Cap persisted tool results before rehydration

[s-salamatov]

Summary

• cap persisted tool and tool_result content when session history is rehydrated into provider-bound ChatMessages
• apply the capped conversion to normal chat, streaming chat, retry-after-compaction, prompt inspection, silent memory turns, and LLM-backed compaction prompts
• keep provider-format conversion unbounded so hook-modified provider payloads are preserved exactly

Testing

• RED: cargo test -p moltis-agents model::tests::convert_caps_persisted_string_tool_result_when_limit_is_set -- --nocapture failed before implementation with missing values_to_chat_messages_with_tool_result_limit
• cargo test -p moltis-agents model::tests
• cargo test -p moltis-agents
• cargo test -p moltis-chat
• cargo fmt --check
• git diff --check

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 318e64e828

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[greptile-apps]

Greptile Summary

This PR adds a values_to_chat_messages_with_tool_result_limit conversion function that caps persisted tool/tool_result content at config.tools.max_tool_result_bytes (default 50 KB) before feeding session history to an LLM provider. All relevant production rehydration callsites — normal chat, streaming, retry-after-compaction, silent memory turns, prompt inspection, LLM-backed compaction, session summary, title generation, and quick actions — are updated. The provider_values_to_chat_messages path is intentionally left unbounded to preserve hook-modified provider payloads.

• Core truncation is delegated to the existing sanitize_tool_result helper (strips base64/hex blobs, then byte-caps at a char boundary), keeping logic DRY.
• Seven new unit tests cover string results, structured JSON results, errors, multi-byte char boundary safety, orphan-filtering preservation, and the provider conversion path.
• Two new integration tests (llm_replace, structured) verify that the provider-level byte cap — not tool_prune_char_threshold — is applied when building LLM summary prompts.

Confidence Score: 5/5

Safe to merge — the change is additive and all production rehydration paths now apply a consistent byte cap with no missed call sites.

Every callsite that converts persisted session JSON to ChatMessages for provider consumption is updated. The core truncation is delegated to the already-tested sanitize_tool_result helper. The intentional exception (provider_values_to_chat_messages) is clearly documented and left unbounded by design. Unit and integration tests cover the new code path end-to-end, including char-boundary safety and orphan-filter preservation.

No files require special attention.

Important Files Changed

Filename	Overview
crates/agents/src/model/convert.rs	Adds values_to_chat_messages_with_tool_result_limit and maybe_limit_tool_result_content; applies capping to both tool and tool_result roles, preserves orphan filtering and char-boundary safety.
crates/agents/src/model/mod.rs	Exports new public API and adds seven well-structured unit tests covering string, JSON, error, char-boundary, orphan-filter, and provider-path scenarios.
crates/chat/src/compaction_run/llm_replace.rs	Threads max_tool_result_bytes through to the summary-input conversion; new integration test correctly distinguishes the provider byte cap from tool_prune_char_threshold.
crates/chat/src/compaction_run/structured.rs	Threads max_tool_result_bytes into the structured summary-prompt build and adds a parallel integration test; recency-preserving path stays JSON-native and is correctly left unchanged.
crates/chat/src/run_with_tools.rs	Applies the capped conversion at both initial chat-history load and retry-after-compaction reload, sourcing the limit from persona config each time.
crates/chat/src/service/chat_impl.rs	Updates silent memory turn and prompt-inspection paths; minor refactor avoids calling discover_and_load() twice by storing the config in a local binding.
crates/chat/src/streaming.rs	Switches the streaming history-to-messages conversion to the capped variant; single-line, low-risk change.
crates/gateway/src/session/summary.rs	Session summary memory write now uses the bounded conversion; straightforward call-site update.
crates/gateway/src/session/title.rs	Title-generation history conversion switched to the capped variant; no logic change, just call-site update.
crates/chat/src/compaction_run/runner.rs	Threads max_tool_result_bytes into structured::run and llm_replace::run; recency_preserving::run correctly omits it since it operates directly on raw JSON values.

_{Reviews (2): Last reviewed commit: "Cap persisted tool results before rehydr..." | Re-trigger Greptile}

[s-salamatov]

Addressed the review feedback in 869fe56:

• llm_replace and structured compaction now use the configured provider-bound tools.max_tool_result_bytes cap for summary input history instead of chat.compaction.tool_prune_char_threshold, so the summarizer no longer sees only the first ~200 bytes by default.
• Threaded max_tool_result_bytes through run_compaction / compact_session so LLM-backed compaction has access to the same operator-configured tool-result cap as normal chat paths.
• Updated the missed LLM-bound history paths to use capped rehydration: periodic memory extraction, session-end summary, /btw quick action context, and auto-title generation.
• Added regression coverage for both LLM-backed compaction modes to verify summary prompts preserve tool-result content beyond the stored-history prune threshold when the provider cap allows it.

Verification:

• cargo test -p moltis-chat -> 118/118
• cargo test -p moltis-gateway -> 445/445 plus 14/14 integration tests
• cargo test -p moltis-agents model::tests -> 58/58
• cargo fmt --check
• git diff --check

@greptile

[s-salamatov]

@greptile Pls review

[s-salamatov]

@codex advise tests set, docs

[chatgpt-codex-connector]

To use Codex here, create an environment for this repo.

[s-salamatov]

@codex advise tests set, docs

[chatgpt-codex-connector]

To use Codex here, create an environment for this repo.