Repository navigation
fix(help): correct input and MCP descriptions, add a first-screen help gate - #3788
Conversation
launchConsole said 'Launch console mode.', but the value is a file path: runIosSimulatorConsoleLaunch writes the launch stdout/stderr to that path (simctl launch --console-pty), and open refuses it off an iOS simulator and for URL opens. jitterPx said 'Randomization radius in pixels.', but pressJitter cycles a fixed PRESS_JITTER table scaled by the amount, so the offsets are deterministic; the --jitter-px flag text already says so.
callHelpTool returns buildUsageText() when the topic is absent, and buildUsageText() renders renderCliHelpOverview(), the root decision card. The description said that it returns the full workflow card.
The workflow and manual-qa topics listed themselves in the shared Related footer. The self-link repeats the topic that the reader is on.
Agents often read only the first screen of a help text, and the help benchmark caps the root help at 30 lines (`--help:first30`). Declare, for the root help and for every topic id, the rules from the current rendering that must stay in lines 1-30: prerequisites, refusals, verification steps, and recovery. A topic id without an entry fails the test. The declared rules quote the main rendering; no help text changed.
gpt-5.4-mini fails with a ChatGPT login on this host. gpt-6-luna is the latest Codex model; a smoke run of one case completes the call and grades the plan without a runner error.
Size Report
Startup median (7 runs, lower is better):
|
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
There was a problem hiding this comment.
All reported issues were addressed across 7 files
You’re at about 97% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.
Reply with feedback, questions, or to request a fix.
View guided diff | Turn on auto-fix | Re-trigger cubic
pressJitter multiplies a pattern that includes diagonal offsets, so a press moves up to sqrt(2) x jitterPx. Align the input description, the response schema, and the flag usage description on the per-axis wording. #3788 (comment)
A substring match let a short generic rule like 'Commands:' pass from inside unrelated prose. A rule now counts only when the trimmed rendered line equals it or starts with it at a word boundary, and each pinned rule quotes its line from its start. Mutation-proven: a rule moved below line 30 and a deleted rule both still fail, and a short rule embedded in prose fails. #3788 (comment)
|
Both cubic P3 threads fixed and replied.
Validation: pnpm format, focused vitest (src/commands/schema, src/mcp, parser, option-field), check:mcp-metadata, then pnpm check:affected --run --head 5d267ec — all runnable checks passed (exactCommit=true). Pushed fast-forward on top of 138be32. No CI changes; no rebase needed. |
|
Reviewed 5d267ec. The code looks good. The corrected Not blocking: no test pins the Related-footer self-omission (https://github.com/callstack/agent-device/blob/5d267ec/src/commands/schema/cli-help.ts#L970), and the router test at https://github.com/callstack/agent-device/blob/5d267ec/src/mcp/__tests__/router.test.ts#L103 still says "workflow card". Could the Smoke Tests is still running, and every other check passes. The diff does not touch the smoke route, so a failure there would most likely be unrelated. There are no conflicts. |
thymikee
added
the
ready-for-human
|
…ntract The gate asserted the pre-#3761 help literal, so the merged pair (#3788 test, #3761 help owner) fails only in combination on main. Replace the stale rule with the current full contract line, including the capture-panel-unidentified refusal. The help text itself is unchanged; the assertion now matches the rendered first screen. CI evidence: Coverage job 117649571201 at 73d562b fails this assertion on unchanged-base integration; it is not caused by this PR's edits and not a flake.
1 participant
Summary
Salvages three non-STE parts of closed #3735. The ASD-STE100 help rewrite is not brought over; main's help wording is unchanged.
fix(help): correct two wrong input descriptions—launchConsoleis a file path for the initial iOS simulator launch stdout/stderr (direct app launches only, not URL opens), not a "launch console mode";jitterPxis deterministic (pressJittercycles a fixed table), not random. Matches the existing flag texts.fix(help): name the root decision card in the MCP help tool description— the no-topic call returnsbuildUsageText()(the root decision card), not the workflow card.fix(help): omit a topic's own entry from its Related footer—manual-qa/workflowno longer self-link.test(help): pin each topic's key rules to its first 30 lines— newcli-help-first-screen.test.ts. Declared rules quote main's exact phrases. The gate passes for the root help and all 18 topics with no rule excluded, so there are no first-screen findings to report.build(bench): default the help bench Codex runner to gpt-6-luna— gpt-5.4-mini fails with a ChatGPT login here.Touched files: 7. No scope expansion beyond the salvaged items.
Validation
138be32b01:pnpm check:affected --run --head 138be32b01— all runnable checks passed; receipt at.tmp/validation/run-X9KO20/receipt.json(exactCommit=true).pnpm vitest run src/commands/schema src/mcp— 294 passed.node scripts/help-conformance-bench.mjs --runner codex:gpt-6-luna --case focused-type-stops-at-successcompletes with no runner error.