Skip to content

CHANGELOG.md

Latest commit

 

History

History
154 lines (120 loc) · 9.47 KB

CHANGELOG.md

File metadata and controls

154 lines (120 loc) · 9.47 KB
Raw

Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.1.13 - 2026-08-09

Added

  • Experimental concurrent child agents for the native coding workflow, including parent/child messaging, steering, follow-up work, waiting, cancellation, and deletion.
  • Live terminal activity for the agent hierarchy and a replay command for the complete root-and-child trajectory.
  • Durable child status, ancestry, model, sandbox, budget, result, usage, and mailbox records that remain available when a root session resumes.

Security

  • Child repository and command actions use the same configured sandbox, approval policy, limits, and audit trail as the root agent.

Known limitations

  • Background children do not continue running after the RLM Code process exits. Resumed sessions recover their hierarchy, completed results, supported Python checkpoints, and undelivered messages; daemon-backed execution and automatic crash recovery remain future work.

0.1.12 - 2026-08-09

Added

  • Experimental rlm-code agent run workflow for repository coding tasks.
  • A persistent Python workspace that keeps variables and working context available between model turns.
  • Repository search, file updates, command execution, and test runs under the selected sandbox and approval policy.
  • Live terminal visibility for Python activity, repository and command actions, approvals, usage, and final results.
  • Usage accounting, cancellation, and configurable turn and time limits.
  • Saved root-agent sessions that can resume their conversation and supported Python state.

Security

  • Repository and command actions are checked by RLM Code's path, sandbox, approval, and audit controls.
  • Docker or Apple Container is recommended for isolation. The local runtime is intended for development in trusted environments.

Known limitations

  • The experimental native agent currently runs as a single root agent.
  • Concurrent child agents, parent/child communication, an agent-tree view, and replay or resume of a complete multi-agent hierarchy are not included in this release.

0.1.11 - 2026-07-20

Added

  • Opt-in repo_evidence and locally in-distribution (lid) Pure RLM harness profiles with focused decomposition guidance.
  • Deterministic mini, evidence, full, and explicit repository context profiles through the public RepositoryContextBuilder API.
  • Opaque, constant-shape root observations; structural root history; and automatic offloading of older history to versioned REPL variables.
  • Root/submodel call attribution, bounded subcall trace previews with hashes, and harness trajectory similarity metrics.
  • Benchmark metadata for explicit context, expected answers, task family, domain, split, and length buckets.
  • API-key-free cross-domain harness proof with an 8× evaluation-length extrapolation.
  • Maintained AI Engineer World's Fair 2026 live probe, prompts, and use-case notes under examples/aie_world_fair_2026.

Changed

  • Pure RLM caller-provided contexts are preserved rather than replaced by automatic runner discovery.
  • Evidence context selection ranks matching files before applying file budgets.
  • Incomplete repo_evidence and lid runs sanitize trajectories before root-model fallback synthesis.
  • Release source distributions now include the examples/ directory.
  • Corrected the RLM paper link to arXiv:2512.24601.

Fixed

  • Pure RLM paradigm comparison now passes its supplied context through the runner and reports measured submodel calls.
  • Configuration parsing now handles absent or mocked harness settings without changing compatibility defaults.

0.1.10 - 2026-06-28

Fixed

  • Accept new-format Google Gemini API keys (prefix AQ.) in addition to legacy AIza keys. Both the connector key validation (models/llm_connector.py) and the config key-format pattern (validation/config_validator.py) now recognize AQ. keys, so /connect gemini ... and CLI connections work with keys issued in Google's current format.

0.1.9 - 2026-06-26

Added

  • Pure RLM runner context initialization from explicit workspace file references in the task, with compact repository snapshot fallback.
  • Context-load events for Pure RLM runs, including loaded file names and total context characters.
  • Runner JSONL replay coverage for action code, observations, success state, token counts, and cumulative reward.

Changed

  • TUI trajectory and replay views now show Pure RLM signals including REPL code, stdout/stderr previews, llm_query counts, executed code blocks, finalization status, and REPL variables.
  • Run visualization now includes richer Pure RLM previews for completed runs.

0.1.8 - 2026-05-01

Added

  • AHE-style layered trace evidence corpus export from TraceStore.
  • New trace_analysis action export_evidence_corpus for writing overview.md, per-trace detail reports, index.json, and optional processed raw JSONL spans.
  • Evidence corpus tests covering direct store export and environment action export.

0.1.7 - 2026-04-30

Added

  • HALO-style trace_analysis RLM environment for diagnosing agent harness failures from one-span-per-line JSONL traces.
  • Trace sidecar indexing with dataset rollups for trace counts, span counts, error traces, services, models, agents, token totals, and sample trace ids.
  • Bounded trace inspection actions: get_dataset_overview, query_traces, count_traces, view_trace, search_trace, and view_spans.
  • Large-trace safeguards: per-attribute truncation, oversized trace summaries, and higher-cap selected-span reads.
  • Tests for trace indexing, querying, searching, selected-span viewing, and trace environment actions.
  • Trace analysis documentation under the Core Engine docs.

Changed

  • /rlm command help now advertises env=trace_analysis for run, chat, and doctor workflows.

0.1.6 - 2026-02-20

Added

  • Harness strategy selector with tool_call (default) and opt-in codemode.
  • CodeMode execution flow in harness: MCP tool discovery (search_tools), typed tool surface prompt, single-program generation, guardrail validation, and MCP chain execution (call_tool_chain).
  • Benchmark support for harness strategy comparison with CodeMode telemetry fields (harness_strategy, codemode_chain_calls, codemode_search_calls, codemode_discovery_calls, codemode_guardrail_blocked).
  • New top-level CodeMode docs section with dedicated pages for quickstart, architecture, guardrails, and evaluation.
  • Release documentation set for CodeMode:
    • quickstart and operator workflow
    • integration architecture and runtime controls
    • provider/bridge separation model (Cloudflare-based, UTCP, custom)
    • CodeMode sandbox responsibility and deployment matrix
    • guardrail policy and safety runbook
    • benchmark evaluation and promotion-gate criteria

Changed

  • /harness run supports strategy=tool_call|codemode and mcp_server=<name>.
  • /rlm bench in mode=harness supports strategy=tool_call|codemode.
  • Harness and benchmark command handling now auto-enables MCP when strategy=codemode is selected.

Security

  • Added explicit CodeMode guardrail policy documentation with blocked API classes and runtime limit defaults.
  • Codemode path remains opt-in; default harness behavior remains strict baseline strategy=tool_call.

0.1.5 - 2026-02-15

Initial public release of RLM Code.

Added

  • Unified Textual TUI with tabs for RLM, Files, Details, Shell, and Research.
  • Recursive execution engine with multiple patterns: pure RLM, harness/code-agent, and direct LLM flows.
  • Research workflows: run tracking, trajectory capture, replay, benchmark presets, compare/report flows.
  • Sandbox runtime layer (Superbox) with profile-driven runtime selection and fallback orchestration.
  • Secure runtime options including Docker and Monty, plus pluggable runtime adapters.
  • LLM integrations for cloud and local model routes, including BYOK workflows and ACP connectivity.
  • Coding harness with optional MCP tool integration for local/BYOK development workflows.
  • Framework adapter surface for RLM-style integrations (including DSPy-native and ADK-oriented paths).
  • Observability integrations (MLflow, LangFuse, Logfire, LangSmith, OpenTelemetry) via sink architecture.
  • Documentation site (MkDocs Material) with onboarding, CLI, TUI, sandbox, integrations, and benchmark guides.

Changed

  • Project identity standardized as RLM Code (legacy inherited naming removed from repository-facing surfaces).
  • Packaging and project metadata prepared for open-source release.
  • License updated to Apache-2.0.

Security

  • Safer sandbox-first runtime guidance in docs and configuration defaults.
  • Unsafe local exec usage preserved only as an explicit, opt-in path for advanced development scenarios.