v0.25.0

Changelog

[0.25.0] — 2026-09-04

Other

• chore(deps): bump dravr-tronc 0.11.0 -> 1.0.0

[0.24.0] — 2026-09-04

[0.23.0] — 2026-09-02

[0.22.6] — 2026-09-01

[0.22.5] — 2026-09-01

Fixed

• fix(acp): the model a turn runs on cannot be changed by a concurrent spawn

Other

• chore(tronc): move to dravr-tronc 0.8.0

[0.22.4] — 2026-08-30

[0.22.3] — 2026-08-30

Other

• perf(acp): the production chat path runs on the pool, and only reads its own session

[0.22.2] — 2026-08-28

Fixed

• fix(ci): clear the expect/panic gate, red on main since db6de60
• fix(acp): derive the session ceiling from measured memory, not a guess

[0.22.1] — 2026-08-28

[0.22.0] — 2026-08-27

Added

• feat(usage): TokenUsage carries the cache counts providers already send
• feat(examples): a probe that asks the model whether it owns a host tool

[0.21.3] — 2026-08-25

Fixed

• fix(tests): assert the parser's output without .expect()
• fix(copilot-headless): an empty turn is retried, not returned

[0.21.2] — 2026-08-22

[0.21.1] — 2026-08-19

[0.21.0] — 2026-08-19

Added

• feat(tool-host): ask the caller's surface per call, not once per session
• feat(copilot): give the non-ACP runner real tools, via MCP
• feat(tool-host): host a caller's tools to an ACP agent over loopback MCP

Fixed

• fix(tool-host): keep the error kind when adapting an McpToolExecutor
• fix(tool-host): survive a poisoned lock, and pin isolation and shutdown
• fix: repair the SessionStart bootstrap guard for an empty .build

Other

• ci(release): bump and publish embacle-tool-host with the rest
• test(tool-host): prove streaming carries the caller's tools
• style(tests): rustfmt and SPDX header for the permission policy test

[0.20.0] — 2026-08-07

Fixed

• fix: clear the warnings the flag removal left behind

[0.19.7] — 2026-07-28

[0.19.6] — 2026-07-13

[0.19.5] — 2026-07-10

Fixed

• fix(claude-code): classify prompt-too-long as ContextLength (400), not 502
• fix(clippy): backtick doc addrs, is_ok_and, unwrap-free auth tests CI clippy --all-targets flagged doc_markdown (127.0.0.0/8, ::1) + map().unwrap_or -> is_ok_and in is_loopback_host; arch-validation counted 3 test unwrap()s — replaced with matches!. Fixes the 52d04d6 CI failure.

Other

• test(copilot_headless): annotate safe test unwraps for arch-validation 15 test-only unwrap sites (tempdir/fs/serde setup) get same-line // Safe: comments; zero flagged remain

[0.19.4] — 2026-06-29

Added

• feat(ci): auto-validate Homebrew install after release release.yml now calls homebrew.yml (reusable workflow_call) as a final gated job once the tap formula is pushed, passing the released version for an exact match; drops the vestigial push:tags trigger (GITHUB_TOKEN tag pushes never fired it and would race the formula update). workflow_dispatch retained for manual runs.

Fixed

• fix(copilot-headless): route ACP sessions to the configured model via settings.json copilot --acp routes by ~/.copilot/settings.json "model" and ignores both the session/new model field and the --model flag (cosmetic label only); write the model into settings.json before each spawn (preserving other keys), track the warm-process model and respawn on change, and log the real routing model — fixes ACP sessions silently falling through to the account GPT/Gemini default while logs claimed the requested model.
• fix(ci): trust dravr-ai tap for Homebrew 6.0 install (ubuntu) Homebrew 6.0 refuses untrusted third-party taps; brew trust dravr-ai/tap (|| true for older brew) before install. Last of 4 Homebrew jobs.
• fix(ci): make Homebrew test harness pass (llvm-nm, linuxbrew PATH, 503 health) macOS FFI symbols via toolchain llvm-nm (Apple nm can't read Rust 1.96 LLVM objects); add linuxbrew to PATH on ubuntu (brew not found); smoke test accepts degraded 503 /health (no provider CLIs in CI) by dropping curl -f. All validated locally.

[0.19.3] — 2026-06-22

Added

• feat(mcp): migrate to dravr-tronc 0.5.3; release v0.19.2 ActiveConfig->interior RwLock (provider/model/multiplex accessors &self async); init protocolVersion now 2025-11-25; -mcp/-server only.
• feat(release): adopt dravr-tronc 0.3 via coordinated republish (embacle-mcp/server 0.19.1) Republish embacle-mcp+embacle-server 0.19.1 on tronc 0.3.1 so the publish graph resolves a single tronc version; root embacle (no tronc dep) stays 0.19.0.

Fixed

• fix(ci): combine publish dry-run into one workspace invocation Per-crate dry-run resolved internal embacle dep vs crates.io (only 0.19.0 there, not the unpublished 0.19.2) and failed; co-publishing all three resolves siblings locally. Verified cargo 1.96.0 exit 0.

Other

• style(lints): enforce error-handling restriction lints; create_runner_with_config -> Result Sole prod fix: ClaudeWeb embedded-config parse propagated via ? (was expect); crate-root cfg_attr(test, allow) + tests/examples allow attrs; --all-targets clippy green.
• Revert "chore(deps): bump dravr-tronc 0.2 -> 0.3" (publish-dry-run version skew) embacle-server publishes against registry embacle-mcp 0.19.0 (tronc 0.2) -> two tronc versions clash; needs a coordinated member-crate release to adopt 0.3.
• chore(deps): bump dravr-tronc 0.2 -> 0.3 (now published to crates.io) Server/mcp crates compile against 0.3.1; picks up otel + NotifyLayer + error-handling fixes.

[0.19.2] — 2026-06-19

Changed

• deps: migrate embacle-mcp and embacle-server to dravr-tronc 0.5.3 (dual-era
  MCP engine). tronc 0.5 hands tools/handlers a shared &Arc<ServerState> with no
  outer RwLock, so the active provider/model/multiplex config moved to a
  per-field interior RwLock<ActiveConfig>; those accessors are now &self async
  (the runner cache was already interior Mutex). All three crates aligned to a
  single workspace version. The core embacle crate is unchanged.

[0.19.0] — 2026-06-07

[0.18.0] — 2026-06-06

Added

• feat(acp): per-request mcp_servers for native Copilot ACP tool calling ChatRequest.mcp_servers flows into session/new; opt-in mcp_tool_calling re-enables SDK_TOOL_CALLING to route tool turns through converse().

Fixed

• fix(example): Duration::from_mins in web_login for clippy 1.96

[0.17.0] — 2026-06-05

Added

• feat(web-ui): browser-driven Claude.ai provider (REST + MCP) via dravr-browser

Other

• test(e2e): skip Copilot headless E2E when backend auth unavailable

[0.16.0] — 2026-06-04

[0.15.15] — 2026-06-02

[0.15.14] — 2026-05-29

Fixed

• fix(tool-sim): strip echoed <tool_result> scaffolding from CLI replies

[0.15.13] — 2026-05-19

[0.15.12] — 2026-05-19

[0.15.11] — 2026-05-19

[0.15.10] — 2026-05-19

[0.15.9] — 2026-05-19

[0.15.8] — 2026-05-19

[0.15.7] — 2026-05-13

Fixed

• fix(cli): surface stdout when stderr is empty claude-code on exit=1 writes the error to stdout while emitting zero bytes to stderr. Prefer stderr, fall back to stdout, truncate at 500 chars.

Other

• build(deps): bump softprops/action-gh-release to v3
• build(deps): bump agent-client-protocol-schema to 0.12

[0.15.6] — 2026-05-08

[0.15.5] — 2026-05-01

Other

• obs: payload logging across CLI runners, ACP path, and openai_api Process::run_cli_command emits info (lengths, latency) + trace (argv, full output) covering all 12 runners; copilot_headless::complete and openai_api gain matching pre/post call info+trace; agent::run logs each multi-turn iteration.

[0.15.4] — 2026-04-23

[0.15.3] — 2026-04-21

[0.15.2] — 2026-04-21

[0.15.1] — 2026-04-20

[0.15.0] — 2026-04-17

Added

• feat: ranked Copilot model catalog with self-heal + reasoning effort Replaces hardcoded default; adds ModelUnavailable kind and 1.95 clippy fixes.

[0.14.11] — 2026-04-10

Other

• build: reduce tokio feature footprint to minimal set

[0.14.10] — 2026-04-03

Fixed

• fix: use backticks for code identifiers in doc comment

[0.14.9] — 2026-03-31

[0.14.8] — 2026-03-31

[0.14.7] — 2026-03-31

Fixed

• fix: repair broken expect annotations in openai_types test code
• fix: add Safe annotation to test expect() calls in openai_types
• fix: resolve error handling violations found by dravr-build-config validation

[0.14.6] — 2026-03-30

[0.14.5] — 2026-03-30

[0.14.4] — 2026-03-30

Fixed

• fix: stop re-injecting system prompt into ACP prompt text, forward max_tokens

Other

• deps: bump opentelemetry and opentelemetry_sdk from 0.28 to 0.31
• deps: bump toml from 0.8.23 to 1.1.0
• deps: bump agent-client-protocol-schema from 0.11.3 to 0.11.4
• deps: bump docker/build-push-action from 6 to 7
• deps: bump docker/login-action from 3 to 4
• deps: bump docker/setup-buildx-action from 3 to 4

[0.14.3] — 2026-03-28

[0.14.2] — 2026-03-26

Other

• deps: bump dravr-tronc to 0.2 with error notification support

[0.14.1] — 2026-03-23

[0.13.5] — 2026-03-18

[0.13.4] — 2026-03-18

Added

• feat: make ACP session and prompt timeouts configurable via env vars

[0.13.3] — 2026-03-18

[0.13.2] — 2026-03-18

[0.13.1] — 2026-03-18

Fixed

• fix: inject system prompt into ACP prompt blocks

[0.13.0] — 2026-03-13

Added

• feat: add C FFI static library for Swift integration
• feat: materialize images to temp files for all 12 CLI runners

[0.12.0] — 2026-03-13

Added

• feat: add vision/image support and wire CopilotHeadless into provider system

Fixed

• fix: resolve clippy and test count for copilot-headless wiring

[0.11.0] — 2026-03-12

Added

• feat: add Kilo Code CLI runner with NDJSON parsing and streaming
• feat: unify embacle-server with embacle-mcp into single binary
• feat: add guardrail middleware and cross-decorator scenario tests GuardrailProvider with 3 built-in guards, ScriptedProvider, 7 scenarios
• feat: add cost tracking, response cache, and OTel metrics export TokenPricing/PricingTable, CacheProvider, otel feature flag with instruments
• feat: add TOML config file loading behind config-file feature flag Declarative provider, fallback, alias setup via embacle.toml or ~/.config/embacle/
• feat: add retry with exponential backoff to FallbackProvider Add ErrorKind::is_transient(), RetryConfig, FallbackProvider::with_retry()
• feat: add Kiro CLI runner as 11th LLM provider

Fixed

• fix: add kilo to --provider help text in server and mcp binaries
• fix: add permissions block to homebrew workflow
• fix: resolve clippy doc_markdown warnings for provider help strings
• fix: update health endpoint test to expect 11 providers

[0.10.2] — 2026-03-11

Added

• feat: add Kiro CLI runner as 11th LLM provider
• feat: add retry with exponential backoff to FallbackProvider
• feat: add TOML config file loading behind config-file feature flag
• feat: add cost tracking, response cache, and OTel metrics export
• feat: add guardrail middleware and cross-decorator scenario tests
• feat: unify embacle-server with embacle-mcp into single binary
• feat: add Kilo Code CLI runner with NDJSON parsing and streaming

Fixed

• fix: resolve clippy doc_markdown warnings for provider help strings
• fix: add permissions block to homebrew workflow (code scanning alert #10)
• fix: filter artifact download to exclude Docker buildx cache

[0.10.1] — 2026-03-11

Fixed

• fix: Docker build uses -p flag for workspace binary targets

[0.10.0] — 2026-03-11

Added

• feat: add MCP Streamable HTTP to embacle-server, Homebrew tap with CI
• docs: add OpenAI API runner section to README

[0.9.0] — 2026-03-11

Added

• feat: add with_client() constructor for HTTP client injection
• feat: add OpenAI-compatible HTTP API client runner (openai-api feature) SSE streaming, tool calling, model discovery, 30 unit tests
• feat: add Docker image with ghcr.io release workflow

Fixed

• fix: resolve 3 CodeQL uncontrolled-allocation-size alerts in stop field Use bounded to_bounded_vec() instead of clone+truncate to avoid copying unbounded user input
• fix: revert accidental version bump in Cargo.toml files
• fix: update OpenAI API default model to gpt-5.4
• fix: resolve 9 CodeQL security alerts (stop bounds, CI permissions)

[0.8.1] — 2026-03-09

[0.8.0] — 2026-03-09

Added

• feat: add Warp terminal (oz) CLI runner as 10th provider

Fixed

• fix: ACP prompt timeout, stdio error propagation, redact log output
• fix: timing-safe auth, dependabot config, remove Box::leak

Other

• refactor: dedup factory/runner code, wire top_p/stop/response_format Move factory+parsing to core, add top_p/stop to OpenAI types, fix READMEs (0.6->0.7, 9 providers)
• Update Dependabot configuration version

[0.7.0] — 2026-03-07

Added

• feat: add InstalledAuthUnverified variant and env var auth checking
• feat: enhance JSON schema validation with recursion, enum, and bounds
• feat: add native tool calling types to core library

Fixed

• fix: remove [Unreleased] section before v0.7.0 release
• fix: use npm prefix -g to locate copilot binary on CI
• fix: use npm bin -g to find copilot binary path on CI
• fix: symlink copilot-linux-x64 binary to copilot on CI PATH
• fix: use env var for secrets conditional in CI workflow
• fix: wire COPILOT_GITHUB_TOKEN secret into CI for headless e2e tests
• fix: CI now tests all workspace crates, fix workspace-wide clippy warnings
• fix: ACP permission response prefers AllowAlways, never falls back to reject
• fix: update AGENTS.md project overview to reflect full workspace scope
• fix: default to claude-opus-4.6-fast, update Copilot fallback model list
• fix: rewrite OpenCode NDJSON parser, add E2E test suite for all runners

[0.6.0] — 2026-03-07

Fixed

• fix: purge stale copilot-sdk references from docs, CI, and release workflow
• fix: complete headless tool calling + usage extraction Add SDK_TOOL_CALLING, ObservedToolCall tracking, converse() API, ACP usage parsing

[0.5.0] — 2026-03-05

Fixed

• fix: streaming downgrade for non-streaming providers, reduce lock scope Permissive mode now falls back to complete() + SSE instead of erroring

Other

• refactor: deduplicate runners, remove as_any, async model discovery CliRunnerBase + macro (-528 lines), #[must_use] on RunnerError

[0.4.1] — 2026-03-04

Added

• feat: fix tool calling for coding-assistant LLMs, add index to ToolCall Code-generation framing for catalogs, ToolCall.index per OpenAI spec, live Copilot tests

[0.3.0] — 2026-03-03

Added

• feat: add agent loop, fallback chains, metrics, quality gate, MCP bridge, and structured output
• feat: add Goose, Cline, and Continue CLI runners Expand embacle from 6 to 9 CLI runners with JSON parsing, session resume, and streaming

Fixed

• fix: restore trailing newlines stripped by LinesStream in SSE deltas

[0.2.1] — 2026-03-02

Added

• feat: add Gemini CLI and Codex CLI runners Add GeminiCliRunner and CodexCliRunner with streaming JSONL support, expanding embacle to 6 CLI tools

Fixed

• fix: add GeminiCli and CodexCli to subcrate runner factories and provider lists

[0.2.0] — 2026-03-02

Added

• feat: add Gemini CLI and Codex CLI runners Add GeminiCliRunner and CodexCliRunner with streaming JSONL support, expanding embacle to 6 CLI tools

[0.1.3] — 2026-03-01

[0.1.2] — 2026-03-01

[0.1.1] — 2026-03-01

[0.1.0] — 2026-03-01

Added

• feat: add embacle-server crate with OpenAI-compatible REST API Axum HTTP server with /v1/chat/completions, /v1/models, /health, SSE streaming, multiplex fan-out, bearer auth
• feat: add embacle-mcp binary crate with MCP server Stdio/HTTP transports, 7 tools, JSON-RPC 2.0, multiplex fan-out, README MCP section
• feat: add Timeout error kind, enhanced logging, and stdout capture on failure
• feat: add tool_simulation module for text-based tool calling Text-based tool loop for CLI runners with catalog generation, XML parsing, and async execution
• feat: address analysis weak spots, expand tests, add CI branch triggers Fix doc drift, propagate max_tokens, add gh-auth check, 44 new tests (18→62)
• feat: make available_models() configurable at runtime Change LlmProvider::available_models() return type from &'static [&'static str] to &[String] so models can be determined at runtime. Each runner stores its model list in a Vec. CopilotRunner and CopilotSdkRunner discover models via gh copilot models at construction time, falling back to static defaults if the command fails.
• feat: default to claude-opus-4.6 for Copilot SDK, Copilot CLI, and Claude Code runners
• feat: add SDK_TOOL_CALLING capability and as_any() for downcasting Add flag for SDK-managed tool loops, as_any() on LlmProvider trait, re-export ToolHandler types
• feat: add Copilot SDK provider behind copilot-sdk feature flag

Fixed

• fix: release workflow — use macos-14, build --workspace, reset versions to 0.0.1
• fix: rename remaining capitalized Embache references to Embacle
• fix: rename crate from embache to embacle
• fix: wire env keys to sandbox, guard streaming child lifecycle, remove dead ExecutionMode
• fix: use RunnerConfig::new in doc examples
• fix: correct rust-toolchain.toml format

[0.1.0] — 2026-03-01

Added

• feat: add embacle-server crate with OpenAI-compatible REST API Axum HTTP server with /v1/chat/completions, /v1/models, /health, SSE streaming, multiplex fan-out, bearer auth
• feat: add embacle-mcp binary crate with MCP server Stdio/HTTP transports, 7 tools, JSON-RPC 2.0, multiplex fan-out, README MCP section
• feat: add Timeout error kind, enhanced logging, and stdout capture on failure
• feat: add tool_simulation module for text-based tool calling Text-based tool loop for CLI runners with catalog generation, XML parsing, and async execution
• feat: address analysis weak spots, expand tests, add CI branch triggers Fix doc drift, propagate max_tokens, add gh-auth check, 44 new tests (18→62)
• feat: make available_models() configurable at runtime Change LlmProvider::available_models() return type from &'static [&'static str] to &[String] so models can be determined at runtime. Each runner stores its model list in a Vec. CopilotRunner and CopilotSdkRunner discover models via gh copilot models at construction time, falling back to static defaults if the command fails.
• feat: default to claude-opus-4.6 for Copilot SDK, Copilot CLI, and Claude Code runners
• feat: add SDK_TOOL_CALLING capability and as_any() for downcasting Add flag for SDK-managed tool loops, as_any() on LlmProvider trait, re-export ToolHandler types
• feat: add Copilot SDK provider behind copilot-sdk feature flag

Fixed

• fix: rename remaining capitalized Embache references to Embacle
• fix: rename crate from embache to embacle
• fix: wire env keys to sandbox, guard streaming child lifecycle, remove dead ExecutionMode
• fix: use RunnerConfig::new in doc examples
• fix: correct rust-toolchain.toml format