convez.ai
MCP ServerDesenvolvimento WebIntermediário
crisnahine

Chameleon

crisnahine/chameleon

Código gerado que se camufla no estilo do seu repo

Ver no GitHub
BomPluginScore · de 0 a 100

MCP server que aprende as convenções do seu repositório e injeta orientações conscientes de arquétipos a cada edição, garantindo que código gerado por IA corresponda ao estilo existente na primeira tentativa. Suporta TypeScript e Ruby on Rails, integrando-se nativamente com Claude Code.

Instalação

$npx convez-ai crisnahine/chameleon

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
8
Forks
3
Acompanhando
0
Saúde
71%
Último commit
há 3 meses
Versão
v4.10.0
Issues abertas
1
Contribuidores
1

Casos de uso

  • Manter consistência de padrões em edições assistidas por IA
  • Reduzir tempo de revisão de código gerado fora do padrão
  • Capturar e enforçar idiomas e convenções específicas do time

Veredito

Ferramenta sólida para times que usam Claude Code e querem garantir que gerações de código respeitem convenções locais. Reduz trabalho de revisão ao entregar código já alinhado. Requer setup inicial (profiling) mas vale o investimento em projetos estabelecidos.

README

A tradução para o português deste README ainda está na fila.

chameleon

"Code that blends in."

License: MIT Claude Code

chameleon learns your repo's actual conventions and injects archetype-aware guidance per-edit, so AI-generated code matches your existing style on the first try. Supports TypeScript and Ruby on Rails.

Quickstart

Before you start, you need uv and Node.js 20+ on your PATH. Ruby 3.0+ is also needed if you edit Rails repos. On a fresh machine, docs/install.md has copy-paste setup for macOS, Linux, and Windows, plus a check for each tool. Skip ahead if you already have them.

1. Install the plugin. In any Claude Code session:

/plugin marketplace add crisnahine/chameleon
/plugin install chameleon@chameleon

Restart Claude Code. Confirm it loaded by asking "What chameleon tools do you have?"

2. Profile a repo. Open a TypeScript or Ruby on Rails repo, then:

/chameleon-init    # build a profile (3-10s for repos under 5k files)
/chameleon-trust   # approve it for your user

After that, every Edit/Write in that repo gets archetype-aware context automatically.

Why

AI-generated code in established codebases routinely violates local conventions: wrong file location, off-pattern naming, missed team idioms, divergent error handling. Reviewer time gets spent on style and shape instead of logic and security.

chameleon clusters your actual code patterns (AST + statistical analysis), captures team-specific idioms (/chameleon-teach), and injects archetype-keyed guidance per-edit so the model writes code that fits.

How it works

  1. Bootstrap. /chameleon-init runs an AST scan, clusters files into archetypes by a 7-tuple signature, picks a canonical example per archetype, and writes .chameleon/profile.json — committed to git, team-shared.

  2. Trust. /chameleon-trust is a per-user, per-repo approval gate. Same mental model as git config --get user.signingkey: the profile lives in the repo, the trust grant lives on your machine.

  3. Per-edit context. Before every Edit/Write/NotebookEdit, the PreToolUse hook calls the chameleon MCP server, which returns the matched archetype's canonical excerpt, rules, and idioms. The hook injects <chameleon-context> into the model's context.

  4. Teach. /chameleon-teach captures idioms an AST can't infer — banned imports, mandatory wrappers, custom HTTP clients, internal conventions. Persisted to .chameleon/idioms.md and surfaced through the trust gate so reviewers see them before granting trust.

  5. Drift detection. Per-edit confidence is tracked in ~/.local/share/chameleon/<repo_id>/drift.db. When the profile no longer matches reality, /chameleon-status escalates and recommends /chameleon-refresh.

Because the skills trigger automatically, you don't need to do anything special after install. Edits in a trusted, profiled repo just blend in.

Install

The Quickstart above has the two commands for Claude Code. For the full guide - per-OS prerequisite setup (macOS, Linux, Windows), verification, updating, uninstall, and troubleshooting - see docs/install.md.

Workflow

  1. /chameleon-init — bootstrap a profile. Runs the AST scan, clusters files into archetypes, selects canonical examples, writes .chameleon/profile.json. Commit the result.

  2. /chameleon-trust — review and approve the committed profile for your user. Asks you to type the repo's basename to confirm. Trust state lives at ~/.local/share/chameleon/<repo_id>/.trust.

  3. Edit normally. The PreToolUse hook looks up the target file's archetype, fetches the canonical excerpt + rules + active idioms, and injects them as <chameleon-context> before the edit lands. The model references the canonical example and writes code that fits.

  4. /chameleon-teach — capture missed patterns as team idioms. Persists to .chameleon/idioms.md and survives /chameleon-refresh.

  5. /chameleon-status — check profile state, drift score, and plugin health. High drift recommends /chameleon-refresh to re-cluster against the current code.

  6. /chameleon-refresh — re-run the bootstrap after meaningful team changes. Existing idioms are preserved; the trust grant must be re-issued because the profile SHA changes.

What's Inside

Slash commands

CommandPurpose
/chameleon-initBootstrap a new profile
/chameleon-refreshRe-analyze and update profile after team changes
/chameleon-statusView profile state, drift score, plugin health
/chameleon-teachCapture a missed pattern as a team idiom
/chameleon-auto-idiomDerive novel team idioms from repo evidence (append-only, deduplicated against the profile)
/chameleon-trustApprove a committed profile for your user
/chameleon-disableSuppress chameleon for the rest of this session
/chameleon-pause-15mPause for 15 minutes (auto-resume)
/chameleon-doctorRun health checks on the chameleon installation
/chameleon-journeyRun the end-to-end journey test harness
/chameleon-pr-reviewReview a branch/PR against the repo's conventions and task intent

using-chameleon is one more skill on top of these — it auto-fires on SessionStart and orients the model.

Hooks

Five hooks drive the runtime:

  • SessionStart — detects the repo, loads the profile, and announces archetype awareness to the model.
  • PreToolUse — fires on Edit/Write/NotebookEdit; injects <chameleon-context> with the archetype's canonical excerpt, rules, and idioms.
  • PostToolUse (recorder) — fires on Bash/Edit/Write/NotebookEdit; records drift signals for /chameleon-status.
  • PostToolUse (verify) — fires on Edit/Write/NotebookEdit; runs archetype conformance lint on the written file.
  • UserPromptSubmit — detects frustration phrases and surfaces disable/pause/teach options.

MCP server

chameleon-mcp (Python, FastMCP, stdio transport) exposes 34 tools, including detect_repo, get_archetype, get_pattern_context, get_canonical_excerpt, get_rules, lint_file, get_drift_status, refresh_repo, bootstrap_repo, list_profiles, merge_profiles, teach_profile, teach_profile_structured, teach_competing_import, get_idiom_coverage, check_idiom_candidates, trust_profile, disable_session, pause_session, propose_archetype_renames, apply_archetype_renames, daemon_status, doctor, plus the review-gate tools (query_symbol_importers, get_crossfile_context, get_duplication_candidates, get_shadow_report, get_override_audit, get_longitudinal_signals, get_review_history, record_review_verdict, explain_edit, dep_audit, get_status). The full table lives in docs/architecture.md.

Opt-out hierarchy

Most-permanent → most-temporary:

.chameleon/.skip          per-repo, all users (committed)
CHAMELEON_DISABLE=1       per-user globally (in your shell rc)
CHAMELEON_VERIFY=0        disable post-edit verification only
/chameleon-disable        this session only
/chameleon-pause-15m      next 15 minutes (auto-resume)

Philosophy

  • Evidence over assumption — cluster what the repo actually does; don't assume a framework's defaults match this team's style.
  • Atomic and reviewable — profile writes use a flock-serialized COMMITTED sentinel; the trust gate surfaces idioms verbatim so reviewers see what they're approving.
  • Fail closed — the canonical-selection pipeline rejects candidates that fail secret, injection, or poisoning scans; the profile is shipped with no example before it is shipped with a poisoned one.
  • Opt-out at every layer — five ways to silence the plugin, from committed repo flag to a 15-minute pause.

See docs/architecture.md for the full design: bootstrap pipeline, cluster signature function, atomic profile commit, drift model, and security mitigations.

Contributing

See .github/CONTRIBUTING.md for dev setup, test workflows, and the change conventions used in this repo. Contributors hacking on the plugin itself should use --plugin-dir, not the marketplace install above.

License

MIT — see LICENSE.

Author

Cris Nahine — crisjosephnahine@gmail.com

Ver o original no GitHub

Bom para

Relacionados