Rho SDK changelog ​
This changelog covers the embeddable rho-sdk crate.
6.0.1 (2026-11-16) ​
Bug Fixes ​
6.0.0 (2026-11-13) ​
⚠ BREAKING CHANGES ​
- sdk: Message::User, Message::Assistant and Message::BoundaryInput hold Arc<[ContentBlock]> instead of Vec<ContentBlock>; Message::ToolResult holds Arc<ToolResult>; Message::EnrichedAssistant and Message::AbortedAssistant hold Arc instead of Box. Build them with
.into()and change a shared payload with Arc::make_mut. - hooks: HookCapability::ExecuteProcess gains a requested_command field and HOOK_SCHEMA_VERSION is now 3.
- sdk: Message gains CompactionSummary, ModelContext, BoundaryInput, and ReasoningShift variants rendered through Message::wire(); ToolCall.execution replaces async marker blocks; ModelEvent gains ToolCallAccepted, Stopped, and ProviderContext with a ReplayScope; ProviderContextBlock.replay replaces the hard-coded Codex replay exception; the session snapshot schema is now 3. Saved 5.x history loads as typed entries.
- sdk: ToolResult carries status: ToolResultStatus and images instead of ok: bool; the tool image supplement message API is removed; ToolCompletion::Completed replaces Success and CompletedFailure; RunEvent::ToolStarted carries execution and replaces ToolDetached; ToolProgress uses new/with_metadata with getters. Saved 2.x sessions still load.
- sdk: ModelProvider has one send_turn_stream(request, ProviderTurnContext) and native_compact takes ModelRequestOptions; provider_steering_channel returns ProviderSteeringSender; StepStarted carries context_estimate; SteeringRetraction gains Delivered; InvalidConfiguration, Persistence, and Interrupted errors are struct variants with an optional source; host input and tool host builders use with_* setters and unsuffixed getters; AfterToolUsePayload carries capability and process.
- sdk: QuestionKind, DefaultSelection, HostInputSource, CompactorCancellationMode, HookSessionFailureKind, and ScriptedToolOutcome are now #[non_exhaustive]; exhaustive matches outside rho-sdk need a wildcard arm. Decision adapters should reject unknown question kinds with InvalidRequest::UnsupportedQuestionKind.
Features ​
- context: timestamp user messages and assistant text sent to the model (#1903) (39a30ad)
- daemon: configure a session's model, reasoning, fast mode and permission mode (#1740) (a6c9e39)
- daemon: run a session's commands with its client's environment (#1733) (0833ae2)
- hooks: show hooks the command the model requested when rtk rewrites it (#1892) (e2e01bd)
- hooks: tell hooks when a turn waits for input (#1721) (2258360)
- sdk: add a session-control capability for acting on other sessions (91b250f)
Bug Fixes ​
Performance Improvements ​
Code Refactoring ​
- sdk: carry tool outcomes and images on tool results (#1693) (4d52a30)
- sdk: collapse provider turn methods and finish builder renames (#1691) (233e1f9)
- sdk: mark host-facing sdk enums non-exhaustive (#1687) (3bac6f5)
- sdk: type compaction summaries, model context, and stop reasons in history (#1695) (4f44921)
5.15.2 (2026-10-31) ​
Bug Fixes ​
5.15.1 (2026-10-30) ​
Bug Fixes ​
- sdk: stop reporting recorded reasoning effort as dropped context on model switches (#1622) (1190306)
5.15.0 (2026-10-25) ​
Features ​
5.14.0 (2026-10-22) ​
Features ​
- acp: report max_tokens and refusal stop reasons from provider finishes (#1599) (20c3f91)
- acp: run shell commands in client terminals (#1602) (12f1f2c)
5.13.0 (2026-10-21) ​
Features ​
- acp: show grep and glob as searches, not file reads (#1583) (20b23ed)
- tui: add "Send for this session" so unchanged native-context warnings stop repeating (#1586) (eb083c1)
5.12.2 (2026-10-20) ​
Bug Fixes ​
5.12.1 (2026-10-19) ​
Performance Improvements ​
5.12.0 (2026-10-18) ​
Features ​
Performance Improvements ​
5.11.1 (2026-10-17) ​
Bug Fixes ​
- acp: show file diffs for edit and write calls (e528ee3)
5.11.0 (2026-10-16) ​
Features ​
- acp: load every mentioned skill in one host batch (#1496) (2cd6c5f)
- sdk: batch host-requested tool calls (#1494) (7281464)
5.10.0 (2026-10-15) ​
Features ​
5.9.0 (2026-10-10) ​
Features ​
- todo: keep task progress accessible across compaction (#1441) (77b31cc)
- tui: show what fills the context window with /context (#1455) (b61af1b)
- tui: tell /compact what to keep with optional instructions (#1458) (a2b9dbf)
5.8.0 (2026-10-07) ​
Features ​
- hooks: give after_tool_use hooks the real exit code and output of shell calls (#1395) (fad8ab5)
- providers: add OpenAI's Decisions API as a decision-model host for the permission screen (#1433) (14688fe)
5.7.0 (2026-10-04) ​
Features ​
- sdk: add a public decision protocol with noul, choice, and score questions (#1387) (7199e39)
- tools: Starlark code_mode scaffold (ToolHost + MCP path) (#1365) (c2d6771)
- workflow: continue crashed agent nodes from step checkpoints (#1376) (c38f12b)
Performance Improvements ​
- codemode: await every call_tools approval at once (#1393) (856713e)
- permissions: stop auto-mode approvals from queueing behind each other (#1391) (6afc173)
5.6.0 (2026-09-29) ​
Features ​
- compaction: cut summary cost by reusing the session prompt cache, and allow a cheaper summarizer (#1313) (07a5780)
- compaction: keep compacting while async tool jobs run (#1317) (2803104)
- compaction: keep summaries from degrading across repeated compactions (#1303) (e56768a)
- compaction: recover from context-window overflows instead of failing the run (#1282) (7e614a7)
5.5.0 (2026-09-21) ​
Features ​
5.4.1 (2026-09-21) ​
Bug Fixes ​
5.4.0 (2026-09-10) ​
Features ​
Bug Fixes ​
- codex: preserve encrypted reasoning across model switches (#1198) (2a37242)
- sdk: keep tool history paired when cancelling agents (#1199) (d8a2e87)
5.3.0 (2026-09-06) ​
Features ​
Bug Fixes ​
- codex: preserve compaction summaries across model switches (#1181) (af9e72e)
- tui: keep agent notifications distinct from human messages (#1182) (ab689b5)
5.2.1 (2026-09-05) ​
Bug Fixes ​
Performance Improvements ​
5.2.0 (2026-09-04) ​
Features ​
- openai: forward mid-turn steering over the Codex websocket on gpt-6-astra (#1150) (3628bda)
- openai: send reasoning changes as configuration_update on gpt-6-astra (#1149) (7928476)
- rho-sdk: run async tool calls as detached jobs (#1151) (ed4c276)
5.1.2 (2026-09-03) ​
Bug Fixes ​
5.1.1 (2026-08-30) ​
Bug Fixes ​
5.1.0 (2026-08-30) ​
Features ​
- sdk: put the requested capability on after_tool_use (#1090) (90f1a0e)
- tui: add /side overlay for frozen-context asides (#1094) (3ebbdf9)
5.0.0 (2026-08-27) ​
⚠ BREAKING CHANGES ​
- rho-sdk removes RunEvent::ProviderActivity and the PROVIDER_ACTIVITY_* constants, collapses ProviderStreamResetReason retryable variants into RetryableFailure { kind, retry_after }, removes ModelCallMetrics::output_tokens_per_second, adds revision to RunEvent::Failed, folds ContextEstimated into StepStarted, and replaces the stringly generation-output-token carriers with a typed ModelCallMetrics::generation_output_tokens metric and ModelEvent::GenerationOutputTokens variant. rho-providers adds ProviderId::OpenAiCompatible for config-defined hosts, folds CustomProviderOptions into CustomProviderSpec (removing the _with_lookup/ _with_options funnels), makes providers() include custom hosts (removing visible_providers), removes deprecated registry API-base re-exports and ProviderRuntime::same_family, and drops the CredentialStore parameter from OpenAiProvider::new_with_identity. rho-tools removes the edit_file tool-name alias, the write_file frontmatter alias, and the EditToolKind type alias.
Code Refactoring ​
4.2.0 (2026-08-21) ​
Features ​
4.1.1 (2026-08-20) ​
Bug Fixes ​
4.1.0 (2026-08-17) ​
Features ​
4.0.1 (2026-08-15) ​
Bug Fixes ​
4.0.0 (2026-08-12) ​
⚠ BREAKING CHANGES ​
- hooks:
AuthorizationDenialKindandApprovalAuditDecisiongain a hook variant. Both are#[non_exhaustive], so only exhaustive matches written before this change need updating. - xai: ModelEvent::WebSearch and RunEvent::WebSearch now carry a name field so hosts can distinguish web_search from x_search.
Features ​
- add advisor mode with a selectable advisor model (#752) (13c1ebb)
- config: mid-session edit tool, advisor, and auto preference (#840) (423d026)
- hooks: add typed lifecycle hooks (#668) (4a69c3d)
- openai: add fast mode (#610) (8c5cd6d)
- permission: rename Auto to Bypass and add classifier Auto (#870) (3192daa)
- providers: add native Google Gemini support (#430) (34ef307)
- providers: add Qwen Token Plan OpenAI-compatible provider (#738) (6aa6df2)
- providers: use OpenAI server-side compaction for codex and api-key (#514) (b18eadd)
- questionnaire: add focused default selection (#530) (47d1853)
- questionnaire: support choice descriptions (#510) (066899c)
- sdk: add embeddable Rho runtime (#262) (6fdac81)
- sdk: execute independent tool calls concurrently (#459) (0bb5a83)
- sessions: add conversation tree navigation (#474) (8abb138)
- subagents: add parent-child plain-text messaging for Rho runtime (#852) (dd25d8e)
- subagents: route background questionnaires to parent (#539) (e0cab31)
- tui: add retractable pending input (#334) (5f293a2)
- tui: redesign questionnaire with tabbed question layout (#369) (a90135a)
- tui: render read file image previews (#393) (52165ec)
- tui: show detailed activity stages (#403) (267a47b)
- tui: show model output token rate (#623) (a5aa688)
- usage: add durable request ledger (#381) (0502b99)
- workflow: add deterministic DAG workflows (#680) (77f36f9)
- xai: support hosted x_search tool (#647) (cd0c897)
Bug Fixes ​
- exclude reasoning tokens from throughput (#819) (d261b5b)
- kimi: use provider-native K3 reasoning (#402) (5453cdc)
- metrics: include reasoning latency in output rate (#632) (7f7fa39)
- openai: align Codex Responses wire contracts (#644) (76cf855)
- poolside: publish final stream usage snapshot (#516) (d51ebab)
- providers: show bounded error diagnostics (#344) (e3fc489)
- providers: surface rate-limit reset time and /limits pointer (#733) (b9371fc)
- sdk: commit turn history on provider failure (#739) (b4158ab)
- sdk: recover failed 1.17.1 release packaging (#587) (224189e)
- sdk: recover failed 1.32.0 release packaging (#792) (a782145)
- sdk: retry retryable provider failures instead of failing the run (#401) (b2867da)
- sdk: serialize identical concurrent approval requests to prompt once (#469) (45b2b0a)
- sdk: stabilize compaction release benchmark (#561) (8364edc)
- skills: enforce manual skill invocation (#453) (4f6f043)
- subagents: run multi-agent batches in parallel (#544) (7dd6706)
- tools: allow file paths outside workspace (#537) (8a3cc24)
- tools: scrub provider credential env vars from child processes (#502) (6d66913)
- tui: open approval prompts at the start and default to deny (#636) (59efc07)
- tui: report generation token throughput (#803) (4772f68)
- tui: report herdr blocked during questionnaire (#500) (1dcdbe9)
- tui: show codex fast mode and report tier fallback (#663) (177043f)
- tui: show hosted x_search tool cards (#662) (4381667)
- tui: stabilize live stream tool cards and markdown previews (#595) (752794f)
Performance Improvements ​
- optimize orchestration and session hot paths (#603) (62aa8f5)
- reduce hot-path allocations and redundant I/O (#280) (c18e582)
- sdk: avoid duplicate history clone during compaction (#276) (79e3926)
3.0.0 (2026-08-11) ​
⚠ BREAKING CHANGES ​
- hooks:
AuthorizationDenialKindandApprovalAuditDecisiongain a hook variant. Both are#[non_exhaustive], so only exhaustive matches written before this change need updating. - xai: ModelEvent::WebSearch and RunEvent::WebSearch now carry a name field so hosts can distinguish web_search from x_search.
Features ​
- add advisor mode with a selectable advisor model (#752) (13c1ebb)
- config: mid-session edit tool, advisor, and auto preference (#840) (423d026)
- hooks: add typed lifecycle hooks (#668) (4a69c3d)
- openai: add fast mode (#610) (8c5cd6d)
- providers: add native Google Gemini support (#430) (34ef307)
- providers: add Qwen Token Plan OpenAI-compatible provider (#738) (6aa6df2)
- providers: use OpenAI server-side compaction for codex and api-key (#514) (b18eadd)
- questionnaire: add focused default selection (#530) (47d1853)
- questionnaire: support choice descriptions (#510) (066899c)
- sdk: add embeddable Rho runtime (#262) (6fdac81)
- sdk: execute independent tool calls concurrently (#459) (0bb5a83)
- sessions: add conversation tree navigation (#474) (8abb138)
- subagents: add parent-child plain-text messaging for Rho runtime (#852) (dd25d8e)
- subagents: route background questionnaires to parent (#539) (e0cab31)
- tui: add retractable pending input (#334) (5f293a2)
- tui: redesign questionnaire with tabbed question layout (#369) (a90135a)
- tui: render read file image previews (#393) (52165ec)
- tui: show detailed activity stages (#403) (267a47b)
- tui: show model output token rate (#623) (a5aa688)
- usage: add durable request ledger (#381) (0502b99)
- workflow: add deterministic DAG workflows (#680) (77f36f9)
- xai: support hosted x_search tool (#647) (cd0c897)
Bug Fixes ​
- exclude reasoning tokens from throughput (#819) (d261b5b)
- kimi: use provider-native K3 reasoning (#402) (5453cdc)
- metrics: include reasoning latency in output rate (#632) (7f7fa39)
- openai: align Codex Responses wire contracts (#644) (76cf855)
- poolside: publish final stream usage snapshot (#516) (d51ebab)
- providers: show bounded error diagnostics (#344) (e3fc489)
- providers: surface rate-limit reset time and /limits pointer (#733) (b9371fc)
- sdk: commit turn history on provider failure (#739) (b4158ab)
- sdk: recover failed 1.17.1 release packaging (#587) (224189e)
- sdk: recover failed 1.32.0 release packaging (#792) (a782145)
- sdk: retry retryable provider failures instead of failing the run (#401) (b2867da)
- sdk: serialize identical concurrent approval requests to prompt once (#469) (45b2b0a)
- sdk: stabilize compaction release benchmark (#561) (8364edc)
- skills: enforce manual skill invocation (#453) (4f6f043)
- subagents: run multi-agent batches in parallel (#544) (7dd6706)
- tools: allow file paths outside workspace (#537) (8a3cc24)
- tools: scrub provider credential env vars from child processes (#502) (6d66913)
- tui: open approval prompts at the start and default to deny (#636) (59efc07)
- tui: report generation token throughput (#803) (4772f68)
- tui: report herdr blocked during questionnaire (#500) (1dcdbe9)
- tui: show codex fast mode and report tier fallback (#663) (177043f)
- tui: show hosted x_search tool cards (#662) (4381667)
- tui: stabilize live stream tool cards and markdown previews (#595) (752794f)
Performance Improvements ​
- optimize orchestration and session hot paths (#603) (62aa8f5)
- reduce hot-path allocations and redundant I/O (#280) (c18e582)
- sdk: avoid duplicate history clone during compaction (#276) (79e3926)
2.0.0 (2026-08-10) ​
⚠ BREAKING CHANGES ​
- hooks:
AuthorizationDenialKindandApprovalAuditDecisiongain a hook variant. Both are#[non_exhaustive], so only exhaustive matches written before this change need updating. - xai: ModelEvent::WebSearch and RunEvent::WebSearch now carry a name field so hosts can distinguish web_search from x_search.
Features ​
- add advisor mode with a selectable advisor model (#752) (13c1ebb)
- config: mid-session edit tool, advisor, and auto preference (#840) (423d026)
- hooks: add typed lifecycle hooks (#668) (4a69c3d)
- openai: add fast mode (#610) (8c5cd6d)
- providers: add native Google Gemini support (#430) (34ef307)
- providers: add Qwen Token Plan OpenAI-compatible provider (#738) (6aa6df2)
- providers: use OpenAI server-side compaction for codex and api-key (#514) (b18eadd)
- questionnaire: add focused default selection (#530) (47d1853)
- questionnaire: support choice descriptions (#510) (066899c)
- sdk: add embeddable Rho runtime (#262) (6fdac81)
- sdk: execute independent tool calls concurrently (#459) (0bb5a83)
- sessions: add conversation tree navigation (#474) (8abb138)
- subagents: route background questionnaires to parent (#539) (e0cab31)
- tui: add retractable pending input (#334) (5f293a2)
- tui: redesign questionnaire with tabbed question layout (#369) (a90135a)
- tui: render read file image previews (#393) (52165ec)
- tui: show detailed activity stages (#403) (267a47b)
- tui: show model output token rate (#623) (a5aa688)
- usage: add durable request ledger (#381) (0502b99)
- workflow: add deterministic DAG workflows (#680) (77f36f9)
- xai: support hosted x_search tool (#647) (cd0c897)
Bug Fixes ​
- exclude reasoning tokens from throughput (#819) (d261b5b)
- kimi: use provider-native K3 reasoning (#402) (5453cdc)
- metrics: include reasoning latency in output rate (#632) (7f7fa39)
- openai: align Codex Responses wire contracts (#644) (76cf855)
- poolside: publish final stream usage snapshot (#516) (d51ebab)
- providers: show bounded error diagnostics (#344) (e3fc489)
- providers: surface rate-limit reset time and /limits pointer (#733) (b9371fc)
- sdk: commit turn history on provider failure (#739) (b4158ab)
- sdk: recover failed 1.17.1 release packaging (#587) (224189e)
- sdk: recover failed 1.32.0 release packaging (#792) (a782145)
- sdk: retry retryable provider failures instead of failing the run (#401) (b2867da)
- sdk: serialize identical concurrent approval requests to prompt once (#469) (45b2b0a)
- sdk: stabilize compaction release benchmark (#561) (8364edc)
- skills: enforce manual skill invocation (#453) (4f6f043)
- subagents: run multi-agent batches in parallel (#544) (7dd6706)
- tools: allow file paths outside workspace (#537) (8a3cc24)
- tools: scrub provider credential env vars from child processes (#502) (6d66913)
- tui: open approval prompts at the start and default to deny (#636) (59efc07)
- tui: report generation token throughput (#803) (4772f68)
- tui: report herdr blocked during questionnaire (#500) (1dcdbe9)
- tui: show codex fast mode and report tier fallback (#663) (177043f)
- tui: show hosted x_search tool cards (#662) (4381667)
- tui: stabilize live stream tool cards and markdown previews (#595) (752794f)
Performance Improvements ​
- optimize orchestration and session hot paths (#603) (62aa8f5)
- reduce hot-path allocations and redundant I/O (#280) (c18e582)
- sdk: avoid duplicate history clone during compaction (#276) (79e3926)
1.17.3 (2026-08-08) ​
Bug Fixes ​
1.17.2 (2026-08-07) ​
Bug Fixes ​
1.17.1 (2026-08-07) ​
Bug Fixes ​
1.17.0 (2026-08-06) ​
Features ​
1.16.0 (2026-08-04) ​
Features ​
1.15.2 (2026-08-04) ​
Bug Fixes ​
1.15.1 (2026-08-03) ​
Bug Fixes ​
1.15.0 (2026-08-02) ​
Features ​
1.14.0 (2026-07-31) ​
Features ​
1.13.1 (2026-07-30) ​
Bug Fixes ​
- tui: show codex fast mode and report tier fallback (#663) (177043f)
- tui: show hosted x_search tool cards (#662) (4381667)
1.13.0 (2026-07-30) ​
Features ​
Bug Fixes ​
1.12.2 (2026-07-29) ​
Bug Fixes ​
1.12.1 (2026-07-28) ​
Bug Fixes ​
1.12.0 (2026-07-28) ​
Features ​
1.11.0 (2026-07-28) ​
Features ​
Performance Improvements ​
1.10.3 (2026-07-27) ​
Bug Fixes ​
1.10.2 (2026-07-27) ​
Bug Fixes ​
1.10.1 (2026-07-26) ​
Bug Fixes ​
1.10.0 (2026-07-25) ​
Features ​
Bug Fixes ​
- subagents: run multi-agent batches in parallel (#544) (7dd6706)
- tools: allow file paths outside workspace (#537) (8a3cc24)
1.9.0 (2026-07-24) ​
Features ​
1.8.0 (2026-07-23) ​
Features ​
- providers: use OpenAI server-side compaction for codex and api-key (#514) (b18eadd)
- questionnaire: support choice descriptions (#510) (066899c)
Bug Fixes ​
1.7.2 (2026-07-22) ​
Bug Fixes ​
- tools: scrub provider credential env vars from child processes (#502) (6d66913)
- tui: report herdr blocked during questionnaire (#500) (1dcdbe9)
1.7.1 (2026-07-22) ​
Bug Fixes ​
1.7.0 (2026-07-22) ​
Features ​
1.6.0 (2026-07-21) ​
Features ​
Bug Fixes ​
1.5.0 (2026-07-20) ​
Features ​
1.4.0 (2026-07-20) ​
Features ​
Bug Fixes ​
- kimi: use provider-native K3 reasoning (#402) (5453cdc)
- sdk: retry retryable provider failures instead of failing the run (#401) (b2867da)
1.3.0 (2026-07-18) ​
Features ​
1.2.0 (2026-07-17) ​
Features ​
- tui: redesign questionnaire with tabbed question layout (#369) (a90135a)
- usage: add durable request ledger (#381) (0502b99)