…tocol#4 write-up

Adds Scenario modelcontextprotocol#4 — `transformers-js-demo` — testing skill loading on
a code-output skill. Agent reads transformers-js SKILL.md and writes
a self-contained browser HTML file; no tool execution, the artifact
is the agent's response. Three graded prescriptions from SKILL.md:
@huggingface/transformers package name, pipeline() API, dispose()
memory-management rule.

Plan evaluator
- DISCOVERY_TOOLS gains agent-internal planning calls (goose's
  todo__todo_write, gemini-cli's update_topic) so the
  skill-read-before-plan fallback gate doesn't fire on housekeeping
  before skill activation
- One new test for the discovery/neutral-tools fallback path

Findings (docs/experimental-findings.md)
- Scenario modelcontextprotocol#4 — only codex passes overall. Two distinct failure modes:
  gemini-cli skipped catalog activation entirely; claude variants
  (haiku via fast-agent, sonnet via goose) read the skill but missed
  the dispose() prescription despite the ⚠️ warning. Suggests
  in-body emphasis isn't sufficient on its own — open question on
  what makes SKILL.md prescriptions stick

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

…rotocol#3/modelcontextprotocol#4 as N=1 samples

Adds a one-line note under each scenario header so readers don't
mistake single-trial outcomes for stable behavioral claims. Re-running
flipped haiku from FAIL to PASS on the dispose-rule criterion in
Scenario modelcontextprotocol#4, which is exactly the noise the caveat warns about.

Pass-rate replication (N=3+ per cell) is the proper fix and pending
in a follow-up — N-trial wrapper in each client harness, then re-run
the matrix and rewrite the findings to report rates rather than
single-sample binaries.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

Drop the "Scenarios modelcontextprotocol#2/modelcontextprotocol#3/modelcontextprotocol#4 are N=1, replication pending" snapshot —
that fact rots as findings get re-run. Keep just the principle:
treat individual outcomes as samples, not stable behavior.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

…extprotocol#3/modelcontextprotocol#4

Three new YAMLs that exercise the plan-kind evaluator path:

- hf-jobs-plan.yaml (modelcontextprotocol#2): probes structured plan output (PEP 723 UV
  script + HF_TOKEN secret forwarding + Trackio instrumentation) on
  hf-mcp-server with HF_JOBS_DRY_RUN=true, so submissions are
  intercepted server-side and graded from the call args
- hf-train-with-monitoring.yaml (modelcontextprotocol#3): cross-skill composition --
  agent must independently activate both huggingface-llm-trainer
  and huggingface-trackio from the catalog, with required phrases
  drawn from both skills' references
- transformers-js-demo.yaml (modelcontextprotocol#4): code-output domain (browser HTML/JS
  artifact). No tool execution; phrase-grep checks @huggingface/transformers
  package + pipeline API + dispose() rule

All three use gate_tools to anchor skill-read-before-plan against
meaningful actions instead of housekeeping calls (todo writes,
catalog enumeration), so agentic planners aren't penalized.