@galaxy9day/executor-adapter 0.10.1 → 0.10.3

package/README.md

@@ -74,19 +74,22 @@ reviewer    = "anthropic/claude-opus-4-7" # pi executor, mode=check / cross-mode
 # fast        = "newapi/gpt-5-mini"
 
 [executor_adapter.codex]
-implementer = "gpt-5.5-codex"            # optional; omit the section to use the codex CLI default model
+implementer = "gpt-5.5"                  # optional; omit the section to use the codex CLI default model
 reviewer    = "gpt-5.5"
 ```
 
 Without this config, pi-executor calls without a fully-qualified `model` parameter throw a friendly error pointing you here. The codex executor never requires config: with no `[executor_adapter.codex]` entry and no `model` parameter it omits `-m` and uses the codex CLI's own configured default model.
 
-You can also pass a fully-qualified Pi route (or codex model name) directly per call:
+You can also pass a fully-qualified Pi route or exact Codex model id directly per call:
 
 ```
 dispatch(model="anthropic/claude-opus-4-7", executor="pi", ...)
-dispatch(model="gpt-5.5-codex", executor="codex", ...)
+dispatch(executor="codex", ...)                  # use this machine's Codex default
+dispatch(model="gpt-5.5", executor="codex", ...) # explicit Codex model id
 ```
 
+Do not pass `model="codex"`: `codex` is the executor backend name, not a model.
+
 ## Executors
 
 Two backends share the same dispatch pipeline (worktree isolation, diff export, post-validation, result classes, channel audit):
@@ -116,7 +119,7 @@ Assemble context, run the selected executor, optionally emit channel events, run
 | `task_dir` | string | Trellis task directory (relative to repo). Omit to auto-resolve via `task.py current`. |
 | `working_directory` | string | Repo root; defaults to CWD. |
 | `executor` | `pi` \| `codex` | Executor backend. Defaults: `default_executor` config, else `implement/custom→codex`, `check→pi`. |
-| `model` | string | Logical name (`implementer` / `reviewer` / custom key; `[executor_adapter.codex]` for codex) or fully qualified route/model name. |
+| `model` | string | Logical name (`implementer` / `reviewer` / custom key; `[executor_adapter.codex]` for codex) or fully qualified route/model name. For Codex, omit it to use `$CODEX_HOME/config.toml`; never use `model="codex"`. |
 | `thinking` | string | Reasoning effort. Default `xhigh` (`--thinking` for pi, `model_reasoning_effort` for codex). |
 | `execution_mode` | string | `review`, `patch`, `worktree`, or `direct`. Defaults to `worktree` for `implement/custom`, `review` for `check`. |
 | `isolate_executor` / `isolate_pi` | boolean | Default `true`. `isolate_executor` is the preferred name; `isolate_pi` remains as a compatibility alias. Pi: disables extensions/skills/prompt templates/context files/session persistence and uses a per-worker Pi home. Codex: `--ignore-rules --ephemeral` while still loading user config for provider/auth settings. |
@@ -188,7 +191,7 @@ Same args as `dispatch` (subset). Renders the prompt without launching an execut
 
 ### `smoke({ model?, mode?, executor? })`
 
-One-shot connectivity test. Verifies the executor binary is on PATH and the resolved model answers a trivial round-trip. `model` accepts either a logical name or a fully qualified route/model name, `mode` (`implement` or `check`) chooses the default logical key when `model` is omitted, and `executor` follows the same routing defaults as `dispatch`. On failure it prints separate stdout and stderr blocks, the resolved model, safe env values, and (pi) the config files copied into the isolated Pi home.
+One-shot connectivity test. Verifies the executor binary is on PATH and the resolved model answers a trivial round-trip. `model` accepts either a logical name or a fully qualified route/model name, `mode` (`implement` or `check`) chooses the default logical key when `model` is omitted, and `executor` follows the same routing defaults as `dispatch`. For Codex, `model="codex"` is rejected for the same reason as `dispatch`. On failure it prints separate stdout and stderr blocks, the resolved model, safe env values, and (pi) the config files copied into the isolated Pi home.
 
 ### `read_report({ log_file?, report_file?, runtime_dir?, worker_id?, lines? })`
 
@@ -246,7 +249,7 @@ Task(subagent_type="trellis-pi-check")
 
 This keeps long dispatch output out of the main context. The sub-agent receives the MCP result, reads reports only when needed, and returns a short summary to the orchestrator.
 
-Use `trellis-codex-implement` for the normal GPT implementation path. Use `trellis-pi-implement` only when you explicitly want Pi provider routing for implementation, and `trellis-pi-check` for read-only cross-model review. The Codex template omits `model` so `codex exec` uses this machine's Codex default from `$CODEX_HOME/config.toml`; the Pi templates resolve logical routes from `~/.pi/config.toml`.
+Use `trellis-codex-implement` for the normal GPT implementation path. Use `trellis-pi-implement` only when you explicitly want Pi provider routing for implementation, and `trellis-pi-check` for read-only cross-model review. The templates set `effort: xhigh` for the Claude Code sub-agent itself; the MCP `thinking` parameter separately defaults to `xhigh` for the Codex/Pi executor it launches. The Codex template omits `model` so `codex exec` uses this machine's Codex default from `$CODEX_HOME/config.toml`; the Pi templates resolve logical routes from `~/.pi/config.toml`.
 
 No Trellis `inject-subagent-context` hook edit is required. `dispatch` assembles Trellis task context itself from the task artifacts (`prd.md`, `design.md`, `implement.md`, `implement.jsonl`, and `check.jsonl` where applicable). For Codex/Pi/other platforms, translate these templates to the platform's native custom-agent syntax; do not put MCP tool names in `.trellis/agents/*.md`, because those files are provider-level channel runtime prompts, not MCP-capability declarations.
 

package/executors.js

@@ -90,8 +90,18 @@ export class ModelResolutionError extends Error {
   }
 }
 
+export class InvalidCodexModelError extends Error {
+  constructor(model) {
+    super(`Invalid Codex model "${model}": "codex" is the executor backend name, not a model. Use executor="codex" and omit model to use the Codex CLI default from $CODEX_HOME/config.toml, or pass a real Codex model id such as model="gpt-5.5".`);
+    this.code = 'INVALID_CODEX_MODEL';
+    this.model = model;
+  }
+}
+
 // ---- Shared subprocess helpers ----
 
+const CODEX_BACKEND_ALIASES = new Set(['codex', 'pi', 'executor-adapter', 'pi-adapter']);
+
 // Explicit env override is authoritative: a configured-but-missing path is an
 // error, not a trigger for silent PATH fallback.
 function makeBinaryFinder(envVar, command) {
@@ -239,6 +249,10 @@ const codex = {
     const defaultKey = mode === 'check' ? 'reviewer' : 'implementer';
     const map = loadAdapterConfig().codex;
     const logicalKey = input || defaultKey;
+    if (input) {
+      const normalized = String(input).trim().toLowerCase();
+      if (CODEX_BACKEND_ALIASES.has(normalized)) throw new InvalidCodexModelError(input);
+    }
     if (map[logicalKey]) return { resolved: map[logicalKey], from: 'config', key: logicalKey };
     if (input) return { resolved: input, from: 'direct', key: null };
     // No config and no explicit model: defer to the codex CLI's own

package/index.js

@@ -45,7 +45,7 @@ import { EXECUTORS, resolveExecutor } from './executors.js';
 
 const SERVER_NAME = 'executor-adapter';
 const LEGACY_SERVER_NAMES = ['pi-adapter'];
-const SERVER_VERSION = '0.10.1';  // keep in sync with package.json
+const SERVER_VERSION = '0.10.3';  // keep in sync with package.json
 const TMP_RUNTIME_DIR = path.join(os.tmpdir(), SERVER_NAME);
 const CHANNEL_ENV_ALIASES = ['TRELLIS_CHANNEL', 'TRELLIS_CHANNEL_NAME'];
 const TRELLIS_BIN_ENV = 'TRELLIS_BINARY';
@@ -1725,7 +1725,7 @@ const TOOLS = [
         task_dir: { type: 'string', description: 'Explicit Trellis task directory (relative to repo root). Omit to auto-resolve.' },
         working_directory: { type: 'string', description: 'Repo root. Defaults to cwd.' },
         executor: { type: 'string', enum: ['pi', 'codex'], description: 'Executor backend. Defaults: [executor_adapter] default_executor in ~/.pi/config.toml, else implement/custom→codex, check→pi.' },
-        model: { type: 'string', description: 'Logical name (implementer / reviewer / custom key from ~/.pi/config.toml [executor_adapter], or [executor_adapter.codex] for codex) or a fully qualified route/model name. Omit to use the default for the mode (codex falls back to the codex CLI default model).' },
+        model: { type: 'string', description: 'Logical name (implementer / reviewer / custom key from ~/.pi/config.toml [executor_adapter], or [executor_adapter.codex] for codex) or a fully qualified route/model name. Omit to use the default for the mode (codex falls back to the codex CLI default model). For the Codex backend, use executor="codex"; do not pass model="codex".' },
         thinking: { type: 'string', default: 'xhigh' },
         execution_mode: { type: 'string', enum: ['review', 'patch', 'worktree', 'direct'], description: 'review=read-only report, patch=final-answer diff, worktree=isolated git worktree + exported diff.patch, direct=legacy in-place execution. Defaults to worktree for implement/custom and review for check.' },
         isolate_executor: { type: 'boolean', default: true, description: 'Preferred isolation flag. When true, Pi gets --no-extensions/--no-skills/etc. plus a per-worker PI_CODING_AGENT_DIR; codex gets --ignore-rules --ephemeral while still loading user config for provider/auth settings.' },
@@ -1772,7 +1772,7 @@ const TOOLS = [
     inputSchema: {
       type: 'object',
       properties: {
-        model: { type: 'string', description: 'Logical name or fully qualified route/model name.' },
+        model: { type: 'string', description: 'Logical name or fully qualified route/model name. For the Codex backend, use executor="codex"; do not pass model="codex".' },
         mode: { type: 'string', enum: ['implement', 'check'], default: 'implement', description: 'Default logical key to resolve when model is omitted.' },
         executor: { type: 'string', enum: ['pi', 'codex'], description: 'Executor backend. Defaults: [executor_adapter] default_executor, else implement→codex, check→pi.' },
         working_directory: { type: 'string' },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galaxy9day/executor-adapter",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "description": "Multi-executor MCP (Codex CLI for GPT implementation, Pi CLI for cross-model review) for Trellis-aware and standalone use. Reads Trellis task artifacts, runs the executor in an isolated subprocess/worktree, and can optionally emit non-invasive Trellis channel audit messages.",
   "type": "module",
   "main": "index.js",

package/templates/claude/agents/trellis-codex-implement.md

@@ -5,6 +5,7 @@ description: >-
   executor via executor-adapter in an isolated worktree, returning only a short
   summary with result_class, changed files, and apply command. Use this as the
   normal GPT implementation path for Trellis tasks.
+effort: xhigh
 tools: mcp__executor-adapter__dispatch, mcp__executor-adapter__preview_prompt, mcp__executor-adapter__read_report, mcp__executor-adapter__smoke, Read, Grep, Glob, Bash
 ---
 
@@ -41,6 +42,9 @@ main orchestrator context. Dispatch only when the task is ready and bounded.
 - Use `mode="implement"`, `executor="codex"`, and
   `execution_mode="worktree"` unless the main session explicitly asks for a
   different safe mode.
+- Do not set `model="codex"`; `codex` is the executor backend name. Omit
+  `model` unless the main session gives an exact Codex model id such as
+  `gpt-5.5`.
 - Provide an explicit `scope` naming the files, directories, or behavior the
   executor may change, and cheap `validation_commands` when the project has
   fast checks.

package/templates/claude/agents/trellis-pi-check.md

@@ -4,6 +4,7 @@ description: >-
   Cross-model code review via the Pi reviewer model through executor-adapter; read-only;
   supplements, but does not replace, the orchestrator's own verification and
   native trellis-check. Returns severity-graded findings only.
+effort: xhigh
 tools: mcp__executor-adapter__dispatch, mcp__executor-adapter__read_report, Read, Grep, Glob, Bash
 ---
 

package/templates/claude/agents/trellis-pi-implement.md

@@ -4,6 +4,7 @@ description: >-
   Dispatch mechanical, well-specified implementation tasks to the Pi executor via
   the executor-adapter MCP only when provider routing outside Codex is explicitly
   needed. For normal GPT implementation, use trellis-codex-implement instead.
+effort: xhigh
 tools: mcp__executor-adapter__dispatch, mcp__executor-adapter__preview_prompt, mcp__executor-adapter__read_report, mcp__executor-adapter__smoke, Read, Grep, Glob, Bash
 ---