feat: reuse Codex and Claude subscriptions for LLM roles

README.md

@@ -178,6 +178,56 @@ commands, expected report fields, acceptance criteria, and troubleshooting.
 - `analyze`
 - `validate`
 
+## Codex and Claude subscription backends
+
+Local Codex and Claude Code logins can power proposer and built-in evaluator
+roles without API keys. Selection is explicit; omitting backend flags preserves
+the existing LiteLLM API behavior.
+
+```bash
+# Install the pinned Codex SDK adapter, then authenticate with ChatGPT.
+uv sync --extra codex
+codex login
+
+# Codex subscription: proposer plus built-in judge.
+optimize-anything optimize seed.txt \
+  --proposer-backend codex --judge-backend codex \
+  --objective "Improve clarity"
+
+# Claude Code subscription. Install/sign in to the claude CLI first.
+optimize-anything optimize seed.txt \
+  --proposer-backend claude --judge-backend claude \
+  --objective "Improve clarity"
+```
+
+Subscription calls are serialized per provider by default. The CLI may switch
+an eligible failure to a same-vendor API only when a matching key and fallback
+model are available, and prints a billing warning first. Use
+`--no-api-fallback` to prohibit that switch. Claude support is local-only and
+experimental; neither adapter initiates login or copies credential contents.
+
+Single-role commands use `--analysis-backend` (`analyze`) or
+`--judge-backend` (`score`). `validate --providers` also accepts `codex`,
+`codex:<model>`, `claude`, and `claude:<model>`. Structured TOML uses role
+tables such as:
+
+```toml
+[model.proposer]
+backend = "codex"
+api_fallback = false
+
+[model.judge]
+backend = "claude"
+api_fallback_model = "anthropic/claude-sonnet-5"
+```
+
+Opt-in live gates consume local subscription quota:
+
+```bash
+OPTIMIZE_ANYTHING_RUN_SUBSCRIPTION_LIVE=1 \
+  uv run pytest tests/test_subscription_live.py
+```
+
 ## Agent Plugins
 
 The Claude Code plugin and Codex plugin share the same `skills/` tree and
@@ -380,7 +430,7 @@ optimize-anything optimize seed.txt \
 
 ### `optimize` flags (complete)
 
-Exactly one evaluator source is required: `--evaluator-command` OR `--evaluator-url` OR `--judge-model`.
+Exactly one evaluator source is required: `--evaluator-command` OR `--evaluator-url` OR a built-in judge selected by `--judge-model`/`--judge-backend`.
 
 | Flag | Description | Default |
 |---|---|---|
@@ -397,7 +447,13 @@ Exactly one evaluator source is required: `--evaluator-command` OR `--evaluator-
 | `--budget <int>` | Max evaluator calls | `100` |
 | `--output, -o <file>` | Write best artifact to file | -- |
 | `--model <model>` | Proposer model (or env fallback) | `OPTIMIZE_ANYTHING_MODEL`, then `openai/gpt-5.6-sol` |
+| `--proposer-backend api\|codex\|claude` | Proposal backend | `api` |
 | `--judge-model <model>` | Built-in LLM judge evaluator model | -- |
+| `--judge-backend api\|codex\|claude` | Built-in judge backend | `api` |
+| `--subscription-concurrency <int>` | Concurrent calls allowed per subscription provider | `1` |
+| `--no-api-fallback` | Prohibit subscription-to-API fallback | `false` |
+| `--openai-api-fallback-model <model>` | Same-vendor API fallback for Codex | -- |
+| `--anthropic-api-fallback-model <model>` | Same-vendor API fallback for Claude | -- |
 | `--judge-objective <text>` | Judge objective override | falls back to `--objective` |
 | `--api-base <url>` | Override LiteLLM API base | -- |
 | `--diff` | Print unified diff (seed vs best) to stderr | `false` |

SKILL.md

@@ -32,6 +32,19 @@ evaluation, invoke `$optimize-prompt` in Claude Code or the namespaced
 `$optimize-anything:optimize-prompt` in Codex. It uses the bundled runtime,
 keeps search output outside the source, and applies only an accepted result.
 
+## Reuse the Host Subscription
+
+When this skill runs in Codex, pass `--proposer-backend codex` and use
+`--judge-backend codex` for built-in judging. When it runs in Claude Code, use
+the corresponding `claude` values. For `analyze`, pass `--analysis-backend`;
+for `validate`, use the reserved provider selector `codex` or `claude`.
+Do not infer a backend in the Python runtime or for an unknown host.
+
+Tell the user before running that subscription calls are serialized by default.
+Eligible availability, authentication, rate-limit, or quota failures may switch
+to a same-vendor API model only when a matching key and fallback model are
+available; pass `--no-api-fallback` when billed fallback is not acceptable.
+
 ## Available Skills
 
 - **optimize-prompt** — Optimize inline prompts, files, embedded regions, or independent batches with fast prompt-quality or rigorous task-output evidence

commands/analyze.md

@@ -2,6 +2,10 @@
 name: analyze
 description: Discover quality dimensions for an artifact and objective
 ---
+When running in Codex, pass `--analysis-backend codex`; in Claude Code, pass
+`--analysis-backend claude`. A subscription backend may omit `--judge-model`.
+Unknown hosts use the existing API model. Announce possible same-vendor billed
+API fallback, or pass `--no-api-fallback` when it is not acceptable.
 
 # analyze
 
@@ -21,6 +25,9 @@ Use an LLM to discover relevant quality dimensions for a given artifact and opti
   --objective "Score for clarity and persuasiveness"
 ```
 
+In Codex or Claude Code, replace `--judge-model ...` with
+`--analysis-backend codex` or `--analysis-backend claude`, respectively.
+
 ## Next Step: optimize with discovered dimensions
 After dimension discovery, proceed directly to optimization using the returned `intake_json`:
 

commands/compare.md

@@ -2,6 +2,10 @@
 name: compare
 description: Side-by-side comparison of two artifacts using composed score calls
 ---
+When comparison uses LLM scoring, select the current host explicitly:
+`--judge-backend codex` in Codex or `--judge-backend claude` in Claude Code.
+Unknown hosts keep API defaults. Announce possible same-vendor billed API
+fallback, or add `--no-api-fallback` to prohibit it.
 Compare two artifacts with the same scoring setup by composing existing `score` calls.
 
 ## Usage

commands/optimize.md

@@ -8,6 +8,12 @@ For inline prompts, embedded prompt regions, independent prompt batches, or
 task-output evaluation with representative examples, invoke the shared
 `$optimize-prompt` skill instead.
 
+When executing this Claude Code command, add `--proposer-backend claude` and,
+when using the built-in judge, `--judge-backend claude`. Announce that
+subscription calls are serialized and that
+eligible failures may use billed same-vendor API fallback; honor a request to
+disable it with `--no-api-fallback`.
+
 ## Step 1: Identify the artifact
 - If the user provided a file argument, use it directly.
 - Otherwise ask: **"What file should I optimize?"**
@@ -33,7 +39,8 @@ Present these options and ask the user to choose one unless they already specifi
 - If evaluator is already specified, proceed.
 - If no evaluator is specified:
   1. Run `analyze` first:
-     - `"${CLAUDE_PLUGIN_ROOT}/scripts/run-optimize-anything" analyze <file> --judge-model <model> --objective "<objective>"`
+     - Subscription mode: `"${CLAUDE_PLUGIN_ROOT}/scripts/run-optimize-anything" analyze <file> --analysis-backend claude --objective "<objective>"`
+     - API mode: `"${CLAUDE_PLUGIN_ROOT}/scripts/run-optimize-anything" analyze <file> --judge-model <model> --objective "<objective>"`
   2. If analyze fails (API key missing, model unavailable): ask the user for their preferred model, or suggest using `--evaluator-command` with a custom script instead.
   3. Ask: **"Should we use LLM judge directly, or do you want a custom evaluator?"**
   4. If custom evaluator is needed, invoke evaluator generation workflow.

commands/quick.md

@@ -4,15 +4,21 @@ description: Zero-config one-shot optimization for fast improvements
 ---
 Run a no-questions-asked fast optimization.
 
+Use Claude Code's subscription explicitly with `--analysis-backend claude`,
+`--proposer-backend claude`, and `--judge-backend claude`. Announce serialized
+subscription use and possible same-vendor billed API fallback before running.
+
 ## Usage
 `/optimize-anything:quick <file> "<objective>"`
 
 ## Behavior (do not ask follow-up questions)
 1. Run analysis to discover dimensions:
-   - `"${CLAUDE_PLUGIN_ROOT}/scripts/run-optimize-anything" analyze <file> --judge-model openai/gpt-5.6-luna --objective "<objective>"`
+   - Subscription mode: `"${CLAUDE_PLUGIN_ROOT}/scripts/run-optimize-anything" analyze <file> --analysis-backend claude --objective "<objective>"`
+   - API mode: `"${CLAUDE_PLUGIN_ROOT}/scripts/run-optimize-anything" analyze <file> --judge-model openai/gpt-5.6-luna --objective "<objective>"`
    - If analyze fails, skip dimension discovery and run optimize with `--judge-model` directly using the objective as-is.
 2. Run optimization using LLM judge with:
-   - `--judge-model openai/gpt-5.6-luna`
+   - Subscription mode: `--proposer-backend claude --judge-backend claude` (omit API model strings)
+   - API mode: `--judge-model openai/gpt-5.6-luna`
    - `--budget 50`
    - `--diff`
    - `--early-stop`

commands/score.md

@@ -2,6 +2,10 @@
 name: score
 description: Score a single artifact with an evaluator
 ---
+For an LLM judge, use `--judge-backend codex` in Codex or
+`--judge-backend claude` in Claude Code; the subscription model may be omitted.
+Do not add a judge backend for command or HTTP evaluators. Announce possible
+same-vendor billed API fallback, or add `--no-api-fallback` to prohibit it.
 
 # score
 

commands/validate.md

@@ -2,6 +2,10 @@
 name: validate
 description: Cross-validate an artifact with multiple LLM judge providers
 ---
+The `--providers` list accepts `codex`, `codex:<model>`, `claude`, and
+`claude:<model>` before ordinary LiteLLM model strings. Prefer the selector for
+the current host when subscription reuse is requested. Announce possible
+same-vendor billed API fallback, or add `--no-api-fallback` to prohibit it.
 Use multiple LLM judges to verify that a quality improvement is not provider-specific.
 
 ## When to use