[skulitom]

Report

ANODE-BUGS.md B-041 (filed 2026-10-15 by compass-anode-fixer from a read-only code hunt in its nightly run): a rejected gamepad_set leaves part of its change pending, so the next set or tap sends it.

Root cause

GamepadManager.Apply (src/Anode/Core/Gamepad/GamepadManager.cs) resolved and set each button, axis and trigger as it read them, then called SubmitReport once at the end. Every pad runs with AutoSubmitReport = false (set in Attach), so a set only changes the pad's pending report. When a later name was unknown, ResolveButton, ResolveAxis or ResolveTrigger threw before SubmitReport, the agent got an error, and the changes made before it stayed pending. The agent's next gamepad_set or gamepad_tap sent them (a gamepad_reset or a detach clears the report first, so those threw them away). Example: gamepad_set {"buttons": {"a": true, "dpad_up": true}} fails on dpad_up, and A goes down with the agent's next call, so the game in the seat sees A held that the agent believes it never pressed. The MCP schema takes any button name, so a guessed name reaches this path.

Two smaller faults in the same names:

• The unknown-button error listed "lt or rt" among the buttons, so gamepad_set buttons {"lt": true} and anode gamepad hold lt were told "Unknown gamepad button 'lt'. Use …, lt or rt."
• gamepad_tap recognised lt, rt, lefttrigger and righttrigger as triggers, but not l2 and r2, which ResolveTrigger accepts; gamepad_tap l2 failed with that same error.

Fix

• Apply resolves every name and value first into a list of changes, then sets them and submits one report under the lock. A refused call changes nothing.
• One TriggerNamed helper now names the triggers for ResolveTrigger, Tap and the button error. A trigger passed as a button hears "'lt' is a trigger, not a button: tap it, or, over MCP, set it from 0 to 1 under gamepad_set's triggers." (the CLI can tap a trigger but not hold one) The unknown-button list no longer offers lt or rt.
• No change to the MCP schema, the CLI or the docs: they already list the buttons without the triggers, and gamepad_tap's "lt or rt" stays true.

Tests

• New quick check gamepad set all or nothing (Cli/GamepadChecks.cs, SetAllOrNothing). The stand-in FakePad now keeps what was set until a report sends it, like ViGEm's pad with AutoSubmitReport off, and keeps each report it sent. Refused sets for an unknown button, an unknown axis after a good button and axis, an unknown trigger after a good axis and trigger, and a trigger named as a button leave nothing for the next report; the unknown-button error no longer offers lt or rt; the next valid set sends only its own change, in one report; a full set sends buttons, axes and triggers together; gamepad_tap l2 pulls the left trigger to 255 and releases it. No controller is plugged in.
• Mutations (each built in Debug and run through selftest --quick): 7 of 7 caught (buttons, axes or triggers set as they are read, which is the old code for each; the trigger message reverted; lt or rt back in the button list; Tap's old trigger test; a valid set never submitted). Every restore byte-identical.
• Second review (a fresh Claude subagent, read-only): no P1. Its P2: the 0.11.3 release notes and validation record ([compass] Anode 0.11.3 candidate: ship the merged fixes, and move the registry off 0.11.0 #47) didn't list this fix; f2ef33f adds it there (and [compass] Fix: a refused Reconnect sets a running seat to error, so display reconnects answer "stopped" and the next disconnect drops the seat #48 did the same for itself), with the quick-check count now 110. Its P3s applied in 714100f (the stand-in, the tap check, a vacuous assertion, the CLI-true trigger message, this description's reset/detach sentence); not taken: Tap resolving the trigger once (kept the diff small).
• Codex GitHub connector: 👍, no findings on 6e0e297.
• Merged with main after [compass] Fix: a refused Reconnect sets a running seat to error, so display reconnects answer "stopped" and the next disconnect drops the seat #48 (f6e4c8c); the mutations were re-run on the final code: 7 of 7 caught.
• FOCUS test command (scripts\build.ps1 -QuickTest -OutputDirectory artifacts\pkg-build): passed on the final head with main merged in: Release publish, 110/110 quick checks, docs consistent.
• Debug selftest --quick: 109/109 before main was merged in.

Manual check in a real seat

None needed for the fix itself: the stand-in models the one ViGEm behaviour it relies on (sets wait for SubmitReport when AutoSubmitReport is off). The full self-test's "virtual gamepad" check (a real ViGEm pad, machine-wide) still runs the same Apply call when someone runs it with a gamepad; agents don't.

CHANGELOG, release notes and the 0.11.3 validation record: one entry each, since 0.11.3 isn't released yet.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-15T22:45:44.940495Z	6e0e297	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.