anti-hall-context-conserve

name	anti-hall-context-conserve
description	Codex-native context and usage conservation mode. Use when the user asks for context conservative mode, limit conservation, cheap-model routing, or reducing context/token burn while keeping work moving.

anti-hall context-conserve for Codex

This is the Codex port of anti-hall limit/context conservation. It combines hook nudges, model routing, and output hygiene; it does not depend on Claude Workflow JS.

Resolve the plugin root

Codex does not expand ${PLUGIN_ROOT} inside a skill's own instructions — that variable is only set for plugin-bundled hook commands (see docs/KB-codex-platform-hooks-plugins.md). Codex does show you this skill's own file path when it selects the skill ("Codex starts with each skill's name, description, and file path" — official Codex Skills doc). Resolve the plugin root from that path before running anything below:

# SKILL_FILE = the absolute path Codex showed you for this SKILL.md.
ANTI_HALL_ROOT="$(cd "$(dirname "$SKILL_FILE")/../../.." && pwd)"
test -f "$ANTI_HALL_ROOT/.codex-plugin/plugin.json" || { echo "anti-hall plugin root not found relative to $SKILL_FILE — aborting" >&2; exit 1; }

Activation

Project-local hooks are installed by anti-hall-activate:

node "$ANTI_HALL_ROOT/codex/install-codex.js"

The installer registers hooks/limit-conserve-inject.js on UserPromptSubmit so conservation instructions are injected when active.

Manual controls:

export ANTIHALL_LIMIT_CONSERVE=on    # force conservation
export ANTIHALL_LIMIT_CONSERVE=off   # disable conservation
export ANTIHALL_LIMIT_THRESHOLD=85   # default threshold

Status check:

node -e "const {isConserving}=require('$ANTI_HALL_ROOT/hooks/limit-conserve.js'); console.log(JSON.stringify(isConserving(), null, 2))"

Main-model downshift (flagship preservation)

When conservation is active and the main/coordinator agent is on gpt-5.6-sol, switch it to gpt-5.6-terra (1M-context variant only) to preserve the flagship weekly bucket.

• Target: gpt-5.6-terra at 1M context — same context window, lower cost.
• NEVER downshift to gpt-5.4-mini (only 400k context; context loss is unacceptable).
• Keep gpt-5.6-sol for delegated hard-reasoning subagents and on-demand escalation.
• Codex cannot self-switch its own model — surface this recommendation to the user or let OMX/orchestration layer set it.
• Long-context cost premium: both gpt-5.6-terra and gpt-5.6-sol incur a confirmed 2× input / 1.5× output multiplier once a request exceeds ~272K input tokens — this applies regardless of which of the two you're on. For a very-large-context task (a big repo dump, a huge diff), consider capping scope or routing to a Claude model instead (no equivalent premium up to 1M there). See docs/KB-token-usage-models.md §2/§7.

Codex routing policy

When conservation is active:

• keep the coordinator concise; do not paste broad command output into the main thread
• use explore / gpt-5.4-mini for codebase lookup (gpt-5.6-luna when 5.6-era capability/cutoff matters; gpt-5.3-codex-spark is a distinct, faster/less-capable model, ChatGPT Pro only)
• use gpt-5.4-mini (default) for mechanical command running
• use gpt-5.6-terra for settled implementation
• reserve gpt-5.6-sol for planning, root-cause, validation, code review, and debate gates
• if a model is unavailable or rate-limited, record the limitation and choose the nearest safe fallback; do not retry-loop

Context hygiene rules

• Query graphify before broad repo search when graphify-out/ exists.
• Delegate noisy tests/builds/checks and require compact summaries.
• Report only changed files, validation evidence, and remaining blockers.
• Do not claim quantitative token savings without a measured benchmark.

What is not hard-enforced in Codex

Codex does not expose Claude's Agent/Task PreToolUse hook surface for hard model-routing blocks. The Codex port enforces this via skill instructions, AGENTS.md, and UserPromptSubmit nudges instead of pretending it can block every spawn.