[senamakel]

Summary

Four related changes that let a host keep a long conversation cheap and keep what the agent knows about the user across compactions.

1. [context] summarizer_model is honoured. The setting (and OPENHUMAN_CONTEXT_SUMMARIZER_MODEL) was parsed but never read: TurnModels::summarizer always built a second instance of the turn's own model. It now builds the configured model on the turn's provider route (route_with_model) and logs [context_compression] summarizer model=<id> source=config|turn. If the configured summarizer fails, the existing breaker falls back to a deterministic trim; it does not fall back to the turn model.
2. Summarizer request shape. New [context] summarizer_tool_rendering (full | markers) and summarizer_max_output_tokens, with env overlays OPENHUMAN_CONTEXT_SUMMARIZER_TOOL_RENDERING / OPENHUMAN_CONTEXT_SUMMARIZER_MAX_OUTPUT_TOKENS. With markers, the summarizer sees the conversation's prose and one [used <tool>: <args>] line per tool call, and no tool output (tinyagents#587). Config.context is now a ContextSettings wrapper that flattens tinyagents' ContextConfig (reachable through Deref), so existing TOML is unchanged.
3. Out-of-band compaction. openhuman_embed::Agent::compact_thread(thread_id, CompactOptions { min_tokens, keep_recent_tokens }) -> CompactOutcome::{Compacted, Skipped, Empty} compacts a thread without a model turn, for hosts that compact while a user is idle.
   • It reuses the turn's compression middleware and CompactionCarry.
   • It holds the session turn lock, so a turn arriving mid-compaction waits for it.
   • It seals the head with begin_generation_from_baseline, so nothing is erased.
   • It writes the successor the same way Session::persist does.
   • The system prompt and tool list stay byte-identical.
4. Memory block and per-turn recall.
   • AgentDefinitionSpec::memory_block(MemoryBlockSpec { max_tokens }) appends a <memory> system message holding the identity's top beliefs and learnings after the prompt. It is re-rendered only when a thread starts or is compacted, and left out of the prompt-change comparison, so a memory change alone never busts the prefix cache or adds PROMPT_UPDATE_NOTE. Logs are [memory_block] rendered|kept|empty|error.
   • MemoryBinding::recall_each_turn(Duration) (config [memory] recall_each_turn_ms) recalls for the message being answered, with a deadline, and adds the hits as a tail note that is not persisted. Hits already in the block are dropped. Previously prefetch::pre_turn only returned the pack completed for the previous message, which had usually expired, so recall almost never reached the model. A late recall is logged as [memory:recall] deadline.

Public API / behaviour changes

• New: Agent::compact_thread, CompactOptions, CompactOutcome, MemoryBlockSpec, AgentDefinitionSpec::memory_block, MemoryBinding::recall_each_turn.
• New config: [context] summarizer_tool_rendering, summarizer_max_output_tokens, and [memory] recall_each_turn_ms.
• Behaviour: summarizer_model now takes effect when set. All other defaults are unchanged.

Depends on

tinyhumansai/tinyagents#587, merged. vendor/tinyagents now points at its merge commit 231d91c59 on tinyagents main. No tinyagents release has been cut since, so Cargo.lock lists the tinyagents crates at main's version, 3.0.0, instead of the 3.1.3 release tag the old pin pointed at. The tag commit only bumped versions, so no code differs.

Validation

• cargo fmt --all -- --check: clean.
• node scripts/ci/check-openhuman-rust-layout.mjs: passed.
• cargo clippy -p openhuman -p openhuman-embed -p openhuman-cli -p openhuman-tinyhumans --all-targets --features <product features> -- -D warnings: clean.
• cargo nextest run -p openhuman -p openhuman-embed --lib over agent::, memory::, config::, inference::, agent_registry and channel memory: 3175 passed, 6 failed. All 6 fail identically on unmodified origin/main (checked in a separate checkout):
  • a_detached_driver_uses_its_carried_actor_for_lifecycle_commits
  • resuming_after_owner_cache_loss_registers_live_task_before_launch
  • failed_driver_resume_marks_paused_session_failed_and_releases_owner
  • spawn_subagent_with_same_task_key_resumes_paused_job_through_shared_driver
  • headless_child_at_depth_cap_gets_readable_spawn_refusal
  • native_host_agent_turn_enforces_the_configured_channel_deadline
• openhuman-embed lib (296), plus the integration tests compact_thread, memory_block, memory_recall_each_turn, session_store, route_headers and harness_embed: pass. The embed lib needs --test-threads=1 under plain cargo test because of process-global slots; under nextest it passes.
• New tests:
  • turn_models_summarizer_tests.rs (wiremock: the request goes to the configured model id);
  • context_wiring_tests::a_compaction_records_the_summarizer_model_it_called;
  • config env/TOML parsing;
  • memory_block_tests.rs (core and session host);
  • prefetch_recall_each_turn_tests.rs;
  • tests/compact_thread.rs: skip below threshold, empty thread, compact above it, a racing turn waits, prompt and tools byte-identical, fresh memory block, sealed generation readable.

Known limits

• The sealing write sits in OpenHuman, because tinyagents-runtime has no public way to write a new generation outside a turn. The cleaner home is a Session::compact_history beside reset_history upstream; that is a follow-up.
• compact_thread's token estimate counts messages, not tool schemas.
• After a compaction where memory changed, the next turn seals one extra, single-turn generation.

Summary by CodeRabbit

• New Features
  • Compact saved conversations without starting a model turn, with options to retain recent history and outcomes for empty, skipped, or compacted threads.
  • Add memory blocks to agent prompts, with configurable token limits and refreshes when a new transcript generation begins.
  • Optionally wait for current-message memory recall before the first model call, up to a configured timeout.
  • Configure summarizer model selection, tool rendering, and output-token limits.
• Documentation
  • Expanded guidance on conversation compaction, memory blocks, per-turn recall, and summarizer settings.

[coderabbitai]

Caution

Review failed

The pull request is closed.

Walkthrough

This change adds configurable summarizer models and settings, idle-thread compaction, memory blocks in session prompts, and optional deadline-bounded memory recall for each turn.

Changes: Agent context and memory

Layer / File(s)	Summary
Summarizer configuration and routing crates/openhuman-core/src/config/schema/*, crates/openhuman-core/src/inference/*, crates/openhuman-core/src/agent/tinyagents/*, docs/gitbooks/en/developing/architecture/agent-harness.md	[context] now includes summarizer rendering and output-token settings. Turn construction selects a configured summarizer model when present and otherwise uses the turn model. Summarizer settings and model identity pass into compression middleware.
Memory blocks in session prompts crates/openhuman-core/src/memory/lifecycle/memory_block*, crates/openhuman-core/src/agent/session_host/runtime_session/*, crates/openhuman-core/src/agent_registry/host_definition.rs, crates/openhuman-embed/src/agent/definition*, crates/openhuman-embed/tests/memory_block.rs	Agents can configure a token-bounded memory block. Session prompt handling renders or retains the block across turns and generations, and excludes it from prompt-change comparison.
Idle-thread compaction crates/openhuman-core/src/agent/session_host/runtime_session/compact.rs, crates/openhuman-core/src/agent/tinyagents/context_wiring.rs, crates/openhuman-core/src/agent_runs/agent_chat/*, crates/openhuman-embed/src/agent/compact.rs, crates/openhuman-embed/tests/compact_thread.rs, docs/gitbooks/en/developing/embed/*	Agent::compact_thread compacts bound history without a model turn. It reports Empty, Skipped, or Compacted; successful compaction seals the current generation and writes a successor.
Deadline-bounded per-turn recall crates/openhuman-core/src/memory/lifecycle/prefetch*, crates/openhuman-core/src/config/schema/memory.rs, crates/openhuman-embed/src/agent/spec.rs, crates/openhuman-embed/tests/memory_recall_each_turn.rs	Memory bindings can set a recall timeout. A recalled pack is returned to the turn when it is available within the deadline; otherwise the turn proceeds without that pack.

Priority: Not assessed

Estimated code review effort: 4 (Complex) | ~60 minutes

Change: Feature

Sequence Diagram(s)

sequenceDiagram
  participant Agent
  participant compact_thread_for
  participant OpenHumanSessionHost
  participant compact_history
  participant TranscriptStore
  participant Summarizer
  Agent->>compact_thread_for: pass thread ID and CompactOptions
  compact_thread_for->>OpenHumanSessionHost: bind thread and request compaction
  OpenHumanSessionHost->>TranscriptStore: read bound transcript under turn lock
  OpenHumanSessionHost->>compact_history: pass history and idle-fold options
  compact_history->>Summarizer: request summary when history can be folded
  compact_history-->>OpenHumanSessionHost: return folded history
  OpenHumanSessionHost->>TranscriptStore: seal current generation and write successor

Loading

sequenceDiagram
  participant TurnPrelude
  participant memory_pre_turn
  participant pre_turn
  participant MemoryEngine
  participant Turn
  TurnPrelude->>memory_pre_turn: read recall_each_turn_ms
  memory_pre_turn->>pre_turn: pass RecallEachTurn timeout
  pre_turn->>MemoryEngine: queue current-message recall
  MemoryEngine-->>pre_turn: return pack or complete after timeout
  pre_turn-->>Turn: return timed pack or None

Loading

Suggested labels: rust-core, agent

Suggested reviewers: m3ga-mind, sanil-23, oxoxdev

Merge Risk: Moderate · up to fb3bb

Turns can stall or slow down when the memory engine is slow or unavailable, because the memory block is read without a deadline and re-read on every turn when memory is empty or failing. Per-turn recall can also be silently skipped during bursts of messages. These should be addressed or explicitly accepted before merge. The remaining summarizer-model edge case is minor.

- [ ] Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{A rabbit checked the context flow,
And watched the idle transcripts grow.
A memory block found its place,
Recall arrived at measured pace.
The summarizer trimmed the tale,
Then hopped away along the trail.}

_{Comment @coderabbitai help to get the list of available commands.}

[tinysweeper]

Tiny Sweeper review

⏳ Reviewing fb3bbeb94d10 now. The last completed report, when available, remains below until this pass finishes.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-28T06:07:39.630182Z We couldn’t complete this review. Comment @codex review to try again.	fb3bbeb	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.