ace-llm 0.32.1 → 0.38.2

data/.ace-defaults/llm/providers/togetherai.yml

@@ -1,26 +1,59 @@
 name: togetherai
-last_synced: 2025-12-05
+last_synced: 2026-04-24
 class: Ace::LLM::Organisms::TogetherAIClient
 gem: ace-llm
-context_limit: 128000  # Default context limit; varies by model
 models:
-  - Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8
-  - deepseek-ai/DeepSeek-V3
-  - moonshotai/Kimi-K2-Instruct
-  - openai/gpt-oss-120b
+- MiniMaxAI/MiniMax-M2.5
+- MiniMaxAI/MiniMax-M2.7
+- Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8
+- Qwen/Qwen3-Coder-Next-FP8
+- Qwen/Qwen3.5-397B-A17B
+- essentialai/Rnj-1-Instruct
+- google/gemma-4-31B-it
+- moonshotai/Kimi-K2.5
+- moonshotai/Kimi-K2.6
+- openai/gpt-oss-120b
+- zai-org/GLM-5.1
+limits:
+  default:
+    context: 262144
+    output: 262144
+  models:
+    MiniMaxAI/MiniMax-M2.5:
+      context: 204800
+      output: 131072
+    MiniMaxAI/MiniMax-M2.7:
+      context: 202752
+      output: 131072
+    Qwen/Qwen3.5-397B-A17B:
+      output: 130000
+    essentialai/Rnj-1-Instruct:
+      context: 32768
+      output: 32768
+    google/gemma-4-31B-it:
+      output: 131072
+    moonshotai/Kimi-K2.6:
+      output: 131000
+    openai/gpt-oss-120b:
+      context: 131072
+      output: 131072
+    zai-org/GLM-5.1:
+      context: 202752
+      output: 131072
 aliases:
   model:
-    qwen: Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8
-    deepseek: deepseek-ai/DeepSeek-V3
+    qwen-coder: Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8
+    deepseek: deepseek-ai/DeepSeek-V3.1
     kimi: moonshotai/Kimi-K2-Instruct
+    kimi-T: moonshotai/Kimi-K2-Thinking
     oss: openai/gpt-oss-120b
 api_key:
   env: TOGETHER_API_KEY
   required: true
   description: Together AI API key
 capabilities:
-  - text_generation
-  - streaming
+- text_generation
+- streaming
 default_options:
   temperature: 0.7
   max_tokens: 16384

data/.ace-defaults/llm/providers/xai.yml

@@ -1,14 +1,27 @@
 name: xai
-last_synced: 2025-12-06
+last_synced: 2026-04-24
 class: Ace::LLM::Organisms::XAIClient
 gem: ace-llm
-context_limit: 131072  # Grok models have 131K context window
 models:
-  - grok-4
-  - grok-4-1-fast
-  - grok-4-1-fast-non-reasoning
-  - grok-4-fast-non-reasoning
-  - grok-code-fast-1
+- grok-4
+- grok-4-1-fast
+- grok-4-1-fast-non-reasoning
+- grok-4-fast-non-reasoning
+- grok-4.20-0309-non-reasoning
+- grok-4.20-0309-reasoning
+- grok-4.20-multi-agent-0309
+- grok-code-fast-1
+limits:
+  default:
+    context: 2000000
+    output: 30000
+  models:
+    grok-4:
+      context: 256000
+      output: 64000
+    grok-code-fast-1:
+      context: 256000
+      output: 10000
 aliases:
   global:
     grok: xai:grok-4
@@ -24,7 +37,7 @@ api_key:
   required: true
   description: x.ai API key
 capabilities:
-  - text_generation
+- text_generation
 default_options:
   temperature: 0.7
   max_tokens: 16384

data/.ace-defaults/llm/providers/zai.yml

@@ -1,18 +1,29 @@
 name: zai
-last_synced: 2026-02-27
+last_synced: 2026-04-24
 class: Ace::LLM::Organisms::ZaiClient
 gem: ace-llm
-context_limit: 128000
 models:
-  - glm-4.7-flashx
-  - glm-4.7
-  - glm-5
+- glm-4.7
+- glm-4.7-flashx
+- glm-5
+- glm-5-turbo
+- glm-5.1
+- glm-5v-turbo
+limits:
+  default:
+    context: 200000
+    output: 131072
+  models:
+    glm-4.7:
+      context: 204800
+    glm-5:
+      context: 204800
 api_key:
   env: ZAI_API_KEY
   required: true
   description: Z.AI API key
 capabilities:
-  - text_generation
+- text_generation
 default_options:
   temperature: 0.7
   max_tokens: 16384

data/CHANGELOG.md

@@ -6,6 +6,154 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
+### Fixed
+- Stubbed the Z.ai endpoint in query command CLI-routing tests so provider fallback no longer triggers blocked real HTTP calls during deterministic suite runs.
+
+## [0.38.2] - 2026-04-24
+
+### Technical
+- Tightened `TS-LLM-002` verification so `ace-llm --list-providers` remains a discovery-only public-surface check instead of implying provider readiness.
+
+## [0.38.1] - 2026-04-24
+
+### Changed
+- Clarified provider-listing documentation so `ace-llm --list-providers` is described as discovery plus setup hints, with setup readiness delegated to `ace-config doctor`.
+
+## [0.38.0] - 2026-04-24
+
+### Added
+- Added shared resolved-model limit lookup so alias, role, and preset-expanded concrete provider/model targets can read merged context and output limits from provider config.
+
+### Changed
+- Updated provider config validation and downstream callers to use the new `limits` schema and the expanded concrete model target instead of provider-wide limit assumptions.
+
+## [0.37.1] - 2026-04-23
+
+
+### Changed
+- Added concurrent provider ping support in query execution and refreshed fallback orchestration behavior used by provider-driven query flows.
+- Updated provider ping coverage so live query failures no longer block local diagnostics and stale fallback assumptions are exercised more explicitly.
+
+
+## [0.37.0] - 2026-04-23
+
+### Added
+- Added `ace-llm --no-fallback` so callers can run ordinary prompts such as `ping` against the exact requested provider/model without fallback routing.
+
+## [0.36.5] - 2026-04-22
+
+### Fixed
+- Stabilized deterministic verification by stubbing CLI query routing away from provider fallback HTTP calls and removing retry sleep from the provider/model fallback test.
+
+## [0.36.4] - 2026-04-22
+
+### Technical
+- Added regression coverage to keep fresh `commit` role defaults from reintroducing stale `codex:gpt-5-mini` mappings while preserving `codex:mini` fallback behavior.
+
+## [0.36.3] - 2026-04-19
+
+### Fixed
+- Fixed `ace-llm --interactive` startup for Codex aliases so the `/as-onboard` and related interactive alias prompts are launched with a valid subprocess command shape instead of aborting with a missing keyword.
+
+## [0.36.2] - 2026-04-19
+
+### Fixed
+- Refreshed Codex model examples in README, usage, getting-started, demo, and help output from stale `codex:gpt-5` forms to the current stable `codex:gpt` alias and updated current provider alias guidance.
+
+## [0.36.1] - 2026-04-16
+
+### Fixed
+- Switched the fresh-default `commit` role to prefer `google:lite` before CLI fallbacks so new `ace-config init` projects keep generated commit messages working when stale CLI aliases or unavailable local CLIs would otherwise break message generation.
+
+## [0.36.0] - 2026-04-15
+
+### Changed
+- Reworked `TS-LLM-001` to public-surface goal style by tightening model-routing evidence checks, adding `TC-003` output-file-contract coverage, and prioritizing end-state/output artifacts over helper-only signals.
+- Added `TS-LLM-002` provider-discovery coverage for `ace-llm --list-providers` so public setup-hint and provider-list output behavior is verified as a first-class E2E journey.
+
+### Technical
+- Updated runtime dependency constraint to `ace-support-models ~> 0.11` after the coordinated support-models minor release.
+
+## [0.35.1] - 2026-04-15
+
+### Fixed
+- Updated the `cli_args` threading feature tests to match the normalized array contract and stub provider preset loading directly, so `ace-test ace-llm` no longer fails on stale scalar expectations or package-local preset lookup assumptions.
+
+## [0.35.0] - 2026-04-15
+
+### Fixed
+- Merged preset CLI args with explicit `--cli-args` so interactive launches keep provider autonomy presets such as `@yolo` while still accepting additional flags like `--no-alt-screen`.
+
+### Technical
+- Threaded the current working directory and subprocess environment explicitly through `ace-llm --interactive` so provider-side startup policy hooks can make launch decisions from the real execution context.
+
+## [0.34.0] - 2026-04-15
+
+### Added
+- Added `ace-llm --interactive` for supported CLI-backed providers so alias resolution, presets, thinking levels, and canonical skill handoff translation can launch the provider's native interactive terminal UI in one call.
+
+### Changed
+- Documented interactive launch usage in the README, getting-started guide, and usage guide, including the tmux fork handoff pattern built around `/as-assign-drive <assignment>@<root>`.
+
+## [0.33.6] - 2026-04-13
+
+### Fixed
+- Switched `TS-LLM-001` basic-query coverage to the explicit `--prompt` form so the scenario executes with canonical prompt input and no longer misclassifies dropped positional input as provider behavior.
+
+## [0.33.5] - 2026-04-13
+
+### Fixed
+- Hardened `TS-LLM-001` basic-query execution by persisting the exact prompt and command artifacts so empty-input/provider-contract failures are distinguished from valid model responses.
+
+## [0.33.4] - 2026-04-13
+
+### Fixed
+- Added the missing `ace-support-models` runtime dependency so `ace-llm --list-providers` no longer fails on fresh setups with `cannot load such file -- ace/support/models`.
+
+### Changed
+- Updated quick-start onboarding guidance to include `ace-llm` in the base install command used before provider verification.
+
+## [0.33.3] - 2026-04-13
+
+### Fixed
+- Added `codex:mini` as a fallback model for the `commit` role after `glite`, so role-based commit-message generation tries a secondary model instead of failing when `glite` is unavailable.
+
+
+### Technical
+- Tightened the package version contract test to require a full semantic-version match instead of accepting partial matches.
+## [0.33.2] - 2026-04-13
+
+### Fixed
+- Fixed provider credential availability checks so role-based model selection no longer crashes when a provider relies on fallback environment keys.
+
+### Technical
+- Added regression coverage for `ClientRegistry#provider_api_key_present?` using Google fallback API key resolution.
+
+## [0.33.1] - 2026-04-13
+
+### Changed
+- Completed the batch i05 migration follow-through for this package and aligned it with the restarted `fast` / `feat` / `e2e` verification model.
+
+### Technical
+- Included in the coordinated assignment-driven patch release for batch i05 package updates.
+
+
+## [0.33.0] - 2026-04-12
+
+### Changed
+- Migrated package tests to the restarted `fast` / `feat` / `e2e` contract:
+  - moved deterministic test suites from legacy top-level folders into `test/fast/`
+  - moved legacy `test/integration/` coverage into `test/feat/`
+  - reduced `TS-LLM-001` E2E scenario from 3 to 2 TCs and added an E2E decision record
+- Updated README and package docs to teach only `ace-test ace-llm`, `ace-test ace-llm feat`, `ace-test ace-llm all`, and `ace-test-e2e ace-llm`.
+
+### Fixed
+- Preserved full fallback target selectors (including provider/model + suffix context) across fallback attempts and rebuilt generation options per target instead of reusing stale primary-provider options.
+
+## [0.32.2] - 2026-04-10
+
+### Changed
+- Updated the shared role catalog for restarted E2E execution with explicit `e2e-runner`, `e2e-verifier`, and `e2e-reporter` roles in the project-owned LLM configuration surface.
 
 ## [0.32.1] - 2026-04-01
 

data/README.md

@@ -18,7 +18,20 @@
 
 ![ace-llm demo](docs/demo/ace-llm-getting-started.gif)
 
-`ace-llm` gives developers and coding agents one command surface for querying any LLM provider. Address models by alias (`gflash`, `sonnet`), explicit `provider:model` notation, or with thinking levels (`codex:gpt-5:high`) and execution presets (`cc@ro`). Pass prompts and system instructions inline or as file paths. Fallback routing and retry behavior keep prompt workflows resilient.
+`ace-llm` gives developers and coding agents one command surface for querying any LLM provider. Address models by alias (`gflash`, `sonnet`), explicit `provider:model` notation, or with thinking levels (`codex:gpt:high`) and execution presets (`cc@ro`). Pass prompts and system instructions inline or as file paths, or start supported CLI providers in native interactive mode with `--interactive`. Fallback routing and retry behavior keep prompt workflows resilient.
+
+## Testing
+
+Package-local verification contract:
+
+```bash
+ace-test ace-llm
+ace-test ace-llm feat
+ace-test ace-llm all
+ace-test-e2e ace-llm
+```
+
+Deterministic coverage lives in `test/fast/` and `test/feat/`. Scenario assets stay in `test/e2e/`.
 
 ## How It Works
 
@@ -30,12 +43,16 @@
 
 **Switch providers with aliases** - use short names like `gflash`, `sonnet`, `opus` instead of full `provider:model` notation. Aliases resolve through versioned YAML in [`.ace-defaults/`](docs/usage.md).
 
-**Control reasoning depth** - append a thinking level (`codex:gpt-5:high`, `claude:sonnet:low`) to tune reasoning budgets. Supported CLI providers: `claude`, `codex` (levels: `low`, `medium`, `high`, `xhigh`).
+**Control reasoning depth** - append a thinking level (`codex:gpt:high`, `claude:sonnet:low`) to tune reasoning budgets. Supported CLI providers: `claude`, `codex` (levels: `low`, `medium`, `high`, `xhigh`).
 
 **Run preset-driven prompts** - apply execution profiles with `@preset` or `--preset`. Built-in presets for CLI providers: `@ro` (read-only), `@rw` (read-write), `@yolo` (full autonomy). Supported by: `claude`, `codex`, `gemini`, `opencode`, `pi`.
 
+**Launch real interactive agents** - use `--interactive` for supported CLI providers so alias resolution, presets, and skill translation still flow through `ace-llm`, but the provider starts its native terminal UI instead of one-shot print mode.
+
 **Build resilient prompt workflows** - configure fallback chains and retry behavior through the [config cascade](.ace-defaults/llm/config.yml) so transient provider issues do not block work.
 
+**Check exact provider reachability** - use `ace-llm gemini:pro "ping" --no-fallback --timeout 15 --max-tokens 4` to verify the requested provider/model without fallback routing.
+
 **Power LLM-enhanced flows in sibling packages** - serve as the execution backend for [ace-git-commit](../ace-git-commit), [ace-idea](../ace-idea), [ace-review](../ace-review), [ace-sim](../ace-sim), [ace-prompt-prep](../ace-prompt-prep), and more.
 
 ---

data/exe/ace-llm

@@ -18,7 +18,8 @@ args = ARGV.empty? ? ["--help"] : ARGV
 
 # Start ace-support-cli single-command entrypoint with exception-based exit code handling (per ADR-023)
 begin
-  Ace::Support::Cli::Runner.new(Ace::LLM::CLI::Commands::Query).call(args: args)
+  exit_code = Ace::Support::Cli::Runner.new(Ace::LLM::CLI::Commands::Query).call(args: args)
+  exit(exit_code) if exit_code.to_i.positive?
 rescue Ace::Support::Cli::Error => e
   warn e.message
   exit(e.exit_code)