Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Roadmap

Development plan, version targets, and future enhancements.

Current Status

The project is in active development. Breaking changes are acceptable — the focus is on getting things right, not backward compatibility.

Completed Phases

Phase 1: Bug Fixes & Dead Code Removal ✅

Phase 2: Documentation Fixes ✅

Phase 3: Test Suite Improvements ✅

In Progress

Phase 4: Reliability Improvements

Focus: reduce LLM hallucinations, improve fix accuracy, move mechanical rules to deterministic checking.

ItemDescriptionStatus
4.1 Structural GuardrailsReject destructive fixes (heading/directive deletion)Planned
4.2 Line-Number AnchoringReplace free-text quoting with line-number targetingPlanned
4.3 Deterministic Checkers~13 mechanical rules via regex (zero hallucination risk)Planned
4.4 Rule ClarityImprove 12 rule descriptions to reduce misinterpretationPlanned
4.5 Scope ReductionReduce noise from overly subjective rulesPlanned
4.6 Prompt ConsolidationMerge 8 identical prompt files into single prompt.mdPlanned
4.7 Extended ThinkingClaude reasons internally → 0% false positivesDone (v0.7.0)

Phase 5: Style Suggestion UX

Focus: make it easy for authors to review and accept/reject suggestions.

ItemDescriptionStatus
5.1 Better FormattingLine numbers, side-by-side tables, cap displayed suggestionsPlanned
5.2 GitHub Suggestion BlocksOne-click “Commit suggestion” via PR review commentsPlanned
5.3 Token/Cost TrackingParse API usage, aggregate by rule/category, report in PRPlanned

Phase 6: New Capabilities

ItemDescriptionStatus
6.1 PR Review ModeAuto-trigger on PR, review only changed filesPlanned
6.2 Checkbox + /apply-styleAuthor selects suggestions, comments to applyPlanned
6.3 Batch ProcessingResume capability, progress reporting, partial failure handlingPlanned
6.4 Rule Confidence ScoringTrack suggestion acceptance rates to guide refinementPlanned
6.5 Interactive CLIqestyle --interactive — step through suggestions like git add -pPlanned

Version Plan

VersionScopeStatus
0.6.0Bug fixes, docs, rule clarity (Phases 1–2)Done
0.6.1Anti-false-positive prompt instructionDone
0.7.0Extended thinking, minimal prompt, qestyle CLI, testsDone
0.7.1Fix circular import (github.pyaction.py)Done
0.7.2Fix temperature default (must be 1 for extended thinking)Done
0.8.0Structural guardrails + rule clarity (4.1, 4.4)Not started
0.9.0Line numbers + deterministic checkers + prompt consolidationNot started
0.10.0Suggestion UX + suggestion blocks (5.1, 5.2)Not started
0.11.0Token tracking + PR mode (5.3, 6.1)Not started
1.0.0Stable release after production validationNot started

Future Enhancements

Large-Context Batch Evaluation

The current architecture makes 49 separate LLM calls per full review (one per rule). With next-generation models (1M+ context), batch evaluation may become reliable — potentially reducing 49 calls to 1–8.

Three progressive strategies to test:

  1. Category-level batching (8 calls) — all rules for one category in a single call

  2. Two-pass detect-then-fix (2+ calls) — detect all violations in one call, fix individually

  3. Full single-call batch (1 call) — maximum speed, maximum risk

Multi-Model Support

Different rules have different complexity requirements. Simple mechanical checks could use cheaper/faster models, while complex stylistic judgments benefit from more capable models.

Interactive CLI Mode

qestyle --interactive — guided editing experience where authors step through suggestions one at a time, similar to git add -p.

Phase Dependencies

Phase 1 (Bugs) → Phase 2 (Docs) → Phase 3 (Tests) → Phase 4 (Reliability) → Phase 5 (UX) → Phase 6 (Features)