chameleon-ai 0.5.12

Transpile a neutral agent configuration into Claude Code, Codex CLI, and other agent-specific formats — and back again.

Chameleon

One neutral config for Claude Code and Codex CLI, with round-trip merge in both directions.

Chameleon keeps a repo-backed neutral YAML file as the source of truth for your AI coding setup, but it does not pretend the live agent configs are read-only. Claude and Codex can still mutate their own files at runtime. On the next merge, Chameleon detects that drift, reconciles it back into neutral, and re-derives every other target.

That makes Chameleon the reconciliation layer between:

• operator-authored neutral config
• Claude Code's live files
• Codex CLI's live files
• login-time background merges

Today Chameleon ships built-in targets for Claude Code and Codex CLI. The plugin system is designed so more targets can join the same loop.

What it manages

Current built-in targets:

• Claude: ~/.claude/settings.json and the mcpServers slice of ~/.claude.json
• Codex: ~/.codex/config.toml and ~/.codex/requirements.toml

Chameleon stores its own state in XDG locations:

• neutral config: ~/.config/chameleon/neutral.yaml
• state root: ~/.local/state/chameleon/
• per-target state repos: ~/.local/state/chameleon/targets/<target>/
• last-known-good neutral snapshot: ~/.local/state/chameleon/neutral.lkg.yaml
• operator notices: ~/.local/state/chameleon/notices/

Each target gets a small local git repo. Those repos are not decoration; they are how Chameleon remembers last-applied state, computes drift, restores a target, and keeps a timeline of what was merged.

Why this exists

Claude and Codex overlap enough that maintaining both configs by hand gets old, but they are not the same system.

• Some settings are truly shared and should propagate across tools.
• Some settings are target-specific and should survive round-trip without being flattened away.
• Some settings conflict and need an operator decision.
• Some merges need to happen unattended at login, without blocking your shell.

Chameleon handles those cases with a neutral schema, per-target codecs, pass-through bags for target-only data, and an explicit conflict model instead of silent overwrite.

Core workflow

graph LR
    A["neutral.yaml"] --> B["merge"]
    C["Claude live files"] --> B
    D["Codex live files"] --> B
    B --> E["state repos"]
    B --> F["re-derived Claude"]
    B --> G["re-derived Codex"]

The merge reads four sources for each logical key:

• N0: last-known-good neutral
• N1: current neutral file
• per-target live values
• any stored prior resolution for the same conflict path

If everything agrees, the merge is silent. If they do not, Chameleon either:

• accepts the uniquely newest changed source with latest
• resolves interactively on a TTY
• applies a non-interactive strategy
• exits non-zero when unattended resolution is unsafe

Quick start

Chameleon is a uv project. Use uv run, not bare python or pytest.

uv sync
uv run chameleon init

init bootstraps neutral.yaml if missing, then runs a merge with the equivalent of "keep target-local drift".

Edit the neutral file:

schema_version: 1

identity:
  reasoning_effort: high
  model:
    claude: claude-sonnet-4-7
    codex: gpt-5.4

directives:
  system_prompt_file: ~/.config/chameleon/AGENTS.md

capabilities:
  mcp_servers:
    memory:
      transport: stdio
      command: npx
      args:
        - -y
        - "@modelcontextprotocol/server-memory"

environment:
  variables:
    CI: "true"

Then apply it:

uv run chameleon merge

Plain chameleon merge defaults to latest: if exactly one newest changed source can be proven, Chameleon takes it. On an interactive TTY, ambiguous conflicts fall through to the resolver prompt. Off a TTY, ambiguity exits non-zero rather than guessing.

What a merge actually does

chameleon merge is not just "render YAML into JSON and TOML".

It:

1. Reads live target files.
2. Disassembles them back into neutral-shaped fragments.
3. Compares those fragments against current neutral and last-known-good neutral.
4. Detects conflicts at the neutral-key level.
5. Resolves them interactively or non-interactively.
6. Re-assembles target files.
7. Writes live files and state repos.
8. Updates last-known-good neutral.
9. Clears stale transaction markers after a clean run.

If nothing changed, the summary is merge: nothing to do.

Conflict behavior

This is the part that matters.

Interactive merge

On a TTY, uv run chameleon merge opens an interactive resolver when needed. For each conflicting key, it shows:

• last-known-good neutral (N0)
• current neutral (N1)
• each changed target value

You can choose:

• neutral
• one target
• last-known-good
• target-specific preservation
• skip

Interactive decisions are stored in neutral.resolutions. If the same conflict shape appears again, Chameleon can replay the prior decision instead of prompting every time. If the values drift enough to invalidate the old decision hash, it re-prompts and shows the old decision as context.

Non-interactive strategies

For unattended runs:

uv run chameleon merge --on-conflict=latest
uv run chameleon merge --on-conflict=fail
uv run chameleon merge --on-conflict=keep
uv run chameleon merge --on-conflict=prefer-neutral
uv run chameleon merge --on-conflict=prefer-lkg
uv run chameleon merge --on-conflict=prefer=claude
uv run chameleon merge --on-conflict=prefer=codex

Meaning:

• latest: take the uniquely newest changed source; fail if ambiguous
• fail: stop on conflict
• keep: leave unresolved cross-target drift in place
• prefer-neutral: current neutral wins
• prefer-lkg: revert to last-known-good neutral
• prefer=<target>: one target wins

adopt <target> is a convenience wrapper for merge --on-conflict=prefer=<target>.

Login and dotfiles integration

Chameleon is built for login-time sync.

The intended model is:

• dotfiles pull/install happens first
• Chameleon runs last
• clean merges produce no output
• ambiguous conflicts explain themselves before prompting on an interactive shell

The repo ships recipes for:

• docs/login/launchd.md
• docs/login/systemd.md
• docs/login/zlogin.md

The unattended form is:

uv run chameleon merge --on-conflict=latest --quiet --no-warn

For a repo-backed neutral in dotfiles, pass it explicitly:

uv run chameleon merge \
  --neutral "$DOTFILES_DIR/chameleon/neutral.yaml" \
  --on-conflict=latest \
  --quiet \
  --no-warn

If a clean merge changes that checked-in neutral file, it is left as a normal dirty working-tree change. Chameleon never commits or pushes for you.

If Git reports a conflict in chameleon/neutral.yaml, resolve the Git conflict first. Then rerun chameleon merge so the merged neutral spreads to Claude and Codex.

If Chameleon itself sees an ambiguous conflict on an interactive login, it prints a short preamble explaining the neutral, Claude, and Codex sources before asking you to choose. In a non-interactive service, the same ambiguity exits non-zero and should be resolved by rerunning chameleon merge from a shell.

CLI

init

uv run chameleon init
uv run chameleon init --dry-run

Creates a minimal neutral file if needed, then runs an initial merge.

merge

uv run chameleon merge
uv run chameleon merge --dry-run
uv run chameleon merge --verbose
uv run chameleon merge --quiet --no-warn
uv run chameleon merge --profile deep-review
uv run chameleon merge --on-conflict=prefer=codex

• --dry-run emits unified diffs for files it would write
• --verbose prints state paths, registered targets, pending notices / transactions, per-target warning counts, and merge id
• --quiet suppresses the summary line and dry-run diff output
• --no-warn suppresses end-of-merge LossWarning errata
• --profile <name> applies a named overlay from profiles

status

uv run chameleon status

Shows whether neutral exists, whether each target is clean or drifting, and whether there are pending notices or unresolved transactions. Exit code is 1 if anything is dirty or pending.

diff

uv run chameleon diff
uv run chameleon diff claude
uv run chameleon diff codex

Shows unified diff from state-repo HEAD to the live target file(s). Exit code matches git-style diff behavior: 0 clean, 1 drift, >1 error.

log

uv run chameleon log claude
uv run chameleon log codex

Shows the per-target state-repo history.

adopt

uv run chameleon adopt claude
uv run chameleon adopt codex

Resolve every conflict in favor of the chosen target, then propagate that back through neutral and out to the other targets.

discard

uv run chameleon discard claude
uv run chameleon discard codex --yes

Restore live target files from the state-repo HEAD. This is intentionally guarded: off a TTY you must pass --yes.

For partial-owned files, Chameleon only rewrites the keys it owns. Today that matters for ~/.claude.json, where it preserves keys outside mcpServers.

validate

uv run chameleon validate

Schema-validates neutral.yaml.

doctor

uv run chameleon doctor
uv run chameleon doctor --notices-only
uv run chameleon doctor --clear-notices

Shows Chameleon version, resolved paths, pending operator notices, and unresolved transaction markers from interrupted merges.

targets

uv run chameleon targets list

Lists built-in and plugin-provided targets discovered through the chameleon.targets entry point.

resolutions

uv run chameleon resolutions list
uv run chameleon resolutions clear
uv run chameleon resolutions clear 'identity.model[claude]' --yes

Inspects or clears stored interactive conflict decisions.

Neutral schema

The neutral schema is centered in src/chameleon/schema/. It has eight top-level domains:

• identity
• directives
• capabilities
• authorization
• environment
• lifecycle
• interface
• governance

Plus:

• profiles: named overlays applied with merge --profile <name>
• targets.<target>.items: pass-through bag for target-only features
• resolutions: stored conflict decisions

The rule is simple: codecs adapt to the neutral schema. They do not redefine it.

Round-trip guarantees

Round-trip is the design center.

For settings Chameleon claims, the goal is:

from_target(to_target(x)) == canonicalize(x)

When a target has features the other target cannot represent, Chameleon does not silently drop them. It uses one of two escape hatches:

• typed LossWarnings for genuinely lossy cross-target translations
• per-target pass-through under targets.<target>.items

Examples in the current implementation:

• plugin and marketplace state is reconciled across targets rather than flattened to one side
• target-specific extras can survive under targets.claude.items or targets.codex.items
• interrupted merges leave transaction markers instead of disappearing into logs

Repository-backed workflow

The local state repos are a core part of the operator workflow:

• status checks drift against them
• diff shows what changed since HEAD
• log gives you a timeline
• discard restores from HEAD
• merges commit new snapshots with a merge id trailer

This gives Chameleon a memory that is stronger than "last file on disk", while still staying local and inspectable.

Plugin targets

New targets register through the chameleon.targets entry point.

At a minimum a plugin provides:

• a Target class
• an assembler with file routing
• eight codec classes, one per neutral domain
• vendored _generated.py models derived from the upstream schema

See docs/plugins/authoring.md.

Development

Install dependencies:

uv sync

Before calling work complete, all four local gates must pass:

uv run ruff check
uv run ruff format --check
uv run ty check
uv run pytest

The default pytest run excludes the longer fuzz suite. Run it explicitly with:

uv run pytest -m fuzz

The repo also carries integration tests that pin the important operator-facing behavior:

• exemplar end-to-end round-trip
• byte-stable idempotency across repeated merges
• login recipe docs staying aligned with the real CLI
• schema drift against vendored upstream models

CI runs the four standard gates on macOS and Linux, across Python 3.12 and 3.13.

Schema sync

Generated codec models under src/chameleon/codecs/*/_generated.py are vendored artifacts. Do not hand-edit them.

When you intentionally refresh an upstream schema:

uv run --group schema-sync python tools/sync-schemas/sync.py claude
uv run --group schema-sync python tools/sync-schemas/sync.py codex

The Codex sync uses the pinned upstream revision from tools/sync-schemas/pins.toml.

Status

Chameleon is pre-1.0 and still tightening parity edges, but the current shape is already practical:

• repo-backed neutral config
• bidirectional Claude/Codex reconciliation
• unattended login-time merge hooks
• interactive conflict resolution with stored decisions
• plugin target expansion path

License

MIT. See LICENSE.