Skip to content

feat: add opt-in plaintext V2 subagent mode - #7376

Merged
lidge-jun merged 74 commits into
devfrom
codex/hybrid-plaintext-v2-20261019
Nov 13, 2026
Merged

lidge-jun merged 74 commits into
devfrom
codex/hybrid-plaintext-v2-20261019

Conversation

@luvs01

@luvs01 luvs01 commented Oct 19, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Add the explicit experimental v2-plaintext subagent mode: V2 catalog pins with the global override off, canonical ChatGPT task compilation, and plaintext-retention acknowledgment at GUI/CLI/API entry. Default-off behavior and the independent legacy flag remain unchanged. Catalog pins affect new chats; enabled compilation also applies to subsequent eligible canonical ChatGPT requests from existing V2 chats. Existing ciphertext is not decrypted, and disabling the mode does not delete retained copies.

  • Retained response IDs require the same admitted caller principal and the request's own thread. Principals are process-local, so a restart requires full client replay. Filtered loopback requests retain only the content-free principal needed for this check.
  • Generic config set and config import cannot newly enter plaintext mode. They direct the operator to ocx v2 mode v2-plaintext --acknowledge-plaintext-agent-messages, which also coordinates the native transition. Already-active restoration, unrelated settings, exits and the independent legacy preference remain supported.
  • Import rejects invalid current authority before creating coordination files, then checks the current mode and saves in one synchronous transaction. Input is read before taking that lock.

The dev redesign moves the dashboard mode switch into the settings row. Its existing capability/consent path is preserved there and in the Subagents panel, including the enabled-plaintext hint. WebSocket upgrade data carries both dev's observational client classification (fifth argument) and the admitted replay principal (sixth); client classification cannot grant replay authority.

The owner accepted the feature and its retention behavior. Current head is e6d38eb1df6b6048bf820e99994e06bd3288aa19, merged with dev aa30ea9d5e without rewriting author history.

Verification

The previous hosted test 2/5 failure was reproduced locally: the newly merged chained-reset fixture called the raw handler without a replay owner, so the next request returned 429 instead of 200. The fixture now uses the existing admitted-owner helper. Its upstream refusal assertions remain intact.

The state.ts merge keeps owner authorization before spill materialization and preserves upstream's distinction between client-carried and proxy-restored history. A new regression covers that distinction and its propagation to an internal body clone.

Bun 1.4.4 validation:

  • Chained-reset, replay-owner storage and combo identity: 12 pass, 0 fail, 50 assertions.
  • Native compaction provenance, store:false state and HTTP caller/thread ownership: 140 pass, 0 fail, 1,103 assertions.
  • New client-carried provenance regression: 1 pass, 0 fail, 7 assertions.
  • bun check --no-script --noEmit: passed, 2,437 files.
  • bun run structure:check and git diff --check: passed.

Full local/GUI suites were not repeated because shared Windows storage was under heavy contention; current-head hosted CI supplies broader coverage and passed. No native app acceptance or security-review approval is inferred from these tests.

Hosted-CI follow-up

e6d38eb1df binds the newly merged eager-catalog HTTP/WS fixtures to the same admitted principal and own thread as their stored replay. It also waits for the synthetic Claude CLI stdout result write before exiting, preventing large JSON replies from being truncated by pipe buffering. Production behavior is unchanged by this follow-up.

  • All five claude-cli-client-http.test.ts cases passed locally after the stdout fix.
  • The two previously failing catalog replay cases and its WebSocket handshake case passed (3 tests, 8 assertions). These real-server checks used a local --timeout 30000 after measured Windows startup exceeded the default 5-second hook budget; no committed test timeout changed.
  • The initial broad local catalog run was interrupted after repeated Windows startup-hook timeouts. It is not claimed as passing.
  • Final typecheck passed for 2,437 files; git diff --check passed. Current-head hosted CI passed: https://github.com/lidge-jun/opencodex/actions/runs/43185802386 (all five test shards and gates).

UI evidence

No visual change in this follow-up. Existing plaintext-mode UI evidence:

Current English mode surface
Current English consent
Current Korean mode surface
Current Korean consent

Final head 13f788ed69152e8cb0348da5550be2636a2412e9 (maintainer integration)

  • Fixed a blocking regression found in final review: the replay owner binding applied in every mode, so default loopback users without an opencodex API key lost local previous_response_id continuation. Replay scoping is now decided by isPlaintextV2AgentMessagesEnabled(config) (mode v2-plaintext or the legacy plaintextV2AgentMessages flag): strict principal + own thread-id binding while plaintext is enabled, dev's parent-thread scope otherwise; state bound while plaintext was on stays bound.
  • Security review fix: the plaintext decision is captured once per logical request (incl. combo children and recovery) and retention refuses compiled plaintext without a valid owner, closing a mode-flip race.
  • Restored nine default-mode test fixtures to dev behaviour; synced English and all eight locales (incl. new Spanish) for the mode roster, enablement paths, consent, config restore, replay identity, Ultra eligibility, same-target status checks and catalog-pin vs compiler scope.
  • Hosted CI: all 34 checks passed on this head (test 3/5 passed on rerun after a client-link orphan-liveness flake that also passes locally on this head and on dev). Independent security review PASS (no PR-owned src/gui change since it); final technical review PASS (608 focused tests, typecheck, structure, docs build). Union with current dev checked locally: replay ownership 40/0, eager catalog 38/0, layout and Lab boundary green, typecheck green; the only docs failure is troubleshooting/update-failed.md Spanish parity, which also fails on dev itself (from fix(update): verify pnpm global-store installs against the store pnpm names #8331) and is being fixed separately.

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.

Select V2 through model catalog pins while enabling the existing plaintext task-message compiler. Require explicit CLI/API/UI acknowledgement, preserve legacy preferences and existing chat history, and verify runtime support and write receipts.

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@github-actions

Copy link
Copy Markdown
Contributor

✅ Deterministic PR hygiene checks passed.

@github-actions github-actions Bot added the enhancement New feature or request label Oct 19, 2026
@github-actions

github-actions Bot commented Oct 19, 2026 •

Copy link
Copy Markdown
Contributor

✅ READY

  • all PR quality gates passed.

Hygiene

✅ Deterministic PR hygiene checks passed.

@coderabbitai

coderabbitai Bot commented Oct 19, 2026 •

Copy link
Copy Markdown
Contributor

Explain Change →

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

  • @coderabbitai resume to resume automatic reviews.
  • @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

  • Resume reviews
  • Trigger review

No actionable comments were generated in the recent review.

Walkthrough

The PR adds the acknowledged v2-plaintext mode for eligible task messages. It also binds local Responses replay state to the admitted caller and request-owned thread. Configuration, CLI, API, GUI, persistence, tests, and documentation are updated.

Changes: V2 plaintext mode

Layer / File(s) Summary
Mode contract and configuration
src/types/*, src/config/*, src/codex/*, src/cli/*, src/server/management/*, structure/*
Configuration, catalog handling, CLI, and API accept v2-plaintext. Entry requires explicit acknowledgement. The mode selects V2 catalog behavior while keeping the global V2 override disabled.
GUI consent and mode selection
gui/src/components/*, gui/src/pages/*, gui/src/subagent-surface.ts, gui/src/i18n/*
The GUI exposes the mode when supported, requires acknowledgement, validates the server receipt, and displays plaintext-mode status.
Mode tests and documentation
tests/cli/*, tests/codex-integration/*, tests/server/*, docs-site/*, structure/*
Tests and documentation cover mode transitions, catalog stamping, compiler selection, retained task text, and transport behavior.

Responses replay ownership

Layer / File(s) Summary
Replay ownership and persistence
src/responses/state.ts, src/responses/state/*, src/responses/spill-store.ts
Resident, spilled, and snapshot response state carries replay-owner metadata. Replay checks the caller and thread before materializing stored state.
Caller and thread identity flow
src/server/auth-cors.ts, src/server/index/*, src/server/ws-bridge.ts, src/server/responses/*
Request handling derives replay ownership from the admitted caller and request thread-id. WebSocket handling retains and revalidates a principal when forwarded-header filtering changes identity.
Replay validation and documentation
tests/server/*, tests/responses/*, tests/codex-integration/*, tests/providers/*, tests/adapters/*, tests/helpers/*, structure/transports/*, docs-site/src/content/docs/guides/codex-integration.md
Tests cover ownership mismatches, credential changes, persistence, and authenticated continuations. Documentation defines the local replay identity requirements.

Estimated code review effort: 4 (Complex) | ~60 minutes

Change: Feature · Severity of issue fixed: Medium

Possibly related PRs

  • lidge-jun/opencodex#8096: Both changes gate previous_response_id continuations with request ownership derived in the Responses request-preparation flow.
Merge Risk: Minimal · up to bbf4c

The identified display, help-text, and test-cleanup concerns are addressed. No remaining issue shown here prevents merging after normal checks.

Pre-merge checks: Passed 4

Check name Status Explanation
Title check Passed The title "feat: add opt-in plaintext V2 subagent mode" matches the main change in the changeset. The raw summary and PR objectives show a new experimental v2-plaintext multi-agent mode, with a CLI …
Linked Issues check Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check Passed Check skipped because no linked issues were found for this pull request.
Description Check Passed Check skipped - CodeRabbit’s high-level summary is enabled.
Full details: Title check

Explanation

The title "feat: add opt-in plaintext V2 subagent mode" matches the main change in the changeset. The raw summary and PR objectives show a new experimental v2-plaintext multi-agent mode, with a CLI command, a GUI consent modal, API acknowledgement, and config and catalog wiring. The title is short, specific, and in a conventional commit style. It omits the replay-ownership hardening, which is a secondary change and does not need to appear in the title.

Finishing Touches

Generate unit tests (beta)

  • Commit to this branch
  • Create a new PR
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

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

Recent review info

Run configuration

Commits

Reviewing files that changed from the base of the PR and between af7ae8a and bbf4cd5.

Files selected for processing: 21

  • gui/src/i18n/de.ts
  • gui/src/i18n/en.ts
  • gui/src/i18n/fr.ts
  • gui/src/i18n/ja.ts
  • gui/src/i18n/ko.ts
  • gui/src/i18n/pt.ts
  • gui/src/i18n/ru.ts
  • gui/src/i18n/tr.ts
  • gui/src/i18n/vi.ts
  • gui/src/i18n/zh-TW.ts
  • gui/src/i18n/zh.ts
  • scripts/test-layout/layout-responses.json
  • src/server/index/websocket-handler.ts
  • src/server/responses/core-combo.ts
  • src/server/responses/request-prepare.ts
  • src/server/ws-bridge.ts
  • structure/cli-management.md
  • tests/helpers/native-injection-fixture.ts
  • tests/responses/ws-native-continuation-send-budget.test.ts
  • tests/responses/ws-native-injection.test.ts
  • tests/responses/ws-steering-stability.test.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 8 remain after this review.

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@Ingwannu

Copy link
Copy Markdown
Owner

Scoped intake at 7609a69bee918a6c23c789dd58e1cbbd2d62c492: this is an explicit new catalog/runtime mode plus GUI/CLI/API consent and linked-home compatibility, not merely another spelling of v2. The current hosted CI passes, and the description correctly labels screenshots/fixtures synthetic and leaves real native-to-routed handoff unverified. Keep Draft pending owner product/consent and task-retention disposition; confirm default/legacy-flag independence, cached downgrade refusal, malformed HTTP-200 acknowledgments and rollback against the same head. No live activation, history rewrite, native session or final GUI approval was performed here. The completed CI jobs are evidence only for the selected platforms, not real handoff acceptance.

@luvs01

luvs01 commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator Author

Confirmed against 7609a69bee918a6c23c789dd58e1cbbd2d62c492; the relevant existing coverage is:

  • Defaults and the independent legacy flag: the mode and standalone flag are tested independently, including the empty/default configuration and the flag remaining enabled after leaving the mode (activation, leaving the mode). The new mode does not silently change the separately stored flag.
  • Cached downgrade refusal: a cached v2-plaintext selection does not establish support on the current server. The test holds the new GET pending and verifies that the fourth option remains unavailable, then checks the older response (test).
  • Malformed HTTP-200 acknowledgments: a receipt with the wrong mode or without ok: true preserves the previous selection, reports an error, and leaves the cache unchanged (behavior test). The shared predicate also rejects missing or weakened fields (predicate coverage).
  • Rollback: the catalog round trip restores recorded origins (test); a failed feature transition returns 502 without changing the in-memory mode or persisted config.json bytes (test). This is configuration/catalog rollback coverage, not a promise to downgrade existing native thread histories.

Hosted CI is now complete with no failed or pending checks; skipped platform jobs remain unverified. These are fixture/unit/DOM-harness results. A fresh native session using catalog-only V2 through a live ChatGPT-to-routed-provider handoff is still unvalidated. The separate V1-to-V2 resume experiment does not establish that handoff contract.

Keeping Draft for the owner product/consent and task-retention disposition and the remaining handoff validation. No live activation, history rewrite, or final GUI acceptance is claimed.

@luvs01

luvs01 commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator Author

Additional bounded native evidence for the catalog selection contract at 7609a69bee918a6c23c789dd58e1cbbd2d62c492:

  • Installed Codex 0.163.0-alpha.10.2 was run with an isolated home, no authentication file, a loopback-only mock Responses provider, plugins disabled, and global multi_agent_v2=false. Effective configuration was read back before permitting a model request.
  • A full gpt-5.5 catalog row, changing only its synthetic slug and multi_agent_version to v2, exposed the actual collaboration/spawn_agent tool with task_name; the native parent TurnContext was v2.
  • A separate one-spawn fixture created a native child. The returned task name matched the linked child's agent path, the child received the task in an agent_message, and both parent and child recorded V2 contexts and completed turns. The parent also recorded the child's SubAgentActivity/completed item. This used three mock model requests (two parent, one child); no real provider model was called.

The first handoff checker timed out because it expected a nonexistent child_task_complete event and a conventional user-message task. Its FAIL receipt is preserved. Read-only verification of the already completed raw rollouts established the claims above using the actual event/message contract; no native turns were replayed to obtain a PASS.

The parent finished before the child's completed activity arrived, so the parent model consuming the child's final answer was not observed. This was native-to-mock validation, not the OpenCodex compiler path or a real ChatGPT-to-external-provider handoff. The child message retained its native encrypted content; no decryption or live-account operation was involved. Classic tool rows also do not establish code_mode_only/Responses Lite behavior. These remain separate validation gaps, so this evidence does not justify marking the PR ready. Live Codex/OpenCodex configuration hashes were unchanged across these runs, and all owned native processes exited normally.

@luvs01

luvs01 commented Oct 20, 2026

Copy link
Copy Markdown
Collaborator Author

The remaining native/compiler/result-delivery path now passes a bounded integration fixture at 7609a69bee918a6c23c789dd58e1cbbd2d62c492.

Installed Codex 0.163.0-alpha.10.2 acted as the real parent/child client against the production OpenCodex handleResponses path under Bun 1.4.2. Both providers terminated at strict loopback mocks; no real provider model was called. Effective configuration was read back before the turn: global V2 off, complete catalog rows pinned to V2, plugins/analytics off, and both request/stream retry budgets explicitly zero.

  • The canonical request exposed the compiled alias catalog; returned spawn/wait calls were restored to native collaboration calls. The mock explicitly returned encrypted_function_args: [].
  • The native child recorded readable task input_text. Its routed wire request contained the task and no agent_message, encrypted-content part, or omission placeholder.
  • After native wait, the next parent request contained the actual child-authored, parent-addressed final answer in agent_message/input_text. This was not inferred from a generic tool-result string. Both native turns completed with V2 contexts.
  • Four mock model requests completed (three canonical, one routed), with no blocked egress or guard violations. Live config/source hashes remained unchanged, both owned processes exited normally, and an independent process/listener readback found no residue.

The earlier failed fixtures are preserved: one lacked the explicit local-provider allowance; the next mismatched the adapter's default /v1/responses path. A construction-only preflight now verifies the actual URL/auth rules before launch, and the new run used fresh homes rather than replaying completed turns.

Still unverified: real ChatGPT acceptance of the alias/plaintext-task contract, a real routed provider, modern code-mode/Responses Lite execution, and final GUI/owner product-consent/task-retention acceptance. This closes the classic native-to-production-compiler-to-mock handoff gap, not the live-provider gap. Keeping Draft; no live activation or history rewrite.

@Ingwannu

Copy link
Copy Markdown
Owner

Follow-up intake at 7609a69bee918a6c23c789dd58e1cbbd2d62c492 against dev 79fa97be6e6ae974ccd31d9aeedc2da96a0baa73. The new same-head native parent/child plus production handleResponses/compiler loopback receipt materially narrows the earlier handoff uncertainty. It demonstrates that particular Codex build/configured catalog/compiler/result-delivery fixture; it is not a paid upstream or arbitrary installed-build guarantee. Preserve default-off explicit consent at CLI/API/GUI, capability rejection for old children, no decryption of historical agent messages, and configured-versus-proven status. This remains Draft for the owner's experimental-mode/UI decision and independent plaintext-retention boundary review; no settings, user transcript or live agent were changed here.

@luvs01

luvs01 commented Oct 22, 2026

Copy link
Copy Markdown
Collaborator Author

Current-head native validation passed at 043fb3d7f960c945088d806f3ddc46e0286403c9, using installed Codex 0.163.0-alpha.11 and Bun 1.4.3.

The fixture rebuilt the production handleResponses path from this head and used isolated homes and strict loopback mock providers. Effective configuration was checked before the turn: global V2 off, V2 catalog rows, plugins/analytics off, HTTP/SSE transport and both retry budgets zero.

  • The real native parent created /root/probe_child, waited, and consumed the child-authored result addressed back to the parent. Both threads recorded V2 and exactly one error-free task completion.
  • The child made one actual native exec call. Its matching output appeared before its final answer; the routed function-call representation was restored to the native custom-tool contract and returned through the production adapter.
  • Five mock model requests completed, including two routed child requests. Current Responses Lite catalogs and the parent alias compiler were observed. This run made zero real provider calls.
  • Source, live configuration and protected input hashes stayed unchanged. Both owned processes exited normally; process handles and the Job closed, with zero active Job processes and no remaining proxy/mock listeners.

The component receipt SHA-256 is d123c2c0f576a8c911c1275ab78854d4a2689696be8425d6d3caf066f50b19ad; the current bundle SHA-256 is 0d41db45efca0da083cb057232be27a302526cbec4a00b52006641428f395224. Earlier harness failures stopped before any native turn and were preserved separately; they were not counted as passing component runs.

This refresh closes the current-head native parent/child/tool/result-delivery fixture gap. Real ChatGPT or routed-provider acceptance, WebSocket transport, and arbitrary installed builds remain separate claims. Keeping Draft for the owner's experimental-mode/UI decision and independent plaintext-retention boundary review, as requested; no live activation or history rewrite.

Distinguish new-session catalog pins from request-time compilation in every registered GUI locale, CLI output and documentation. Cover an existing V2 request before, during and after activation, including retained continuation messages after opt-out. Runtime policy and compiler behavior are unchanged.
@luvs01

luvs01 commented Oct 23, 2026 •

Copy link
Copy Markdown
Collaborator Author

Current head is a3f2e7eeedfce041f4a66fddc6e29a7b6267e422, pushed as a normal fast-forward; Draft is unchanged. This test-only follow-up replaces the stale CLI wording assertion found by automatic CI at 9b18cfb08b and explicitly distinguishes hybrid v2 from acknowledged v2-plaintext, alongside their common request-scope and retention explanation.

Pinned Bun 1.4.3 ran both complete CLI files: 120 passed, with one existing native root-slot translation case exceeding the default 5-second timeout. That case alone passed under a 15-second limit (1 passed, 20 assertions, 7.19 seconds). No assertion failure remained. Exact-head automatic CI is tracked in the PR checks. Production/GUI/docs source is identical to 9b18cfb08b; prior 65 checks, gates and screenshots were not repeated.

The retention review found that “applies to new sessions” conflated two scopes. Catalog pins affect new sessions; the enabled compiler also processes subsequent eligible canonical ChatGPT requests from existing V2 sessions. All 11 registered GUI locales, CLI and docs now make that distinction and state that disabling does not delete retained copies. Existing ciphertext is not decrypted and stored history is not rewritten; runtime policy is unchanged.

The new production-handler regression reuses an existing-V2 request through OFF → ON → OFF, checks the actual outbound declarations and retained continuation content, and uses a strict fetch stub. The current correction passed 65 focused checks plus type/structure/privacy/metadata/ratchet, GUI build and docs build. Four installed-browser headless layout/initial-state cases also passed; current PNGs remain local, unuploaded. The old a1dd3a7 screenshot is explicitly marked historical.

This resolves the reported disclosure-scope finding, not the separately reserved owner product/UI decision or real ChatGPT/provider acceptance. Earlier native fixtures remain attributed to their original heads; the installed OpenCodex 2.89.0 Devin observation remains separate. Keeping Draft under the owner's disposition. No manual CI, live configuration change, paid native/provider call or Security Cloud scan was performed.

@lidge-jun

Copy link
Copy Markdown
Owner

Owner decision: accepted — the opt-in plaintext V2 subagent mode is approved as a product feature, including the documented retention behavior (disabling it does not delete copies already retained). Remaining before merge: update the branch onto current dev (it conflicts now), refresh Verification for the new head, and complete the readiness boxes; a maintainer will then review the changed CI/package.json surface and approve PR CI.

lidge-jun and others added 2 commits October 29, 2026 06:29
Preserve admitted configuration writes and extracted catalog construction.

Co-authored-by: Epinephrine <27862058+luvs01@users.noreply.github.com>
Use the shared mode resolver while preserving retained native priorities.
The focused regression fails with the pre-fix resolver and passes after.

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

Copy link
Copy Markdown
Owner

Maintainer carry pushed normally to 9216be9.

  • f99e03c99188b69c0202aa85ef6c1bc528066e96 merges dev@b16bc0978b2eef1799f38042084809b5a0cd11bf, resolving six conflicts while retaining configuration admission/revalidation, read-only live catalog construction, retained native priorities, dev's upload fixture, and explicit plaintext consent.
  • 9216be92789099f7927f0d0d2e65d5f30be3cb1b fixes the integration-specific live catalog mode omission with the shared resolver and adds its registered regression. Both commits carry Co-authored-by: Epinephrine <27862058+luvs01@users.noreply.github.com>.
  • Red → green: bun test --isolate --timeout 60000 tests/server/codex-model-list-plaintext.test.ts produced 1 pass / 1 fail before the resolver fix (missing plaintext V2 pin; V1 control passed). After fixing, the regression plus tests/codex-integration/multi-agent-v2-plaintext.test.ts produced 15 pass / 0 fail / 71 assertions. A first fixture-construction error was corrected before recording the behavioral red result; assertions were not weakened.
  • The six remaining root files listed in the refreshed Verification ran with bun test --isolate --timeout 60000: 379 pass / 0 fail / 1,872 assertions. The four listed GUI files: 43 pass / 0 fail / 174 assertions. Selected layout tooling guards: 2 pass / 0 fail / 4 assertions, 25 unrelated cases filtered out. Total: 439 pass / 0 fail / 2,121 assertions. Bun 1.4.4, frozen dependencies, OCX_TEST_NO_QUEUE=1, one test process at a time.
  • bun run typecheck: exit 0, 2,228 files. bun run --cwd gui build, bun run lint:gui, bun run structure:index, bun run structure:check, bun run privacy:scan, bun run skill:surface:check, bun scripts/file-size-ratchet.ts, and git diff --check: exit 0. Docs bun run build: exit 0, 593 pages and 85,241 internal links. Existing duplicate-key/chunk-size/docs warnings are recorded in Verification.
  • git diff --quiet origin/dev -- .github package.json: exit 0. No PR-specific .github/ or package.json hunks remain after integration; current dev's CI policy/package metadata were imported unchanged. No workflow was dispatched or rerun, no labels changed, no fork CI approved, and no merge/rebase/force-push performed.
  • GraphQL returned 0 review threads, with no further page: 0 replies needed, 0 threads resolved. The fetched dev tip is contained (0 behind), the local worktree is clean, and the owner acceptance is linked in the single refreshed description. Readiness attests local validation/review suitability, not CI or security approval.

Independent maintainer security review and successful exact-head hosted CI are still required before merge. No older-head checks are promoted to current-head evidence.

NOT RUN: complete root/GUI suites (owner-requested resource exception for parallel Windows lanes; left to CI), current-head native Codex handoff, live ChatGPT/routed-provider acceptance, macOS/Linux runtime proof, full native matrix, manual CI dispatch, independent maintainer security review.

Authorize direct and combo continuation reads before spill materialization.
Preserve owner bindings through snapshots and spills and refuse foreign-id
replacement. Add handler and storage regressions and migrate replay fixtures
to real configured admission without weakening their behavior assertions.

The final shared-state compatibility run retains one unresolved 3000 ms
shutdown-child watchdog timeout; the PR remains draft for that validation
failure and independent exact-head security acceptance.

Co-authored-by: Epinephrine <27862058+luvs01@users.noreply.github.com>
t and others added 21 commits November 13, 2026 11:01
Co-authored-by: Epinephrine <luvs01@hanmail.net>
…t-impl

# Conflicts:
#	tests/fixtures/claude-efficiency-cli.ts
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
…ompiled plaintext without an owner

Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
# Conflicts:
#	docs-site/src/content/docs/guides/subagent-v1-default.md
#	docs-site/src/content/docs/ja/reference/configuration/agents.md
#	docs-site/src/content/docs/ko/guides/subagent-v1-default.md
#	docs-site/src/content/docs/ko/reference/configuration/agents.md
#	docs-site/src/content/docs/ru/reference/configuration/agents.md
#	docs-site/src/content/docs/tr/guides/subagent-v1-default.md
#	docs-site/src/content/docs/tr/reference/configuration/agents.md
#	docs-site/src/content/docs/zh-cn/guides/subagent-v1-default.md
#	docs-site/src/content/docs/zh-cn/reference/configuration/agents.md
#	docs-site/src/content/docs/zh-tw/guides/subagent-v1-default.md
#	docs-site/src/content/docs/zh-tw/reference/configuration/agents.md
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Epinephrine <luvs01@hanmail.net>
@lidge-jun

Copy link
Copy Markdown
Owner

Maintainer judgment — merge of #7376 at 13f788ed69152e8cb0348da5550be2636a2412e9

Value. Adds the explicit, consent-gated v2-plaintext sub-agent mode so a native ChatGPT V2 parent can hand tasks to routed children across the encrypted-task boundary. The owner accepted the feature and its retention behaviour (comment 6477493523).

UX change. Entering the mode requires acknowledgement (CLI ocx v2 mode v2-plaintext --acknowledge-plaintext-agent-messages, management API plaintextAgentMessagesAcknowledged, GUI consent); generic config set/import cannot newly enter it. Catalog pins apply to new sessions, but the compiler follows the current setting, so enabling it also affects later eligible requests in existing V2 sessions. Retained plaintext copies are not deleted when the mode is turned off.

Existing-user regression risk. Default (v1/default/v2) users keep dev's replay behaviour exactly — this was a blocking regression in the original head and is fixed with regressions for direct, combo and pre-upgrade state. Users who already set the legacy plaintextV2AgentMessages flag now get strict replay identity (admitted principal + own thread-id); clients without it fall back to full-history resend. That is an intended hardening of plaintext-bearing state, documented in every locale.

Windows impact. No Windows-specific code; Windows CI jobs passed on this head. Not exercised on a native Windows host locally.

Review and verification. Fresh independent security review PASS (owner-bound retention, single per-request plaintext decision, WS per-frame decision, replay-before-spill/Link ordering; PR-owned src/gui unchanged since). Final technical review PASS on this exact head. Hosted CI 34/34. Local union with current dev green except the pre-existing update-failed.md Spanish parity failure that dev itself has (#8331), fixed separately. Merging with --admin because only an approving review is missing.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ci:full Run the four full Ubuntu test shards on this pull request enhancement New feature or request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants