70 to 100% fewer issues left.
With the LSP and its auto-fix hook, coding agents leave far fewer Robot Framework lint issues. On Haiku and Sonnet that costs about 10 percent more than no lint, and less than the robocop MCP on every model.
Measured 2026-11-14 on one task: write two Robot Framework suites with keywords, then run them. Every run in every arm ended with passing tests and a clean dry run. Issues left = robocop 9.4.0 default findings in the final files. Plugin version c84eb65.
The hook made the difference.
Haiku, 25 runs per arm, default effort. Issues left per run.
No LSP
18.5Robocop MCP
13.2LSP v2
13.0LSP v4
16.5LSP v5
16.5LSP v6
4.7| Version | What changed | Cost / run | Turns | Issues left |
|---|---|---|---|---|
| No LSP | Baseline | $0.028 | 16 | 18.5 |
| Robocop MCP | Lint as MCP tools | $0.036 | 27 | 13.2 |
| LSP v2 | Deprecations shown as warnings | $0.036 | 20.7 | 13.0 |
| LSP v4 | Grouped findings, fix commands | $0.034 | 21.4 | 16.5 |
| LSP v5 | Ungrouped, more fix commands | $0.033 | 19.8 | 16.5 |
| LSP v6 | Auto-fix hook after edits | $0.034 | 18.9 | 4.7 |
In v4 and v5 agents read the fix commands but never ran one, in 56 runs. The v6 hook applies the same fixes without spending a turn.
Cheaper than the MCP on every model.
3 runs per arm, effort medium. A small sample: read these as direction, not exact figures.
Haiku
Sonnet
Opus
| Model | Cost none | Cost MCP | Cost LSP | Turns none | Turns MCP | Turns LSP |
|---|---|---|---|---|---|---|
| Haiku | $0.030 | $0.040 | $0.033 | 19 | 26 | 21 |
| Sonnet | $0.171 | $0.319 | $0.186 | 8.7 | 27 | 11.7 |
| Opus | $0.251 | $0.443 | $0.395 | 11 | 18.3 | 17 |
Bars show issues left; the figure after the bar is issues left and cost per run. On Sonnet the LSP cost 42% less than the MCP. On Opus it cost 58% more than no lint.
Edit, fix, report.
Agent edits
The agent writes or edits a .robot or .resource file.
PostToolUseAuto-fix hook
Robocop's safe fixes run on the file. It prints nothing.
didChangeLSP lints
Robocop defaults plus the RFL rules, warnings and up, at most 20.
short linesDiagnostics pushed
One line per finding, 120 characters at most, with a fix hint.
next turnAgent fixes the rest
Only what robocop cannot fix reaches it, such as long keywords.
Before you trust it on your repo.
- Run the same task on your own suites. One task is one data point.
- Confirm the hook runs:
claude --debugshowsPostToolUsewithrobotframework-lint-autofixand status 0. - Count leftover issues with
robocop checkon the final files, not from the chat.