fix: enforce max_steps on codex/claudecode/hermes harnesses

[Wanli-Lee]

Summary

Only the openclaw harness actually enforced max_steps (via a step-cap watchdog). codex/claudecode/hermes accepted the parameter but never used it — codex's "translated to model_max_turns" comment was never implemented, hermes's --effort path doesn't exist in 0.14.0, and claudecode hardcoded MAX_STEPS_REACHED=0. Those three runtimes were bounded only by the wall-clock timeout, breaking cross-harness comparability.

• claudecode: pass the native --max-turns N flag on both the first-turn and resume invocations; detect the error_max_turns result subtype to emit a real MAX_STEPS_REACHED marker.
• codex: background step-cap watchdog (parity with openclaw) counting "type":"function_call" events across rollout.jsonl, pkills codex at max_steps * (1 + max_refusal_retries).
• hermes: same watchdog, counting "role":"assistant" in the incrementally-written session_*.json — the same unit openclaw counts.

All three guard the retry loop with [ -f $STEPS_CAPPED ] && break so a watchdog kill isn't misread as a transient failure and retried. A step-capped task keeps its deliverables and is still judged (parity with openclaw): hitting the cap is a real "model looping" signal, not a fake zero.

Verification (done on real VMs, not just unit tests)

• claudecode --max-turns: verified accepted by claude 2.1.76 in-VM. It's undocumented in --help but functional — a real flag returns "Not logged in" (parsed OK) while a fake flag returns "error: unknown option".
• hermes counting: verified in-VM that hermes writes session_*.json (NOT request_dump_*.json — an earlier assumption, disproven and corrected). The "role":"assistant" grep count matched the JSON-parsed assistant count exactly (29 == 29) on a live session.
• codex counting: find … -exec grep "type":"function_call" cross-file accumulation + empty-dir/no-file edge cases tested.

Test plan

• pytest tests/ — 73 passed (added tests/test_harness_max_steps.py: cap formula, --max-turns injection on both branches, function_call/assistant counting, anti-re-retry break, step markers)
• bash -n on all three rendered runner scripts
• claude step-decision python snippet unit-checked against real stream-json result events (error_max_turns / normal / boundary / empty / malformed)

🤖 Generated with Claude Code