UltronCore Adaptive Codex Utilization
Claude thinks. Codex executes. Choose automatically based on task type and usage pressure.
Routing Decision
Is this reasoning/judgment/architecture?
YES → Claude handles it
Is this execution/mechanical/terminal-based?
YES → use Codex
Mixed?
Claude plans, Codex executes
Usage Pressure Mode
| Pressure | Behavior |
|---|---|
| Low (early in cycle) | Claude quality first; Codex for clearly mechanical work |
| Moderate (mid-cycle) | Balance — offload execution-heavy tasks to Codex |
| High (near limit) | Aggressively route to Codex; Claude for reasoning only |
| Reset imminent | Best quality via Claude again |
Do not ask the user which mode. Infer from context (session length, explicit mentions, prior messages).
Codex Task Types
Route to Codex when the task is:
- Terminal commands, shell scripts
- File read/write/move/organize (including Obsidian vault file ops)
- Repo scanning, grep, directory listing
- Repetitive or batch code edits
- Dependency installs, env setup
- Running tests, lint, typecheck
- Log parsing, structured transformations
- Debugging implementation issues (not architectural ones)
Claude Task Types
Keep in Claude when:
- Architecture or design decisions
- Ambiguity is high or instructions are unclear
- Output requires judgment, nuance, or synthesis
- Planning what Codex should do next
Codex Execution Template
When routing to Codex, define clearly:
Task: [exact action]
Goal: [expected result]
Scope: [specific files/directories]
Constraints:
- no unnecessary dependencies
- no unrelated edits
- preserve architecture and style
- no secret exposure
Checks: [tests/lint/typecheck if available]
Then invoke:
codex "<task description>"
# or with scope:
codex -d /path/to/scope "<task description>"
Validation
After Codex runs:
- Confirm result matches goal
- Check for obvious breakage
- Return concise summary: what was done, outcome, any issues
Failover
Codex fails → retry once with tighter scope → if still failing, resolve in Claude → return to Codex for remaining execution.
Efficiency Rules
- Don't use Codex for trivial one-liners (Claude handles inline)
- Don't use Claude for repetitive mechanical execution (Codex wins)
- Keep outputs concise — no unnecessary explanation in either tool
~/.agents/skills/is Codex's skill directory — this skill is installed there too