proto-change

name	proto-change
description	buf-validated Protocol Buffers authoring — write or modify .proto files with pre-edit analysis of wire-compat constraints, style conventions, and coordinated codegen + downstream consumer updates. Use for any non-trivial .proto edit (new message/field/rpc, type change, deprecation).

Proto Authoring Skill

Write and modify .proto files as a structured workflow — analogous to go-modify. Treat the .proto as the source of truth and generated Go/TS bindings as derived artifacts that must stay in lock-step.

When to Use This Skill

Use when:

• Adding or removing a message, field, enum value, or RPC
• Changing a field's type, label, or number
• Splitting or merging messages
• Introducing a new .proto file or package
• Any change that will regenerate *.pb.go / *_grpc.pb.go / generated TS types

Skip when:

• Comment-only edits
• Formatting / import ordering with no semantic change

Prerequisites

• buf, protoc, or the project's codegen task available (check buf.gen.yaml, Makefile, mise.toml, package.json)
• Generated output directories are writable and not in .gitignore if the repo commits generated code

PHASE 1: Pre-Edit Analysis

Read the entire target .proto file (and its import chain one level deep) before any edit.

File-level constraints checklist

• syntax — proto2 vs proto3 changes defaults (optional/required, zero values)
• package + option go_package / option java_package — generated path is derived from these
• Existing field numbers — note the highest number; reserved ranges; never reuse a removed field number or name (use reserved)
• Naming conventions in the file: snake_case field names, PascalCase messages, SCREAMING_SNAKE_CASE enum values, enum zero value suffix (_UNSPECIFIED)
• Wire-compat history — if this proto is on a network or persisted, deletions/renumberings are breaking. Mark removed fields with reserved <num>; and reserved "name";
• Custom options / annotations — google.api.http, validate.rules, buf.validate — follow the existing pattern
• Streaming RPCs — note stream on either side; clients regenerate differently
• Well-known types — prefer google.protobuf.Timestamp, Duration, Empty, Struct over reinventing

Codegen surface checklist

Identify every consumer before editing:

• Generator config (buf.gen.yaml plugins → output dirs)
• Go package(s) importing the generated *.pb.go
• TS package(s) importing the generated client (often @<scope>/<api>-ts)
• Any non-Go/TS consumer (Python, Java)

PHASE 2: Planning

1. Classify the edit:
   • Additive (new field with new number, new optional message, new RPC) → wire-compatible
   • Renaming (field/message name only, number unchanged) → wire-compatible but breaks source consumers
   • Breaking (type change, number change, removal, label change) → requires migration plan
2. For breaking changes: write the migration in .notes/ before editing (new field + dual-write + flip + remove).
3. Pick field numbers explicitly. Use 1–15 for hot fields (1-byte tag), 16+ for cold.
4. List the exact codegen command(s) to run after the edit.

Execution Rule

Never rm -rf a generated directory. Re-run codegen in place. If output looks stale, delete only files the generator owns and re-emit. Generated files belong in the same commit as the .proto change.

Never pause mid-task to ask "shall I proceed?" once intent is clear. Only stop on destructive ambiguity (e.g., the field number you want is in a reserved range).

PHASE 3: Implementation

Step 3.1: Edit the .proto

Use Edit tool. Match existing style. For removals, add reserved entries in the same edit.

Step 3.2: Lint

buf lint            # or: buf lint <module>
buf breaking --against '.git#branch=main'   # if the project tracks wire compat

Fix all findings. buf breaking failures are not optional unless the migration plan from Phase 2 explicitly accepts the break.

Step 3.3: Regenerate

Run the project's codegen task (look it up — do not guess):

buf generate                    # most projects
mise run gen                    # if a mise task exists
# or the Makefile target the repo defines

Verify the generator wrote new files. If output is empty, the plugin config is wrong — fix buf.gen.yaml, do not delete the output dir.

Step 3.4: Update Consumers

For each downstream Go/TS package identified in Phase 1:

• Apply the same naming change (rename, new field accessor)
• For new optional fields: existing call sites compile unchanged
• For removed/renamed fields: every reference must be updated in this same change

Use mcp gopls go_symbol_references (Go) and ast-grep / ccc (TS, polyglot) to find call sites.

Step 3.5: Build & Test

go build ./...
go test ./<affected>/...
pnpm -r build      # or the project's TS build

PHASE 4: Commit Hygiene

One commit contains:

1. The .proto change
2. All regenerated files
3. All consumer updates required to compile

Never split .proto from its generated artifacts across commits — intermediate states won't build.

Required Tools

• buf (lint + breaking + generate) or project codegen task
• Edit tool for the .proto
• mcp gopls go_symbol_references for Go consumers
• ast-grep / ccc for cross-language reference tracking

Final Report

• Edit class: additive / renaming / breaking
• Field numbers used / reserved
• buf lint and buf breaking results
• Generated files: list paths
• Consumers updated: files + symbols
• Build/test: results per package

Autoresearch rules

Eval checklist:

1. Did the agent read the full .proto and its imports before editing?
2. Was buf lint (or equivalent) run and clean?
3. Was buf breaking run for projects with wire-compat history?
4. Were generated files committed alongside the .proto?
5. Were zero rm -rf operations performed on generated dirs?
6. Did all downstream Go/TS consumers build after the change?

Test inputs:

• "Add a repeated string tags = 4 field to Tree message"
• "Rename GetNode RPC to FetchNode across api.proto and TS client"
• "Deprecate legacy_id field and add id as bytes"

Can change: pre-edit checklist, lint commands, consumer-discovery steps, generator invocation Cannot change: no-rm -rf rule, same-commit rule for proto + generated, lint-before-generate order Min sessions before eval: 5 Runs per experiment: 3