| name | spawn-codex |
| description | Spawn project-local Codex CLI instances from an agent, pinned to GPT-5.5, with explicit reasoning effort and speed controls. |
Spawn Codex Instances
Use this skill when an agent needs to launch another Codex CLI instance for a bounded task in this repository.
Command
Use codex exec for non-interactive agent runs:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
"inspect the selected area and report findings"
Use stdin when the prompt is easier to compose as a block:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
-
Then pass the full prompt on stdin.
Permissions
Default to read-only for exploration and review:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
--sandbox read-only \
"review this branch for correctness risks"
Allow workspace edits only when the child instance is explicitly expected to modify files:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
--sandbox workspace-write \
"make the smallest code change needed to fix the described issue"
Do not use danger-full-access for spawned instances unless the parent task explicitly requires broad filesystem access and the environment is externally controlled.
Model
Always use GPT-5.5 for spawned Codex instances in this project:
-m gpt-5.5
Do not choose mini models for spawned instances in this repository.
Reasoning Effort
Set reasoning effort with model_reasoning_effort.
Use medium for straightforward implementation, focused review, and summarization:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
-c model_reasoning_effort='"medium"' \
"summarize the rendering pipeline"
Use high for debugging, geometry changes, code review where false negatives are costly, or tasks spanning multiple modules:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
-c model_reasoning_effort='"high"' \
"find the likely cause of this road-geometry rendering bug"
Use xhigh only for hard design/debugging tasks where latency and token cost are acceptable:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
-c model_reasoning_effort='"xhigh"' \
"analyze transition morph geometry and propose a minimal robust fix"
Supported effort values are minimal, low, medium, high, and xhigh. Prefer medium or high unless the prompt gives a reason to do otherwise.
Speed
Set speed tier with service_tier.
Use standard/default speed by omitting service_tier.
Use Fast mode when the task is latency-sensitive and the extra credit cost is acceptable:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
-c service_tier='"fast"' \
-c model_reasoning_effort='"medium"' \
"quickly inspect the changed files and list likely issues"
Use Flex when lower-priority/background execution is acceptable:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
-c service_tier='"flex"' \
-c model_reasoning_effort='"high"' \
"perform a deeper background review of road geometry edge cases"
Do not use service_tier = "priority"; this CLI expects fast or flex.
Reusable Custom Agent
For reusable spawned agents, create TOML files under .codex/agents/ or ~/.codex/agents/.
Example:
name = "citynista-reviewer"
description = "GPT-5.5 reviewer for Citynista correctness, geometry, and test-risk checks."
model = "gpt-5.5"
model_reasoning_effort = "high"
service_tier = "fast"
sandbox_mode = "read-only"
developer_instructions = """
Review code like an owner.
Prioritize correctness bugs, behavior regressions, rendering geometry risks, and missing tests.
Use Citynista's AGENTS.md instructions.
Do not suggest broad refactors unless required to fix the issue.
"""
Then ask the parent Codex session to spawn it:
Spawn a citynista-reviewer agent to inspect the renderer changes, wait for it, and summarize its findings.
Output
For scriptable runs, add --json to stream JSONL events:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
--json \
"review the current diff"
For a final response file:
codex exec \
-C ~/projects/citynista \
-m gpt-5.5 \
--output-last-message /tmp/citynista-codex-result.md \
"summarize the current diff"