heimdall-agent 1.2.0 → 1.4.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -43,17 +43,10 @@
43
43
  ### Phase 3: Integration & Playwright E2E Verification
44
44
  ### Phase 4: Review, Purge & Handoff
45
45
 
46
- ## 5. Symmetric & Tri-Directional Cross-Model Validation Loop
47
- - In Herdr (`HERDR_ENV=1`), validate tasks directly via peer agent prompts:
48
- - Tasks executed by Claude -> Validate with Gemini or Codex:
49
- `herdr agent prompt gemini-reviewer "Review git diff for security regressions and type errors." --wait --timeout 120000`
50
- `herdr agent prompt codex-reviewer "Review git diff for security regressions and boundary conditions." --wait --timeout 120000`
51
- - Tasks executed by Gemini -> Validate with Claude or Codex:
52
- `herdr agent prompt claude-reviewer "Review git diff for security regressions and type errors." --wait --timeout 120000`
53
- `herdr agent prompt codex-reviewer "Review git diff for security regressions and boundary conditions." --wait --timeout 120000`
54
- - Tasks executed by Codex -> Validate with Claude or Gemini:
55
- `herdr agent prompt claude-reviewer "Review git diff for security regressions and type errors." --wait --timeout 120000`
56
- `herdr agent prompt gemini-reviewer "Review git diff for security regressions and type errors." --wait --timeout 120000`
46
+ ## 5. Symmetric Cross-Family Validation Loop
47
+ - In Herdr (`HERDR_ENV=1`), validate tasks using a **different model family** from the authoring model (e.g., Anthropic ↔ Google ↔ OpenAI/Codex ↔ Zhipu/GLM ↔ Tencent/Hunyuan ↔ Xiaomi/MiMo) selected from `config/models.json` / `playbooks/06-model-registry-matrix.md`:
48
+ - `herdr agent prompt <cross-family-reviewer> "Review git diff for security regressions, boundary conditions, and type errors." --wait --timeout 120000`
49
+ - Select any live reviewer agent/model from a distinct family rather than restricting validation to Claude/Gemini/Codex alone.
57
50
 
58
51
  ## 6. Orchestrator Heartbeat
59
52
  - Orchestrator monitors progress every 3 minutes.
@@ -103,4 +96,4 @@
103
96
  - Every task must explicitly declare one of the registered models and an allowed effort level (`low`, `medium`, `high`, `xhigh`) defined in `playbooks/06-model-registry-matrix.md`.
104
97
  - Tasks requiring deep architectural reasoning, type safety, or complex algorithms must use `high` or `xhigh`.
105
98
  - Fast scans, boilerplate scaffolding, or rapid audits may use `low` or `medium`.
106
- - When Gemini 4 Argon reaches quota cooldown, immediately fallback to Gemini 3.8 Flash, Gemini 3.1 Pro, or GhostCLI (Claude Sonnet 5 / GPT-6 Astra).
99
+ - When Gemini 4 Argon reaches quota cooldown, fallback in order: Gemini 4 Hydrogen → Gemini 3.8 Flash → GhostCLI models (Claude/GPT).
@@ -10,8 +10,8 @@ Models are partitioned into two architectural access channels:
10
10
 
11
11
  - **Channel 1: GhostCLI Unified Gateway (`https://ghostcli.cloud`)**
12
12
  - Powers **Claude Code CLI**, **OpenAI Codex CLI**, and **Hermes Agent**.
13
- - Serves all Anthropic models (`claude-*`), OpenAI models (`gpt-6-astra`), and Zhipu AI models (`z-ai/glm-5.3`, `z-ai/glm-5.3-flash`).
14
- - Because GhostCLI exposes multi-protocol translation (`/v1/messages`, `/v1/responses`, `/v1/chat/completions`), Claude Code can run GPT-6 Astra or GLM directly (`claude --model gpt-6-astra`), and Codex can run Claude Sonnet directly (`codex --model claude-sonnet-5`).
13
+ - Serves all GhostCLI-proxied model families (Anthropic `claude-*`, OpenAI `gpt-*`, Zhipu `z-ai/glm-*`, Tencent `tencent/Hy4-*`, Xiaomi `XiaomiMiMo/*`, and any newly discovered models in `config/models.json`).
14
+ - Because GhostCLI exposes multi-protocol translation (`/v1/messages`, `/v1/responses`, `/v1/chat/completions`), any CLI harness can run any GhostCLI model directly (e.g. `claude --model z-ai/glm-5.3`, `claude --model tencent/Hy4-preview`, `codex --model claude-sonnet-5-5`).
15
15
  - When coordinating between harnesses in Herdr (`HERDR_ENV=1`), use native Herdr multiplexing (`herdr agent start/prompt`) directly.
16
16
 
17
17
  - **Channel 2: Google Gemini Direct API (`gemini.googleapis.com`)**
@@ -22,19 +22,25 @@ Models are partitioned into two architectural access channels:
22
22
 
23
23
  ## 2. GhostCLI Models (via `https://ghostcli.cloud`)
24
24
 
25
- All GhostCLI models support text and image input:
26
-
27
- | Model Name | Canonical Flag | Context Window | Max Output | Parallel Tools | Search | Allowed Effort Levels | Optimal Specialization |
28
- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
29
- | **Claude Opus 5.5** | `claude-opus-5-5` | 256K / 1M extended | 128K | ✅ | ✅ | `low`, `medium`, `high`, `xhigh` | Deepest architectural reasoning, complex multi-system debugging, critical security audits, knotty algorithms. |
30
- | **Claude Sonnet 5** | `claude-sonnet-5` | 256K / 1M extended | 128K | ✅ | ✅ | `low`, `medium`, `high`, `xhigh` | Primary full-stack coding driver, type-safe API schemas, database mutation logic, Playwright E2E test authoring. |
31
- | **Claude Fable 5.1** | `claude-fable-5.1` | 256K / 1M extended | 128K | ✅ | ✅ | `low`, `medium`, `high`, `xhigh` | Visual storytelling, high-polish frontend design, copycrafting, UI aesthetic alignment, interaction flows. |
32
- | **GLM 5.3** | `z-ai/glm-5.3` | 1M | 128K | ❌ | ❌ | `low`, `medium`, `high`, `xhigh` | Massive context analysis, long-document processing, Zhipu AI reasoning (upstream may experience upstream queue delays). |
33
- | **GLM 5.3 Flash** | `z-ai/glm-5.3-flash` | 1M | 128K | ❌ | ❌ | `low`, `medium`, `high` | Fast responses for code tasks, rapid reasoning, lightweight context queries. |
34
- | **GPT-6 Astra** | `gpt-6-astra` | 1M | 128K | ✅ | ✅ | `low`, `medium`, `high`, `xhigh` | OpenAI reasoning, broad general-purpose tasks, alternative perspective for cross-validation. |
35
-
36
- > [!WARNING]
37
- > **Naming Asymmetry Guardrail**: Opus is version **5.5** (`claude-opus-5-5`), while Sonnet is strictly version **5** (`claude-sonnet-5`). Do NOT hallucinate `claude-sonnet-5-5` or `sonnet-5-5`. There is no 5.5 release for Sonnet.
25
+ <!-- heimdall:ghostcli-models:start -->
26
+ All GhostCLI models are discovered dynamically via `heimdall sync` / `heimdall sync-models` and stored in `config/models.json`:
27
+
28
+ | Model Name | Canonical Flag | Family | Context Window | Allowed Effort Levels | Specialization / Description |
29
+ | :--- | :--- | :--- | :--- | :--- | :--- |
30
+ | **Claude Opus 5.5** | `claude-opus-5-5` | anthropic | 1M | `low`, `medium`, `high`, `xhigh` | Melhor equilíbrio entre capacidade e custo |
31
+ | **Claude Sonnet 5** | `claude-sonnet-5` | anthropic | 1M | `low`, `medium`, `high` | O mais rápido sem abrir mão de qualidade |
32
+ | **Claude Sonnet 5.5** | `claude-sonnet-5-5` | anthropic | 1M | `low`, `medium`, `high` | Rápido e mais capaz que o Sonnet 5 |
33
+ | **Claude Fable 5.2** | `claude-fable-5.2` | anthropic | 1M | `low`, `medium`, `high`, `xhigh` | Máxima capacidade disponível |
34
+ | **GLM 5.3** | `z-ai/glm-5.3` | z-ai | 1M | `low`, `medium`, `high`, `xhigh` | Código agêntico com contexto de 1M |
35
+ | **GLM 5.3 Flash** | `z-ai/glm-5.3-flash` | z-ai | 1M | `low`, `medium`, `high` | Respostas rápidas para código |
36
+ | **Hunyuan Hy4 Preview** | `tencent/Hy4-preview` | tencent | 1M | `low`, `medium`, `high`, `xhigh` | Raciocínio e código |
37
+ | **DeepSeek V4.1 Flash** | `deepseek-ai/DeepSeek-V4.1-Flash-1017` | deepseek-ai | 1M | `low`, `medium`, `high` | Respostas rápidas para código |
38
+ | **MiMo V2.6 Pro** | `XiaomiMiMo/MiMo-V2.6-Pro` | xiaomi | 1M | `low`, `medium`, `high`, `xhigh` | Código e raciocínio |
39
+ | **GPT-6 Astra** | `gpt-6-astra` | openai | 1M | `low`, `medium`, `high`, `xhigh` | Trabalho agêntico longo com janela de 1M |
40
+
41
+ > [!NOTE]
42
+ > **Dynamic Model Registry**: Always resolve live model IDs from `config/models.json` or this table after `heimdall sync` / `heimdall sync-models`. Route tasks by **capability tier** across all available families (Anthropic, Google, OpenAI, GLM, Hunyuan, MiMo, etc.), not only Claude/Gemini/Codex.
43
+ <!-- heimdall:ghostcli-models:end -->
38
44
 
39
45
  ---
40
46
 
@@ -44,8 +50,10 @@ Accessed natively via Google Antigravity / `agy` CLI (never through GhostCLI):
44
50
 
45
51
  | Model Name | Canonical Flag | Allowed Effort Levels | Optimal Specialization |
46
52
  | :--- | :--- | :--- | :--- |
47
- | **Gemini 4 Argon** | `gemini-4-argon-{effort}` (e.g. `gemini-4-argon-high`) | `low`, `medium`, `high`, `xhigh` | Next-generation flagship Gemini reasoning, complex multi-file architecture, deep cross-workspace synthesis, high-velocity full-stack implementation. |
48
- | **Gemini 3.8 Flash** | `gemini-3.8-flash-{effort}` (e.g. `gemini-3.8-flash-high`) | `low`, `medium`, `high` | Fast scans, workspace audits, boilerplate generation, unit test scaffolding, rapid prototyping, responsive styling. |
53
+ | **Gemini 4 Hydrogen** | `gemini-4-hydrogen-{effort}` (e.g. `gemini-4-hydrogen-xhigh`) | `low`, `medium`, `high`, `xhigh` | Efficient frontier reasoning with speed and long decodes up to 1M tokens. Supports xhigh effort tier. Replaces Flash branding. |
54
+ | **Gemini 4 Argon** | `gemini-4-argon-{effort}` (e.g. `gemini-4-argon-high`) | `low`, `medium`, `high` | Efficient Gemini 4 frontier reasoning, faster responses and lower token spend. Subject to quota cooldowns. |
55
+ | **Orchestrator (Argon)** | N/A (model dropdown option in Antigravity) | Auto-managed | Dynamic mode that splits tasks between Carbon (planning/review) and Hydrogen (implementation) within one session. Uses ~24% less quota and finishes ~19% faster than Carbon alone. Requires Google AI Ultra subscription; not a standalone model. |
56
+ | **Gemini 3.8 Flash** | `gemini-3.8-flash-{effort}` (e.g. `gemini-3.8-flash-high`) | `low`, `medium`, `high` | Stable previous-generation Flash for software engineering, autonomous agents, and complex enterprise workflows. Current default workhorse. |
49
57
 
50
58
  ---
51
59
 
@@ -72,12 +80,12 @@ The `/model` picker in Claude Code shows a hardcoded list that may include model
72
80
 
73
81
  To dispatch any GhostCLI model — including those invisible in the picker — pass its canonical flag via the Agent tool's `model` parameter when spawning subagents (e.g. `model: "gpt-6-astra"`, `model: "z-ai/glm-5.3"`). The API call routes through GhostCLI regardless of what `/model` displays.
74
82
 
75
- Gemini models (`gemini-4`, `gemini-3.8-flash`, `gemini-3.1-pro`) are never reachable through GhostCLI; they are accessed natively via Google Antigravity (`agy` CLI in Herdr).
83
+ Gemini models (`gemini-4-hydrogen`, `gemini-4-argon`, `gemini-3.8-flash`) are never reachable through GhostCLI; they are accessed natively via Google Antigravity (`agy` CLI in Herdr). `Orchestrator (Argon)` is a dynamic mode that splits tasks between planning and implementation within one session — it is not a standalone model. Gemini 4 Carbon exists but is unavailable on current plan tier.
76
84
 
77
85
  ---
78
86
 
79
87
  ## 6. Effort Constraint & Fallback Rules
80
88
 
81
89
  - **Effort Constraint Rule**: Each task in a plan must explicitly declare one of the registered models and an allowed effort level matching the tables above.
82
- - **Quota Fallback Rule**: When Gemini 4 Argon reaches quota cooldown, immediately fallback to Gemini 3.8 Flash (for fast scans/reviews) or Claude Sonnet 5 / GPT-6 Astra via GhostCLI.
90
+ - **Quota Fallback Rule**: When Gemini 4 Argon reaches quota cooldown, fallback in order: Gemini 4 Hydrogen → Gemini 3.8 Flash → GhostCLI models (Claude/GPT).
83
91
  - **Headless Execution Limitation**: The `agy` CLI requires a TTY for all execution modes including `--print`. Non-interactive/headless invocations from scripts, CI, or piped environments will hang indefinitely. Use Herdr split/tab/window modes instead, or run `agy` interactively.