docs(plan): spike 3 run labels F0-F5 per the controller's notes; untrack the gitignored gate report
This commit is contained in:
1 parent
3a589f9fa4
commit
e1c102ab03
2 files changed
+1
-14
No files matched your search
@@ -1,13 +0,0 @@
|
||||
# GATE 1 round 1 report
|
||||
|
||||
Reload canary script: /home/bchanot/.cache/claude-tmp/p5/t12/reload-one.sh (coproc, claude -p --model haiku --max-turns 1 --tools "" --input-format stream-json --output-format stream-json --verbose). Log: /home/bchanot/.cache/claude-tmp/p5/t12/reload-one-session.jsonl
|
||||
|
||||
Result lines (session_id, subtype, result):
|
||||
- ee65507c-a958-411b-9580-8aca8105e326 success `/seo` isn't available in this session, so it didn't run. Your message had no other text, so I don't know what you want checked. Tell me the goal and I'll do it with the tools I have. For example: -
|
||||
- ee65507c-a958-411b-9580-8aca8105e326 success Reloaded: 3 plugins · 2 skills · 32 agents · 7 hooks · 0 plugin LSP servers
|
||||
- ee65507c-a958-411b-9580-8aca8105e326 success # /seo — parallel SEO + GEO dispatcher
|
||||
- ee65507c-a958-411b-9580-8aca8105e326 success # /seo — parallel SEO + GEO dispatcher
|
||||
|
||||
Stderr: Warning: claude.ai MCP servers blocked by enterprise policy: claude.ai higgsfield, claude.ai Higgsfield AI (deny profile minimal).
|
||||
--tools "" did not break /reload-plugins. Turn 1 gave no literal 'Unknown skill': the model answered that /seo isn't available.
|
||||
Verify: doctrine-citers PASS=23, pack-census PASS=33, pack-census.sh PACKS OK 9 packs 55 skills.
|
||||
@@ -45,7 +45,7 @@ r2 2026-11-02 after three lenses (correctness FATAL(11), robustness FATAL(10), s
|
||||
- **Spike 1c, typed short slash.** Method: `-p '/<s>'` on a pack skill with canary reply. Result: the typed short `/<skill>` on a pack skill expands like `/<pack>:<skill>`.
|
||||
- **Spike 2, `/reload-plugins` mid-session.** Method: two-turn session (`--input-format stream-json`): turn 1 a canary prompt, a `.claude/settings.local.json` edit between turns, turn 2 `/reload-plugins` then the canary again. Result: `/reload-plugins` works in `-p` and reloads a pack's skills + hooks after a local edit; a pack `.mcp.json` server needs a new session.
|
||||
- **Spike 2b, reload under `--settings`.** Method: same two-turn session launched with `--settings <flag file>` setting a pack false, local edit true, `/reload-plugins`. Result: the flag value is read once; `/reload-plugins` re-reads neither the flag file nor a local true over a flag false, so a pack OFF by flag only comes back through a relaunch.
|
||||
- **Spike 3, claude.ai connectors per project (F0-F5).** Method: fresh `-p` sessions in the scratch project, init `mcp_servers[]` and `tools[]` counted per run. F0 baseline: connectors listed. F1 `.claude/settings.local.json` `deniedMcpServers:[{serverName:"claude.ai <Name>"}]` for Gmail + Higgsfield AI: 199 -> 192 tools. F2 `disableClaudeAiConnectors:true`: all nine dropped, 26 tools, model-router kept. F3 the same two keys via `--settings`: same result per conversation. F4 `deniedMcpServers` with a `serverUrl` wildcard: matches no connector (no URL exposed). F5 `--strict-mcp-config`: drops connectors and project `.mcp.json` servers per conversation. `disableClaudeAiConnectors` is any-source-true. The native `/mcp` toggle writes `disabledMcpServers` per project into `~/.claude.json` (untested, interactive only). Docs, not measured: `/reload-plugins` re-reads `enabledPlugins` from every settings file; `--plugin-dir` loads a pack as `<name>@inline` for one session; the skills-dir id is the manifest `name`.
|
||||
- **Spike 3, claude.ai connectors per project (F0-F5).** Method: fresh `-p` sessions in the scratch project, init `mcp_servers[]` and `tools[]` counted per run. F0 baseline: connectors listed. F1 `.claude/settings.local.json` `deniedMcpServers:[{serverName:"claude.ai <Name>"}]` for Gmail + Higgsfield AI: 199 -> 192 tools. F2 `disableClaudeAiConnectors:true`: all nine dropped, 26 tools, model-router kept. F3 `deniedMcpServers` via `--settings`: the same two connectors dropped per conversation. F4 `disableClaudeAiConnectors:true` via `--settings`: all nine dropped per conversation. F5 `deniedMcpServers` with a `serverUrl` wildcard: matches no connector (no URL exposed). Earlier runs E1-E5 (not levers): `disabledMcpjsonServers`, `enableAllProjectMcpServers:false` + `enabledMcpjsonServers:[]`, a synced plugin true leave the nine connectors; E4 `--strict-mcp-config` drops the connectors and the project `.mcp.json` servers per conversation (199 -> 26 tools). `disableClaudeAiConnectors` is any-source-true. The native `/mcp` toggle writes `disabledMcpServers` per project into `~/.claude.json` (untested, interactive only). Docs, not measured: `/reload-plugins` re-reads `enabledPlugins` from every settings file; `--plugin-dir` loads a pack as `<name>@inline` for one session; the skills-dir id is the manifest `name`.
|
||||
- **Spike 4, higgsfield MCP in a pack `.mcp.json`.** Method: a canary pack `spike-pack` whose `.mcp.json` is `{"mcpServers":{"spike-hf":{"type":"http","url":"https://mcp.higgsfield.ai/mcp"}}}`; fresh `-p` session reading init `mcp_servers[]`, and `claude mcp list` in the scratch project. Result: the server is registered as `plugin:<pack>:<server>` and `claude mcp list` shows "Needs authentication"; while unauthenticated it is absent from the init `mcp_servers`; the OAuth login is interactive (`/mcp`). Today higgsfield = two claude.ai connectors (`claude.ai higgsfield`, `claude.ai Higgsfield AI`), so a pack server would duplicate their tools.
|
||||
- Connectors, summary: `deniedMcpServers:[{serverName:"claude.ai <Name>"}]` drops the named ones (199 -> 192 tools); `disableClaudeAiConnectors:true` drops all nine (26 tools), any-source-true. Both work via `--settings`. `--strict-mcp-config` drops connectors and project `.mcp.json` servers.
|
||||
- model-router (`mods/model-router/hooks/register.ts`): `skillRow` matches `e.skill` by EXACT key in `routing.json` `skills`; `typedSkill` takes the text after `/` up to a space, so a typed `/pack-seo:seo` yields the key `pack-seo:seo` whatever the Skill tool reports. `/route show` prints `main: skill <phase> …`, never the skill name. Consequence: long-name rows are added unconditionally (Task 10); no router canary.
|
||||
|
||||
Reference in new issue
Block a user