estack-drive-cli-agent: document exact per-run model + reasoning/effort controls

[ElliotDrel]

Problem

skills/estack-drive-cli-agent already documents how to launch Codex CLI and Claude Code non-interactively, and it already documents explicit model selection for both providers. It does not currently document the exact per-run controls for reasoning/effort.

That leaves delegated sessions nondeterministic when the parent task specifies both a model and a reasoning level, for example:

• "Run Codex with gpt-5.6-sol at xhigh reasoning"
• "Run Claude Code with a specific model at high effort"

The delegation skill should teach agents to pass both dimensions explicitly rather than inherit global defaults.

Verified current syntax

Codex CLI

Codex uses -m / --model for the model and a per-run config override for reasoning effort:

codex exec \
  -m "<model-id>" \
  -c model_reasoning_effort=xhigh \
  --skip-git-repo-check -a never \
  "<prompt>" < /dev/null

Equivalent persistent config is:

model = "<model-id>"
model_reasoning_effort = "xhigh"

For delegation, prefer the per-run -c model_reasoning_effort=<level> override so the spawned session is deterministic and does not depend on or mutate global config.

OpenAI has explicitly documented/confirmed the codex exec -c model_reasoning_effort=<level> pattern in the Codex CLI issue tracker and CLI documentation references. Supported reasoning levels are model-dependent; do not hard-code one universal set for every Codex model. For example, current GPT-5.6 models support none, low, medium, high, xhigh, and max, while older Codex-specialized models may expose a different subset.

Claude Code

Claude Code has a first-class --effort flag for a single session:

claude -p "<prompt>" \
  --model "<model-id-or-alias>" \
  --effort xhigh \
  --output-format json \
  --allowedTools "Read,Grep,Glob" \
  --permission-mode dontAsk

Anthropic's current CLI reference documents:

--effort <level>

with session-level options low, medium, high, xhigh, and max; available levels depend on the selected model. --effort overrides the configured effortLevel for that session and does not persist. max is session-only rather than a persistent settings value.

Requested change

Update estack-drive-cli-agent and both reference files so an orchestrating agent knows how to set both dimensions explicitly:

1. exact model
2. exact reasoning/effort level

Specifically:

• Add Codex -c model_reasoning_effort=<level> to the main invocation patterns and references/codex-exec.md.
• Add Claude --effort <level> to the main invocation patterns and references/claude-headless.md.
• Show at least one combined model + reasoning/effort example for each CLI.
• State clearly that supported effort values depend on the selected model/provider.
• Prefer per-run overrides over changing global config for delegated work.
• Add a rule: when the caller specifies a model or reasoning/effort level, the delegated CLI command must pass it explicitly rather than silently inherit defaults.
• Re-verify exact supported values against current official provider docs whenever this skill is updated, because both CLIs and model catalogs move quickly.

Acceptance criteria

• A reader can launch Codex with an exact model and exact reasoning effort from one command.
• A reader can launch Claude Code with an exact model and exact effort from one command.
• The main SKILL.md examples expose these controls rather than hiding them only in references.
• references/codex-exec.md documents model_reasoning_effort as a per-run -c config override.
• references/claude-headless.md documents --effort and notes that available levels are model-dependent.
• The skill tells delegating agents not to silently inherit model/reasoning defaults when the caller specified them.

Sources verified 2026-08-08

• OpenAI Codex CLI usage/maintainer confirmation: codex exec -c model_reasoning_effort=<level> ...
• OpenAI current model documentation for model-specific reasoning levels
• Anthropic Claude Code CLI reference: --model and --effort
• Anthropic Claude Code model configuration docs: effort precedence and model-dependent availability