skills

2026-06-02 04:18:55 +07:00
parent 4e28c8db5d
commit e6bb5d6514
8 changed files with 373 additions and 0 deletions
@@ -0,0 +1,29 @@
+# Workflow Skill Docs Mirror Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Add a repo-local `wf-workflow` skill with agent-facing references and expose those same skill files through the MCP documentation source.
+
+**Architecture:** `docs/` remains project-facing documentation. `skills/wf-workflow/` becomes a self-contained agent-facing operating manual with distilled references. `src/wf_mcp/documentation.py` exposes the skill files as documentation resources under `wf://skills/...` without changing existing `wf://docs/...` resources.
+
+**Tech Stack:** Markdown skills, `wf_mcp.documentation`, `wf_platform.DocumentationResource`, pytest.
+
+---
+
+## Tasks
+
+- [ ] Add a failing docs resource test for `wf://skills/wf-workflow/SKILL.md` and one reference.
+- [ ] Create `skills/wf-workflow/SKILL.md`.
+- [ ] Create `skills/wf-workflow/references/workflow-lifecycle.md`.
+- [ ] Create `skills/wf-workflow/references/capabilities-and-wrappers.md`.
+- [ ] Create `skills/wf-workflow/references/draft-workspaces.md`.
+- [ ] Create `skills/wf-workflow/references/troubleshooting.md`.
+- [ ] Update `src/wf_mcp/documentation.py` to load the skill files as resources.
+- [ ] Verify focused docs tests and lint.
+
+## Guardrails
+
+- Do not replace the existing `skills/wf-cli` skill.
+- Do not make `SKILL.md` a link-only file.
+- Do not symlink skill references to `docs/`; checked-in distilled files are more portable.
+- Do not remove existing `wf://docs/...` resources.
@@ -0,0 +1,70 @@
+---
+name: wf-workflow
+description: Use when an agent needs to discover workflow capabilities, create or patch draft workspaces, save workflow or wrapper artifacts, deploy, run, resume, debug, or troubleshoot workflows in this repository.
+---
+
+# wf Workflow
+
+Use this skill for the workflow lifecycle, regardless of front door:
+
+- MCP tools exposed by `wf.workflow.*`
+- repo-local `wf` CLI commands
+- future `wf_api` callers
+
+Prefer small discovery calls, draft workspaces, validation, and bounded trace
+reads. Do not write raw workflow plans unless the user explicitly asks for the
+low-level escape hatch.
+
+## Workflow Lifecycle
+
+1. Discover sources with `wf.admin.list_sources` or `wf cap list`.
+2. Discover workflow-ready capabilities with `wf.workflow.list_capabilities`.
+3. Inspect one candidate with `wf.workflow.inspect_capability`.
+4. Create a patchable draft workspace with
+   `wf.workflow.create_draft_workspace_from_capability`.
+5. Patch targeted fields with focused helpers or JSON Patch.
+6. Validate with `wf.workflow.validate_draft_workspace`.
+7. Save with `wf.workflow.create_artifact_from_workspace` or
+   `wf.workflow.create_wrapper_from_workspace`.
+8. Save a deployment with `wf.workflow.save_deployment`.
+9. Validate with `wf.workflow.validate_deployment`.
+10. Run with `wf.workflow.run_deployment`.
+11. Inspect stopped runs with `wf.workflow.inspect_run`; read bounded trace
+    slices only when debugging.
+
+## Rules
+
+- Use workflow capabilities, not raw MCP tools, when building graphs.
+- Treat wrapper hints as scaffolding, not semantic truth.
+- Use draft workspaces for iterative authoring; avoid rewriting full drafts.
+- Use explicit source bindings at deployment time.
+- Keep traces bounded with `trace_range`.
+- Do not expect a newly saved workflow to become a new MCP tool mid-session.
+- For MCP content blocks, add explicit extraction/wrapper steps instead of
+  pretending `content` is always plain text.
+
+## References
+
+Read only the reference needed for the current task. Start with a small preview
+or search hit, then open the relevant section; do not dump every reference into
+context.
+
+Useful patterns:
+
+```bash
+# PowerShell preview
+Get-Content skills/wf-workflow/references/workflow-lifecycle.md -TotalCount 40
+
+# Targeted search
+rg -n "trace|resume|binding_missing|wrapper_hints" skills/wf-workflow/references
+```
+
+- `references/workflow-lifecycle.md`: end-to-end lifecycle and tool order.
+- `references/capabilities-and-wrappers.md`: raw capability vs workflow
+  capability, wrapper artifacts, and content-block handling.
+- `references/draft-workspaces.md`: draft shape, mappings, patches, and save
+  path.
+- `references/troubleshooting.md`: missing source/capability, unrunnable
+  deployment, run/trace issues.
+
+For shell-first workflows, also use the existing `wf-cli` skill.
@@ -0,0 +1,52 @@
+# Capabilities And Wrappers Reference
+
+Use this when deciding whether to call a capability directly, wrap it, or place
+it in a workflow.
+
+## Core Split
+
+```text
+raw MCP tool != workflow capability
+```
+
+A raw MCP tool may be callable and still be awkward for workflows. Workflow
+capabilities are graph-facing contracts with schemas and outcomes.
+
+## Discovery Order
+
+1. Use `wf.admin.list_sources` to understand owners.
+2. Use `wf.workflow.list_capabilities` for graph-ready capabilities.
+3. Use `wf.workflow.inspect_capability` for one full contract.
+4. Use `wf.workflow.call_capability` to test a single workflow-facing contract.
+
+## Wrapper Artifacts
+
+Use a wrapper artifact when the raw provider shape needs normalization:
+
+- provider status strings should become workflow outcomes
+- raw envelopes should be narrowed
+- `isError` or error blocks need explicit outcome routing
+- provider inputs are too broad or unstable for graph use
+
+Saved wrappers appear as workflow capabilities under source id `workflow`, with
+names like `workflow.echo_wrapper.v1`.
+
+## Wrapper Hints
+
+`inspect_capability` returns `wrapper_hints`. Treat them as authoring
+scaffolding:
+
+- `confidence=high`: simple shape, likely safe to validate first
+- `confidence=medium`: usable but review candidates
+- `confidence=low`: patch missing decisions before saving
+
+`next_actions` is guidance, not validation authority.
+
+## MCP Content Blocks
+
+MCP tools often return `content: [{type, text, ...}]`. Do not map this list
+into a string state field. Filter/extract text explicitly, or write a wrapper
+that decides what content types are acceptable.
+
+If a result exposes a convenience `text` field, inspect the capability schema
+and wrapper hints before using it. Do not assume every content block is text.
@@ -0,0 +1,72 @@
+# Draft Workspaces Reference
+
+Use draft workspaces for iterative workflow authoring. They are mutable and
+revisioned; artifacts are immutable and versioned.
+
+## Draft Shape
+
+A draft has:
+
+- `name`
+- `input_schema`
+- `state_schema`
+- `output_schema`
+- `start`
+- `steps`
+- `routes`
+- optional top-level `output`
+
+`steps` are keyed by stable ids. `routes` map outcomes to another step id or
+`__end__`.
+
+## Mapping Rules
+
+Step input bindings read graph values into node-local input:
+
+```json
+{
+  "target": { "root": "local", "parts": ["text"] },
+  "path": { "root": "input", "parts": ["text"] }
+}
+```
+
+Step output bindings write node-local output into workflow state:
+
+```json
+{
+  "source": { "root": "local", "parts": ["echoed"] },
+  "target": { "root": "state", "parts": ["echoed"] }
+}
+```
+
+Top-level workflow output uses `path` / `target`, not step-level
+`source` / `target`:
+
+```json
+{
+  "path": { "root": "state", "parts": ["echoed"] },
+  "target": { "root": "local", "parts": ["echoed"] }
+}
+```
+
+## Workspace Flow
+
+1. Create workspace from capability.
+2. Get workspace with `include_draft=true` before patching.
+3. Patch with current `revision`.
+4. Validate workspace.
+5. Save artifact or wrapper from workspace.
+
+If a patch returns `revision_conflict`, fetch the workspace again and retry
+against the latest revision.
+
+## Focused Helpers
+
+Prefer focused helpers over JSON Patch for common edits:
+
+- `set_draft_name`
+- `set_draft_route`
+- `set_step_input_map`
+- `set_step_output_map`
+
+Use JSON Patch for structural edits the helpers do not cover.
@@ -0,0 +1,48 @@
+# Workflow Troubleshooting Reference
+
+Use this when a source, capability, artifact, deployment, or run is missing,
+unrunnable, or surprising.
+
+## Missing Capability
+
+Check in this order:
+
+1. `wf.admin.list_sources`
+2. `wf.admin.inspect_source`
+3. `wf.workflow.list_capabilities`
+4. `wf.workflow.inspect_capability`
+
+Remember: MCP control tools are not workflow capabilities. They appear in
+MCP `tools/list`, not `wf.workflow.list_capabilities`.
+
+## Unrunnable Deployment
+
+Run `validate_deployment` before `run_deployment`.
+
+Common diagnostics:
+
+- `binding_missing`: deployment lacks a logical-to-concrete source binding.
+- `source_missing`: bound concrete source does not exist or is disabled.
+- `capability_missing`: required node/reducer is not available.
+- `schema_changed`: saved snapshot no longer matches current source.
+- `source_unreachable`: live check could not contact an upstream source.
+
+Use `live_check=true` only when you intentionally want to contact upstream
+sources. It may spawn stdio servers or perform network I/O.
+
+## Run Debugging
+
+If a run fails:
+
+1. Read `status`, `error`, `diagnostics`, and `trace_count`.
+2. Use `inspect_run` for stored summary.
+3. Use `read_run_trace` with a bounded range.
+
+Do not request full traces unless the user explicitly asks and the trace is
+known to be small.
+
+## Harness Problems
+
+Some LLM harnesses do not refresh `tools/list` mid-session. Do not rely on new
+saved workflows becoming new tools. Use `run_deployment` and `call_capability`
+instead.
@@ -0,0 +1,48 @@
+# Workflow Lifecycle Reference
+
+Use this when an agent needs to go from available capabilities to a saved,
+validated, runnable deployment.
+
+## Primary Path
+
+1. List sources.
+   - MCP: `wf.admin.list_sources`
+   - CLI: `wf cap list --format ids`
+2. List workflow capabilities.
+   - MCP: `wf.workflow.list_capabilities`
+   - CLI: `wf cap list`
+3. Inspect one capability.
+   - MCP: `wf.workflow.inspect_capability`
+   - CLI: `wf cap inspect <name>`
+4. Bootstrap a draft workspace.
+   - MCP: `wf.workflow.create_draft_workspace_from_capability`
+   - CLI: `wf draft create-from-capability <workspace_id> <capability>`
+5. Inspect/patch/validate the workspace until valid.
+   - Prefer focused helpers when available.
+   - Use JSON Patch only for general edits.
+6. Save an artifact.
+   - Full workflow: `create_artifact_from_workspace`
+   - Reusable wrapper: `create_wrapper_from_workspace`
+7. Save and validate a deployment.
+8. Run the deployment.
+9. Inspect the run summary first; read bounded traces only when needed.
+
+## Object Model
+
+- **Source**: owner of capabilities, such as `wf.std` or `everything.default`.
+- **Workflow capability**: graph-ready `NodeSpec` or saved wrapper artifact.
+- **Artifact**: immutable saved workflow or wrapper.
+- **Deployment**: mutable binding from artifact version to concrete sources.
+- **Run**: durable stopped execution record with status, output, and trace count.
+
+## Result Handling
+
+`run_deployment` returns compact status by default. Capture `run_id` even for
+completed or failed runs. Use:
+
+- `inspect_run` for compact stored result.
+- `read_run_trace` for explicit debug slices, for example
+  `{"start": 0, "limit": 25}`.
+- `resume_run` only for interrupted runs.
+
+Do not ask for unbounded traces.
@@ -13,6 +13,7 @@ from wf_platform import (
 def build_local_documentation_source(repo_root: Path) -> CapabilitySource:
    """Load project manuals into a provider-neutral local documentation source."""
    docs_dir = repo_root / "docs"
+    skills_dir = repo_root / "skills"
    return build_documentation_source(
        [
            _markdown_resource(
@@ -53,6 +54,50 @@ def build_local_documentation_source(repo_root: Path) -> CapabilitySource:
                title="Workflow Drafts",
                description="Draft and workspace authoring model for workflows.",
            ),
+            _markdown_resource(
+                path=skills_dir / "wf-workflow" / "SKILL.md",
+                name="wf.skills.wf_workflow.skill",
+                uri="wf://skills/wf-workflow/SKILL.md",
+                title="wf-workflow Skill",
+                description="Agent-facing workflow lifecycle skill entrypoint.",
+            ),
+            _markdown_resource(
+                path=skills_dir
+                / "wf-workflow"
+                / "references"
+                / "workflow-lifecycle.md",
+                name="wf.skills.wf_workflow.workflow_lifecycle",
+                uri="wf://skills/wf-workflow/references/workflow-lifecycle.md",
+                title="wf-workflow Workflow Lifecycle Reference",
+                description="Agent-facing workflow lifecycle and tool order.",
+            ),
+            _markdown_resource(
+                path=skills_dir
+                / "wf-workflow"
+                / "references"
+                / "capabilities-and-wrappers.md",
+                name="wf.skills.wf_workflow.capabilities_and_wrappers",
+                uri=("wf://skills/wf-workflow/references/capabilities-and-wrappers.md"),
+                title="wf-workflow Capabilities And Wrappers Reference",
+                description=(
+                    "Agent-facing raw capability, workflow capability, and "
+                    "wrapper guidance."
+                ),
+            ),
+            _markdown_resource(
+                path=skills_dir / "wf-workflow" / "references" / "draft-workspaces.md",
+                name="wf.skills.wf_workflow.draft_workspaces",
+                uri="wf://skills/wf-workflow/references/draft-workspaces.md",
+                title="wf-workflow Draft Workspaces Reference",
+                description="Agent-facing draft workspace authoring guidance.",
+            ),
+            _markdown_resource(
+                path=skills_dir / "wf-workflow" / "references" / "troubleshooting.md",
+                name="wf.skills.wf_workflow.troubleshooting",
+                uri="wf://skills/wf-workflow/references/troubleshooting.md",
+                title="wf-workflow Troubleshooting Reference",
+                description="Agent-facing workflow troubleshooting ladder.",
+            ),
        ],
        prompts=[
            DocumentationPrompt(
@@ -28,6 +28,8 @@ def test_server_exposes_platform_documentation_resources() -> None:
            assert "wf://docs/operator-manual" in uris
            assert "wf://docs/workflow-capabilities" in uris
            assert "wf://docs/workflow-drafts" in uris
+            assert "wf://skills/wf-workflow/SKILL.md" in uris
+            assert "wf://skills/wf-workflow/references/workflow-lifecycle.md" in uris

            result = await client.read_resource("wf://docs/operator-manual")
            assert isinstance(result[0], mcp_types.TextResourceContents)
@@ -44,6 +46,13 @@ def test_server_exposes_platform_documentation_resources() -> None:
                or "Do not use `source` at top level" in drafts_result[0].text
            )

+            skill_result = await client.read_resource(
+                "wf://skills/wf-workflow/SKILL.md"
+            )
+            assert isinstance(skill_result[0], mcp_types.TextResourceContents)
+            assert "name: wf-workflow" in skill_result[0].text
+            assert "Workflow Lifecycle" in skill_result[0].text
+
    asyncio.run(run_proxy())