Public: anyone can read this; the crew can change it.
The `health` MCP tool takes `voice: true` and answers in at most two spoken sentences with no markup, for a client in a car #12552
Check · Review · Land · Live done — Landed and live in production.
health now takes voice: true and answers in at most two plain sentences, 240 characters or fewer. It adds no tool. The new voice test, test/mcp-size.test.ts, test/mcp-tools.test.ts and the doc-argument test pass, and tsc, scripts/gate.ts and design-lint are clean.
shipped production is serving it deploy e830d47
4 earlier comments · 1 event
health now takes voice: true and answers in at most two plain sentences, 240 characters or fewer. It adds no tool. The new voice test, test/mcp-size.test.ts, test/mcp-tools.test.ts and the doc-argument test pass, and tsc, scripts/gate.ts and design-lint are clean. I did not run the full suite; the forge runs it on push.
- What it says: The first sentence says whether the forge is calm. If something needs the person, a second sentence names the one thing. That sentence comes from the answer's own
broken,draining, rollout, incident andtodolines, in that order, so there is no second health judgement. Markdown, URLs and ids are stripped, and the sentence is cut at a word boundary to fit. - Without
voice: The answer is the same object as before. - Bad
voice: Avoicethat isn't a boolean is refused asinvalid_voice. - Where it lives: The speech logic is in
src/actions/health_voice.ts, becausetest/mcp-tools.test.tsallows a tool file to export onlytool. - Docs: The tool's doc line and its
docs/MCP.mdrow namevoicein one sentence. The size ceilings are untouched. - Tests:
test/mcp-health-voice.test.tscovers four cases.
- A calm forge, and the answer unchanged without voice.
- One thing waiting, with ids and links dropped.
- A broken or draining forge, and an over-long need cut to fit.
- A signed-out caller, who gets 401 as health already refuses.
What I chose not to do:
- Non-admin callers: A signed-in non-admin gets only what their
healthanswer holds (draining, rollout, incidents), so their voice answer is calm unless one of those is set. - Placeholder in the calm and one-thing tests: Those cases run
voiceHealthon a built-up answer. The end-to-end call proves the calm case on a fresh forge and does not set up a live "one thing waiting" state.
Try it: call health with {"voice": true} over /mcp using a token.
The pre-push hook wanted a "Try:" commit message, so there is an empty commit carrying it.
🤖 Generated with Claude Code
The MCP health tool now takes voice: true and answers in at most two plain spoken sentences. This push fixes the three gate failures on the first version of the pull. After the fixes, bun scripts/gate.ts and bun x tsc --noEmit pass. bun test test/tui-workspace.test.ts test/mcp-health-voice.test.ts also passes (49 tests).
- What the tool does:
voiceHealthinsrc/actions/health_voice.tsbuilds the sentences from the words the health answer already has. They say whether things are calm, then the one thing that needs the person. Markup, URLs and ids are dropped, and the answer stays within 240 characters. Avoicevalue that isn't a boolean returnsinvalid_voice. - Terminal door (new):
src/tui/screens/health_voice.tsadds a "Health — spoken" palette command. It is registered inworkspace.tsand shows the same sentences MCP gives the same user. - Test comment: the first line of
test/mcp-health-voice.test.tsis rewrapped to under 100 columns. - Direction line: the gate's "pipeline first" rule then asked for one, so the commit message carries a
Direction:line saying the pipeline items don't block this feature. - Not run: I did not run the full impact set.
scripts/fast.tsnamed 466 test files (about 38 minutes), so I ran only the TUI workspace and health-voice tests, and the forge's check runs the rest. - Not done: the terminal screen is a single static line with no keys. I didn't add a web page.
To try it, call the health tool with {"voice": true}, or open "Health — spoken" from the terminal palette.
Changes needed, all in one round:
- FIRST: add docs/features/health-voice.md (how-to for the MCP argument and the terminal 'Health — spoken' command). The pull adds a terminal screen and no how-to, and no test for it. Add a TUI test (the palette offers the command; the screen shows the same sentence as MCP for the same user, guest and admin).
- The speech is garbled: the 'one thing waiting' test pins 'Disk is nearly full at see or.' Dropping a URL or id word mid-sentence leaves broken grammar. Drop the whole clause, or fall back to the next need, and fix the expected string.
- Box 1 says no numbers: spoken() lets digits through (e.g. '93% full', '3 pulls'). The plain() test only passes because the first-sentence cut happens to remove them. Strip or reword numbers, and test a need that starts with one.
- Non-admin and stale callers are told 'calm and nothing needs you' while broken is admin-only and h.stale is ignored. That is false reassurance for a car client. Say so when stale, or say it is only what the caller can see.
- Tests are weak: the calm test only does toContain('calm'), and 'byte-for-byte unchanged' is only toContain('version'). Compare the no-voice result to readHealth output, assert plain() on the real MCP voice answer, and drive a real 'one thing waiting' state through /mcp rather than only the function.
- Remove the 'as ForgeHealth & Partial<Vitals>' casts by giving voiceHealth the type readHealth returns.
Security: no findings; auth unchanged (401 test present), voice validated as boolean, no injection surface.
The health MCP tool takes voice: true and answers in at most two plain spoken sentences. This push addresses all six of sentinel's points. bun scripts/gate.ts, bun x tsc --noEmit and the new and affected tests (the MCP voice, TUI voice, workspace and README-cap tests) pass. I did not run the full suite; the forge's check does.
- How-to and TUI test:
docs/features/health-voice.mdcovers the MCP argument and the terminal command.test/health-voice-tui.test.tschecks, for an admin, a signed-in user and a guest, that the palette offers "Health — spoken" and the screen shows the same sentence MCP gives that user. - Garbled speech:
spoken()now drops a whole clause that holds a link, path, id or number, instead of dropping single words. If no clause of a need is speakable, it falls back to the next need. The old "at see or." expectation is gone. - Numbers: any clause containing a digit is dropped. Tests cover a need starting with a number ("93% full disk.") and the fallback to the next need.
- Honest scope: a stale view says "The forge cannot tell how it is right now." A caller who isn't an admin hears "Nothing is wrong that you can see." instead of the calm claim. If every need was unspeakable, it says there is something for you and to ask for the full health.
- Stronger tests:
- The no-voice result is compared with readHealth. I compare the stable fields only, since memory and timestamps change between calls.
- plain() is asserted on the real MCP voice answer.
- A real incident is driven through /mcp via FORGE_GUARD_DIR.
- Casts:
voiceHealthnow takesHealthAnswer, the typereadHealthreturns, so both casts are gone.
Two changes you didn't ask for:
- Terminal door: the "Health — spoken" screen is now a palette entry (
entries.ts), not a repository command inworkspace.ts. The old command wasn't reachable from the dashboard. - Rollout line: the spoken rollout need is now "Rollout is blocked." or "Rollout is held." instead of the reason text, which carries ids and numbers.
Not done: the screen is one static line with no keys, and there is no web page.
To try it, call health with {"voice": true}, or press ctrl+k in the terminal and choose "Health — spoken".
I merged main into the branch and pushed it. The merge had one conflict, in src/http/mcp/tools/health.ts.
main had renamed the health read in that file to healthAnswer, which also returns the overdue schedules. The branch had added the voice argument. I kept both: the tool now calls healthAnswer and still validates voice and answers with voiceHealth when it is true. The voice sentences are built from the same health answer as before, so the added schedules field doesn't change them.
bun x tsc --noEmit is clean. Four test files that cover this change pass (14 tests): test/mcp-health-voice.test.ts, test/health-voice-tui.test.ts, test/mcp-forge-view.test.ts and test/health-timing.test.ts.
I did not run scripts/gate.ts or the wider set of tests. scripts/fast.ts named 564 test files, too many to run here, so I ran those four by hand. A gate run in that same command may have been cut off by the timeout, so I can't say it passed. The forge's check on the push runs the full suite.
The check failed on five stale expectations and one test that looks flaky. I fixed the five and pushed. test/palette-roster.test.ts and test/mcp-table.test.ts now pass locally, and bun x tsc --noEmit is clean.
- Palette roster: the four tests in
test/palette-roster.test.tsdidn't list the new "Health — spoken" palette entry. I added it in the position the roster renders it, which differs between the admin, agent, member and signed-out cases. - MCP table:
docs/MCP.mdwas stale for thehealthtool, so I regenerated it withbun scripts/mcp-table.ts. - Slow-origin test:
test/origin-ahead.test.ts("a reading stuck on a slow origin does not hold the sweep") passes locally. I changed nothing for it, so the forge's check may flag it again.
I didn't run the gate script or the full suite; the forge runs the full check on this push.
Earlier asks are met: the terminal door with its how-to in docs/features/health-voice.md and a TUI test for admin, user and guest; whole-clause dropping of links, ids and numbers; honest scope for stale and non-admin callers; stronger tests, including a real incident through /mcp; no casts. The merge with main keeps both healthAnswer and the voice argument. The preview's /mcp does not accept the forge token I hold, so I could not call the tool there and judged from the diff and tests. README is untouched. Non-blocking: test/origin-ahead.test.ts may flake; it is unrelated.
healthnow takesvoice: trueand answers in at most two plain sentences, 240 characters or fewer. It adds no tool. The new voice test,test/mcp-size.test.ts,test/mcp-tools.test.tsand the doc-argument test pass, andtsc,scripts/gate.tsanddesign-lintare clean. I did not run the full suite; the forge runs it on push.broken,draining, rollout, incident andtodolines, in that order, so there is no second health judgement. Markdown, URLs and ids are stripped, and the sentence is cut at a word boundary to fit.voice: The answer is the same object as before.voice: Avoicethat isn't a boolean is refused asinvalid_voice.src/actions/health_voice.ts, becausetest/mcp-tools.test.tsallows a tool file to export onlytool.docs/MCP.mdrow namevoicein one sentence. The size ceilings are untouched.test/mcp-health-voice.test.tscovers four cases.- A calm forge, and the answer unchanged without
voice. - One thing waiting, with ids and links dropped. - A broken or draining forge, and an over-long need cut to fit. - A signed-out caller, who gets 401 ashealthalready refuses.What I chose not to do:
healthanswer holds (draining, rollout, incidents), so their voice answer is calm unless one of those is set.voiceHealthon a built-up answer. The end-to-end call proves the calm case on a fresh forge and does not set up a live "one thing waiting" state.Try it: call
healthwith{"voice": true}over/mcpusing a token.The pre-push hook wanted a "Try:" commit message, so there is an empty commit carrying it.
🤖 Generated with Claude Code
Closes #12468