default-config-registry — Workgraph live mirror

Metadata

Status	abandoned ‖ paused
Created	2026-04-26T12:52:50.010651414+00:00

Description

Today wg init -x claude writes a partial config (sets [agent] model, leaves [tiers] empty). wg setup is interactive-only, hard to escape, and shows confusing labels. wg config reset doesn't exist. All four problems share one fix: a per-executor default-config registry that everything calls into.

Spec

defaults module (e.g. src/config/defaults.rs):
- pub fn for_executor(executor: Executor) -> Config returning a complete, working Config.
- For claude: agent.model=claude:opus, tiers={fast=claude:haiku, standard=claude:sonnet, premium=claude:opus}, models.evaluator=claude:haiku, models.assigner=claude:haiku, coordinator.model=claude:sonnet (orchestration tier), no model_registry entries that aren't claude.
- For amplifier, codex, shell, nex: equivalent complete defaults appropriate to each.
- Unit-tested: every executor returns a Config that round-trips through TOML and has no empty required tables.
wg init -x uses defaults::for_executor() instead of hand-built partial config.
- Add --dry-run flag that prints the would-be config and exits without writing.
- On success, print a one-line summary: 'Wrote .wg/config.toml: executor=claude, model=claude:opus, tiers=haiku/sonnet/opus'.
wg setup rework:
- After provider/executor pick, show a diff-style preview: current config vs proposed defaults.
- Prompt: [a]ccept / [c]ustomize / [q]uit. Quit leaves existing config untouched.
- Add --non-interactive mode with required --executor flag; uses defaults::for_executor() with no prompts.
- Provider menu labels disambiguated: 'Anthropic API (direct, needs ANTHROPIC_API_KEY)' vs 'Claude Code CLI (uses local claude binary)' — and they map to different executor configs (or same executor with different auth).
- Print '(Ctrl-C to exit; nothing saved until final confirm)' at the top.
wg config reset (new command):
- Calls defaults::for_executor(current_executor_or_claude).
- --keep-keys preserves [[llm_endpoints.endpoints]] entries and api_key_file paths.
- --dry-run prints diff vs current.
- --yes skips confirmation.
- Always writes timestamped backup config.toml.bak-YYYY-MM-DDTHHMMSS.

Validation

Failing tests written first: test_default_config_claude_complete, test_default_config_amplifier_complete, test_init_dry_run_no_write, test_setup_non_interactive_writes_defaults, test_config_reset_keep_keys_preserves_endpoints, test_config_reset_creates_backup
Implementation makes all tests pass
cargo build + cargo test pass with no regressions
Manual smoke (in scratch dir): wg init -x claude produces config with populated [tiers]; wg setup --non-interactive --executor claude works without prompts; wg config reset --dry-run prints diff; wg setup interactive shows escape hint and disambiguated provider labels

## Description
Today `wg init -x claude` writes a partial config (sets [agent] model, leaves [tiers] empty). `wg setup` is interactive-only, hard to escape, and shows confusing labels. `wg config reset` doesn't exist. All four problems share one fix: a per-executor default-config registry that everything calls into.

### Spec

1. **defaults module** (e.g. `src/config/defaults.rs`):
   - `pub fn for_executor(executor: Executor) -> Config` returning a complete, working Config.
   - For claude: agent.model=claude:opus, tiers={fast=claude:haiku, standard=claude:sonnet, premium=claude:opus}, models.evaluator=claude:haiku, models.assigner=claude:haiku, coordinator.model=claude:sonnet (orchestration tier), no model_registry entries that aren't claude.
   - For amplifier, codex, shell, nex: equivalent complete defaults appropriate to each.
   - Unit-tested: every executor returns a Config that round-trips through TOML and has no empty required tables.

2. **wg init -x <exec>** uses defaults::for_executor() instead of hand-built partial config.
   - Add `--dry-run` flag that prints the would-be config and exits without writing.
   - On success, print a one-line summary: 'Wrote .wg/config.toml: executor=claude, model=claude:opus, tiers=haiku/sonnet/opus'.

3. **wg setup** rework:
   - After provider/executor pick, show a diff-style preview: current config vs proposed defaults.
   - Prompt: [a]ccept / [c]ustomize / [q]uit. Quit leaves existing config untouched.
   - Add `--non-interactive` mode with required `--executor` flag; uses defaults::for_executor() with no prompts.
   - Provider menu labels disambiguated: 'Anthropic API (direct, needs ANTHROPIC_API_KEY)' vs 'Claude Code CLI (uses local claude binary)' — and they map to different executor configs (or same executor with different auth).
   - Print '(Ctrl-C to exit; nothing saved until final confirm)' at the top.

4. **wg config reset** (new command):
   - Calls defaults::for_executor(current_executor_or_claude).
   - `--keep-keys` preserves [[llm_endpoints.endpoints]] entries and api_key_file paths.
   - `--dry-run` prints diff vs current.
   - `--yes` skips confirmation.
   - Always writes timestamped backup config.toml.bak-YYYY-MM-DDTHHMMSS.

## Validation

- [ ] Failing tests written first: test_default_config_claude_complete, test_default_config_amplifier_complete, test_init_dry_run_no_write, test_setup_non_interactive_writes_defaults, test_config_reset_keep_keys_preserves_endpoints, test_config_reset_creates_backup
- [ ] Implementation makes all tests pass
- [ ] cargo build + cargo test pass with no regressions
- [ ] Manual smoke (in scratch dir): `wg init -x claude` produces config with populated [tiers]; `wg setup --non-interactive --executor claude` works without prompts; `wg config reset --dry-run` prints diff; `wg setup` interactive shows escape hint and disambiguated provider labels

Depends on

(none)

Required by

(none)

Log

2026-04-26T12:52:50.010526600+00:00 Task paused
2026-04-26T13:00:39.501121104+00:00 Task abandoned