feat: add human-like smooth typing with optional typo injection

[ulziibay-kernel]

Summary

• Adds smooth and typo_chance fields to POST /computer/type for human-like typing via xdotool
• When smooth=true, text is typed in word-sized chunks with variable intra-word delays ([50, 120]ms) and inter-word pauses ([80, 200]ms, 1.5x at sentence boundaries)
• When typo_chance is set (0.0-0.10), realistic typos are injected using geometric gap sampling (O(typos) random calls, not O(chars)) and corrected with backspace after a "realization" pause

New package: server/lib/typinghumanizer

• SplitWordChunks -- word chunking with trailing delimiters attached to preceding word
• UniformJitter -- random duration in a range, clamped to minimum
• IsSentenceEnd -- sentence boundary detection for longer pauses
• AdjacentKey -- QWERTY neighbor lookup from static [26][]byte array, O(1)
• GenerateTypoPositions -- geometric gap sampling for typo placement

Typo types (weighted distribution)

Type	Weight	Mechanism
Adjacent key	60%	QWERTY neighbor substitution
Doubling	20%	Character typed twice
Transpose	15%	Swap with next character
Extra char	5%	Random adjacent key inserted

Related: #169 (plan document)

Demo

Test plan

• All 17 typinghumanizer tests pass (word chunking, adjacency, typo generation, distribution)
• Builds clean, no lint issues
• Manual test with smooth: true on a running instance
• Manual test with smooth: true, typo_chance: 0.03 to verify typo/correction behavior

Made with Cursor

---

Note

Medium Risk
Changes the POST /computer/type behavior by defaulting to a new smooth-typing path that introduces randomized delays and optional typo/backspace correction, which could affect automation timing and determinism. Also modifies the generated OpenAPI client/server glue via new post-generation patching, which can subtly impact request/response handling if the codegen output format changes.

Overview
Adds human-like typing to POST /computer/type via new request fields smooth (defaults to true) and typo_chance (0–0.10), enabling variable per-chunk delays, inter-word pauses, and optional typo injection with backspace correction.

Introduces server/lib/typinghumanizer (with tests) to chunk text, generate jittered timings, and compute realistic typo positions/adjacent-key substitutions, and wires this into the API implementation.

Updates OpenAPI (openapi.yaml + regenerated lib/oapi/oapi.go) and the server/Makefile to run an additional codegen patch step (patch_strict_optional_json) to tolerate empty JSON bodies for optional-body endpoints and restore omitempty tags; adds a small demo page + Python script to record side-by-side typing behavior.

^{Reviewed by Cursor Bugbot for commit d63ab1d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jarugupj]

Really cool stuff, the typinghumanizer package is super clean. The realization pause before backspace is a great detail, and the typo types cover the common cases really well -- adjacent key, doubling, transpose, extra char feels like it captures how people actually mess up.

One small thing I noticed -- in typeChunkWithTypo, the transpose branch and the else branch do the same thing:

if typo.Kind == typinghumanizer.TypoTranspose && typoLocalPos+1 < len(chunkRunes) {
    correctText = string(chunkRunes[typoLocalPos:])
} else {
    correctText = string(chunkRunes[typoLocalPos:])
}

Could probably just be correctText = string(chunkRunes[typoLocalPos:]) without the if/else?

[ulziibay-kernel]

@jarugupj Thanks! Good catch — both branches were identical, leftover from when I considered different correction behavior for transposition. Simplified to a single expression in 99e2e79.

[masnwilliams]

nit: "Requires smooth to be true" — this wording implies the server would return an error, but it actually silently ignores it. consider "Ignored when smooth is false" to match the actual behavior. no schema-level constraint enforces this either, so aligning the docs avoids confusion.

[masnwilliams]

nit: comment says "geometric gap sampling" but the implementation uses uniform gap sampling — halfGap + Intn(avgGap) produces a uniform distribution over [halfGap, halfGap+avgGap). geometric sampling would use exponentially distributed gaps (e.g. rng.ExpFloat64()). the behavior is fine, just the label is inaccurate.

[masnwilliams]

this only asserts TypoAdjacentKey (~60%). the other three kinds (TypoDoubling ~20%, TypoTranspose ~15%, TypoExtraChar ~5%) are counted but never checked — if two kinds got accidentally merged the test would still pass. consider asserting all four buckets.

[masnwilliams]

solid feature — clean separation between the humanizer library and the integration layer. main notes are nits on docs wording and test coverage. the identical if/else branches in typeChunkWithTypo were already flagged.

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit c32989e. Configure here.}

[masnwilliams]

looks good — all first-round comments addressed. one thing to follow up on: phantom corrections when a typo lands on a non-letter character (AdjacentKey returns the char unchanged, so the backspace-correct cycle fires with no visible typo). not blocking — typo_chance defaults to 0 so it only hits opt-in users. nice work on the multi-typo rewrite.