompt is
built as before. On:
- prompt time (`{{current_datetime}}` and friends, the web search runtime context) is
anchored to the conversation start, and a date change is announced by a
`<context_update>` block at the start of the first user message sent after it;
- the conversation-stable part of `additional_instructions` is merged into the system
instructions for Anthropic and OpenRouter with prompt caching, while memory, RAG and
one-turn queued file context stay in `additional_instructions`, where the SDK moves
them next to the newest user message;
- dynamic tool context goes after the additional instructions;
- the user message is inserted with the turn start as `createdAt`, so later turns render
the same stamp.
`promptMessageTimestamps` (default off) additionally prefixes each user message with the
time it was sent.
The history notes, the instruction split and the lever live in `packages/api`; the
controller only calls them. `applyContextToAgent` accepts the run context as parts and
records which are volatile, so no text is split by matching.
…assified parts Producers (project, memory guard, artifacts, skills catalog, RAG, memory, agent context files, always-apply skills of isolated subagents) contribute keyed parts through contributeContext instead of concatenating strings into additional_instructions. One registry fixes the canonical order and classifies every part, tool context map entries included, as static, conversation or turn. A runtime key the registry does not know counts as turn, so a new producer cannot break the stable prefix by default. assembleSystemPrompt builds the strings handed to the SDK. Once parts exist they are the only source of the tail; a direct write to additional_instructions is ignored with a warning. Without promptCache.stableContext it reproduces the previous request; with it, createRun moves only the turn parts with the prompt-cache tail and merges the rest into the instructions. The assembled text is byte-identical to string concatenation for both values of the lever, guarded by a snapshot recorded on the previous implementation. Isolated subagents with an always-apply skill place the skill ahead of the tool runtime context when stableContext is on.
…m prompt per segment Behind `endpoints.agents.promptCache.stableContext`; off, nothing here runs and the prompt is built as before. A single send-time step (turnContext) decides what the primary agent's model sees. Context sources (project, artifacts mode, memory tool guard, skills catalog, tool runtime context, memory, date, send time, retrieved context, queued upload files) are sampled once when the user sends a message; only the net difference from what the model already knows along the parent chain is stored on that user message (message.contextUpdates) and rendered as one <context_update> block at the start of its prompt copy. Toggles flipped back and forth before sending leave no trace, and stored values render the same bytes on every later turn, whatever the timezone. A regeneration sends no message: it renders what the message stored and writes nothing. The system prompt is rendered from the segment's anchor instead of the live parts, so it stays byte-identical until the next compaction. The anchor of the first segment is stored on the conversation, the anchor of a segment opened by a compaction on the message carrying the summary (contextAnchor); a conversation without one is anchored to its first turn after the upgrade. A re-anchor keeps the previous order, drops what is off and appends what is new, and dates itself, so the first message after the boundary announces nothing. Memory is a state source: an update lists the entries added, changed and removed (by key where the memory has keys) instead of the whole memory, and no longer rewrites the system prompt; memory switched off is withheld from the whole request. Retrieved context is stored with the user message it was retrieved for, like the file context bound to the message that carried the files; isolated subagents still take it in their system prompt. Queued upload files describe their own message only. A resumed turn (tool approval or an answered question) renders its system prompt from the paused turn's anchor, not from the live state. The planning, the anchor storage and the token accounting live in packages/api; the controller calls them. The conversation anchor is read with the conversation the ingress already loads, and a resume reads the messages once with a projection. contextUpdates and contextAnchor are server-written only, select: false, and stripped from client input, so they stay out of reads, shares and forks.
…d the SDK dispatcher Behind `endpoints.agents.promptCache.stableContext`, which also stays off, with one warning, while the installed @librechat/agents has no tool dispatcher. Providers serialize `tools` ahead of the system prompt, so switching an MCP server, an action or an agent tool mid-conversation reset the prompt cache from position 0. - Tool groups (`tool:mcp:<server>`, `tool:action:<domain>`, `tool:agent:<toolkit>`) are a turn-context source. The segment anchor stores each group's schemas and MCP instructions; `tools` and the MCP instructions are rendered from it until the next compaction. - The initial set orders MCP servers and their tools by name, other groups as the agent lists them, after the ungrouped tools. A compaction keeps the anchored order, drops groups that are off and appends the new ones. - A group enabled mid-segment is deferred and announced once in the user message with its schemas; the model calls it through the SDK dispatcher (`toolDispatch`), which resolves the call before hooks, approval and remembered approvals, so they see the real tool. An anchored group that is off keeps its schemas with `disabled`; an unreachable one is `unavailable` without an update. - A schema a server changes mid-segment is announced once, in the message sent with the change: the changed parameters, tools that are new (reached through the dispatcher) and tools that are gone. `tools` itself changes at the next anchor. - Tools found by `tool_search` stay deferred and are called through the dispatcher; the host search returns their schemas. - The formatter tool set keeps past dispatcher calls; isolated subagent copies bind the tools as loaded. - MCP servers and their tools load in name order, and MCP instructions are listed by server name; both only with the lever. The instructions of the named servers are read in packages/api, so MCPManager keeps its API. - UI: `resolveDisplayToolCall` routes a dispatcher call to the card and the group header of the tool it ran; the stream reducer keeps `dispatched`. The package depends on the tool dispatcher of @librechat/agents, which is not in a release yet; `package.json` moves to the release that has it.
…lare prefix resets Behind `endpoints.agents.promptCache.stableContext`, like the rest of the turn context. Switching web search, the code environment, memory or another built-in capability mid-conversation still changed `tools` and the system prompt, and changes the dispatcher cannot absorb (another model, an edited agent, provider tools, tool options that rewrite many schemas) mutated the anchored prompt silently. - Built-in tools are tool groups (`tool:builtin:<capability>`): web_search, file_search, the whole code environment, programmatic tool calling, skills, memory, ask_user_question, background tasks and tool_search. A new anchor binds them first, in a fixed order. Like other groups they are bound from the anchor, deferred and announced when enabled mid-segment, kept with `disabled` when switched off, and "available again" when switched back. - A built-in's instructions (`toolContextMap`) belong to its group: the system prompt states the anchored ones, a mid-segment group announces its own. - The file tools the code environment and skills both register are owned by the code group when the agent has a code environment, else by skills; an anchored file tool stays enabled while either group has it. - Declared resets: the anchor records the model, a digest of the agent's setup (instructions, handoffs, subagent targets) and the provider tools, serialized with sorted keys so the same setup built in another order never differs. A sent message that changes one of them, gains or loses a schema-profile parameter (tool intents, background runs) on an anchored tool, or enables a group the dispatcher cannot reach (the in-graph ask_user_question) re-anchors the segment on itself, as a compaction would: anchored order kept, deferred groups appended, switched-off ones dropped. Anchors recorded before the profile never reset. A regeneration sends no message, so it neither resets nor rewrites an anchor. - When the provider has a prompt cache (Anthropic, OpenRouter, Bedrock with promptCache), the reply stores the reasons in `metadata.promptCacheReset` and the chat shows a one-line notice that earlier context was reloaded. Nothing is recorded or shown for providers without one. - A switched-off ask_user_question stays bound in the graph and is refused through the registry; a deferred one is not bound. - The ask_user_question group header reads a dispatched call by the tool it ran.
…y them for restored tool rounds An OpenAI Responses tool round carries ids, status and encrypted reasoning that the next turn must send back as they were, or the request stops being a prefix of the previous one. The agents SDK reports them per finished round (`ON_PROVIDER_OUTPUT_ITEMS`) and restores a stored round from `tool_call.providerItems`. This depends on the SDK release that has the event; without it the handler is not registered and rounds are restored as before. - The handler adds the items to the run's content parts. `saveMessage` moves them out of `content` into a `providerItems` field of the message, keyed by tool call id, with `select: false`: no read, share, fork, export or search sees them unless it asks for the field, so there is no list of places that must remember to hide them. The history read of the agent client asks for it and puts the items back on their tool calls before formatting; the stored row is not changed. - Live content (the sync and final stream events, `/chat/status`) is cleaned on the known paths: `resumeState.aggregatedContent`, `aggregatedContent` and `responseMessage.content`. - Only the server writes them: a client message, an import or a top-level field never reaches the stored field. A fork or a copy reads without the field, so a forked conversation, including one forked from a shared link, starts without items and restores its rounds as before. Rounds sent through the OpenAI-compatible and Responses routes of this server keep their previous behavior: those routes have their own handlers and store no items.
With `endpoints.agents.promptCache.stableContext` and Anthropic, OpenRouter or Bedrock with promptCache, everything that is stable within the conversation (the anchored context, tool runtime context, the agent's own guidance) goes to the agents SDK as `conversation_instructions`. The SDK keeps that block right after the instructions with a breakpoint of its own and never moves it, and carries a compaction summary as a stable message of its own. Only the per-turn parts (memory, retrieved context, files queued for this turn) stay in `additional_instructions`, which the SDK relocates next to the newest user message or, on Bedrock, appends after the cached blocks. Before, the stable part was merged into `instructions`. A summary created in a run and the same summary restored by the next run now render identically, so a compaction costs one prefix rebuild instead of two. `conversation_instructions` needs the SDK release that has it; without it the lever stays off with a warning. Bedrock was not covered before and is not verified against a real endpoint.
codegraph-librechat
Bot
added
🗺️ Backend Infra
Shabonasar
marked this pull request as draft
October 16, 2026 06:21
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Two things still cost cache after the earlier PRs. An OpenAI Responses tool round carries ids, status and encrypted reasoning that the next turn must send back unchanged, and the stable per-conversation context was merged into
instructions, so a compaction summary created in a run and the same summary restored by the next run rendered differently and every compaction cost two prefix rebuilds. Withendpoints.agents.promptCache.stableContext(defaultfalse) provider output items are stored beside the message and replayed for restored tool rounds, and with Anthropic, OpenRouter or Bedrock withpromptCacheeverything stable within the conversation goes to the SDK asconversation_instructions, which keeps it at a fixed position with its own breakpoint.Related to #17102. Stacked on #17148. Depends on LibreChat-AI/agents#613 (
conversation_instructions) and LibreChat-AI/agents#615 (ON_PROVIDER_OUTPUT_ITEMS, restore of tool rounds); without them the handler is not registered and the lever stays off with a warning.How it works
providerItemsare written by the server only: a client message, an import or a top-level field never reaches the stored field; forks and copies read without it and restore rounds as before. Live content (sync/final events,/chat/status) is cleaned on the known paths.Measurements
Run: group G3 measured on Qwen3.8-27B (vLLM behind NVIDIA Dynamo, through LiteLLM), OpenAI chat and Anthropic Messages, dev / lever off / lever on; the same self-hosted model answers both formats, the real OpenAI and Anthropic APIs were not measured; full method in #17102. After a compaction Anthropic (30.2%) is higher than OpenAI chat (20.1%) because OpenAI chat does not use the conversation block, so its summary still renders in two places.
Note: the backend caches in 1600-token blocks, so a request reuses cache only in whole blocks and the first turns of a short conversation (prompt under two blocks) cannot hit; shares are therefore lower on short chats than on long ones.
Verification
npm run lighthouse(3 cold navigations per run, Mongo delay 250 ms per query),devagainst the head of the stack with the lever off and on; ranges over the valid runs.All results are within the CI budgets and the differences are within run-to-run noise; the lever only adds
+contextAnchorto an existing select, so message loading does the same queries in all three configurations.Type of change
Testing
Enable the lever with Anthropic or OpenRouter prompt caching, hold a conversation across a compaction, and check that the request after the compaction starts with the bytes of the summary request. For Responses, run two tool rounds with a reasoning model and check that turn N+1 starts with the bytes of turn N.
Automated tests: the test files this PR adds or changes, run per workspace with
npx jest <files>(mongod-backed suites against a real mongod): packages/api 546, packages/data-schemas 310, api 541 tests passing. Files:callbacks.spec.js,client.test.js,abortMiddleware.spec.js,messages-get.spec.js,messages-retention.spec.js,streamTenant.spec.js,importers.spec.js,promptPrefixStability.test.ts,providerOutputItems.test.ts,systemParts.snapshot.test.ts,toolGroupStability.test.ts,run.spec.ts,systemParts.spec.ts,turnContext.spec.ts,message.spec.ts,message.spec.ts,share.test.ts,providerItems.spec.ts.Static checks:
tsc --noEmitin packages/api, packages/data-schemas; eslint and prettier on the touched files.Screenshots / recordings
No user-facing change.
Risk / compatibility
New non-selected
providerItemsfield on messages (the SDK caps the size of a stored round); default off stores nothing.@librechat/agentsmoves to the release that has both SDK changes.Checklist