ma-agents 3.14.0 → 3.14.1

package/lib/templates/instruction-block-onprem.template.md

@@ -1,86 +1,86 @@
-## On-Prem / Local-LLM Guardrails
-
-These rules apply ONLY when this project is installed with `profile: on-prem`.
-They are appended to the universal block by the composer in `lib/installer.js`.
-Local LLMs (Nemotron, Qwen, DeepSeek, Llama-3, etc. served via vLLM, Ollama, or
-TGI) fail in patterns cloud LLMs rarely exhibit — the rules below pin those
-failure modes down explicitly. Keep these rules verbatim in every response
-context where tool use is possible.
-
-### Reasoning mode: `/no_think` on planning-phase prompts
-
-Local reasoning-capable models (Nemotron-variants and similar) default to
-chain-of-thought reasoning that bloats planning-phase prompts with internal
-deliberation the operator does not need to read. Prepend the literal token
-`/no_think` as the first line of any planning-phase system prompt or user turn
-you compose. The token is consumed by the serving layer and suppresses the
-model's reasoning trace on that turn.
-
-- Planning-phase turns (PM, Architect, Tech Lead): begin the turn with
-  `/no_think` on its own line. Reasoning-mode OFF.
-- Implementation-phase turns (Dev): omit `/no_think`. Reasoning-mode ON is
-  desirable for stepwise code synthesis and debugging.
-- Review / QA turns: omit `/no_think` when the review benefits from explicit
-  reasoning (root-cause analysis). Include `/no_think` for mechanical checks
-  (lint, style, formatting).
-
-If the downstream serving layer does not recognize `/no_think`, it is a no-op
-text token — safe to include unconditionally on planning turns.
-
-### No writes to `~/.claude/` or any user home directory
-
-Local LLMs frequently hallucinate paths under `~/.claude/`, `~/.cache/`,
-`~/Library/`, or `%APPDATA%` — imitating patterns learned from Claude Code and
-Cursor training data. These paths are OUTSIDE the project and cross-contaminate
-other projects on the same machine.
-
-- NEVER create, write, or modify files under `~/.claude/`, `~/.cache/`,
-  `~/Library/`, `~/AppData/`, `%APPDATA%`, or any path that resolves outside
-  the current project directory.
-- All project artifacts — code, configuration, logs, scratch notes, and agent
-  state — MUST land under the current working directory (the project root) or
-  an explicitly-named subdirectory thereof.
-- When a tool call appears to target a home-directory path, refuse the write
-  and respond in text explaining the violation. Ask the user for an explicit
-  in-project path before proceeding.
-
-### No `str_replace_editor` or Claude Code-specific tools
-
-Local LLMs hallucinate Anthropic-proprietary tools — most commonly
-`str_replace_editor`, `text_editor_20241022`, and `computer_use_preview` — that
-do NOT exist outside the Anthropic API. Calling them against a local-LLM
-serving layer produces a tool-not-found error or, worse, a silent no-op.
-
-- Do NOT emit tool calls named `str_replace_editor`, `text_editor_*`,
-  `computer_use_*`, or any other tool whose name includes `str_replace_editor`
-  or matches Anthropic-specific tool schemas.
-- Use only tools enumerated in the active tool manifest (`MANIFEST.yaml`) or
-  the IDE's native tool surface (Roo Code, Cline, OpenCode native tools).
-- When you want to edit a file, use the native file-write tool of the active
-  agent — not `str_replace_editor`. If unsure what tool is available, list
-  available tools or ask the user before emitting a tool call.
-
-### Per-phase reasoning and sampling guidance
-
-Local LLMs require tighter sampling control than cloud LLMs. Use these defaults
-unless the serving layer overrides them.
-
-- **Planning phase (PM, Architect, Tech Lead):**
-  - Reasoning: OFF (`/no_think` prepended).
-  - Temperature: low (0.0 – 0.3). Planning artifacts should be deterministic
-    and reproducible.
-  - Top-p: 0.9 or unset. Top-k: unset.
-  - Max tokens: generous (8k+) — planning documents are long.
-- **Implementation phase (Dev):**
-  - Reasoning: ON (omit `/no_think`).
-  - Temperature: moderate (0.3 – 0.6). Code synthesis benefits from controlled
-    exploration but not creative rewriting.
-  - Top-p: 0.95 or unset. Top-k: unset.
-  - Max tokens: generous (8k+) — full-file rewrites are common.
-- **Review / QA phase:**
-  - Reasoning: ON for root-cause analysis; OFF for mechanical checks.
-  - Temperature: low (0.0 – 0.2). Reviews should be deterministic.
-
-If the serving layer applies its own sampler defaults, the per-phase guidance
-above is advisory — but the phase boundary and `/no_think` placement are
-load-bearing and MUST be honored on every turn.
+## On-Prem / Local-LLM Guardrails
+
+These rules apply ONLY when this project is installed with `profile: on-prem`.
+They are appended to the universal block by the composer in `lib/installer.js`.
+Local LLMs (Nemotron, Qwen, DeepSeek, Llama-3, etc. served via vLLM, Ollama, or
+TGI) fail in patterns cloud LLMs rarely exhibit — the rules below pin those
+failure modes down explicitly. Keep these rules verbatim in every response
+context where tool use is possible.
+
+### Reasoning mode: `/no_think` on planning-phase prompts
+
+Local reasoning-capable models (Nemotron-variants and similar) default to
+chain-of-thought reasoning that bloats planning-phase prompts with internal
+deliberation the operator does not need to read. Prepend the literal token
+`/no_think` as the first line of any planning-phase system prompt or user turn
+you compose. The token is consumed by the serving layer and suppresses the
+model's reasoning trace on that turn.
+
+- Planning-phase turns (PM, Architect, Tech Lead): begin the turn with
+  `/no_think` on its own line. Reasoning-mode OFF.
+- Implementation-phase turns (Dev): omit `/no_think`. Reasoning-mode ON is
+  desirable for stepwise code synthesis and debugging.
+- Review / QA turns: omit `/no_think` when the review benefits from explicit
+  reasoning (root-cause analysis). Include `/no_think` for mechanical checks
+  (lint, style, formatting).
+
+If the downstream serving layer does not recognize `/no_think`, it is a no-op
+text token — safe to include unconditionally on planning turns.
+
+### No writes to `~/.claude/` or any user home directory
+
+Local LLMs frequently hallucinate paths under `~/.claude/`, `~/.cache/`,
+`~/Library/`, or `%APPDATA%` — imitating patterns learned from Claude Code and
+Cursor training data. These paths are OUTSIDE the project and cross-contaminate
+other projects on the same machine.
+
+- NEVER create, write, or modify files under `~/.claude/`, `~/.cache/`,
+  `~/Library/`, `~/AppData/`, `%APPDATA%`, or any path that resolves outside
+  the current project directory.
+- All project artifacts — code, configuration, logs, scratch notes, and agent
+  state — MUST land under the current working directory (the project root) or
+  an explicitly-named subdirectory thereof.
+- When a tool call appears to target a home-directory path, refuse the write
+  and respond in text explaining the violation. Ask the user for an explicit
+  in-project path before proceeding.
+
+### No `str_replace_editor` or Claude Code-specific tools
+
+Local LLMs hallucinate Anthropic-proprietary tools — most commonly
+`str_replace_editor`, `text_editor_20241022`, and `computer_use_preview` — that
+do NOT exist outside the Anthropic API. Calling them against a local-LLM
+serving layer produces a tool-not-found error or, worse, a silent no-op.
+
+- Do NOT emit tool calls named `str_replace_editor`, `text_editor_*`,
+  `computer_use_*`, or any other tool whose name includes `str_replace_editor`
+  or matches Anthropic-specific tool schemas.
+- Use only tools enumerated in the active tool manifest (`MANIFEST.yaml`) or
+  the IDE's native tool surface (Roo Code, Cline, OpenCode native tools).
+- When you want to edit a file, use the native file-write tool of the active
+  agent — not `str_replace_editor`. If unsure what tool is available, list
+  available tools or ask the user before emitting a tool call.
+
+### Per-phase reasoning and sampling guidance
+
+Local LLMs require tighter sampling control than cloud LLMs. Use these defaults
+unless the serving layer overrides them.
+
+- **Planning phase (PM, Architect, Tech Lead):**
+  - Reasoning: OFF (`/no_think` prepended).
+  - Temperature: low (0.0 – 0.3). Planning artifacts should be deterministic
+    and reproducible.
+  - Top-p: 0.9 or unset. Top-k: unset.
+  - Max tokens: generous (8k+) — planning documents are long.
+- **Implementation phase (Dev):**
+  - Reasoning: ON (omit `/no_think`).
+  - Temperature: moderate (0.3 – 0.6). Code synthesis benefits from controlled
+    exploration but not creative rewriting.
+  - Top-p: 0.95 or unset. Top-k: unset.
+  - Max tokens: generous (8k+) — full-file rewrites are common.
+- **Review / QA phase:**
+  - Reasoning: ON for root-cause analysis; OFF for mechanical checks.
+  - Temperature: low (0.0 – 0.2). Reviews should be deterministic.
+
+If the serving layer applies its own sampler defaults, the per-phase guidance
+above is advisory — but the phase boundary and `/no_think` placement are
+load-bearing and MUST be honored on every turn.

package/lib/templates/instruction-block-universal.template.md

@@ -1,29 +1,29 @@
-# AI Agent Skills - Planning Instruction
-
-You have access to a library of skills in your skills directory. Before starting any task:
-
-1. Read the skill manifest at {{MANIFEST_PATH}}
-2. Based on the task description, select which skills are relevant
-3. Read only the selected skill files
-4. Then proceed with the task
-
-Always load skills marked with always_load: true.
-Do not load skills that are not relevant to the current task.
-
-## Respond in TEXT vs. create FILES
-
-Choose your response medium deliberately. Defaulting to file creation when the user asked a question is a common failure mode — especially for coding agents running in web UIs.
-
-- **Create or modify FILES when the user's request contains file-action keywords:** `create`, `write`, `generate`, `build`, `implement` (and obvious synonyms such as `add`, `produce`, `refactor`, `fix`, `update <file>`). These signal a concrete artifact is expected.
-- **Respond in TEXT when the request contains text-response keywords:** `what do you think`, `how should we`, `discuss`, `opinion` (and obvious synonyms such as `explain`, `why`, `should I`, `compare`, `recommend`). These signal that a conversation is expected, not a deliverable.
-- **If unsure, respond in TEXT.** A text answer can always be followed by file creation on confirmation; an unwanted file cannot be cleanly undone.
-- **Never create `response.md`, `output.md`, or any similarly named scratch file as a reply.** A reply belongs in the chat transcript, not on disk.
-- **Confirm file paths before writing.** When you are about to create or modify a file whose path the user has not explicitly named, state the intended path in text and wait for confirmation, unless the path is unambiguous from the task context.
-
-## BMAD phase discipline
-
-BMAD-METHOD organizes work into declared phases (analysis, planning, architecture, story-creation, implementation, review). Respect the currently declared phase.
-
-- **Do not skip ahead to implementation during planning.** If the project is in a planning phase — or the user has asked for requirements, architecture, or a story — produce planning artifacts, not code.
-- **Do not retroactively plan after you have already coded.** If implementation has already started, flag the gap instead of fabricating back-dated planning documents.
-- The declared phase is established by the active skill, the story status, or an explicit statement from the user. When none of these is available, ask before assuming.
+# AI Agent Skills - Planning Instruction
+
+You have access to a library of skills in your skills directory. Before starting any task:
+
+1. Read the skill manifest at {{MANIFEST_PATH}}
+2. Based on the task description, select which skills are relevant
+3. Read only the selected skill files
+4. Then proceed with the task
+
+Always load skills marked with always_load: true.
+Do not load skills that are not relevant to the current task.
+
+## Respond in TEXT vs. create FILES
+
+Choose your response medium deliberately. Defaulting to file creation when the user asked a question is a common failure mode — especially for coding agents running in web UIs.
+
+- **Create or modify FILES when the user's request contains file-action keywords:** `create`, `write`, `generate`, `build`, `implement` (and obvious synonyms such as `add`, `produce`, `refactor`, `fix`, `update <file>`). These signal a concrete artifact is expected.
+- **Respond in TEXT when the request contains text-response keywords:** `what do you think`, `how should we`, `discuss`, `opinion` (and obvious synonyms such as `explain`, `why`, `should I`, `compare`, `recommend`). These signal that a conversation is expected, not a deliverable.
+- **If unsure, respond in TEXT.** A text answer can always be followed by file creation on confirmation; an unwanted file cannot be cleanly undone.
+- **Never create `response.md`, `output.md`, or any similarly named scratch file as a reply.** A reply belongs in the chat transcript, not on disk.
+- **Confirm file paths before writing.** When you are about to create or modify a file whose path the user has not explicitly named, state the intended path in text and wait for confirmation, unless the path is unambiguous from the task context.
+
+## BMAD phase discipline
+
+BMAD-METHOD organizes work into declared phases (analysis, planning, architecture, story-creation, implementation, review). Respect the currently declared phase.
+
+- **Do not skip ahead to implementation during planning.** If the project is in a planning phase — or the user has asked for requirements, architecture, or a story — produce planning artifacts, not code.
+- **Do not retroactively plan after you have already coded.** If implementation has already started, flag the gap instead of fabricating back-dated planning documents.
+- The declared phase is established by the active skill, the story status, or an explicit statement from the user. When none of these is available, ask before assuming.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ma-agents",
-  "version": "3.14.0",
+  "version": "3.14.1",
   "description": "NPX tool to install skills for AI coding agents (Claude Code, Gemini, Copilot, Kilocode, Cline, Cursor, Roo Code)",
   "main": "index.js",
   "bin": {