composez-aider-fork 0.86.2

Composez aider fork to support pluggability and novel editing

Aider (Composez Fork)

This is fork of https://github.com/Aider-AI/aider, modified to allow changing the command set and to add new features in support of the composez novel editing features.

You can find the composez repo at https://github.com/kmewhort/composez.

Modifications to aider

Autonomy system

The fork replaces the old architect and agent edit formats with a new autonomy layer (aider/coders/autonomy.py) that is orthogonal to the edit mode. Three autonomy levels are available:

Level	Description
direct	Single-turn — one prompt, one response, edits applied.
compose	Two-phase — an LLM plans changes, then a separate editor coder implements them (replaces the old architect mode).
agent	Multi-step — the LLM produces a YAML plan that is executed via slash commands (requires composez_core.agent_runner).

Autonomy is selected via /chat compose, /chat agent, or /chat direct and is tracked on each coder instance via coder.autonomy_strategy.

New coders

• SelectionCoder (aider/coders/selection_coder.py) — operates on a user-highlighted text range using LSP-style Range coordinates. The LLM receives context lines around the selection and responds with a replacement block delimited by --[REPLACEMENT TEXT START]--> / <--[REPLACEMENT TEXT END]--.
• AgentCoder (aider/coders/agent_coder.py) — produces structured YAML plans for multi-step orchestration (used by the agent autonomy strategy).
• QueryCoder (aider/coders/query_coder.py) — replaces AskCoder. Read-only queries against the project; renamed to better fit the novel-editing workflow.

Composez integration hooks

The coder creation path (Coder.create()) includes soft-dependency hooks into composez_core:

• Novel mode activation — calls composez_core.novel_coder to apply novel-specific prompts and constraints based on the coder type.
• Role-based model resolution — reads a .composez config file to map (edit_format, autonomy) combinations to six named model roles: admin_model, query_model, edit_model, selection_model, compose_model, agent_model.
• Pluggable path validator — coder.edit_path_validator callback to enforce file-editing constraints (e.g. narrative file rules).
• Novel commands — Commands.novel_commands lazily loads composez_core.NovelCommands, which can add, override, or hide slash commands.

All composez_core imports are guarded by try/except ImportError so the fork runs standalone as a regular aider instance when composez_core is not installed.

Command and branding changes

• All user-facing strings rebrand "aider" to "composez" (args.py, commands.py, repo.py, etc.).
• /ask renamed to /query; /code aliased as /edit.
• /chat now shows edit modes (help, query, edit, selection) and autonomy levels (direct, compose, agent) as separate categories.
• /help simplified to show the command list (removed the interactive RAG-based help system).
• /report command removed.
• New --auto-context CLI flags.

Other changes

• Auto-context — when enabled, mentioned files are added to the chat automatically without prompting.
• Auto-create files — files in coder.auto_create_fnames are created without confirmation (e.g. SUMMARY.md, PROSE.md).
• ConfirmGroup for multi-file edits — apply_edits and allowed_to_edit now use ConfirmGroup so users can approve/skip all files at once.
• Lint skipped for query/selection — auto-lint is bypassed when the edit format is query or selection.

License

Licensed under Apache Version 2.0. See `license.txt' for details.