Skip to content

docs: record the CS-RG suspension handoff note #77

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
novelKR merged 2 commits into from
Sep 27, 2026
Merged

docs: record the CS-RG suspension handoff note #77

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
dde8501
docs: record the CS-RG suspension handoff note
novelKR Sep 27, 2026
6a4461e
docs: address the independent review of the CS-RG suspension handoff …
novelKR Sep 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
113 changes: 113 additions & 0 deletions .github/notes/pr75-cs-rg-suspension-handoff.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,113 @@
# PR #75 CS-RG suspension handoff

This note hands off the CodeSpace side of the 2026-09-27 session that suspended the planned CS-RG managed-execution design and revalidated the CodeSpace–DevGuard integration boundary. It is a dated record for agents and reviewers, not part of the documentation site, which avoids past-session narrative. It approves no architecture, starts no work and authorizes no merge. Re-query GitHub and git before acting.

- Living trackers: [#76](https://github.com/novelKR/CodeSpace/issues/76) (CodeSpace) and [novelKR/DevGuard#14](https://github.com/novelKR/DevGuard/issues/14) (primary, cross-repository).
- Full session record: [DevGuard #15](https://github.com/novelKR/DevGuard/pull/15), file `docs/handoff/2026-09-27-session-close.md`. It defines the terms used below (N1, N2, P1 to P12, K1 to K9).
- Reasoning: the DevGuard [revalidation analysis](https://github.com/novelKR/DevGuard/blob/1af1921fb72bb969b4f19ff859bd59fdb361ca55/docs/handoff/2026-09-27-cs-rg-boundary-revalidation-analysis.md) in DevGuard #13.

## Status at close (2026-09-27 15:05 UTC)

| Item | State |
| --- | --- |
| `main` | `794867ef52f530be6bc0d91aa10416d5195367b7`, the merge of #73. CI run 36308518615 and documentation run 36308519199 passed |
| Codex pin | `6b9826e3aa83b1a5947db50f4332cb9c65f1b340` (rust-v0.154.0), unchanged |
| [#75](https://github.com/novelKR/CodeSpace/pull/75) | Hold notices, head `1bee230595698b0974df43561bccfce67d7e8cb9`. Open and CLEAN; 5 checks passed and 6 were not selected by the CI plan (runs 36317763003, 36317763516). Waits for the owner's exact-head merge approval |
| This note's PR | Open; see #76 for its number and head |
| Last scheduled full run | 36275459398 on `b6e7ed2`, before the ETXTBSY fix: failure. No scheduled run on `794867e` has been read yet |
| Kept branches | `codex/ci-fixture-docs-only` and `codex/ci-fixture-pty-only` (closed #71 and #72). `codex/diag-etxtbsy-fixture`, used only for diagnostic run 36302744023; never to be merged. All other remote task branches |

## What is suspended

CS-RG ("CodeSpace consumption and control protection") is a milestone whose owner is CodeSpace; its plan lives in DevGuard's `docs/planning/milestones/CS-RG.md`. The owner suspended its units C00, C03 and C09 as implementation directives. C00 is the managed PTY through DevGuard's `HelperCommand`. C03 is a runner-owned transport outside the Codex adapter. C09 is converging the legacy backends. No replacement is approved.

In CodeSpace, #75 marks the matching sections as suspended:

- `docs/codex-reuse.md`
- `docs/devguard-integration.md`
- `docs/execution-substrate.md`
- `docs/architecture.md`
- `docs/upstream-update.md`
- the Korean counterparts of these five pages

Until #75 merges, those sections on `main` still read like a plan to implement. Nobody implements from them.

## Constraints a CodeSpace agent follows

CodeSpace's product semantics, responsibility boundaries and selective Codex delegation (N1) are joint hard constraints with DevGuard's safety invariants (N2). DevGuard is the opt-in that adapts. For CS-RG this means:

- **Execution ownership**
- No CodeSpace-owned PTY stack, Unix transport or process backend.
- No governed PTY built on Codex `ProcessDriver`.
- PTY stays delegated to `codex-utils-pty`.
- **Unaffected paths stay as they are**
- No rewrite or move of the Tokio pipe path for DevGuard or for symmetry.
- No wrapping of CodeSpace spawns in DevGuard's `spawn_guard`.
- **The `off` path**
- No change to it.
- No mandatory DevGuard service, credential or dependency.
- **Needs the owner's approval**: a Codex pin change, a DevGuard client dependency, an MCP contract change.
- **Allowed adapter scope**: a thin adapter only. That covers admission and identity binding, error and result translation, capability negotiation, outcome submission, the `resources=off|required` setting, and consuming generic backend events. It never takes over spawn, PTY, reaping or lifecycle to create an event.

## CodeSpace facts from the review

These are source facts at `794867e`; outside tests, the code is identical to `b6e7ed2`.

- **Spawns that DevGuard's guard does not cover.** Three spawn sites neither take DevGuard's `spawn_guard` nor close unintended descriptors in the child:
- the pipe spawn ([`process.rs` L266-L300](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/crates/runner/src/process.rs#L266-L300));
- the patch helper ([`patch_helper.rs` L50-L74](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/crates/runner/src/patch_helper.rs#L50-L74));
- the Linux sandbox probe ([`linux_sandbox.rs` L60-L146](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/crates/runner/src/linux_sandbox.rs#L60-L146)).

Codex PTY children do close those descriptors. This is how DevGuard's client session-socket window (D6) becomes reachable from CodeSpace processes. Exposure has not been demonstrated.
- **Both backends reap at once.** The pipe path polls `try_wait` every 20 ms ([`process.rs` L676-L698](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/crates/runner/src/process.rs#L676-L698)), and Codex PTY reaps inside `spawn_blocking(child.wait())`. No existing event fires before the reap. Under DevGuard's current contract on macOS, a survivor that was not adopted before the reap keeps its reservation charged until reboot.
- **Codex primitives.**
- Codex exposes no PTY child PID at the pin or on its `main`.
- Close-on-exec descriptor attachment (`ChildFds::Attached`, openai/codex#47797) exists only on Codex `main` and in prereleases, not in stable rust-v0.157.1.
- No checked Codex revision offers exit observation before reap for PTY children.
- **Rust floor.** CodeSpace's floor is Rust 1.88 and DevGuard's is 1.95. Any DevGuard client dependency would have to stay optional so that the `off` build keeps 1.88.

## Unresolved CodeSpace-only issues

These are outside CS-RG. Each needs its own proposal and the owner's approval, and none is justified by CS-RG.

- **OX-01.** The MCP tool text says losing the worker connection or shutting down the gateway "terminates the owned subtree" ([`mcp.rs` L123-L125](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/crates/server/src/mcp.rs#L123-L125)). The pipe path, however, sets no process group.
- **OX-02.** Completed slots are evicted after `DEFAULT_COMPLETED_TTL`, which is 15 minutes ([`process.rs` L33](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/crates/runner/src/process.rs#L33), [L208-L211](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/crates/runner/src/process.rs#L208-L211)). The dropped Codex `ProcessHandle` then calls `terminate()`, which signals the stored numeric process group (`codex-rs/utils/pty/src/process.rs` L273-L277 at the pin). By then that group may have been reused.
- **OX-03.** If UDS worker setup fails before the handshake, the worker process and its directory are left behind ([`runtime.rs` L51-L64](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/crates/server/src/runtime.rs#L51-L64)).

None of these was run; they are source readings.

## Known intermittent CI failures

| Failure | Cause, at its evidence level | Status |
| --- | --- | --- |
| `linux_sandbox::tests::prepare_unread_large_stdin_times_out` (Rust / Integration) in scheduled run 36275459398 | **Suspected** ETXTBSY: a concurrently forked test child holds a write handle to a freshly written fixture script. A direct experiment showed the mechanism: 42 of 1600 starts failed with an in-process writer, none with a child-process writer. The failing run's holder was not captured | Fix merged in #74 (`701e2b1`). **Not yet confirmed** by a scheduled run |
| `hashFiles('**/Cargo.lock')` template error (runs 36218741807 attempt 1, 35628255763) | The directory walk failed intermittently | Fixed by #67 (`339ae8e`); no recurrence recorded since |

## Next steps and procedures

0. **Approvals and evidence.**
- Owner approvals arrive as the owner's direct instruction to the agent doing the work. The PRs and comments in both repositories are written through the same `novelKR` account that the agents use, so a GitHub comment alone does not prove an approval. When you act on an approval, record it on #76 or novelKR/DevGuard#14 with its time and wording.
- Raw evidence, its manifests and the collector script exist only in the persistent DevGuard checkout on the owner's host (`<DEVGUARD_CHECKOUT>/evidence/`), and they are never committed. DevGuard #15, section 8.3, lists plain `gh` commands for an agent without that checkout.
- Runner modes: the default `InProcessRunner`, or a `UdsRunner` worker that runs the in-process runner ([architecture](https://github.com/novelKR/CodeSpace/blob/794867ef52f530be6bc0d91aa10416d5195367b7/docs/architecture.md)).
1. **Merge only on the owner's approval naming the exact head.**
- Right before merging, re-check head and base.
- Merge with `gh pr merge <n> --repo novelKR/CodeSpace --merge --match-head-commit <sha>`. Never use `--admin`, `--auto` or `--delete-branch`.
- Check the merged state, the merge commit's parents and its tree.
- Afterwards, read the post-merge CI and the documentation publication once.
- Preserve both in the DevGuard checkout's evidence directory, `<DEVGUARD_CHECKOUT>/evidence/codespace-delivery/`, with a manifest.
- Only then remove the task worktree. Remote branches are kept.
2. **Read the first scheduled full run on `794867e` once.** It runs daily at `43 19 * * *` UTC; GitHub may start it hours late. If a fixture exec failure recurs, first classify it from the job history and the other legs: the change, a known flaky recurrence, or the environment. Then propose a fix. Apply, rerun or push nothing without confirmation.
3. **Do no CS-RG CodeSpace work before the owner's decisions** on the DevGuard architecture track (novelKR/DevGuard#14). Four choices are open:
- D6 direction;
- reap-first operability;
- upstream proposals;
- UDS mode timing.
4. **Changing this note triggers a full CI run.** A change under `.github/**` runs all legs by `scripts/ci-policy.json`; a `docs/**` change runs no Rust leg.

## References

- Hold notices: #75, novelKR/DevGuard#12.
- Revalidation record and analysis: novelKR/DevGuard#13.
- Session record: novelKR/DevGuard#15.
- Trackers: #76, novelKR/DevGuard#14.
- Earlier CodeSpace changes of the same day: #73 (CSP-D04), #74 (ETXTBSY fixture race).
Loading