A talk build is reachable through MCP: `say` starts it, reads its versions and diffs, restores and publishes #11079

say now drives a talk build through /mcp, and the work is pushed. test/mcp-size.test.ts passes with the ceilings unchanged, as do the new test, the other say/talk and MCP-doc tests, bun scripts/gate.ts and tsc --noEmit. I did not use the keep_plan tool, so no plan is posted on the issue.

What changed

• New actions: say takes action: build | build_state | build_diff | build_restore | build_publish, plus two optional arguments, template and version. It is still one tool with an action enum; no tool was added.
• Same functions as the API: each action calls the same function in src/actions/build.ts that /api/v1 calls. Errors come back as the BUILD_WORDS sentence under the same code, so the per-hour limit and the unpublished-builds cap are the API's too.
• Input checks: the checks on prompt, template and version are copied from conversation_build.ts, so the MCP and API reject the same input.
• Access: build_state and build_diff are reads, so an ro token may call them. build, build_restore and build_publish need rw. Another person's conversation answers not_found.
• The forge's own agent: it is refused these actions because they sit in the existing branch of say that already refuses its via token.
• Docs and tests: docs/MCP.md is re-rendered with bun scripts/mcp-table.ts. test/mcp-say-build.test.ts drives every action through /mcp on a seeded forge and covers ro refusals, a stranger's not_found, bad input, no-such-version and the hourly limit.

What I left out

• conversation is required for every action except build. It does not default to your latest conversation, because the API takes an id too.

To try it

• Call say with {action:"build", text:"a tip calculator", template:"owner/name"}.
• Then call say with {action:"build_state", conversation:<id>}.

The first push was refused by the pre-push hook, which wants a "Try:" commit. I added an empty commit with that line.

🤖 Generated with Claude Code

Closes #11071

say now drives a talk build through /mcp, and the work is pushed. test/mcp-size.test.ts passes with the ceilings unchanged, as do the new test, the other say/talk and MCP-doc tests, bun scripts/gate.ts and tsc --noEmit. I did not use the keep_plan tool, so no plan is posted on the issue.

What changed

• New actions: say takes action: build | build_state | build_diff | build_restore | build_publish, plus two optional arguments, template and version. It is still one tool with an action enum; no tool was added.
• Same functions as the API: each action calls the same function in src/actions/build.ts that /api/v1 calls. Errors come back as the BUILD_WORDS sentence under the same code, so the per-hour limit and the unpublished-builds cap are the API's too.
• Input checks: the checks on prompt, template and version are copied from conversation_build.ts, so the MCP and API reject the same input.
• Access: build_state and build_diff are reads, so an ro token may call them. build, build_restore and build_publish need rw. Another person's conversation answers not_found.
• The forge's own agent: it is refused these actions because they sit in the existing branch of say that already refuses its via token.
• Docs and tests: docs/MCP.md is re-rendered with bun scripts/mcp-table.ts. test/mcp-say-build.test.ts drives every action through /mcp on a seeded forge and covers ro refusals, a stranger's not_found, bad input, no-such-version and the hourly limit.

What I left out

• conversation is required for every action except build. It does not default to your latest conversation, because the API takes an id too.

To try it

• Call say with {action:"build", text:"a tip calculator", template:"owner/name"}.
• Then call say with {action:"build_state", conversation:<id>}.

The first push was refused by the pre-push hook, which wants a "Try:" commit. I added an empty commit with that line.

🤖 Generated with Claude Code

Both doors now take their build input checks from one new module, src/actions/build_input.ts, and the change is pushed. tsc, scripts/gate.ts, the new say build test, talk-build, mcp-size and talk-conversations all pass.

• Shared module: it exports BUILD_ID, BUILD_TEMPLATE and the prompt, template and version checks. Each check returns the refusal sentence, or nothing when the input is fine.
• Callers: src/http/api/conversation_build.ts and src/http/mcp/tools/say.ts both import it, and the local copies of those constants and checks are gone.
• Wording: the prompt sentence takes a field name, so the API still says prompt and MCP says text. The API's error codes and sentences are otherwise unchanged.
• Test: the ro refusals in test/mcp-say-build.test.ts now assert forbidden instead of toBeDefined.

I did not re-render docs/MCP.md, since the tool's arguments and descriptions did not change in this commit.

🤖 Generated with Claude Code