---

kind: infra workflow_fix_target: .claude/rules/context-hygiene.md fingerprint: codex-composer-roles-have-no-lean-twin-autocompact-ladder-terminates

Autocompact-thrash ladder has no lean twin for the codex-* composer roles

Goal

Close the escalation gap in .claude/rules/context-hygiene.md's autocompact-thrash ladder (Class 2, fixed-overhead subagent-window pressure) for the codex-* prompt-composer roles, which currently have no lean twin and therefore fall off the ladder one rung early.

The gap

context-hygiene.md's Class-2 recipe says: respawn micro-scoped on the default model; if that ALSO thrashes, "escalate ONCE (same v<n>, no counter increment) to the role's LEAN TWIN with the same brief", and it enumerates six: analyzer-lean, planner-lean, critic-lean, experiment-implementer-lean, code-reviewer-lean, consistency-checker-lean.

There is no lean twin for ANY codex-* role — not codex-code-reviewer, codex-critic, codex-interpretation-critic, codex-clean-result-critic, or codex-follow-up-critic. So when a Codex composer thrashes, the ladder runs out after the micro-scoped default-model respawn, and the only remaining moves are the fail-loud terminal or refusal rung (c)'s inline composition by the orchestrator.

Why it matters

Inline composition is sanctioned but expensive in a specific way: it breaks the review site's independence guarantee, which is why it REQUIRES the [epm-inline-fallback] role=<role> round=<n> reason=<one-line> progress marker (#2062). Every Codex composer thrash that reaches this point converts a doubled review site into a single-Claude one. The lean-twin rung exists precisely to avoid paying that price, and the composer roles — the ones most likely to thrash, since a composer inlines a full lens reference or SPEC into its prompt — cannot reach it.

The composers are also the CHEAPEST roles to lean-twin: they are thin prompt-composers whose whole job is Read + Write + return a path. They need no MCP tools at all, and an all-tools spawn loads ~138K tokens of MCP schemas (measured 2026-07-03) that a composer never uses.

Observed

Two Codex composers in one /issue 2291 Step 5 round died degenerate at ~135k tokens with ZERO tool calls — the Class-2 signature (no oversized tool result to bound; fixed overhead alone exhausted the window). The round recovered because the Claude reviewer's verdict stood and a later composer attempt succeeded, so no inline fallback was needed — but the ladder had no rung left to offer.

Scope

• Add codex-*-lean twins for the five composer roles, following the existing lean-twin pattern: restricted tool list (Bash, Read, Write, Edit, Grep, Glob — no mcp__*, no skills:), deferring to the full composer spec by reference rather than restating it.
• Resolve them user-global in ~/.claude/agents/ with repo symlinks, matching the #2072 placement rule (agent types register at session start from the session cwd + user-global dirs, never mid-session).
• Extend context-hygiene.md's lean-twin roster to name them, so the ladder text and the available agent set agree.
• Cross-check .claude/rules/codex-ensemble-review.md: its composer-death handling should point at the new rung before the no-show fallback.

Non-goals

• Do NOT lean-twin the Codex DISPATCH path (scripts/codex_task.py) — that is the orchestrator's bg-Bash call, not a subagent, and has no window of its own.
• Do NOT pin a smaller model as the fix. That is the INVERSE of refusal rung (b2) and is explicitly banned for thrash (#1090: default-model spawns compacted fine while 3/6 sonnet spawns thrashed).

Provenance: surfaced during /issue 2291 Step 5 (two degenerate Codex composer spawns); held during the round to avoid a snapshot-drift race with the live reviewers, filed at Step 10 close. Prose follow-up per .claude/rules/workflow-fix-on-bug.md.