PR Reply Assistant

Draft context-aware, well-reasoned replies to GitHub pull request comments directly in VS Code/Cursor using the native Language Model API (vscode.lm) and your existing GitHub Copilot access.

Features

1) PR Comment Button

• Adds a Draft PR Reply icon button ($(comment-discussion)) to every PR comment's title toolbar.
• Reads the vscode.Comment text plus the parent CommentThread file and line range.
• Optionally prompts for extra one-off guidance before drafting.
• Runs the full multi-agent quality pipeline for Auto strategy (see Pipeline).
• Copies the final draft with metadata (model, tone, strategy, confidence, token usage) to clipboard.

2) Chat Participant (@prreply)

• Registers a sticky Copilot Chat participant named @prreply.
• Uses the model you select in the Chat UI (request.model).
• Accepts # file/location references for stronger code grounding.
• Infers tone and strategy intent from your prompt text (e.g. "push back firmly").
• Streams a formatted draft with metadata back into the chat response.

3) Multi-agent quality pipeline

The Auto strategy runs a structured, evidence-based pipeline:

Safety gates run after the pipeline:

• Anchor gate — if the host did not supply an exact comment line range, falls back to a conservative best-effort reply.
• Safety gate — overrides a naive "agree" when imperative mutation evidence is present in symbols touched by the diff.

4) Deep context retrieval

The Planner has access to a suite of context tools it calls on demand:

• code_context_around_comment — code lines + diff hunks + related symbol snippets around the comment anchor
• read_file_range / read_full_file — arbitrary workspace file reads
• git_diff_file / git_diff_pr / git_log — file-level and PR-wide diffs and commit history
• symbol_evidence — structured write/read/mutation analysis for symbols in scope
• comprehensive_context — full-file content, page diff, reference-impact analysis, and PR change summary combined
• web_search — lightweight DuckDuckGo summary for external documentation when needed

Context depth setting controls pre-seeding:

• Standard — Planner gathers context on demand. Fast.
• Deep — PR-wide diagnostics and detailed diffs are pre-seeded into the evidence pack before the pipeline starts. Slower but broader.

5) Tone and strategy presets

Tone presets (set in Settings or onboarding):

Strategy presets:

6) Token usage reporting

• Prompt / completion / total token counts included in clipboard output metadata and the extension Output panel.

Requirements

• VS Code or Cursor with engines.vscode: ^1.102.0.
• GitHub Copilot with Chat access enabled.
• Signed in to GitHub/Copilot inside VS Code.

Onboarding

On first activation a 3-step setup wizard runs automatically:

1. Language — choose or type the language for draft replies (e.g. English, Türkçe, Español).
2. Model — pick the Copilot language model to use (persisted to Settings).
3. Context depth — Standard or Deep.

Re-run at any time with the command PR Reply Assistant: Run Setup.

Usage

Draft from a PR comment

1. Open a pull request in a VS Code UI that surfaces comment threads (e.g. the GitHub Pull Requests extension).
2. Click the Draft PR Reply icon in the comment header.
3. If askForExtraInstructions is enabled, add optional guidance (or leave blank).
4. The pipeline runs, then the draft is copied to clipboard with metadata.
5. Paste into the PR response box.

Use @prreply in Copilot Chat

Open Copilot Chat and start your message with @prreply:

@prreply Draft a reply to this comment:
"This function should be split — it does too many things at once."

Attach a file or range with # for stronger grounding:

@prreply #src/utils.ts Push back firmly — the complexity is necessary here.

Extension Settings

Architecture

src/
├── extension.ts          # activate / deactivate — wires commands and chat participant
├── agents.ts             # multi-agent quality draft pipeline (Planner, Decider, Critic, Arbiter, Writer)
├── constants.ts          # shared string/numeric constants
├── presets.ts            # tone, strategy, and context-mode preset data and helpers
├── settings.ts           # readUserSettings()
├── errors.ts             # user-facing error message formatting
├── llmClient.ts          # collectResponseWithUsage, token counting
├── modelResolver.ts      # model selection and fallback logic
├── onboarding.ts         # 3-step setup wizard
├── utils.ts              # escapeRegExp, truncateText, execFileAsync
├── context/
│   ├── anchor.ts         # buildAnchorEvidence — seed evidence from comment location
│   ├── chatRequest.ts    # resolveChatRequestContext, prompt inference helpers
│   ├── code.ts           # getCodeContext, buildCodeContextFromDocumentRange
│   ├── comment.ts        # extractCommentData, getThreadConversationContext
│   ├── comprehensive.ts  # getComprehensiveContext — full-page, ref-impact, PR changes
│   ├── evidence.ts       # getSymbolEvidenceContext — symbol write/read/mutation analysis
│   ├── git.ts            # git diff and deep context helpers
│   └── web.ts            # fetchDuckDuckGoSummary
└── pipeline/
    ├── draft.ts          # runAutoDraftPipeline, runForcedDraftPipeline, buildForcedStrategyPrompt
    ├── format.ts         # formatSingleDraftOutput, humanizeStrategy
    ├── gates.ts          # applyAnchorGate, applySafetyGate
    ├── progress.ts       # humanizeProgressMessage
    ├── tools.ts          # buildContextToolRegistry — tool specs for the Planner
    └── types.ts          # AutoDecisionResult, PipelineDetailSummary

Release Notes

0.0.1

• Initial extension scaffolding.

0.1.0

• Added PR comment action + @prreply chat participant.
• Added diff-aware and workspace-aware code context retrieval.
• Added tone and strategy presets with Auto decision mode.
• Added agentic decision flow (judge / critic / writer).
• Added optional web enrichment for uncertain decisions.
• Added token usage reporting in output channel and clipboard metadata.

0.2.0

• Replaced single-pass judge/critic/writer flow with a full multi-agent quality pipeline: Planner (iterative evidence gathering with tool calls), Decider, Critic, Arbiter, and Writer stages.
• Added anchor gate and safety gate post-processing.
• Added Deep context depth mode — pre-seeds PR-wide diagnostics and detailed diffs before the pipeline starts.
• Added 3-step onboarding wizard (language, model, context depth).
• Added comprehensive context tool: full-file content, page diff, reference impact, and PR change summary.
• Added structured symbol evidence: write/read/mutation-signal analysis for symbols in scope.
• Added askForExtraInstructions setting — optional per-draft guidance input.
• Added language setting — replies can be drafted in any language.
• Added model persistence and fallback by family/vendor.
• Refactored codebase into focused modules under src/context/ and src/pipeline/.