Skip to content

fix(security): combine adapter argv and upstream-body hardening - #5517

Closed
luvs01 wants to merge 12 commits into
stack/c-sec-mgmtfrom
stack/c-sec-adapter
Closed

luvs01 wants to merge 12 commits into
stack/c-sec-mgmtfrom
stack/c-sec-adapter

Conversation

@luvs01

@luvs01 luvs01 commented Sep 22, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Combine adapter and external-input hardening from #5315 and #5336 into the top layer of the security stack.

Verification

  • Current head: 260a87bfad854ed2c89af2ed38051b8d357aa283. This follow-up changes only documentation and the specific test consumers listed below; production source is byte-identical to the preceding published head 3d3ec2cb6051504af9af0fad5e4cd5afd7303eb3.
  • Hosted CI exposed structure/runtime.md growing to 602 lines when the branch was combined with newer dev. Two overlapping Remote Workspace descriptions were consolidated without losing their contracts or links. The same patch was verified against both exact failing merge trees (602 to 600) and dev (600 to 598), without raising the 600-line budget. Final combined structure checks passed.
  • Hosted CLI/status and local-management failures were consumers still expecting unbounded health-version strings. The middle-layer fixtures now retain valid semver identity values, project invalid/null/control/overlong inputs to unknown, and verify that raw values are not reflected in JSON or human output. The production security filter is unchanged. Original exact two-file run: 70 passed / 3 failed; corrected run: 76 passed / 0 failed / 385 assertions / 77.72 seconds, pinned Bun 1.4.0. Final fixture bytes match the tested f9ae99bf3579ae24f5bd2c0fa160a54d05cdf2ee commit.
  • Devin discovery now consumes one OAuth snapshot carrying both token and account host; the fixture is aligned to that existing contract. It preserves measured modalities, sidecar rules, tenant-host/account switching, real protobuf/cache parsing and network-zero assertions. Original 2 passed / 6 failed; final 8 passed / 0 failed / 24 assertions / 2.16 seconds, pinned Bun 1.4.0. Final fixture bytes match tested b6cb8a426decf5ff0baa0b65381631d786177dba.
  • Each corrected C2 layer passed typecheck, structure, privacy and file-size checks. These are distinct runs and are not added together. Earlier interrupted whole-auth/four-file runs remain incomplete; current-head hosted results are required before readiness.
  • Full-suite/cross-platform completion, exact updated-head hosted CI and independent security review remain incomplete. No test timeout or safety boundary was weakened. This PR stays draft.

Checklist

  • Scope stays focused and avoids unrelated cleanup.

  • Docs or release notes were updated when needed.

  • Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

Review readiness

  • Local CI is green on this exact head.

  • Branch is based on the current dev commit at preparation time.

  • All correct Codex and CodeRabbit findings are fixed on the combined head.

  • Ready-for-review confirmation.

@luvs01
luvs01 added this pull request to stack #5518 September 22, 2026 09:51
@coderabbitai

coderabbitai Bot commented Sep 22, 2026

Copy link
Copy Markdown
Contributor

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

  • Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@github-actions

Copy link
Copy Markdown
Contributor

✅ Deterministic PR hygiene checks passed.

@github-actions

github-actions Bot commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

✅ READY

  • all PR quality gates passed.

Hygiene

✅ Deterministic PR hygiene checks passed.

@lidge-jun

Copy link
Copy Markdown
Owner

리뷰 · 우선순위 24 / 80

이 PR은 보안·격리 쪽에 가깝게 손본 스택 조각이다. Qoder/CodeBuddy는 시스템·개발자 프롬프트를 자식 프로세스 argv에 넣지 않고, 임시 파일(권한 0600)에 쓴 뒤 경로만 넘긴다. 스테이징이 실패하면 spawn 전에 막고, 턴이 끝나면 임시 디렉터리를 지운다. Devin 쪽은 라이브 모델 목록을 가져올 때 OAuth 스냅샷에 묶인 테넌트 API 호스트를 쓰도록 해서, 한 계정의 키와 레지스트리 기본 호스트(또는 다른 계정 호스트)가 섞이지 않게 한다. xAI 웹검색은 실패 응답 본문을 바이트 한도로 읽고, 넘치면 끊는다. Grok sparse-terminal 관련해서는 tool_choice 밖으로 도구 호출을 다시 만들지 않는 테스트를 보강한다. 베이스는 dev가 아니라 stack/c-sec-mgmt이고, 헤드 stack/c-sec-adapter인 드래프트 스택 PR이다.

라인 - src/adapters/codebuddy/adapter.ts — 예전 --append-system-prompt(기본 프롬프트에 덧붙임)를 --system-prompt-file(기본을 통째로 바꿈)으로 바꿨다. 주석으로 의도한 동작이라고 적혀 있지만, CodeBuddy 기본 시스템 프롬프트에 있던 안전/도구 관련 문구가 빠질 수 있다. Qoder는 --append-system-prompt-file이라 덧붙임이 유지된다.
라인 - src/adapters/qoder/adapter.ts vs codebuddy/adapter.ts — 같은 “임시 파일로 프롬프트 스테이징”인데 writeFile 옵션(flag: "wx" 유무), 에러 code/message가 어댑터마다 다르다. 동작은 비슷해 보이지만 실패 관측·디버깅 때 헷갈릴 수 있다.
라인 - PR 본문 — Summary/Verification/Checklist가 템플릿 그대로다. 커밋 메시지에는 의도가 잘 적혀 있으니, 리뷰어용으로 본문만 채워 두면 좋다.
라인 - tests/responses/responses-snapshot-repair.test.ts — 이 PR diff에는 Grok 구현 파일 변경이 없고 테스트만 있다. 구현·배선은 부모 스택(stack/c-sec-mgmt)에 이미 있는 상태로 보이며, 커밋 메시지의 “Grok sparse-terminal repair …”는 이 헤드 범위만 보면 과장되어 읽힐 수 있다.

메인테이너의 판단이 필요한 지점
CodeBuddy를 “파일로 기본 시스템 프롬프트를 교체”하는 쪽으로 가도 되는지(지원 CLI 버전에서 --system-prompt-file이 실제로 있는지, 교체로 빠지는 기본 지침이 없는지). 부모 스택 stack/c-sec-mgmt와 머지 순서·충돌을 어떻게 묶을지. Grok tool_choice 보강이 이 PR에 테스트만 올리는 구성으로 충분한지, 아니면 구현 커밋과 같은 단위로 보이게 정리할지.

너의 추천
보안 방향(argv 노출 제거, Devin 호스트 핀, xAI 본문 바이트 한도)은 유지하는 편이 좋다. CodeBuddy 플래그/교체 의미만 지원 바이너리와 한 번 맞춰 확인하고, PR 본문을 커밋 요약 수준으로 채운 뒤 부모 스택 위에 올리면 된다. types.ts/config.ts 쪼개기나 프리뷰 배포 이야기는 이 PR과 무관하다. 무효·중복 이슈가 보이면 그쪽으로 닫는 쪽이 낫다.

이 댓글은 grok-bot이 작성했습니다

@luvs01 luvs01 changed the title stack/c sec adapter fix(security): combine adapter argv and upstream-body hardening Sep 22, 2026
luvs01 and others added 3 commits September 22, 2026 19:57
…instead of argv

The folded system+developer prompt was embedded in the child process argument list (--append-system-prompt <text>), which is visible to other local users via process listing. Stage it in a private temp file (mkdtemp, mode 0600) and pass the path: --append-system-prompt-file for Qoder, --system-prompt-file for CodeBuddy. Staging failure fails closed before spawn; the temp directory is removed after the turn.

(cherry picked from commit fefc4c3)
…coped transports honest

- xAI web search: read a non-OK sidecar body through the bounded byte reader instead of an unbounded res.text(), so a hostile or broken upstream cannot grow the retained error body without limit.

- Grok sparse-terminal repair: thread the outbound request body into the rewrite so reconstructed client tool calls stay inside the request's effective tool_choice instead of fabricating calls the caller forbade.

- Devin live-model discovery: resolve the tenant host from the OAuth credential snapshot's allowlisted apiBaseUrl rather than pairing a durable account key with the registry default host.

(cherry picked from commit 6a24ddb)
… contract

Both adapters stage the folded prompt in a private temp file, but they diverged on the exclusive-create flag and on the failure code and message. Give Qoder the same flag:wx write and the shared system_prompt_staging_failed code so a staging failure reads the same way on either adapter.
A test-only writeFile seam on QoderAdapterDeps drives the real exclusive-create failure: no spawn, a 500 upstream_error with the shared system_prompt_staging_failed code, non-retryable, no prompt text in the error, and the fresh temp dir removed.
Reuse the owned management server fixture for the non-loopback settings test,
keeping its real 0.0.0.0 listener, LAN Host/Origin, missing-token rejection,
authenticated response and CORS assertion.

Prepare real token/ACL state before the HTTP body, isolate only unrelated host
diagnostic projections, track cancellation and consume both response bodies.
Teardown settles the body and actual listener/spend owner before restoring homes.
Rename the helper to its shared management-server role without changing policy,
timeouts, assertions or the file-size baseline.
@lidge-jun

Copy link
Copy Markdown
Owner

Carried into #5610 as f2c5ce8 (squash of this PR's own diff at head 260a87b, authorship kept). A review follow-up in #5610 (275076b) keeps the Copilot host overlay out of Devin snapshots and keys Devin's catalog authority by the validated tenant URL. Closing as superseded by #5610. Thanks @luvs01.

@lidge-jun lidge-jun closed this Sep 22, 2026
lidge-jun added a commit that referenced this pull request Sep 23, 2026
)

* fix(service): combine startup ownership, token binding, and slot retention

Carries #5512 by @luvs01 (head a12b2ad), which
consolidates #5477, #5306 and #5357:

- bind the service API token to its owning state, canonicalize qualified-localhost
  binds, and carry WSL ownership state honestly (#5477);
- take a fresh task listing for the second startup ownership decision (#5306);
- retain workflow slots for streaming turns (#5357);
- own server-auth fixture lifetime and project a current-schema config for it.

Squashed from the PR's own diff (origin/dev...a12b2ad) onto current dev.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(server): self-heal a replaced package tree via drain-and-restart

Carries #5513 by @luvs01 (head 4d168f1), which
consolidates #5393 and its scheduler follow-up: detect a replaced installed package
tree, degrade health honestly, and drive a timer-driven, retryable drain-and-restart
whose verify step is deferred past scheduler re-entry. The guard factory lives in
src/server/index/package-tree-guard.ts.

Squashed from the PR's own diff (a12b2ad...4d168f1) onto the #5512 carry.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(security): combine install discovery, credential, and transport hardening

Carries #5515 by @luvs01 (head 843f299), which
consolidates #5359, #5285 and #5322:

- keep selected Codex installation discovery off network filesystems, probe
  oversized wrappers through a held-handle prefix read, and stop a PATH scan at a
  refused probe (#5359);
- exclude npm candidates inside the launch directory subtree (#5285);
- refuse plaintext remote hub origins, fail closed on POSIX chmod for credential
  files, and skip the frame-log write when descriptor hardening fails (#5322).

Squashed from the PR's own diff (origin/dev...843f299) onto the chain carry.
Integration: structure/runtime.md wording reflowed by two lines so the combined
service and security stacks stay within the 600-line structure budget.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(security): combine management-auth and boundary hardening

Carries #5516 by @luvs01 (head 245d542), which
consolidates #5326, #5312, #5363 and #5317:

- harden pairing redemption, agent roster intake, and SOCKS5 decoding (#5326);
- guard gh resolution, anchor the grok managed-region fences to whole lines, and
  bound provider-controlled text (#5312);
- harden management-auth admission and provenance (#5363);
- bound the /healthz version before it reaches diagnostics (#5317).

Squashed from the PR's own diff (843f299...245d542) onto the #5515 carry.
Integration: both stacks rewrote the shared server-auth test fixtures. The carry
keeps the #5512 current-schema fixture projection and config helper (including
its 4 KiB boundary case) and adds this PR's Aside sync capability assertions.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(security): combine adapter argv and upstream-body hardening

Carries #5517 by @luvs01 (head 260a87b), which
consolidates #5315 and #5336:

- stage Qoder and CodeBuddy system prompts in private files instead of
  child-process argv, with exclusive creation and owned cleanup (#5315);
- bound upstream error bodies and resolve account-scoped transports (Copilot,
  Devin) from the same OAuth snapshot as the bearer (#5336).

Squashed from the PR's own diff (245d542...260a87b) onto the #5516 carry.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* feat(codebuddy): integrate capture-only tools with private prompt staging

Carries #5582 by @luvs01 (head 3061ef9), which
integrates the capture-only CodeBuddy tool bridge from #5148 by @mdwsk88 with the
private prompt staging from #5517. Requests with a tool catalog advertise only the
allowed tools through an isolated MCP server that captures calls without executing
them; the client keeps approval, sandboxing and execution. Pre-init, undeclared,
excessive or incomplete calls are rejected, streamed malformed tool arguments are
suppressed, bridge staging failures return a fixed message, and an opt-in live
acceptance harness is included. Design context: #5146.

Squashed from the PR's own diff (260a87b...3061ef9) onto the #5517 carry.

Co-authored-by: mdwsk88 <924038395@qq.com>

* fix(client): bound total hub catalog response lifetime

Carries #5252 by @luvs01 (head 779ef91): give the
hub catalog body read an overall deadline (24x the inactivity window, capped at
120 s) on top of the inactivity window, and release refused, HTTP-error and 304
bodies without awaiting their cancellation.

Squashed from the PR's own diff (origin/dev...779ef91).

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(grok): reserve model aliases only when the written config stays valid

Reimplements #5281 by @luvs01. A user sub-table such as [model.ocx-mine.extra]
only creates an implicit parent, so it no longer forces the generated table to a
suffixed alias. The alias choice is now checked against the bytes actually
written: the unsuffixed alias is used only when the final config (after
model-reference rewriting) parses; otherwise the conservative choice that also
reserves deeper headers is used, and a valid user file for which neither choice
parses is refused without writing. Malformed user TOML keeps the previous
conservative reservation.

The original change reserved only exact two-segment headers, which could emit a
duplicate [model.x] table when the user defines model.x through dotted keys.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(codex-auth): scope Codex OAuth cancellation to the originating flow

Reimplements #4923 by @luvs01 on the current login-state layout (in-flight
controllers moved to src/oauth/login-flow-state.ts in #5220). Cancelling a Codex
login was keyed only by provider, so a stale modal posting an old flowId could
abort a newer attempt, and a cancel without a flowId expired every pending flow.

- Each in-flight controller records the flowId that started it; a cancel whose
  flowId does not match the active attempt is refused before anything aborts.
- POST /api/codex-auth/login/cancel requires a non-empty flowId, rejects unknown
  or non-pending flows with 400 without touching any row, and expires only that
  flow. Provider-wide cancellation through /api/oauth/login/cancel is unchanged.
- ocx account cancel requires --flow for Codex providers and sends no request
  without it.

The dashboard's 409 recovery keeps its code; its ownerless cancel is now refused,
so it ends in the existing "already in progress" message instead of superseding a
flow it does not own.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(socks5): bound compressed event streams by expansion, not total size

Review follow-up to the #5516 carry. The 32 MiB decoded-body cap applied to every
gzip/deflate response, so a long, normally compressed SSE stream through the
SOCKS5 tunnel was cut once its cumulative output crossed the cap. Buffered
responses keep the absolute cap; event streams may continue while decoded bytes
stay within the greater of 32 MiB or 128x the coded bytes consumed, which still
stops high-ratio bombs.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(codex): keep scanning PATH past a missing Windows candidate

Review follow-up to the #5515 carry. The held-handle reader reported a missing
file or directory as open-refused, so the default existence probe stopped the
PATH scan at the first absent PATHEXT candidate (for example codex.com) before it
reached an installed codex.cmd. NtCreateFile's object-name-not-found and
object-path-not-found statuses now map to a distinct not-found result that lets
the scan continue; every other failure still refuses.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(server): require Windows ACL hardening before a frame-log append

Review follow-up to the #5515 carry. On Windows the frame log ignored a failed
permission change and appended anyway. Each append now hardens the target with
the required Windows ACL helper and checks that the path still names the opened
file before writing; any failure writes nothing.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(devin): bind catalog authority to the tenant destination

Review follow-up to the #5517 carry.

- The observe-only OAuth snapshot applied the Copilot-validated apiBaseUrl to
  every provider, so a crafted Devin credential could carry a Copilot host that
  the snapshot claimed as its own. The overlay now applies only to github-copilot.
- Devin's live roster, stale fallback and cooldown were keyed by the token alone
  while discovery also depends on the validated tenant URL. The catalog authority
  and the matching routing-cache resolver now fingerprint the token together with
  the validated destination URL.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(codebuddy): fail closed on unverified bridge turns and staging collisions

Review follow-up to the #5582 carry.

- With the capture-only tool bridge armed, a successful terminal event is no
  longer accepted unless the CLI's system/init frame confirmed the bridge server;
  a turn that ends without it fails with tool_bridge_init_missing.
- A tool_use block that arrives only in the complete assistant message, without
  the partial tool events the bridge captures, now fails the turn instead of
  being dropped silently; partial captures are deduplicated by id.
- The catalog and MCP config staging files are created exclusively (wx, 0600),
  like the prompt file, so a pre-existing file fails before spawn.
- The history-argument repair for a missing JSON object prefix is documented and
  tested as a provider-agnostic contract; other malformed strings keep {}.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>
Co-authored-by: mdwsk88 <924038395@qq.com>

* fix(service): keep service-command ownership bound to the recorded home

Review follow-up to the #5512 carry. On WSL with CODEX_HOME unset, the carried
allowance treated a legacy Linux ~/.codex install record as owned when discovery
now selects the Windows profile, so service stop could stop the Linux-home
service and then restore native Codex in the Windows home, and repair could
rewrite the recorded home. Service commands again require the exact recorded
home and name it in the refusal; the unattended startup inspector reaches the
same foreign verdict.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(server): veto a package-tree restart when its server stops or loses ownership

Review follow-up to the #5513 carry.

- A package-tree restart accepted by the guard stayed scheduled after an explicit
  server.stop(), so the drain-and-respawn could reopen a server the caller had
  stopped. The caller that accepted a pending restart now receives a veto, and
  the guard uses it on dispose.
- When running as a supervised service child, the automatic path checks service
  home ownership when accepting and again before the handoff; a mismatch keeps
  the 503 fence and skips the restart.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(security): resolve gh from fixed paths and look up pairing grants by digest

Review follow-ups to the #5516 carry.

- On Windows the automatically polled star-status route derived gh.exe roots from
  ProgramFiles and LOCALAPPDATA, so a process environment could select any
  absolute directory. Windows candidates are now the fixed system install paths,
  and the child PATH is only the resolved executable's directory. Other installs
  report gh as unavailable, which only hides the sidebar star state.
- Pairing redemption looked each guess up by scanning every live grant; the map
  is keyed by the grant digest, so the lookup is now a direct get. A valid grant
  still redeems behind a throttled source.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* test(server): cover the one-shot Aside sync capability end to end

Review follow-up to the #5516 carry, which added a one-shot, HMAC-bound
capability for the default ocx sync path without exercising it. A real listener
now proves single use, refusal on replay, wrong path, query, method, pid or port,
expiry and a bad MAC, and that the CLI default path performs the attestation and
a bodyless POST (through a narrow transport seam).

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* test: register the review follow-up test files in the layout maps

Adds the three new test files from the L4 review follow-ups to both
scripts/test-layout/layout.json and tests/fixtures/test-layout-expected.json.

* test(grok): pin re-injection and strip for a nested user model table

Review follow-up to the #5281 reimplementation: two injections are byte
identical, every intermediate file parses, and strip restores the exact user
content.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(server): harden a Windows frame log once per file identity

Re-review follow-up: requiring Windows ACL hardening on every append spawned
icacls for every relayed frame and could stall the realtime relay. The hardened
file identity (device and inode) is now remembered for the log path; an
unchanged file skips the respawn, and a replaced file at the same path is
hardened again before any write.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* docs(structure): describe the package-tree restart veto and ownership recheck

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(server): stop an automatic restart from handing off after an explicit shutdown

Security review follow-up to the #5513 carry. Once an automatic package-tree
restart entered its drain, an operator shutdown (signal or management stop)
could still be followed by the restart handoff, because the drain cannot tell
its own listener stop from an independent one. Explicit shutdown paths now mark
the process, and an admission-bound restart checks that mark before every
handoff step. Manually requested restarts keep their behavior.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix(server): mark a management stop before its asynchronous teardown

Security re-review follow-up: the management stop route marked the explicit
shutdown only after awaiting the shared teardown, so an automatic restart
draining concurrently could reach its handoff in that window. The mark now
precedes the first await after the stop is accepted.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* test(server): allow post-lookup pruning in the pairing digest regression

The digest-lookup regression trapped every iteration of the grant map, so a
valid redemption failed once session minting pruned expired grants after the
lookup (hosted CI test 4/4). The trap now fails only on a scan that precedes the
digest lookup, which is the regression it guards.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

* fix: repair standalone bridge and restart ownership

Use the compiled CLI as the capture-only MCP entrypoint, release automatic restart fences on veto, align Devin discovery, and tighten Windows and local transport handling. Apply the documented Qoder prompt environment for both regions and update focused regressions and operator docs.

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>

Co-authored-by: mdwsk88 <924038395@qq.com>

---------

Co-authored-by: luvs01 <27862058+luvs01@users.noreply.github.com>
Co-authored-by: mdwsk88 <924038395@qq.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

bug Something isn't working

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants