Improve aqa & testgen

docs/web/docs/testgen-flow.md

@@ -35,8 +35,8 @@ Prepare the minimum inputs that materially affect output quality:
 
 - A Jira ticket key or Jira ticket URL.
 - Confluence page URLs if you already know the relevant pages. This lets the workflow skip weak auto-search results.
-- Access to the Jira and Confluence retrieval path used in your project.
-- If you want Phase 6, access to TestRail and the ability to provide or create the target section and share its `section_id`. Project and suite details may also be needed when your setup cannot detect them from the current ticket and user profile.
+- Access to the Jira and Confluence retrieval path used in your project **when Phase 1 is guided**; otherwise be ready to paste ticket and wiki content (questionnaire path).
+- If you want Phase 6 with **guided** TestRail export, TestRail MCP access and the ability to provide or create the target section and share its `section_id`. For **questionnaire** export, be ready to supply `project_id`, `suite_id`, and `section_id`, then import manually or paste back TestRail case IDs.
 - Time to review every phase result, especially `questions.md`, `requirements.md`, and `test-scenarios.md`.
 
 Workflow-specific preparation that improves results:
@@ -48,6 +48,16 @@ Workflow-specific preparation that improves results:
 
 For shared Rosetta setup and project-context customization, see [Usage Guide](/rosetta/docs/usage-guide/).
 
+## MCP capability file and questionnaire path
+
+Some workspaces do not attach Jira, Confluence, or TestRail MCP servers to the agent. TestGen (and AQA integrated Path A) still run if you add **`agents/mcp-capability.yaml`** at the repo root (copy the shape from [`mcp-capability.example.yaml`](https://github.com/griddynamics/rosetta/blob/main/instructions/r2/core/templates/mcp-capability.example.yaml)) and optionally **`agents/user-instructions/mcp-guidance.md`** for stable IDs and naming rules.
+
+- **`mcp.mode: absent`** — use structured questionnaires in Phase 1 and Phase 6; no live MCP calls.
+- **`mcp.mode: capable`** with per-integration `true`/`false` — mix guided and questionnaire per system (for example Jira guided, Confluence paste-only).
+- **User override** in the task text (for example “no MCP” or “use MCP”) wins for that run over the YAML file.
+
+Workflow details live in [`mcp-capability-interaction.md`](https://github.com/griddynamics/rosetta/blob/main/instructions/r2/core/workflows/mcp-capability-interaction.md) (routing), [`mcp-tool-resolution.md`](https://github.com/griddynamics/rosetta/blob/main/instructions/r2/core/workflows/mcp-tool-resolution.md) and [`agent-action-resolution.md`](https://github.com/griddynamics/rosetta/blob/main/instructions/r2/core/workflows/agent-action-resolution.md) (placeholder binding), [`testgen-flow-data-collection.md`](https://github.com/griddynamics/rosetta/blob/main/instructions/r2/core/workflows/testgen-flow-data-collection.md), and [`testgen-flow-test-case-export.md`](https://github.com/griddynamics/rosetta/blob/main/instructions/r2/core/workflows/testgen-flow-test-case-export.md).
+
 ## How To Start
 
 Example prompts that match the workflow:

gain.json

@@ -46,7 +46,11 @@
     "vfs": "Virtual File System, after stripping release and org prefix from the physical file path",
     "bundling": "Multiple documents at the same VFS resource path are merged into a single XML response"
   },
+  "rosettaInstructionUpdates": {
+    "aqaFlowPhase1ExecutionMode": "AQA workflow: new Phase 1 Task 0 locks execution mode to integrated (TestRail+Confluence Path A) or minimal-input / agent-led (Path B). Until the user message contains explicit mode keywords (integrated, minimal-input, Integrated AQA, or minimal-input / agent-led) or the user answers the one-time mode question, the agent must not create agents/plans/aqa-*.md, must not write **AQA execution mode**, and must not mark Phase 1 complete in agents/aqa-state.md. Inferring minimal-input from missing TestRail or local HTML alone is forbidden. Phase 2 and Phase 6 instructions aligned; docs/web/docs/aqa-flow.md updated; mirrored under plugins (Cursor, Copilot, Codex).",
+    "mcpCapabilityBPlusATestgenAqa": "Shared fragment mcp-capability-interaction.md: workspace agents/mcp-capability.yaml (see instructions/r2/core/templates/mcp-capability.example.yaml) plus user-message overrides route guided MCP vs questionnaire. AQA Path A and TestGen Phases 1 and 6 acquire this fragment; TestGen adds Step 4Q (Confluence questionnaire), Phase 6 Step 0/1Q/5Q for TestRail without MCP; plugin mirrors updated."
+  },
   "versions": {
-    "rosetta": "2.0.19"
+    "rosetta": "2.0.20"
   }
 }

instructions/r2/core/templates/mcp-capability.example.yaml

@@ -0,0 +1,20 @@
+# Copy to workspace root as: agents/mcp-capability.yaml
+# See: instructions/r2/core/workflows/mcp-capability-interaction.md
+# Placeholder binding (when workflows use tokens): mcp-tool-resolution.md, agent-action-resolution.md
+
+mcp:
+  # "capable" = agent may call MCPs for true integrations after reading mcp-guidance.md
+  # "absent" = use questionnaire + user paste; do not call external MCPs
+  mode: capable
+
+  # Per-integration toggles (optional). If omitted, agent may still try when mode is capable
+  # and workflow needs that integration—unless workflow says otherwise.
+  jira: true
+  # Confluence: one product, two YAML key names used by different workflows (not two layers):
+  # - TestGen (`testgen-flow-data-collection.md`, etc.) reads `confluence`.
+  # - AQA (`aqa-flow-data-collection.md`) reads `atlassian_confluence` (Atlassian MCP naming).
+  # If you use both workflow families, keep both booleans in sync. If you use only one family,
+  # you may omit the key the other family reads (omitted + capable → guided per that workflow).
+  confluence: true
+  testrail: true
+  atlassian_confluence: true

instructions/r2/core/workflows/agent-action-resolution.md

@@ -0,0 +1,18 @@
+---
+name: agent-action-resolution
+description: Maps `{agent-action:…}` tokens to host IDE or agent-runtime file tools (read, write, patch).
+alwaysApply: false
+baseSchema: docs/schemas/phase.md
+---
+
+# `{agent-action:…}` resolution (host file tools)
+
+Use when a workflow uses **`{agent-action:read-file}`**, **`{agent-action:write-file}`**, or **`{agent-action:patch-file}`**. These are **not** remote MCP calls and are **not** covered by **`mcp-tool-resolution.md`**.
+
+| Token | Intent |
+|-------|--------|
+| `{agent-action:read-file}` | Read a file from the workspace via the host agent’s file-read tool |
+| `{agent-action:write-file}` | Create or overwrite a workspace file via the host agent’s file-write tool |
+| `{agent-action:patch-file}` | Apply targeted edits via the host agent’s patch / search-replace tool |
+
+**Resolution:** Map each token to the host IDE’s or agent runtime’s **documented** file operation (names differ: e.g. `Read`, `Write`, `StrReplace`, `apply_patch`). If the host exposes no matching capability, **STOP**, tell the user, and **WAIT** — do not substitute an MCP tool.