ROADMAP-19 — P2 — AGENTS.md compliance cleanups (os.getenv, silent suppress, hardcoded retry-after, arg leakage) #111

Series: ROADMAP — filed from the 2026-10-27 whole-project assessment · Backlog task: M16-T06 · Priority: P2 · Effort: small

Problem

Four places drift from the project's own AGENTS.md rules:

1. openai.py reads os.getenv("OPENAI_API_KEY") directly, bypassing agentkit/config.py (rule: all configuration via Settings)
2. Trace recording is wrapped in contextlib.suppress(Exception) with no log — swallow-without-logging contradicts the "never swallow silently" rule
3. retry_after=60 hardcoded in the OpenAI 429 mapping instead of a settings member
4. failure_reason embeds raw tool arguments (agent.py:358) — prompt-ish content leaks into a client-visible field, while the codebase elsewhere keeps goals out of INFO logs

Evidence

• agentkit/llm/openai.py:37; agentkit/openai mapping :129
• agentkit/core/agent.py:394–395 (suppress), :358 (args in failure_reason)

Suggested fix / acceptance criteria

• Both adapters build from Settings only; factory passes explicit keys
• Trace failure logged (WARNING, run_id/step_no) then continues
• RETRY_AFTER_FALLBACK_S (or equivalent) setting used
• failure_reason carries tool name + error category, not argument text (test pins this)

Suggested branch / commit

fix/agents-md-deviations · fix: align agent/llm code paths with AGENTS.md rules