v0.6.0-alpha.1

LocalAgent v0.6.0-alpha.1

LocalAgent v0.6.0-alpha.1 improves the core coding-agent loop with structured planning, read-only LSP code intelligence, replayable tool-result artifacts, replay artifact inspection, and bounded validator-driven repair.

This is an alpha release focused on making LocalAgent more effective for real coding tasks while preserving its safety model: explicit trust gates, approvals, replayability, auditability, and bounded continuation.

Highlights

Stronger agent planning

• Added update_plan for structured task progress.
• Plan updates are tracked through the runtime/event path.
• The agent can now maintain clearer task state during multi-step coding work.

Cleaner model-facing tool observations

• Built-in tool results are now summarized into compact openagent.observation.v1 observations.
• Observations include high-signal evidence such as status, command, exit code, duration, stdout/stderr tails, diagnostics, changed files, failure class, and deterministic next-action hints.
• Raw tool output is no longer dumped directly into the model transcript for built-in tools.

Replayable raw tool-result artifacts

• Built-in raw tool results are preserved as run artifacts.

• Added raw artifact refs for:

  • raw_result_ref
  • stdout_ref
  • stderr_ref

• Added run-scoped tool-result manifests.

• Added a safe resolver for artifact refs with traversal protections.

Replay artifact inspection CLI

Added commands for inspecting persisted tool-result artifacts:

localagent replay artifacts <run_id> [--json]
localagent replay artifact <run_id> <ref> [--path-only] [--json]

These commands make it easier to inspect compact observation refs and recover the raw evidence behind a run.

Read-only LSP tools

Added gated read-only LSP tools behind:

--lsp-provider <rust|typescript>

New tools:

• lsp.diagnostics
• lsp.document_symbols
• lsp.goto_definition
• lsp.find_references
• lsp.hover

These tools are workdir-scoped, read-only, classified as filesystem-read, and routed through the existing gate/audit path.

Bounded validator-driven repair

LocalAgent can now run inferred Rust/Cargo validation after effective Rust/Cargo writes and allow one extra repair turn when validation fails.

Current validator behavior is intentionally conservative:

• Rust/Cargo-focused.
• Uses the existing shell/gate/audit path.
• Allows one bounded repair turn.
• Stops after the repair budget is exhausted.
• Preserves trust gates, approvals, replay behavior, MCP behavior, and continuation bounds.

Validation

This release passed:

cargo fmt --check
cargo clippy -- -D warnings
cargo test
cargo test --test tool_call_accuracy_ci
python scripts\ci_release_readiness.py
git diff --check

Known limitations

• This is an alpha release.
• The release tag is v0.6.0-alpha.1, while crate metadata still reports v0.5.0.
• Validator-driven repair is currently Rust/Cargo-focused.
• Rust validation currently uses a chained command, so later checks do not run if an earlier check fails.
• LSP tools are read-only and require --lsp-provider <rust|typescript>.
• TUI artifact surfacing is not wired yet.

Notes

This release is focused on improving LocalAgent’s coding-agent loop without weakening the project’s core safety model. It does not add unbounded autonomous continuation, permissive default writes, or background agent behavior.

LocalAgent v0.6.0-alpha.2

LocalAgent v0.6.0-alpha.2 Release Notes

Release date: 2026-07-04

Highlights

• Split inferred Rust validation into ordered validator outcomes:
  • cargo fmt --check
  • cargo test
  • cargo clippy -- -D warnings
• Validator steps now record pass, fail, or skipped outcomes.
• Rust validation stops after the first failed validator and marks later validators as skipped.
• Preserved the existing one-repair budget, shell/gate/audit path, and bounded continuation behavior.

Release Process

• Updated crate metadata to 0.6.0-alpha.2.
• Release readiness now supports validating that a release tag matches Cargo.toml package metadata.

v0.5.0

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.

v0.4.0

LocalAgent v0.4.0 Release Notes

Release date: 2026-03-03

Highlights

• Completed Tool Call Accuracy delivery through PR12 with stricter runtime/tool contracts and deterministic behaviors.
• Added native glob and grep builtins for safe, non-shell file discovery/search workflows.
• Improved shell-call robustness with better compatibility handling and one-shot auto-repair for common program not found failures.
• Upgraded TUI runtime flow with improved approvals behavior and a dedicated Reasoning side pane (Ctrl+4).
• Added a dedicated manual TUI testing pack and archived historical planning/spec docs for cleaner ongoing documentation.

Included Changes

Tool Runtime and Contract Accuracy (PR2-PR12)

• Added tool-call repair gating and repeat-loop guardrails.
• Added runtime record persistence and versioned runtime tool contract prompting.
• Added deterministic tool-call-accuracy CI harness.
• Normalized provider sampling/controls (temperature, top_p, max_tokens, and seed handling).
• Implemented agent-mode build/plan runtime semantics.
• Implemented JSON run-output event projection mode.

New Builtin Tools and Safe Defaults

• Added glob builtin with scoped path behavior and deterministic match handling.
• Added grep builtin with bounded, deterministic text-search behavior.
• Enabled safer default trust policy posture for tool execution.

Shell and Approval Flow Improvements

• Improved shell tool compatibility and error classification for clearer fallback behavior.
• Added one-shot shell auto-repair for command-not-found style failures.
• Improved approvals UX with auto-refresh behavior, pending indicators, and auto-open refinements.

TUI UX

• Added a dedicated right-side Reasoning pane for live/last-run model reasoning visibility.
• Added global Ctrl+4 pane toggle handling (including terminal control-character fallback).
• Reasoning pane is hidden on banner and auto-shows after first prompt submission.

Documentation and Testing Assets

• Added PR8-PR12 spec docs and a manual TUI testing pack.
• Archived historical tool-call accuracy specs (PR2-PR12) under docs/archive/.
• Aligned docs index and references after archive/reorg.

Compatibility Notes

• glob/grep are additive builtins and do not remove existing tooling paths.
• Tool-call runtime behavior is stricter/more deterministic; integrations should continue ignoring unknown JSON fields for forward compatibility.
• Shell execution behavior may differ by trust mode and environment policy; blocked or unavailable shell runs now provide clearer error semantics.

v0.3.1

LocalAgent v0.3.1 Release Notes

Release date: 2026-02-28

Highlights

• Simplified Learn Overlay flow for faster capture/review/promote operation.
• Improved TUI input ergonomics with caret visibility and arrow-key cursor navigation.
• Aligned docs to shipped /learn behavior with explicit workflow and output contracts.

Included Changes

Learn Overlay UX

• Removed the write-arm step from overlay submit flow.
• Capture now saves directly on Enter.
• Promote now publishes directly on Enter when required fields are present.
• Promote controls remain intentionally beginner-focused (target + force).
• Advanced promote options remain available through typed /learn promote ... or CLI:
  • --check-run
  • --replay-verify
  • --replay-verify-run-id <RUN_ID>
  • --replay-verify-strict

TUI Reliability and Input Handling

• Added blinking caret for active text fields.
• Added arrow-key cursor navigation in overlay text inputs.
• Prevented Learn capture row-index panic.
• Removed overlay letter shortcuts that interfered with typing.
• Improved active-run slash handling and pane separation behavior.
• Bounded/normalized overlay paste and improved long-text wrapping.

Documentation and Reference Alignment

• Added and aligned /learn reference docs:
  • docs/reference/LEARN_WORKFLOW_REFERENCE.md
  • docs/reference/LEARN_OUTPUT_CONTRACT.md
• Updated README and CLI reference for current Learn Overlay behavior.
• Reorganized docs and archived historical scope documents under docs/archive/.

Compatibility Notes

• No intentional breaking CLI changes in this patch release.
• Learn artifacts/events remain additive and deterministic; consumers should continue to ignore unknown fields for forward compatibility.

v0.3.0

latest

LocalAgent v0.3.0 Release Notes

Release date: 2026-02-27

Highlights

• Shipped the full learn roadmap from capture through explicit promotion targets.
• Added assisted learning capture with provenance metadata and preview-first write gating.
• Delivered Chat TUI /learn support (Phase A + Phase B) with CLI-parity behavior.

Included Changes

Learning Store and Promotion

• Added learn capture, learn list, learn show <id>, and learn archive <id>.
• Added learn promote <id> --to check --slug <slug>.
• Added learn promote <id> --to pack --pack-id <pack_id>.
• Added learn promote <id> --to agents.
• Added deterministic promotion events (openagent.learning_promoted.v1) including target file hash metadata.

Assisted Capture and Chaining

• Added assisted draft flow: learn capture --assist (preview-only by default) and --write to persist.
• Added provenance metadata fields for assisted captures.
• Added one-shot promote+validate chaining options (--check-run and replay verify paths).

Chat TUI

• Added /learn help, /learn list, /learn show <id>, /learn archive <id>.
• Added /learn capture ... and /learn promote ... passthrough with deterministic quoting/tokenization handling.
• Added busy-state rejection behavior for slash commands during active run/tool work.

Reliability Fixes

• Fixed cancellation-channel lifetime handling so chat/TUI runs no longer terminate immediately as cancelled due to early sender drop.

Compatibility Notes

• Changes are additive at the CLI/behavior layer for existing users.
• Learning artifacts/events added new fields and event kinds; consumers should continue to ignore unknown fields for forward compatibility.

v0.2.1

v0.2.1

Patch release following v0.2.0 with CI/readiness fixes only.

Included fixes

• Applied rustfmt normalization across newly extracted runtime modules.
• Resolved cargo fmt --check CI failure.
• Fixed strict clippy CI (cargo clippy -- -D warnings) by adding a targeted allow on:
  - prepare_tools_and_qualification in src/run_prep.rs
• No intended runtime behavior changes.

Impact

• Improves build/release reliability for contributors and CI.
• Safe upgrade from v0.2.0 for all users.

Upgrade

cargo install --path . --force
localagent version

v0.2.0

LocalAgent v0.2.0 Release Notes

Date: 2026-02-25

Highlights

• Runtime architecture refactor: major decomposition of main.rs into focused runtime modules.
• Startup UX improvement: automatic .localagent/ initialization on first project use.
• Planner/chat/task execution paths hardened with preserved deterministic behavior and full regression coverage.
• Documentation alignment for current behavior, especially init flow and in-chat timeout controls.

What Shipped

Runtime Modularization

The runtime surface was split into dedicated modules to reduce coupling and improve maintainability:

• agent_runtime
• chat_tui_runtime
• chat_repl_runtime
• tasks_graph_runtime
• startup_bootstrap
• startup_detect
• startup_init
• runtime_wiring
• runtime_paths
• runtime_events
• runtime_flags
• task_apply
• task_eval_profile
• ops_helpers
• session_ops
• approvals_ops
• instruction_runtime
• planner_runtime

main.rs now primarily acts as command orchestration and entrypoint wiring.

Startup / Init Behavior

• LocalAgent now auto-initializes .localagent/ on first command use in a project when missing.
• localagent init remains available for explicit, deterministic scaffold generation.

Chat and Timeout UX

• Slash-command timeout behavior is documented and aligned with runtime behavior:
  • /timeout <seconds|+N|-N|off>
  • /timeout off disables request/stream-idle timeout (connect timeout unchanged)

Docs and Guidance Alignment

• Updated docs to reflect:
  • auto-init behavior
  • .localagent/instructions.yaml as the canonical profile path
  • current timeout command semantics

Behavior Notes

• No intentional breaking CLI flag removals in this release.
• Runtime internals were substantially reorganized; integrations depending on internal module layout should update references accordingly.

Verification Summary

• cargo check passes.
• cargo test --workspace passes.
• Regression harness remains green, including MCP/protocol guard tests.

Upgrade

cargo install --path . --force
localagent version

Recommended Post-Upgrade Checks

1. Run localagent in a fresh project dir to confirm auto-init behavior.
2. Validate provider connectivity via localagent doctor --provider <provider>.
3. Run one coding-task prompt in chat TUI and verify tool/approval/log panes behave as expected.
4. If using instruction profiles, verify path/config resolution under .localagent/instructions.yaml.

v0.1.3

LocalAgent v0.1.3 Release Notes

Date: 2026-02-24

Highlights

• Runtime control-loop hardening for planner/worker execution.
• MCP lifecycle visibility expanded with progress, cancellation, and pin enforcement telemetry.
• Deterministic boundedness tightened across runtime/tool/schema paths.
• Trace and eval gating strengthened to catch regressions earlier.

What Shipped

Runtime Enforcement

• Enforced planner control envelope and separated control-plane output from user-facing output.
• Enforced typed step-status transitions and stronger step invariants in planner-worker loops.
• Added controlled replan flow and deterministic retry behavior for tool failures.

Boundedness and Reliability

• Added hard runtime budgets with explicit budget-exceeded exits.
• Added runtime tool-call budgets and bounded schema-repair retry before tool execution.
• Added eval retry/failure-class metrics and regression gates.

MCP Lifecycle and Drift Controls

• Added MCP progress and cancellation lifecycle events.
• Added MCP runtime trace persistence and lifecycle continuity checks.
• Added MCP tool catalog snapshotting and live drift verification.
• Added MCP pin metadata, drift visibility, and configurable pin enforcement modes:
  • hard
  • warn
  • off
• Surfaced MCP enforcement mode and diagnostics in the TUI status views.

TUI Operator Visibility

• Added MCP lifecycle/cancellation/progress badges and stall timers.
• Improved guardrail/reason taxonomy visibility and step-level runtime hints.
• Added two-phase cancel visibility and clearer cancellation completion signaling.

Breaking / Behavior Changes

• Planner-worker completion is now runtime-gated; model outputs no longer imply completion on their own.
• Budget and policy failures terminate with deterministic, categorized exit reasons.
• MCP pin enforcement may block drifted tool catalogs when enforcement is set to hard.

Upgrade

cargo install --path . --force
localagent --help

Recommended Post-Upgrade Checks

1. Verify effective policy and approval behavior in your environment.
2. Validate MCP pin enforcement mode (hard, warn, or off) for your deployment.
3. Run your eval profile/baseline comparison before promoting to production.

Notable Internal Tracking

Primary implementation landed across commits from 4a512f0 through 02cbb7f on main, including runtime boundedness, planner envelope enforcement, MCP lifecycle instrumentation, and pin enforcement diagnostics.

v0.1.2

LocalAgent v0.1.2 Release Notes

Date: 2026-02-22

Highlights

• Startup UI refresh with compact Mode + Provider panes.
• Chat UI refresh with improved header/footer and boxed input row.
• New in-chat mode and timeout controls for faster iteration on local models.

Startup UX

• Updated startup layout:
  • compact Mode + Provider panes
  • centered footer controls
  • provider refresh/details controls (R, D)
• Improved custom flow:
  • selecting Custom opens a compact submenu with Back + custom toggles
  • menu fits default terminal sizes more reliably

Chat UX

• Header now shows mode label (Safe, Code, Web, Custom) and right-justified ?.
• Footer now shows explicit cwd: <absolute path> and right-justified connection status.
• Input area is now a boxed prompt row above footer/overlay.
• Status line above input uses animated wave + rotating thinking/working phrases.
• Keybind overlay (?) rows are aligned uniformly.

New Slash Commands

• /mode <safe|coding|web|custom>
  • switches chat runtime mode in-session
• /timeout
  • shows current timeout settings and prompts for next numeric input
• /timeout <seconds|+N|-N>
  • sets or adjusts request/stream-idle timeout in-session
• /dismiss
  • clears active timeout notification

Timeout Guidance

• Provider timeout-style failures now emit a guidance notice suggesting /timeout.
• Notice can be cleared with /dismiss.

Install / Upgrade

cargo install --path . --force
localagent version