Skip to content

v0.5.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 14 Mar 23:06
v0.5.0
0420112

LocalAgent v0.5.0 Release Notes

Release date: 2026-03-14

v0.5.0 is a major runtime and architecture hardening release. It refactors large parts of LocalAgent's core runtime/tooling surfaces into smaller focused modules, strengthens runtime-owned completion and validation behavior for coding tasks, and expands the coding/eval workflow with broader benchmark coverage, local-model investigation assets, and new tooling integrations.

Highlights

  • Reworked the runtime around checkpoint-backed phases, runtime-owned completion/finalization, explicit coding-task contracts, and tighter validation/exact-closeout enforcement.
  • Completed a broad internal refactor across the runtime, tools, eval, learning, and TUI layers so the codebase is split into smaller focused modules instead of relying on a few oversized orchestration files.
  • Expanded LocalAgent’s coding-task surface with stronger edit/repair paths (str_replace, edit-path normalization, wrapped tool-call recovery) and broader eval/benchmark coverage through common_coding_ux.
  • Added TypeScript LSP context support plus server/runtime harness work, making the repo materially stronger for tool-assisted coding and backend-driven execution flows.
  • Switched one-shot run / exec defaults to isolated ephemeral state and sessionless execution unless the operator opts into persistent state/session settings.
  • Added substantial manual evaluation and local-model investigation assets to support repeatable model/runtime comparisons.

Release Scope

  • Commits since v0.4.0: 239
  • Diff size since v0.4.0: 257 files changed, about 64,073 insertions and 19,884 deletions
  • Primary themes:
    • runtime and architecture hardening
    • coding-task contract enforcement and validation recovery
    • coding/eval benchmark expansion
    • local-model investigation and tooling integration

Included Changes

Runtime Architecture and Completion Semantics

  • Introduced checkpoint-backed runtime phases and stricter runtime-artifact/checkpoint consistency validation.
  • Moved completion/finalization behavior further into runtime-owned control flow, with clearer decision points for:
    • validated completion
    • exact final-answer collection
    • post-write follow-on turns
    • repair/retry boundaries
    • approval/cancel/budget terminal states
  • Formalized explicit runtime validation contracts and exact-final-answer/output contracts for authored tasks and checks.
  • Tightened runtime invariants around verified writes, post-write verification, validation sequencing, and one-tool-call-per-step behavior.
  • Repaired validation-phase shell handoff when a model emits prose-only or wrong-tool turns after a verified edit.
  • Improved inline exact-closeout inference for prompts that say reply with exactly ....
  • Added stricter runtime timeout handling and nonzero timeout defaults in runtime-owned modes.

Tooling, Editing, and Provider Robustness

  • Added str_replace as a small-model-friendly file-edit path and aligned eval assertions/trust defaults around it.
  • Tightened validation flow and edit alias normalization so existing-file coding tasks recover more reliably.
  • Improved malformed wrapped tool-call recovery and LM Studio/OpenAI-compatible message normalization.
  • Improved shell compatibility, shell error classification, and validation-only shell-shape repair behavior.
  • Added TypeScript LSP context support and improved diagnostics robustness for TypeScript/JS coding tasks.

Eval, Benchmarking, and Investigation Workflow

  • Expanded common_coding_ux into the active benchmark decision surface for LocalAgent coding-task improvement work.
  • Added broader task coverage, UX metric rows, baseline artifacts, bundle/report/compare improvements, and helper scripts for coding-task evaluation.
  • Added focused comparison harnesses for closeout behavior and validation-discipline follow-up work.
  • Added substantial manual testing/investigation assets under manual-testing/, including canonical T/D packs, runbooks, logs, and result templates for local-model sweeps.

Runtime Defaults, TUI, and Operator Flow

  • One-shot run / exec now default to ephemeral temp state when --state-dir is not provided.
  • One-shot run / exec also default to --no-session unless session settings are explicitly requested.
  • Added multiline input mode to the TUI chat flow.
  • Improved TUI tool-row reconciliation, verified-write rendering, and post-run status cleanup for complex runs.
  • Closed operator resume/replay gaps and improved runtime/tui coordination around terminal events.

Server, Docs, and Repo Structure

  • Added server runtime foundation and associated harness coverage.
  • Reorganized the repo documentation into clearer architecture/reference/policy/operations/guides sections, with older material archived explicitly.
  • Added and updated runtime-target, runbook, policy, and local-model-improvement docs to match the current codebase.
  • README and release docs now point more directly to the release-notes index for operator-facing release history.

Internal Refactor Scope

  • Split large runtime orchestration paths into focused modules under src/agent/ and src/agent_runtime/.
  • Split tool execution/catalog/schema logic into focused modules under src/tools/.
  • Split eval runner/reporting/output logic into smaller surfaces under src/eval/.
  • Split learning capture/promotion/render/store helpers into focused modules under src/learning/.
  • Continued TUI/chat decomposition so input, overlays, transcript, rendering, approvals, and runtime glue are easier to reason about and test.

Compatibility Notes

  • No intentional breaking schema or artifact version removals were introduced in this release.
  • Runtime/event/task metadata additions remain additive; downstream consumers should continue ignoring unknown fields for forward compatibility.
  • Operators relying on persistent one-shot artifacts should now pass --state-dir <path> explicitly, because the default one-shot path is ephemeral in v0.5.0.
  • Coding-task behavior is stricter in v0.5.0: validation, exact-closeout, and write-verification flows are enforced more explicitly, so models that previously drifted through loosely enforced completion paths may now be asked to repair or continue instead of being accepted early.
Assets 8
Loading