v3.1.0 - Document spawn_agent Model Limitation

v3.1.0

Known Limitation Documented

• spawn_agent only supports built-in Codex models: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.2
• Custom models from Codex++ proxy (e.g., mimo-v2.5) cannot be used as model overrides
• This is a Codex app limitation, not a skill issue
• Codex++ has a modelWhitelistUnlock feature that patches Statsig config and App Server requests, but it may not cover spawn_agent model validation

Current Behavior

• All subagents inherit parent model (no model override)
• Step 0 detection is for capability awareness only (whether to attempt visual checks)
• Step 0 runs once per conversation, result is cached

Installation

git clone https://github.com/bauerelizabeth07139/plan-guardian-skill.git cp -r plan-guardian-skill ~/.codex/skills/.system/plan-guardian cp plan-guardian-skill/AGENTS.global.md ~/.codex/AGENTS.md

Files Changed

• SKILL.md: Added known limitation note
• README.md: Added Known Limitations section, version 3.1.0

v3.0.0 - Full Rewrite: Workers vs Verifiers, No Model Override

v3.0.0 - Breaking Change

Core Fix

• All subagents inherit parent model. Removed ALL model override logic (model="<detected_model>" does not work with spawn_agent).
• Workers and verifiers are now clearly separated:
  • Workers: receive task + deliverable only. Do implementation.
  • Verifiers: receive artifact + acceptance criteria only. Check the work. Memoryless.
• Step 0 runs once per conversation. Result cached for reuse. Capability awareness only (whether to attempt visual checks), NOT model selection.
• Subagents never run 7-step plan. Removed all model override templates that caused confusion.

Files Rewritten

• SKILL.md: 66% rewritten - clean 7-step workflow, no model override
• AGENTS.global.md: 90% rewritten - 7 rules, all inheriting parent model
• AGENTS.md: 82% rewritten - clear worker vs verifier policy
• README.md: Updated model selection, subagent roles

What Changed

Before	After
model="<mimo-v2.5>" in spawn_agent	No model param (inherit parent)
Workers got acceptance criteria	Workers get task + deliverable only
Verifiers got task history	Verifiers get artifact + criteria only
Step 0 on every message	Step 0 once per conversation
Subagents ran 7-step plan	Subagents do assigned task directly

Credential Resolution (unchanged)

1. Command-line arguments
2. Environment variables
3. ~/.mimo2codex/.env
4. ~/.codex/config.toml + ~/.codex/auth.json

v2.12.0 - Step 0 Runs Once Per Conversation

v2.12.0

Change

• Step 0 (model detection) runs ONLY on the first message of a conversation
• Result is cached and reused for all subsequent messages
• No redundant API probing on follow-up messages

Behavior

• First message: run detect_multimodal.py, get full model inventory, cache result
• Subsequent messages: skip Step 0, use cached selected model and status directly
• This applies to the main planning agent; subagents never run Step 0

Files Changed

• SKILL.md: Step 0 header updated with "ONCE per conversation" + caching rule
• SKILL.md: Trigger Rule updated to mention caching
• AGENTS.global.md: RULE 6 Step 0 updated with caching instruction

v2.11.0 - Full Model Inventory with Multimodal Status

v2.11.0

Change

• Step 0 now probes ALL available models and returns a complete inventory with multimodal status for each
• Previously only returned the first multimodal model found

Output Format

json { "mode": "auto", "used_base": "http://127.0.0.1:57321/v1", "models": [ {"id": "mimo-v2.5", "multimodal": true}, {"id": "mimo-v2.5-pro", "multimodal": false}, {"id": "mimo-v2.5-asr", "multimodal": false} ], "selected": "mimo-v2.5", "status": "MULTIMODAL" }

Usage

• selected: use this model for visual workers and verifiers
• models[]: full inventory showing which models support image input
• status: overall MULTIMODAL / NOT_MULTIMODAL / UNKNOWN

v2.10.0 - Fix Model Selection for Visual Workers and Verifiers

v2.10.0

Fix

• Step 4: Visual workers (UI, frontend, beautification) now correctly use model="" when Step 0 = MULTIMODAL
• Step 5: All verifiers now use model="" when Step 0 = MULTIMODAL
• Step 0: Agent must run detect_multimodal.py script directly, no tool/skill guessing
• Subagents: All spawn_agent calls include "Do NOT run 7-step plan-guardian workflow" instruction

Model Selection Summary

Subagent Type	Step 0 = MULTIMODAL	Step 0 = NOT_MULTIMODAL / UNKNOWN
Step 0 detection	Parent model	Parent model
Text/code worker	Inherit parent	Inherit parent
Visual worker	Detected multimodal model	Inherit parent
All verifiers	Detected multimodal model	Inherit parent

Credential Resolution

1. Command-line arguments
2. Environment variables
3. ~/.mimo2codex/.env
4. ~/.codex/config.toml + ~/.codex/auth.json

v2.9.0 - Step 0 Must Run Script Directly

v2.9.0

Fix

• Step 0 simplified: Agent must run detect_multimodal.py via exec_command or spawn_agent
• No more tool/skill guessing: Removed ability to declare MULTIMODAL based on environment inspection
• Script probes the actual model API with a real image payload

What Changed

• SKILL.md: Step 0 now shows a simple bash command instead of a complex spawn_agent template
• SKILL.md: CRITICAL warning says "Do NOT declare based on tools/skills/environment"
• AGENTS.global.md: RULE 6 Step 0 updated to match

Credential Resolution (unchanged)

1. Command-line arguments
2. Environment variables
3. ~/.mimo2codex/.env
4. ~/.codex/config.toml + ~/.codex/auth.json

v2.8.1 - Step 0 Must Run Detection Script

v2.8.1

Fix

• Step 0 now requires running detect_multimodal.py - Agent can no longer declare MULTIMODAL based on tool/skill availability alone
• The script probes the actual model API to verify image input support

v2.8.0 Changes (included)

• detect_multimodal.py auto-reads ~/.codex/auth.json and ~/.codex/config.toml
• Step 0 detection subagent uses parent model (no override)
• Visual/verification subagents use multimodal model when detected
• Subagents no longer run the 7-step plan-guardian workflow
• Codex++ dynamic port support via config.toml

Credential Resolution Order

1. Command-line arguments
2. Environment variables (OPENAI_BASE_URL, OPENAI_API_KEY, MIMO_BASE_URL, MIMO_API_KEY)
3. ~/.mimo2codex/.env
4. ~/.codex/config.toml + ~/.codex/auth.json

v2.8.0 - Codex Config Auto-Detection

Changes

• detect_multimodal.py: Auto-reads ~/.codex/auth.json and ~/.codex/config.toml for credentials
• Step 0 detection subagent: Always uses parent model (no override)
• Visual/verification subagents: Use multimodal model when detected (MULTIMODAL)
• Subagents: No longer run the 7-step plan-guardian workflow
• Codex++ proxy: Dynamic port support via config.toml
• Model selection: Text/code workers inherit parent; visual tasks use detected multimodal model when available

Credential Resolution Order

1. Command-line arguments
2. Environment variables
3. ~/.mimo2codex/.env
4. ~/.codex/config.toml + ~/.codex/auth.json (NEW)

v2.7.0

Changes

• Remove provider-specific assumptions from skill and documentation.
• Keep Codex++ integration provider-agnostic: environment/local proxy first, model-specific endpoint validation when known, plus Codex++ key-file fallback.
• Update README/SKILL wording to avoid tying detection logic to a single provider.

v2.6.0

Changes

• Auto-read Codex++ key from ~/.mimo2codex/.env when environment variables are missing.
• Keep Codex++ proxy default and dual-endpoint validation (environment then model).
• Update documentation to reflect Codex++ key-file fallback.