feat(codex): add manual native compaction for OAuth

[cucurigoo]

Summary

• add CLI-only /native-compact (/compact-native) for eligible ChatGPT Codex OAuth sessions
• call OpenAI’s standalone responses.compact endpoint immediately rather than waiting for the automatic threshold
• keep Hermes’ complete readable transcript; persist the encrypted checkpoint only as a route-scoped replay projection

Design

This deliberately differs from the rejected Responses half of #54633 / #33130. Current main already supports opt-in automatic native checkpoints; this command reuses that existing projection and cross-issuer drop path. The portable local history stays authoritative, so provider/model switching simply drops the checkpoint instead of losing the compacted prefix.

The command is intentionally CLI-only and requires the existing native gate: official Codex OAuth route, gpt-5.6, compression.enabled: true, and compression.codex_responses_native: true. It does not change automatic compaction behavior. Repeated manual compaction replaces the prior checkpoint on its retained carrier, preserving the carrier’s encrypted reasoning after the new checkpoint.

Verification

• scripts/run_tests.sh tests/agent/test_manual_native_compaction.py tests/cli/test_native_compact.py tests/cli/test_slash_dispatch_table.py tests/run_agent/test_native_compaction.py — 82 passed
• ruff check on all changed Python files — passed
• GitHub CI passed: Python tests/lints, docs, Windows/macOS-only tests, Nix, Docker build, and supply-chain checks

Related

• [Feature]: Use OpenAI Codex Compaction API instead of internal Compress for context compaction. #33130
• feat(agent): add Codex-native compaction paths #54633
• feat(compression): native OpenAI Responses server-side compaction for gpt-5.6 #81747

[Enough1122]

AI code review — automated review for reference; please use your judgment.

Summary

Adds manual native compaction for Codex OAuth sessions (agent/native_compaction.py::manual_native_responses_compaction): issues an immediate responses.compact call over the message prefix and stores the returned checkpoint on an existing retained assistant turn (sidecar), leaving the readable local transcript untouched. Strict eligibility gates (Responses transport, Codex backend, model/config support, SDK capability, 401-refresh retry).

Findings

No blocking issues.

• Transcript safety: _manual_native_compaction_carrier requires a completed assistant turn strictly before the protected tail and reuses it as carrier rather than synthesizing a turn — preserves role alternation. Failure modes raise RuntimeError with actionable messages instead of half-applying. Good.
• Request hardening parity: the compact call rebuilds api_kwargs through the same sanitize + preflight_kwargs path as ordinary Codex requests (surrogate/ASCII/Harmony handling), and only forwards a fixed allowlist of keys (model, input, instructions, prompt_cache_key, timeout, extra_headers). Thoughtful — no prompt-injection or credential widening via the new call.
• Non-blocking: confirm the sidecar checkpoint is actually dropped on provider/model switch as the docstring claims (projection logic in the Responses adapter isn't in this diff); if a stale checkpoint ever attached to a non-Codex request, the provider would likely reject unknown items — fail-visible, but worth a test if not already covered elsewhere.

Verdict

Careful implementation. Non-blocking note on switch-drop coverage.