docs(memory): add 260224-wishlist-custodian — project registration UX

7 items for streamlining project-to-custodian registration. Baseline
~27K tokens; target ~1.5K (18x) after W1+W6+W3+W2 implemented.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-02-24 21:42:49 +01:00
parent 880de1870a
commit 715a7169ba

View File

@@ -0,0 +1,200 @@
# Custodian Wishlist — Project Registration UX
_Date: 2026-02-24 | Context: first project registration (markitect_project)_
---
## Baseline: Token Cost of This Session
Registering markitect to the custodian required significant exploratory work.
Rough token estimates (input + output combined):
| Activity | ~Tokens | Avoidable? |
|---|---|---|
| Failed `settings.json` edit → full schema dump in error | 12,000 | ✅ Yes |
| Reading `settings.local.json` (irrelevant historical allow rules) | 4,000 | ✅ Yes |
| Reading `mcp_server/server.py` in full to understand tools | 3,000 | Partially |
| Exploring API via OpenAPI spec + state/summary | 2,500 | Partially |
| Discovering `~/.claude.json` as user-scope MCP store | 1,500 | ✅ Yes |
| Manual `cwd` fix after `claude mcp add-json` dropped it | 800 | ✅ Yes |
| Small bash outputs, listings, failed `mcp list/get` calls | 2,000 | ✅ Yes |
| Actual registration work (CLAUDE.md write, progress event) | 1,500 | — |
| **Total** | **~27,300** | |
| **Irreducible minimum** (if everything was documented) | **~1,500** | |
| **Potential reduction** | **~18×** | |
The dominant waste was a single failed tool call (wrong config file) that returned
~12K tokens of JSON schema. Second largest: reading files that were not relevant
to the registration task.
---
## Wishlist Items
### W1 — Document user-scope MCP config location in global CLAUDE.md (quick win, ~5 min)
**Problem:** The global CLAUDE.md instructs `get_state_summary()` but says nothing
about where the MCP server is registered or how an agent should verify/bootstrap it.
An agent unfamiliar with Claude Code internals will try `settings.json` first
(wrong), triggering a full schema dump.
**Fix:** Add one paragraph to `~/.claude/CLAUDE.md`:
> The `state-hub` MCP server is registered at user scope in `~/.claude.json`
> (key: `mcpServers`). To check: `python3 -c "import json; d=json.load(open('$HOME/.claude.json')); print(d.get('mcpServers', {}).keys())"`.
> To register from scratch: `claude mcp add-json -s user state-hub '<json>'`.
> The JSON config is in `/home/worsch/the-custodian/.mcp.json`.
> The `cwd` field must be added manually after registration (known claude bug).
**Impact:** Saves the schema dump alone (~12K tokens).
---
### W2 — `make register-project` in state-hub (high value, ~1 hr)
**Problem:** Registration requires: check API → get topic ID → verify MCP →
write CLAUDE.md → log progress event. These are always the same steps.
**Fix:** A Makefile target (or small shell script `scripts/register_project.sh`):
```bash
make register-project PROJECT_PATH=/home/worsch/railiance DOMAIN=railiance
# or:
scripts/register_project.sh railiance /home/worsch/railiance
```
The script should:
1. Verify API is reachable (fail fast with `make api` hint)
2. Fetch the topic ID for `DOMAIN` from `/topics/?status=active`
3. Check if state-hub is in `~/.claude.json`; register if missing (include cwd fix)
4. Write `$PROJECT_PATH/CLAUDE.md` from the template (W3)
5. POST a progress event marking the registration
**Token impact:** Reduces agent-driven registration to a single `Bash` call.
An agent simply invokes the script — no discovery needed.
---
### W3 — CLAUDE.md template for new projects (quick win, ~15 min)
**Problem:** The project CLAUDE.md was written from scratch. Its content is
identical for every project except topic ID and domain name.
**Fix:** Store a template at `state-hub/scripts/project_claude_md.template`:
```
# {PROJECT_NAME} — Claude Code Instructions
## Custodian State Hub Integration
This project is tracked as the **{DOMAIN}** domain in the Custodian State Hub.
Hub topic ID: `{TOPIC_ID}`
### Session Protocol
[... static content ...]
```
Used by `register_project.sh` (W2) with `sed` substitution.
---
### W4 — List topic IDs in canon project charters (low effort, future-proof)
**Problem:** An agent needs a live API call to get the topic ID, which requires
the API to be running and adds latency/tokens.
**Fix:** Add a frontmatter line to each `canon/projects/<domain>/project_charter_v0.1.md`:
```yaml
---
custodian_topic_id: 5571d954-0d30-4950-980d-7bcaaad8e3e2
domain: markitect
---
```
An agent can then `grep` for the topic ID without touching the API. Also acts
as a canonical cross-reference between the charter and the live state.
---
### W5 — MCP server tool reference card (medium effort, ~30 min)
**Problem:** An agent reads the full `server.py` to understand what tools are
available. The file is ~350 lines and includes HTTP plumbing, not just tool signatures.
**Fix:** Add `state-hub/mcp_server/TOOLS.md`:
```markdown
## Available Tools
| Tool | Key Args | When to use |
|------|----------|-------------|
| get_state_summary() | — | Session start; orientation |
| add_progress_event(summary, event_type, topic_id, detail) | summary required | Session end; any notable event |
| create_workstream(topic_id, title, owner) | | New workstream for a domain |
| create_task(workstream_id, title, priority) | | New task under a workstream |
| update_task_status(task_id, status) | status: todo/in_progress/blocked/done | |
| record_decision(title, decision_type, topic_id) | | Decisions made or pending |
| resolve_decision(decision_id, rationale, decided_by) | | Close a decision |
```
An agent can read this 20-line card instead of the full server file.
Saves ~2,500 tokens per session that needs to discover the tools.
---
### W6 — Note `cwd` drop as known issue in `.mcp.json` or Makefile (quick fix, 5 min)
**Problem:** `claude mcp add-json` silently drops the `cwd` field, causing the
server to fail if run from a different working directory. This required a manual
JSON patch to `~/.claude.json`.
**Fix (option A):** Add a comment/warning in the state-hub Makefile or README:
> ⚠ `claude mcp add-json` drops `cwd`. After running, patch `~/.claude.json`
> manually: `python3 scripts/patch_mcp_cwd.py`
**Fix (option B):** Avoid `-m mcp_server.server` altogether — use the full
absolute path to `server.py` and add the state-hub dir to `PYTHONPATH` in env:
```json
"command": "/home/worsch/the-custodian/state-hub/.venv/bin/python",
"args": ["/home/worsch/the-custodian/state-hub/mcp_server/server.py"],
"env": { "PYTHONPATH": "/home/worsch/the-custodian/state-hub", "API_BASE": "http://127.0.0.1:8000" }
```
This works without `cwd` and is what the custodian's `.mcp.json` already does
with `-m` — switching to the absolute path + PYTHONPATH env eliminates the bug entirely.
---
### W7 — Idempotent `check-custodian` hook or command in project sessions (longer term)
**Problem:** The global CLAUDE.md says "start every session with get_state_summary()"
but there's no enforcement. An agent that hasn't read the global CLAUDE.md won't know.
**Fix:** Add a `SessionStart` hook to `~/.claude/settings.json`:
```json
"hooks": {
"SessionStart": [{
"hooks": [{
"type": "command",
"command": "curl -sf http://127.0.0.1:8000/state/health && echo 'State Hub: OK' || echo 'State Hub: OFFLINE — run: cd ~/the-custodian/state-hub && make api'",
"statusMessage": "Checking Custodian State Hub…"
}]
}]
}
```
This doesn't call get_state_summary() (which requires MCP), but at least surfaces
whether the API is up, so the agent knows before trying to use the tools.
---
## Priority Order
| # | Item | Effort | Token Saving | Do When |
|---|------|--------|--------------|---------|
| W1 | Document MCP config location in global CLAUDE.md | 5 min | 12K / registration | Now |
| W6 | Fix `cwd` drop (use absolute path + PYTHONPATH) | 10 min | 800 / registration | Now |
| W3 | CLAUDE.md template | 15 min | 500 / registration | Next custodian session |
| W2 | `register_project.sh` script + `make register-project` | 1 hr | 25K / registration | Before next project |
| W5 | `TOOLS.md` reference card | 30 min | 2,500 / session | Next custodian session |
| W4 | Topic IDs in canon charters | 20 min | 200 / registration | Low urgency |
| W7 | SessionStart hook for API health | 30 min | Indirect | After W2 |