Files
isphere-ai-bridge/docs/go-mcp-minimal-plan.md
2026-07-06 20:46:49 +08:00

924 lines
21 KiB
Markdown

# Go MCP Minimal Plan Nodes
> **Process rule:** This is the single active plan for the next phase. The final MCP direction is Go. Python MCP, Python tests, offline-lab/native-lab Python layers are not part of the active path.
**Goal:** Build a minimal Go MCP server that exposes the existing C# `ISphereWinHelper.exe` read-only operations as MCP tools.
**Architecture:** Go is the MCP/service layer. C# is the Windows/UI Automation execution layer. Go calls C# through one-shot stdin/stdout JSON using `isphere.helper.v1`.
**Tech Stack:** Go, MCP stdio, C# .NET Framework WinHelper, PowerShell verification scripts.
---
## 1. Node Tree
```text
N0 Baseline cleanup and direction lock
-> N1 C# WinHelper baseline verification
-> N2 Go module skeleton
-> N3 Go helper protocol contract
-> N4 Go helper process client
-> N5 Helper client live verification
-> N6 MCP library decision
-> N7 Go MCP server skeleton
-> N8 Four read-only MCP tool handlers
-> N9 MCP stdio smoke verification
-> N10 Operator runbook and process scripts
-> N11 Commit/review checkpoint
-> N12-pre Offline evidence intake
-> N12 Real logged-in UIA capture gate
```
Hard rule: **no node may start until the previous node's acceptance conditions are met.**
---
## 2. Node Details and Acceptance Conditions
## N0: Baseline cleanup and direction lock
**Purpose:** Ensure the project is no longer split between Python MCP and Go MCP.
**Preconditions:**
- Current branch is `codex/isphere-win-helper-first-phase` unless explicitly changed by the user.
- Python MCP temporary layer has been removed.
**Actions:**
- Confirm tracked files only include current docs/prompts/C# helper/scripts/evidence placeholders.
- Confirm no active `src/isphere_ai_bridge`, Python MCP package, or Python test suite is tracked.
- Confirm `docs/go-csharp-helper-boundary.md` remains the source boundary document.
**Expected outputs:**
- Clean project direction: C# helper now, Go MCP next.
- No Python MCP fallback path.
**Acceptance conditions:**
```powershell
git status --short
```
Pass if:
- No unexpected modified or untracked files exist except the active plan file while editing.
- `git ls-files` includes `native/ISphereWinHelper/*`.
- `git ls-files` does not include `src/isphere_ai_bridge/*`.
- `git ls-files` does not include `tests/test_*.py`.
**Stop conditions:**
- Python MCP files reappear.
- C# helper files are missing.
- Worktree has unrelated dirty files.
**Next node:** N1.
---
## N1: C# WinHelper baseline verification
**Purpose:** Prove the existing C# helper still works before writing Go code.
**Preconditions:**
- N0 accepted.
- `scripts/build-win-helper.ps1` exists.
- `scripts/verify-win-helper.ps1` exists.
**Actions:**
Run:
```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\verify-win-helper.ps1
```
**Expected outputs:**
- `runs/win-helper/ISphereWinHelper.exe` is built locally.
- `version`, `self_check`, `scan_windows`, and `dump_uia` all run through JSON.
**Acceptance conditions:**
Pass if output contains:
```json
{"ok":true}
```
And confirms:
- `helper` path points to `runs\win-helper\ISphereWinHelper.exe`.
- `version` is `0.1.0`.
- `scan_window_count` is a number.
- `uia_available` is `true` or a valid boolean field.
- Temporary WinForms `dump_uia` succeeds.
**Stop conditions:**
- Build fails.
- Helper stdout is not JSON.
- `dump_uia` cannot capture the temporary WinForms window.
- Helper requires admin rights.
**Next node:** N2.
---
## N2: Go module skeleton
**Purpose:** Create the smallest Go project shell.
**Preconditions:**
- N1 accepted.
- Go toolchain is available on this machine.
**Actions:**
Create:
```text
go.mod
```
Initial module name:
```text
isphere-ai-bridge
```
Do not create server code yet.
**Expected outputs:**
- A Go module exists.
- No dependencies are added unless required by standard Go tooling.
**Acceptance conditions:**
Run:
```powershell
go version
go list ./...
```
Pass if:
- `go version` exits 0.
- `go list ./...` exits 0 or reports only the root module with no packages before packages are created.
- No Python files are added.
**Stop conditions:**
- Go is not installed.
- Module name conflicts with repository layout.
- Toolchain tries to create unrelated files outside the repo.
**Commit boundary:** optional; can be combined with N3.
**Next node:** N3.
---
## N3: Go helper protocol contract
**Purpose:** Model `isphere.helper.v1` in Go before process execution.
**Preconditions:**
- N2 accepted.
- `docs/go-csharp-helper-boundary.md` is available.
**Actions:**
Create:
```text
internal/helperclient/contract.go
internal/helperclient/client_test.go
```
Define minimal structs:
```text
HelperRequest
HelperResponse
HelperError
```
Required fields:
```text
protocol
request_id
op
timeout_ms
args
ok
data
error
warnings
```
**Expected outputs:**
- JSON marshaling generates `isphere.helper.v1` requests.
- JSON unmarshaling accepts C# helper success and error responses.
**Acceptance conditions:**
Run:
```powershell
go test ./internal/helperclient -run Contract -v
```
Pass if tests prove:
- `version` request marshals with `protocol="isphere.helper.v1"`.
- success response unmarshals with `ok=true` and `data.helper_name="ISphereWinHelper"`.
- error response unmarshals with `ok=false` and `error.code="WINDOW_NOT_FOUND"`.
- unknown data fields are preserved as JSON maps or raw JSON.
**Stop conditions:**
- Contract field names differ from C# helper output.
- Tests require a real helper process; N3 must be pure JSON only.
**Commit boundary:**
```powershell
git add go.mod internal/helperclient/contract.go internal/helperclient/client_test.go
git commit -m "feat: add go helper contract"
```
**Next node:** N4.
---
## N4: Go helper process client
**Purpose:** Let Go call `ISphereWinHelper.exe --json` with one request and one response.
**Preconditions:**
- N3 accepted.
- C# helper verification from N1 passes.
**Actions:**
Create:
```text
internal/helperclient/client.go
```
Implement:
```text
Client
Client.Call(ctx, op, args)
```
Client responsibilities:
- Resolve helper exe path.
- Start helper with `--json`.
- Write request JSON to stdin.
- Read stdout.
- Capture stderr.
- Enforce context timeout.
- Return structured Go error for missing helper, timeout, bad JSON, non-zero exit.
**Expected outputs:**
- Go can call C# helper without MCP.
- Process execution logic lives only in `internal/helperclient`.
**Acceptance conditions:**
Run:
```powershell
go test ./internal/helperclient -run Client -v
```
Pass if tests prove:
- Missing helper path returns a structured error.
- Bad helper output returns a structured error using a test stub.
- Timeout returns a structured error using a test stub.
- Real `version` call returns `ok=true` when helper is built.
**Stop conditions:**
- Client leaks helper processes after timeout.
- Client requires admin rights.
- Client writes logs/files by default.
- Client duplicates C# UIA logic.
**Commit boundary:**
```powershell
git add internal/helperclient/client.go internal/helperclient/client_test.go
git commit -m "feat: add win helper process client"
```
**Next node:** N5.
---
## N5: Helper client live verification
**Purpose:** Verify Go-to-C# works against the real helper on this machine.
**Preconditions:**
- N4 accepted.
- `scripts/verify-win-helper.ps1` passes.
**Actions:**
Run:
```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\verify-win-helper.ps1
go test ./internal/helperclient -v
```
**Expected outputs:**
- C# helper works alone.
- Go helper client works against C# helper.
**Acceptance conditions:**
Pass if:
- PowerShell verification exits 0.
- Go tests exit 0.
- Test output includes a real `version` call.
- No tests click/type/send/upload/download.
**Stop conditions:**
- Tests only pass with mocked helper and never call the real C# helper.
- C# helper passes alone but Go client cannot call it.
**Commit boundary:** no separate commit unless tests/scripts changed.
**Next node:** N6.
---
## N6: MCP library decision
**Purpose:** Choose the Go MCP server library before writing server code.
**Preconditions:**
- N5 accepted.
- Network/package access is available or a local dependency strategy is chosen.
**Actions:**
Evaluate one current Go MCP library or official SDK option.
Record the decision inside the implementation commit or a short docs section. Do not create a large architecture document.
Decision must answer:
```text
package/module name
stdio support yes/no
tool registration style
JSON schema support
maintenance risk
why this is enough for four tools
```
**Expected outputs:**
- One MCP server package is selected.
- No competing MCP framework remains in the codebase.
**Acceptance conditions:**
Pass if:
```powershell
go list -m all
```
shows exactly one MCP-related Go dependency, or no external MCP dependency if a minimal stdio implementation is deliberately chosen.
Also pass if:
- The decision does not add HTTP server dependencies.
- The decision does not require Python or Node runtime.
- The decision supports stdio.
**Stop conditions:**
- MCP package cannot build on Windows.
- Package requires a background daemon.
- Package forces HTTP-only transport.
- Multiple MCP packages are added without a reason.
**Commit boundary:** can be combined with N7.
**Next node:** N7.
---
## N7: Go MCP server skeleton
**Purpose:** Create a Go binary that can start as an MCP stdio server.
**Preconditions:**
- N6 accepted.
**Actions:**
Create:
```text
cmd/isphere-mcp/main.go
```
Optional only if it keeps `main.go` small:
```text
internal/mcpserver/server.go
```
Do not register real tools yet if the library supports a clean empty server test. If the library requires tools immediately, register only placeholder-free real tool definitions from N8.
**Expected outputs:**
- `go build ./cmd/isphere-mcp` creates a binary.
- Server starts in stdio mode.
**Acceptance conditions:**
Run:
```powershell
go test ./...
go build ./cmd/isphere-mcp
```
Pass if:
- Build exits 0.
- No Python runtime is required.
- No C# helper is launched during simple server construction tests.
**Stop conditions:**
- Server starts an HTTP listener.
- Server invokes C# helper during initialization.
- Build requires unrelated services.
**Commit boundary:** can be combined with N8.
**Next node:** N8.
---
## N8: Four read-only MCP tool handlers
**Purpose:** Expose the first four read-only WinHelper operations through Go MCP.
**Preconditions:**
- N7 accepted.
- `internal/helperclient` accepted.
**Actions:**
Create:
```text
internal/tools/winhelper.go
internal/tools/winhelper_test.go
```
Register exactly these tools:
```text
win_helper_version
win_helper_self_check
win_helper_scan_windows
win_helper_dump_uia
```
Tool behavior:
```text
win_helper_version -> helper op version
win_helper_self_check -> helper op self_check
win_helper_scan_windows -> helper op scan_windows
win_helper_dump_uia -> helper op dump_uia
```
**Expected outputs:**
- Four tool definitions exist.
- Tool handlers call `internal/helperclient`.
- No handler performs UI logic directly.
**Acceptance conditions:**
Run:
```powershell
go test ./internal/tools -v
go test ./...
```
Pass if tests prove:
- Exactly four WinHelper tools are registered for this phase.
- Tool names match the list above.
- Tool schemas include only needed parameters:
- `include_all_visible` for scan.
- `hwnd`, `max_depth`, `include_text`, `max_children` for dump.
- No send/search/file tools exist.
**Stop conditions:**
- A fifth real action tool is added.
- A handler clicks, types, sends, uploads, downloads, or edits files.
- Tool code bypasses `internal/helperclient`.
**Commit boundary:**
```powershell
git add cmd/isphere-mcp internal/mcpserver internal/tools go.mod go.sum
git commit -m "feat: add minimal go mcp server"
```
**Next node:** N9.
---
## N9: MCP stdio smoke verification
**Purpose:** Verify the Go MCP binary can run over stdio and expose the four tools.
**Preconditions:**
- N8 accepted.
**Actions:**
Create if useful:
```text
scripts/verify-go-mcp.ps1
```
The script should:
- Build C# helper.
- Build Go MCP binary.
- Start Go MCP over stdio or use the MCP library test harness.
- List tools.
- Confirm the four expected tools exist.
- Call at least `win_helper_version`.
**Expected outputs:**
- A repeatable local verification command.
**Acceptance conditions:**
Run:
```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\verify-go-mcp.ps1
```
Pass if output includes:
```json
{"ok":true}
```
And confirms:
- Four tool names are present.
- `win_helper_version` returns `ISphereWinHelper`.
- No real iSphere login is required.
- No message/file/search action is present.
**Stop conditions:**
- Verification needs manual interaction.
- Verification needs Python.
- Tool listing cannot be automated.
**Commit boundary:**
```powershell
git add scripts/verify-go-mcp.ps1
git commit -m "test: add go mcp verification script"
```
**Next node:** N10.
---
## N10: Operator runbook and process scripts
**Purpose:** Make the current tool usable by a human operator without reading code.
**Preconditions:**
- N9 accepted.
**Actions:**
Create:
```text
docs/go-mcp-runbook.md
```
Update only if needed:
```text
README.md
```
Runbook must include:
- Build C# helper.
- Verify C# helper.
- Build Go MCP.
- Verify Go MCP.
- Configure MCP client command.
- Allowed tools.
- Explicit non-goals.
- Real-login UIA capture procedure.
**Expected outputs:**
- A user can run the first phase without asking for hidden context.
**Acceptance conditions:**
Pass if runbook includes exact commands for:
```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\verify-win-helper.ps1
go test ./...
go build ./cmd/isphere-mcp
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\verify-go-mcp.ps1
```
And includes these boundaries:
```text
no login automation
no send message
no send file
no receive/download file
no process injection
no hook
no memory reading
```
**Stop conditions:**
- Runbook says Python MCP is required.
- Runbook implies sending is implemented.
- Runbook omits approval boundaries for future sends.
**Commit boundary:**
```powershell
git add docs/go-mcp-runbook.md README.md
git commit -m "docs: add go mcp runbook"
```
**Next node:** N11.
---
## N11: Commit and review checkpoint
**Purpose:** Freeze the first Go MCP phase before real iSphere data is introduced.
**Preconditions:**
- N10 accepted.
**Actions:**
Run:
```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\verify-win-helper.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\verify-go-mcp.ps1
go test ./...
go build ./cmd/isphere-mcp
git status --short
```
**Expected outputs:**
- All local verification passes.
- Worktree is clean after commits.
**Acceptance conditions:**
Pass if:
- `git status --short` has no output.
- Latest commits are small and tied to N3/N4/N8/N9/N10.
- No Python MCP files exist.
- No search/send/file tools exist.
**Stop conditions:**
- Any verification fails.
- Unrelated generated files are staged.
- Worktree contains ignored build outputs outside `runs/`.
**Next node:** N12-pre if the current environment cannot log in; otherwise N12 may start only when the real logged-in current-environment preconditions below are met.
---
## N12-pre: Offline evidence intake
**Purpose:** Define and validate offline evidence copied from a separate logged-in test sandbox when the current repository environment cannot log in to iSphere.
**Preconditions:**
- N11 accepted.
- Current environment cannot manually log in to iSphere / IMPlatformClient.
- Business manager approves offline evidence intake as a pre-gate only.
**Actions:**
- Follow `docs/offline-evidence-intake-plan.md`.
- Prefer mode A, **Full sandbox artifact bundle**. The user does not need to understand or generate JSON; they only copy folders from the logged-in test sandbox.
- Recommended full-bundle shape:
```text
runs/offline-evidence-intake/<evidence-id>/
copy-notes.txt
raw/
install/
user-profile/
programdata/
shortcuts/
temp-importal-localcache/
other-candidates/
metadata/
notes/
```
- Common candidate source paths include install directories, `%APPDATA%`, `%LOCALAPPDATA%`, `%PROGRAMDATA%`, `Documents`, Desktop/Start Menu shortcuts, and `%TEMP%\importal` or `localcache` folders.
- Pay attention to historical iSphere anchors such as `importal\localcache`, `LoginLog`, `IMPP.Util.dll`, `IMPlatformClient`, and `iSphere`.
- Optional mode B, **Minimal redacted UIA package**, is only for advanced operators who already know how to produce JSON:
```text
runs/offline-evidence-intake/<evidence-id>/
manifest.json
scan_windows-redacted.json
dump_uia-main-redacted.json
```
- For optional mode B, confirm the package states `n12_status = offline_pre_only_not_n12_pass` and uses only:
```text
version
self_check
scan_windows
dump_uia
```
**Expected outputs:**
- Full sandbox artifact bundle for read-only offline inventory, or optional minimal UIA JSON package for advanced users.
- No claim that the current environment has a logged-in iSphere window.
- No claim that N12 passed.
**Acceptance conditions:**
Pass if:
- In mode A, the copied bundle preserves directory structure and includes at least one useful `raw/` source area plus `copy-notes.txt` when available.
- In mode A, repository-side handling is read-only inventory: do not directly run unknown binaries and do not modify original copied files.
- In mode B, required JSON files exist and parse as JSON, and the manifest records offline pre-evidence only, not N12 pass.
- For both modes, the route does not design or perform automatic login, send message, send file, receive file, injection, hook, memory reading, or endpoint-security bypass.
**Stop conditions:**
- Any step requires the user to understand JSON before they can provide the recommended full bundle.
- Any step directly runs unknown binaries from the copied bundle.
- Any step modifies original copied files.
- Package is used to design automatic login, sending, file transfer, injection, hook, memory reading, or endpoint-security bypass.
- Any reader treats the offline package as a replacement for true N12.
**Next node:** N12 remains blocked until the current environment can manually log in and perform a live current-environment UIA capture.
---
## N12: Real logged-in UIA capture gate
**Purpose:** Capture real iSphere UIA evidence after the user provides/login context.
**Preconditions:**
- N11 accepted.
- Optional N12-pre offline evidence, if used, has been validated only as pre-gate context and not as N12 pass evidence.
- User manually opens and logs in to iSphere.
- iSphere main window is visible.
**Actions:**
Use Go MCP tools:
```text
win_helper_scan_windows
win_helper_dump_uia
```
Save output under:
```text
runs/real-loggedin-lab/uia-dumps/
```
**Expected outputs:**
- One or more real iSphere UIA dump JSON files.
- Window metadata showing process name/title/handle.
**Acceptance conditions:**
Pass if:
- Candidate iSphere/IMPlatform window is found.
- UIA root node is a window/control tree.
- Dump includes enough fields to identify controls:
- `name`
- `control_type`
- `automation_id`
- `class_name`
- `children`
- No clicking, typing, sending, uploading, receiving, or downloading occurs.
- No passwords/tokens are captured.
**Stop conditions:**
- User is not logged in.
- No candidate window appears.
- UIA tree is empty or inaccessible.
- Captured data contains sensitive secrets that need redaction before reporting.
- Current environment cannot log in; in that case, use only N12-pre offline analysis and do not mark N12 passed.
**Next node:** Future selector design. Do not implement selectors until N12 passes.
---
## 3. Global Acceptance Rules
Every node must satisfy these rules:
1. No Python MCP code.
2. No real send/search/file action before N12.
3. No process injection, hook, credential extraction, token extraction, or memory reading.
4. No broad `git add -A`; stage exact paths.
5. Every code node uses tests first.
6. Every completion claim needs a fresh verification command.
7. Every failed gate stops the plan and requires a short report before continuing.
---
## 4. Current Next Node
Current guidance after N0-N11 completion is:
```text
N12-pre if the current environment cannot log in; N12 only after current-environment manual login is possible.
```
Immediate next action:
```text
If login is unavailable here, collect/validate the full sandbox artifact bundle described in docs/offline-evidence-intake-plan.md. Minimal manifest/scan/dump JSON is optional and only for advanced operators.
```
Do not treat offline evidence as N12 pass. Do not start selector/action design until true N12 passes in the current environment.