enh: add feature-flagged server-side compaction by cooper-oai · Pull Request #13978 · openai/codex

cooper-oai · 2026-03-08T16:12:20Z

summary

add a default-off server_side_compaction flag for openai responses auto-compaction
send context_management on eligible /responses requests instead of auto-calling the legacy /compact path
apply streamed compaction checkpoints locally so follow-up turns reuse the inline summary without losing current-turn state
keep manual compaction and non-openai providers on the existing client-side flow

implementation notes

this only changes auto-compaction on openai-compatible responses requests
compact_prompt is ignored for inline auto-compaction. this is not a regression: codex already ignores it for openai auto-compaction today, and manual /compact still uses the existing prompt-based endpoint
codex buffers streamed compaction items and rewrites history on response.completed so later streamed items keep their order and retries do not commit partial checkpoints
when inline server-side compaction is available, we skip previous-model preflight compaction so openai auto-compaction uses a single path

rollout

land behind server_side_compaction, default off
validate first against internal openai responses upstreams and app-server flows
watch codex.compaction and codex.compaction_checkpoint_applied
if stable, enable for a small internal cohort; manual compaction and non-openai providers continue on the existing flow

testing

ran unit tests
tested with the local binary pointed at an openai responses api upstream

cooper-oai · 2026-03-08T18:31:09Z

Known fast-follow: while testing this branch against a Responses API upstream configured via env_key, I hit a pre-existing request-compression/auth-gating bug that already exists on main.

Today, request compression is enabled based on cached ChatGPT auth state plus provider.is_openai(), while the actual outgoing bearer token can come from provider.env_key. On some Responses API upstreams that reject compressed /v1/responses bodies, that leads to invalid_json unless --disable enable_request_compression is set.

This is not introduced by the server-side compaction change; this branch just makes the issue easier to hit when validating against non-default OpenAI-shaped upstreams. I’m treating compression-capability/auth-source alignment as a fast-follow rather than part of this PR.

codex-rs/core/src/client.rs

codex-rs/core/src/stream_events_utils.rs

codex-rs/core/tests/suite/compact_remote.rs

codex-rs/core/src/codex.rs

## Summary Foundation PR only (base for PR #3). This PR contains the SDK runtime foundation and generated artifacts: - pinned runtime binary in `sdk/python/bin/` (`codex` or `codex.exe` by platform) - single maintenance script: `sdk/python/scripts/update_sdk_artifacts.py` - generated protocol/types artifacts under: - `sdk/python/src/codex_app_server/generated/protocol_types.py` - `sdk/python/src/codex_app_server/generated/schema_types.py` - `sdk/python/src/codex_app_server/generated/v2_all/*` - generation-contract test wiring (`tests/test_contract_generation.py`) ## Release asset behavior `update_sdk_artifacts.py` now: - selects latest release by channel (`--channel stable|alpha`) - resolves the correct asset for current OS/arch - extracts platform binary (`codex` on macOS/Linux, `codex.exe` on Windows) - keeps runtime on single pinned binary source in `sdk/python/bin/` ## Scope boundary - ✅ PR #2 = binary + generation pipeline + generated types foundation - ❌ PR #2 does **not** include examples/integration logic polish (that is PR #3) ## Validation - Ran: `python scripts/update_sdk_artifacts.py --channel stable` - Regenerated and committed resulting generated artifacts - Local tests pass on branch

Addresses #13586 This doesn't affect our CI scripts. It was user-reported. Summary - add `wiremock::ResponseTemplate` and `body_string_contains` imports behind `#[cfg(not(debug_assertions))]` in `codex-rs/core/tests/suite/view_image.rs` so release builds only pull the helpers they actually use

codex-rs/core/src/codex.rs

Replace the Unix shell lookup path in `codex-rs/core/src/shell.rs` to use `libc::getpwuid_r()` instead of `libc::getpwuid()` when resolving the current user's shell. Why: - `getpwuid()` can return pointers into libc-managed shared storage - on the musl static Linux build, concurrent callers can race on that storage - this matches the crash pattern reported in tmux/Linux sessions with parallel shell activity Refs: - Fixes #13842

There are some bug investigations that currently require us to ask users for their user ID even though they've already uploaded logs and session details via `/feedback`. This frustrates users and increases the time for diagnosis. This PR includes the ChatGPT user ID in the metadata uploaded for `/feedback` (both the TUI and app-server).

Summary - document output types for the various tool handlers and registry so the API exposes richer descriptions - update unified execution helpers and client tests to align with the new output metadata - clean up unused helpers across tool dispatch paths Testing - Not run (not requested)

## Summary - run the split stdout/stderr PTY test through the normal shell helper on every platform - use a Windows-native command string instead of depending on Python to emit split streams - assert CRLF line endings on Windows explicitly ## Why this fixes the flake The earlier PTY split-output test used a Python one-liner on Windows while the rest of the file exercised shell-command behavior. That made the test depend on runner-local Python availability and masked the real Windows shell output shape. Using a native cmd-compatible command and asserting the actual CRLF output makes the split stdout/stderr coverage deterministic on Windows runners.

We already have a type to represent the MCP tool output, reuse it instead of the custom McpHandlerOutput

…_files] Keep current-turn inputs in local inline compaction checkpoints and remember known backend incompatibilities after a compat downgrade so later turns skip the failed inline request path. Co-authored-by: Codex <noreply@openai.com>

…nged_files] Co-authored-by: Codex <noreply@openai.com>

Preserve current-turn history when inline compaction downgrades fail and replace prior same-turn compaction checkpoints instead of stacking them. Tests: - cargo test -p codex-core codex::tests::build_server_side_compaction_replacement_history_keeps_current_turn_inputs -- --exact - cargo test -p codex-core codex::tests::build_server_side_compaction_replacement_history_replaces_prior_same_turn_summary -- --exact - cargo test -p codex-core codex::tests::downgrade_known_inline_compaction_error_restores_current_turn_when_fallback_fails -- --exact Co-authored-by: Codex <noreply@openai.com>

…hanged_files] Co-authored-by: Codex <noreply@openai.com>

Ignore compact_prompt for OpenAI inline auto-compaction, remove the legacy compat downgrade path, and keep /compact on the point-in-time endpoint. Also skip previous-model preflight remote compaction when inline server-side compaction is available.\n\nCo-authored-by: Codex <noreply@openai.com>

Handle repeated inline compactions on turns that started from empty history by stripping leading compaction items after prefix calculation, and add regression coverage for the fresh-session case. Co-authored-by: Codex <noreply@openai.com>

…ci changed_files] Co-authored-by: Codex <noreply@openai.com>

…ged_files] Co-authored-by: Codex <noreply@openai.com>

…_files] Co-authored-by: Codex <noreply@openai.com>

…ed_files] Co-authored-by: Codex <noreply@openai.com>

…g [ci changed_files] Co-authored-by: Codex <noreply@openai.com>

…iles] Co-authored-by: Codex <noreply@openai.com>

Co-authored-by: Codex <noreply@openai.com>

…les] Co-authored-by: Codex <noreply@openai.com>

Co-authored-by: Codex <noreply@openai.com>

- reduce the server-side compaction test matrix to the highest-signal cases - add comments around the deferred checkpoint rewrite and inline/preflight split Co-authored-by: Codex <noreply@openai.com>

Remove the redundant inline compaction request trace and clarify that streamed server-side compaction rebuilds replacement history on response.completed from the checkpoint snapshot. Co-authored-by: Codex <noreply@openai.com>

…les] Co-authored-by: Codex <noreply@openai.com>

Co-authored-by: Codex <noreply@openai.com>

…iles] Co-authored-by: Codex <noreply@openai.com>

Co-authored-by: Codex <noreply@openai.com>

github-actions · 2026-03-26T06:26:40Z

Closing this pull request because it has had no updates for more than 14 days. If you plan to continue working on it, feel free to reopen or open a new PR.

cooper-oai changed the title ~~Add feature-flagged server-side compaction~~ enh: add feature-flagged server-side compaction Mar 8, 2026

cooper-oai marked this pull request as ready for review March 8, 2026 17:59

etraut-openai added the oai PRs contributed by OpenAI employees label Mar 9, 2026