fix(truncation): honor configured output cap on truncated tool-call retry

[hashbender]

Summary

• Preserve custom provider max_output_tokens/max_tokens in config normalization and provider conversion
• Pass runtime output caps through CLI agent setup and route signatures
• Retry truncated tool-call responses using the configured/requested cap instead of the 4k/8k ladder
• Persist a synthetic assistant marker when truncated tool-call retries are exhausted

Validation

• Live Hermes session on custom:hyperspace-responses / gpt-5.5 / codex_responses wrote a 16,247-byte Day 4 briefing validation file via write_file on first try; no truncation ladder or Response truncated failure
• bash scripts/run_tests.sh tests/test_output_truncation_runtime.py
• bash scripts/run_tests.sh tests/test_ctx_halving_fix.py tests/cli/test_cli_provider_resolution.py -q

---

Mirror-of: NousResearch#56619
NousResearch#56619

[tenki-reviewer]

Review Complete

Risk: 🟠 High (62/100) — 1 medium finding · 182 LOC across 4 files

Files Reviewed: 4
Findings: 1

By Severity:

• 🟡 Medium: 1

---

Output truncation retry infrastructure changes with two issues: missing camelCase aliases for new config keys, and oneshot mode bypassing configured output caps.

Files Reviewed (4 files)

agent/conversation_loop.py
hermes_cli/cli_agent_setup_mixin.py
hermes_cli/config.py
tests/test_output_truncation_runtime.py