Skip to content

README.md

Latest commit

 

History

History
91 lines (76 loc) · 5.13 KB

README.md

File metadata and controls

91 lines (76 loc) · 5.13 KB
Raw

Native Codex gate

The gate runs the actual Codex 0.160.0 client against a local scripted Responses provider. Each client receives a private HOME and CODEX_HOME, a fake API key, and an explicit provider configuration. No production provider is contacted.

export SESSIONKIT_CODEX_BIN="$(./scripts/install-codex.sh)"
make native-gate

Go, curl, tar, a SHA-256 utility, and the sqlite3 CLI must be installed. The installer fetches the native payload from the exact platform package of @openai/codex@0.160.0 into a new scratch prefix. It verifies both the npm archive and the extracted binary before printing the binary path. An optional first argument selects a new, nonexistent prefix. Keep or delete that prefix after use. The gate independently hashes the binary. Both an absent binary and a mismatched digest are failures, including when go test -tags=native is invoked directly.

The scripted tool call only prints a fixed nonce. The gate uses --dangerously-bypass-approvals-and-sandbox for exec, and explicit never / danger-full-access app-server settings. These settings prove rollout format compatibility; they do not prove sandbox enforcement or authorization policy.

The handoff test captures source A, installs B, verifies the effective app-server resume response, continues B, restarts with codex exec resume <uuid>, captures B, and installs and resumes C beside workspace A. It checks the original role order and multiplicity, tool name, call ID, arguments and output. It also checks unchanged rollout prefixes, increasing ordinals, appended workspace and permission settings, and native creation of the paginated SQLite row. SQLite is opened only by the read-only CLI in this test, never by the library.

exec resume cannot return the effective configuration or accept an explicit runtimeWorkspaceRoots parameter. The first destination resume therefore uses app-server with those roots, cwd, provider, approvals reviewer, approval policy, and sandbox passed explicitly, then checks the response. The later restart and return exercise the explicit UUID CLI command with destination configuration.

The original A is resolved with native thread/read, checking its exact path, preview, and paginated profile while its digest remains unchanged. Sending a new prompt through exec resume in A would append to it. Paginated thread/read does not hydrate model history or materialize item projection on its own.

Separate cases resume a complete native compaction, reject a legacy rollout before contacting the client, and try to capture while a live native writer is blocked at the local provider. The compaction request must contain the replacement summary and must omit the original tool pair and assistant reply.

TestNativeInterruptedHandoff blocks a real turn at that provider, sends turn/interrupt with the exact thread and turn IDs, and checks the native interrupted completion and matching turn_aborted record with reason interrupted. After native unload and process exit, it captures the stopped source, opens the bundle, installs into a fresh home, and resumes through app-server. The actual resumed provider request must contain the original user, tool call, tool output, assistant reply, interrupted user input, and new input exactly once and in order. The undelivered scripted reply must be absent. The test also verifies effective destination settings, appended ordinals and reply, SQLite backfill, and unchanged source bytes and mtime.

TestNativeKilledTurnRejected waits until the blocked turn's task_started record is persisted, then sends SIGKILL without interrupt or unload. The stopped rollout must have no matching completion or abort. Capture must fail closed with in_flight_turn at that start ordinal, leave no bundle, preserve source bytes and mtime, and make no further provider request. It does not repair or resume an unfinished turn. These cases generate records with the pinned client at runtime; no fixture is edited and no supported format or recovery rule changes. Other accepted abort reasons remain covered by parser tests, not by this native gate.

Run just these cases with the same digest gate:

SESSIONKIT_CODEX_BIN="$(./scripts/install-codex.sh)" \
  go test -count=1 -tags=native -timeout=3m ./native/codex \
  -run 'TestNative(InterruptedHandoff|KilledTurnRejected)$' -v

The client source is pinned to a956835d020762cb2b570053af06f643a11c0ecc. The SSE events follow codex-rs/core/tests/common/responses.rs; app-server parameters follow codex-rs/app-server-protocol/src/protocol/v2/thread.rs and turn.rs.

The npm artifacts were downloaded and hashed on 2026-10-02. These are binary digests; the corresponding archive digests are also pinned in the installer.

Platform Binary SHA-256
Linux x64 12eb3e81114588aca3b7998f4f19e8997b056aca08e57a7ca7c8a3ec8c652aad
Linux arm64 50b06603bdcdac39b714f5c3e68583c002b8ad8779ebfdaaf4932ff016b379c0
macOS x64 5383ef71dd1bd8d2f3658c04a219e2cf165c7969aebd0cceced1bc9f0f68877f
macOS arm64 112fae7a5a1223e673c8a1791d32338f37df8b527ff1159bb8adac6c4dbf1b4b