pull-requests-as-training - SKILL.md Agent Skill

name: "pull-requests-as-training" description: "Apply the Clean-PR agentless repo-level code editing protocol: decompose issues into file localization, fine-grained navigation via AST, and minimal Search/Replace patch generation. Triggers: 'fix this issue across the repo', 'edit multiple files for this bug', 'agentless code edit', 'search replace patch', 'repo-level fix', 'locate and patch this bug'"

Pull-Request-Driven Repo-Level Code Editing

This skill implements the agentless three-stage protocol from the Clean-PR paper for repository-level code editing. Instead of relying on iterative agent loops with tool calls, it decomposes an issue into three deterministic steps: (1) file localization from the repo tree, (2) fine-grained navigation to the exact function/class using AST-level reasoning, and (3) generation of minimal Search/Replace edit blocks that use unique context anchors rather than fragile line numbers. This approach produces precise multi-file patches with fewer hallucinations and without complex scaffolding.

When to Use

When the user describes a bug or feature request and expects edits across multiple files in a large repository
When asked to "fix this issue" or "implement this change" and the relevant files are not yet identified
When generating patches that must apply cleanly — the Search/Replace format guarantees deterministic application
When working on SWE-bench-style tasks: issue description in, code patch out
When the user wants a systematic approach to repo-wide changes rather than ad-hoc file-by-file edits
When editing a codebase with thousands of files and only 1-3 need modification (needle-in-a-haystack localization)

Key Technique

The Clean-PR method treats real-world GitHub pull requests as structured training signals. Raw PR diffs are noisy — 38% contain non-core changes (docs, configs), 25% are bot-generated, and 24% were never merged. The pipeline filters these, then converts surviving diffs into minimal Search/Replace edit blocks: the shortest contiguous span of original code needed to uniquely anchor an edit location within a file, paired with the replacement code. This avoids line-number fragility and ensures patches apply deterministically. A round-trip verification step discards any example where re-applying the generated blocks doesn't bit-wise match the ground-truth post-PR state.

At inference time, the protocol is agentless — no tool-use loops, no retry scaffolding. It decomposes the task into three sequential stages: (1) File Localization maps the issue description + repo tree to a ranked list of file paths, (2) Fine-Grained Navigation uses AST parsing to map within each file to the enclosing function or class definition that needs editing, and (3) Patch Generation produces Search/Replace blocks scoped to the localized context. Error-driven data augmentation improves robustness by training on hard negatives: semantically similar but incorrect files (Stage 2) and distractor code regions near the bug site that require no editing (Stage 3).

Step-by-Step Workflow

Gather the issue description. Extract the bug report, feature request, or change specification. If linked issues exist, concatenate their titles and descriptions for fuller context.
Build the repo tree. Generate a flat listing of all source files in the repository (exclude non-code artifacts like .md, .txt, .json, images, and config files unless they are explicitly relevant). This is the haystack for file localization.
File Localization. Given the issue text and the repo tree, identify the 1-5 source files most likely to require modification. Rank them by relevance. Filter aggressively — average real-world issues touch only 1-2 files even in repos with 3,000+ files.
Read candidate files. Retrieve the full content of each localized file. If a file exceeds practical context limits, apply context windowing centered on the most relevant sections (imports, class definitions matching the issue keywords).
Fine-Grained Navigation via AST reasoning. For each file, identify the specific function(s) or class definition(s) that enclose the code needing modification. Use structural reasoning (indentation, class/function boundaries, decorator patterns) to narrow to the exact scope. Include hard-negative awareness: note nearby code regions that look related but should NOT be edited.
Generate Search/Replace edit blocks. For each edit site, produce a block in this format:
- Search: The minimal unique contiguous span of existing code that unambiguously identifies the edit location within the file. Must match exactly one location.
- Replace: The corrected or new code that replaces the search span. Keep search spans as short as possible while remaining unique. Never use line numbers as anchors.
Round-trip verify. Mentally (or actually) apply each Search/Replace block to the original file content. Confirm the search string matches exactly one location. Confirm the result matches intent. If ambiguous, expand the search context until unique.
Assemble the patch. Combine all Search/Replace blocks into a single structured patch, organized by file path. Present in a clear, deterministic format.
Validate the patch against the issue. Re-read the original issue and confirm every requirement is addressed. Check that no unrelated code was modified.
Handle multi-language repos. The technique works across Python, Java, TypeScript, Go, C/C++, Ruby, Rust, PHP, Scala, Swift, and Kotlin. Use language-appropriate AST reasoning for navigation.

Concrete Examples

Example 1: Bug fix in a Django web application

User: "The /api/users/ endpoint returns a 500 error when the email field is null. Fix it."

Approach:

Build repo tree — identify views.py, serializers.py, models.py, tests/ under the users/ app
File Localization — rank users/serializers.py and users/views.py as top candidates (serializer validation is the likely null-handling site)
Read users/serializers.py — find UserSerializer class with email = serializers.EmailField()
Fine-Grained Navigation — the email field declaration and any validate_email method are the edit scope
Generate Search/Replace block:

File: users/serializers.py

<<<< SEARCH
email = serializers.EmailField()
==== REPLACE
email = serializers.EmailField(required=False, allow_null=True, allow_blank=True)
>>>>

Verify — the search string email = serializers.EmailField() is unique in the file. The replacement adds null tolerance. Round-trip confirmed.

Example 2: Multi-file refactor — rename a utility function

User: "Rename compute_hash to calculate_content_hash everywhere in the repo."

Approach:

Grep for compute_hash across the repo tree to find all usage sites
File Localization — identify utils/hashing.py (definition), services/upload.py, services/cache.py, tests/test_hashing.py (usages)
For each file, navigate to the exact function call or definition site
Generate Search/Replace blocks for each file:

File: utils/hashing.py

<<<< SEARCH
def compute_hash(content: bytes) -> str:
==== REPLACE
def calculate_content_hash(content: bytes) -> str:
>>>>

File: services/upload.py

<<<< SEARCH
from utils.hashing import compute_hash
==== REPLACE
from utils.hashing import calculate_content_hash
>>>>

File: services/upload.py

<<<< SEARCH
    file_hash = compute_hash(file_content)
==== REPLACE
    file_hash = calculate_content_hash(file_content)
>>>>

File: services/cache.py

<<<< SEARCH
from utils.hashing import compute_hash
==== REPLACE
from utils.hashing import calculate_content_hash
>>>>

File: services/cache.py

<<<< SEARCH
    key = compute_hash(payload.encode())
==== REPLACE
    key = calculate_content_hash(payload.encode())
>>>>

Verify — each search string is unique within its respective file. All import and call sites are covered.

Example 3: Feature addition across backend and frontend

User: "Add a last_login timestamp field to the User model and display it on the profile page."

Approach:

File Localization — models/user.py (model definition), api/serializers.py (API serialization), templates/profile.html or frontend/src/components/Profile.tsx (display)
Navigate to User class in model, serializer fields list, and profile template/component
Generate Search/Replace blocks:

File: models/user.py

<<<< SEARCH
    created_at = models.DateTimeField(auto_now_add=True)
==== REPLACE
    created_at = models.DateTimeField(auto_now_add=True)
    last_login = models.DateTimeField(null=True, blank=True)
>>>>

File: api/serializers.py

<<<< SEARCH
    fields = ["id", "username", "email", "created_at"]
==== REPLACE
    fields = ["id", "username", "email", "created_at", "last_login"]
>>>>

File: frontend/src/components/Profile.tsx

<<<< SEARCH
        <p>Member since: {user.created_at}</p>
      </div>
==== REPLACE
        <p>Member since: {user.created_at}</p>
        <p>Last login: {user.last_login ?? "Never"}</p>
      </div>
>>>>

Verify uniqueness of each search span. Confirm the model field, serializer, and UI are all consistent.

Best Practices

Do: Keep search spans minimal but unique. Include just enough surrounding context (a line above or below) to disambiguate if the target line appears multiple times.
Do: Preserve exact whitespace and indentation in search strings. A single space mismatch causes the patch to fail.
Do: Process files in dependency order — model changes before serializer changes before view changes — so the patch tells a coherent story.
Do: Use AST-level reasoning (function boundaries, class scope) rather than keyword grep to navigate to the right edit site. Semantic neighbors that look similar but need no change are common distractors.
Avoid: Using line numbers as anchors. They shift with every edit and across branches.
Avoid: Including unchanged boilerplate in your search span. The search block should be the minimal anchor, not the entire function body.
Avoid: Editing files that don't need changes. If a file is "close" to the bug but not causally involved, leave it alone. Over-editing is a major source of patch failure.

Error Handling

Search string matches zero locations: The code may have changed since the user's description. Re-read the file and regenerate the search span to match current content exactly.
Search string matches multiple locations: Expand the search context by including one or more surrounding lines until the match is unique. Never guess which occurrence is correct.
Patch applies but introduces a syntax error: Re-check indentation in the replace block. Python and YAML are especially sensitive. Verify bracket/brace balance.
File localization misses a required file: If the initial patch doesn't fully resolve the issue, re-scan the repo tree with broader keywords. Check for indirect dependencies (imports, config references, test files).
Context window exceeded for large files: Center the window on the function or class identified during fine-grained navigation. Include imports at the top of the file separately if needed.

Limitations

This agentless protocol does not iterate. If the first patch attempt is wrong, it requires human feedback or a separate retry — there is no built-in tool-use loop to self-correct.
Works best when the issue description is specific. Vague requests like "make it faster" produce weak file localization signals.
Cannot handle changes that require runtime information (e.g., "fix the bug that occurs when I click the button" without a stack trace or reproduction steps).
Large-scale mechanical changes (hundreds of files) are better served by traditional refactoring tools (sed, codemod) than by generating individual Search/Replace blocks.
The technique assumes the repository is in a consistent, buildable state. If the codebase has pre-existing conflicts or broken syntax, localization and patching reliability drops.

Reference

Paper: Pull Requests as a Training Signal for Repo-Level Code Editing (Zhu et al., 2026) Key takeaway: The three-stage agentless protocol (file localization → AST-guided navigation → minimal Search/Replace patch generation) with unique-context anchoring and round-trip verification produces deterministic, high-quality multi-file patches without iterative agent scaffolding.