kapral18/fix-cursor-cli

Fix Cursor CLI

Diagnose first, then apply targeted repair.

Use when

• cursor-agent shows macOS policy/quarantine native-module load failures
• /model is empty or shows No matches while cursor-agent models still lists available models
• /model shows thinking/non-thinking variants as duplicate entries with no Thinking suffix to distinguish them
• the prompt footer briefly shows a thinking model and then drops the Thinking suffix while the underlying model slug is unchanged

Do not use

• generic cursor-agent failures without quarantine or picker-state evidence
• account/permission/model-routing issues not caused by local bundle state
• routine model selection with no runtime symptoms
• non-Cursor CLI issues unrelated to cursor / cursor-agent

Workflow

1. Reproduce and classify (do not fix yet):
   • command -v cursor cursor-agent
   • cursor-agent --version
   • cursor-agent models
   • if needed, reproduce interactive picker behavior (/model) in a fresh session
2. Read-only state check (preferred first step on any machine):
   • bash ~/.agents/skills/fix-cursor-cli/scripts/fix_cursor_cli.sh --reason check
   • Reports picker state per active bundle (v2-good / display-collapses-thinking / v1-collapses-variants / old-empty-picker / unknown-shape) and exits non-zero if any active bundle needs a fix.
3. Choose a targeted fix:
   • if evidence points to fresh-process startup failure with local bundle regression/quarantine signatures, run:
     • bash ~/.agents/skills/fix-cursor-cli/scripts/fix_cursor_cli.sh --reason startup-failure
   • if evidence is interactive-only (/model empty, duplicate-variant entries, or missing Thinking suffixes in /model/footer while cursor-agent models succeeds), run:
     • bash ~/.agents/skills/fix-cursor-cli/scripts/fix_cursor_cli.sh --reason empty-picker
   • --reason auto (default) performs diagnosis (including bundle-state classification) and no-ops on healthy machines; prefer this when symptom classification is incomplete.
   • if the script is missing, run chezmoi apply --no-tty and retry
   • only use manual bundle edits when the bundled script is unavailable or provably insufficient
4. Re-verify end-to-end:
   • cursor-agent models succeeds
   • --reason check reports Check OK
   • a fresh interactive session launched with a thinking model shows the Thinking suffix in the footer
   • /model shows entries and distinguishes thinking/non-thinking variants with a Thinking suffix
5. If the script reports unknown shape and dumps anchor context:
   • Cause: upstream cursor-cli bundle shape changed since this skill was last updated.
   • Recovery (do not retry the script blindly):
     1. Read the printed anchor-context block. It shows the current minified code around the picker call (useModelParameters:!0,doNotUseMarkdown:!0).
     2. Identify the as-shipped buggy form (an A.some((...parameterDefinitions...))?A:void 0 expression, where A is whichever single-letter var the minifier chose for the filtered-models list).
     3. For display-only regressions, inspect the buildParameterizedModelMap(...) method. The known bad form stores model metadata directly in parameterizedModelMap, leaving thinking slugs with duplicate clientDisplayName / inputboxShortModelName values.
     4. Update the regex patterns in scripts/picker_patch.py:
        • The existing PATTERN_OLD / PATTERN_V1 / PATTERN_V2 are already var-name-agnostic and use named groups (A/B/C/D/E/F). Only update them if the structural shape of the code (not the variable letters) has changed.
        • The display-name patch (V3) is var-name-agnostic: it adds Thinking suffixes and context-window tags (1M, 272k, etc.) from tooltip markdown. Update only when buildParameterizedModelMap changes shape.
        • Add a new pattern alongside the existing ones if upstream introduced a new arrangement.
        • The _build_v2() helper handles synthesizing the patched form from the captured group names; reuse it when adding new patterns.
     5. Re-run with --reason check first to confirm classification, then --reason empty-picker.
   • Both scripts/fix_cursor_cli.sh and scripts/picker_patch.py document the bug semantics in their headers; read them before editing.
6. Share learning:
   • state what symptom was detected, what fix was chosen, and what bundle state the script reported (v2-good / display-collapses-thinking / v1-collapses-variants / old-empty-picker / unknown-shape)

Assets

• scripts/fix_cursor_cli.sh: diagnose-first orchestrator. Probes cursor-agent, checks quarantine attrs, invokes picker_patch.py for picker-state classification and patching, and reports a clear per-bundle summary. Supports --reason check for a read-only audit that exits non-zero if anything needs fixing.
• scripts/picker_patch.py: var-name-agnostic regex classifier and migrator for the model-picker bundle. Handles the known states (old-empty-picker -> V2, v1-collapses-variants -> V2, display-collapses-thinking -> display-normalized V2, v2-good no-op) across all webpack chunks containing the picker/display code, and dumps anchor context for unknown shapes.

Output requirements

• Lead with diagnosis (including the picker bundle state per active file) and confidence level.
• Include evidence for the chosen path (commands/results, not guesses).
• State whether fix_cursor_cli.sh changed anything and what worst-state was observed before/after (v2-good / display-collapses-thinking / v1-collapses-variants / old-empty-picker / unknown-shape).
• If interactive verification cannot be automated, ask the user to confirm /model after restart.

Constraints

• Evidence-first: verify before and after each remediation step using --reason check.
• Do not run repair steps preemptively on healthy machines.
• Prefer the bundled skill assets (fix_cursor_cli.sh + picker_patch.py) over ad-hoc manual edits.
• Treat fixes as machine-local and temporary; re-evaluate on each host/update.