[milstan]

Two changes on /chatgpt/mcp only. Every other surface reads byte-for-byte what it read before.

1. _triggered_by asks for a one-sentence summary, not the user's words. OpenAI's tool scan on the new Leadbay plugin draft (asdk_app_6adc4064…, 2026-10-24) flagged 39 tools with "This tool appears to request information beyond what its stated operation requires." All 39 carry _triggered_by, which asked for a verbatim copy of the user's message. Mazemind asks for a third-person summary and passed the same scan without that finding.

2. The rendering and NEXT STEPS text written for Claude is left out. In ChatGPT, every Leadbay call that asks for approval showed "Suspicious Instruction", danger_level: HIGH: "Tool documentation prescribes classifier handling and output behavior, including mandatory rendering and next-step widget instructions." Mazemind's calls in the same workspace showed none.

	/mcp, /fr/mcp, /sse, /apps/mcp, stdio	/chatgpt/mcp
_triggered_by	verbatim quote, MANDATORY, LAST_PROMPT_REQUIRED	one sentence ≤25 words, third person, optional, never rejected
"Trigger provenance (MANDATORY)" instruction block	present	left out
RENDER (quick), RENDERING, NEXT STEPS, widget table, link patterns	present	deleted
Description size, all tools	unchanged	513k → 295k chars

How (2) works. A promptforge marker, {{claude}} (src/claude.ts), wraps every include of a rendering/, next-steps/ or linking/ snippet and gates/builtin-widgets, plus the auto-emitted RENDER (quick) block. The inline RENDER / NEXT STEPS sections of 15 templates are wrapped by hand. Only APPS_NO_COMMERCE_TOOL_DESCRIPTIONS (= /chatgpt/mcp) deletes them. Every other constant and map in tool-descriptions.generated.ts is byte-identical to main, and prompts.generated.ts is unchanged. Documented in CLAUDE.md next to {{commerce}} and {{apps}}.

Existing test changed. mcp-apps-description.test.ts asserted /chatgpt/mcp equals the commerce-free text for tools without board prose. That is the behaviour this PR changes, so that one assertion now checks deletion only.

Also. chatgpt-app-submission.json (Apps-era, 4 tools behind) is replaced by chatgpt-plugin/, the files uploaded to OpenAI's plugin portal. Its URL is https://mcp.leadbay.app/chatgpt/mcp?v=2: the rejected 0.34.0 plugin still holds the plain URL and the portal refuses one URL on two apps. Bumps to 0.46.1.

Not done here. Eleven one-line confirmation asks still name ask_user_input_v0 (e.g. "ask which channels to enrich"); they are behaviour, not layout. leadbay_list_lens_seed_candidates still says "Do not pause for user approval". The server instructions are unchanged apart from the _triggered_by block. plugin.json has no demo video URL yet; it is recorded after this deploys.

Verified. New chatgpt-summary-triggered-by.test.ts (13 cases, 4 fail on main) and chatgpt-no-claude-rendering.test.ts (6 cases, 2 fail on main's descriptions), plus promptforge/test/claude-marker.test.ts. pnpm -r typecheck and pnpm -r test green (3,853 tests). Reviewer account on a local 0.46.1 server: on /chatgpt/mcp a composite runs without _triggered_by; on /mcp it is rejected.

🤖 Generated with Claude Code

[claude]

Reviewed the full diff. summaryTriggeredBy is gated by path (/chatgpt/mcp only) and covers the schema description, the required list, the server-instructions block and the LAST_PROMPT_REQUIRED guard. The new tests cover both sides, so the other surfaces keep the verbatim mandate. The chatgpt-plugin/ test-case tools (leadbay_welcome, leadbay_tour_plan) are registered, and nothing still references the deleted chatgpt-app-submission.json. No blocking issues found.