Use native compaction and refine GPT workflows

[ItayCohen-Prog]

What changed

• replaces custom context-pressure pruning with Claude Code native auto-compaction at a fixed 270,000-token window
• removes GPT-5.3-Codex-Spark from the native model picker because its 128,000-token window is incompatible
• preserves native model, effort, subagent, and Workflow behavior through Codex OAuth
• adds downstream heartbeats and preserves nested upstream error messages
• documents context behavior and fallback model metadata

Why

Claude Code assumed a much larger context window than the GPT backends, allowing Workflow agents to exceed the backend limit. The adapter should inform Claude Code's native compactor instead of rewriting or deleting conversation history.

Verification

• 83/83 tests passed
• package contents verified
• fresh global-install smoke test passed
• gptcode doctor confirmed Codex OAuth, isolation, Claude Code compatibility, and the 270,000-token native compaction window
• live Codex model discovery confirmed Spark is omitted and all exposed models report 272,000-token windows

Secret protection

• full reachable Git history scanned locally with Gitleaks 8.30.1: zero findings
• GitHub secret-scanning alerts: zero
• GitHub provider-secret scanning and push protection enabled
• pinned Gitleaks workflow runs on every push and pull request
• protected main requires the Gitleaks check through a pull request; force-pushes and deletion are disabled