fix(gateway): bridge Codex approvals to messaging platforms

[Fewmanism]

Summary

• add a reusable gateway approval helper for operations that already know they need user approval
• bridge Codex app-server exec/apply-patch approval requests into Hermes' existing gateway approval queue
• cover both exec and file-change approvals without a CLI callback, matching Telegram/Discord/etc. gateway sessions

Verification

• ./venv/bin/python -m pytest tests/agent/transports/test_codex_app_server_session.py -q -n0
• ./venv/bin/python -m pytest tests/gateway/test_approve_deny_commands.py tests/tools/test_approval_heartbeat.py tests/tools/test_approval_plugin_hooks.py -q -n0
• ./scripts/run_tests.sh tests/agent/transports/test_codex_app_server_session.py tests/gateway/test_approve_deny_commands.py tests/tools/test_approval_heartbeat.py tests/tools/test_approval_plugin_hooks.py

[teknium1]

Thanks for tackling the missing interactive gateway path. The premise still holds: current main declines Codex exec and apply-patch requests when no CLI callback exists (agent/transports/codex_app_server_session.py:882, :923). The merged opt-out routing change is complementary, not a replacement (agent/codex_runtime.py:347-381).

Problems

• The proposed request_gateway_approval() copies the gateway queue wait loop, but omits the interrupt handling current main added in _await_gateway_decision() (tools/approval.py:2491-2507). A /stop or /new during a Codex approval could therefore wait for the full approval timeout.
• Current main already centralizes queue insertion, heartbeat polling, hooks, cleanup, and interruption in _await_gateway_decision() (tools/approval.py:2419-2534); the new helper would duplicate that stateful logic.

Suggested changes

• Port the fallback onto current main by using _await_gateway_decision() and mapping its resolved choice with _approval_choice_to_codex_decision().
• Add a Codex gateway-approval interrupt regression alongside the existing approval interrupt coverage.

Automated hermes-sweeper review.

[teknium1]

Please avoid a second gateway approval wait implementation here. Current main centralizes queue lifecycle and, critically, /stop//new interruption in _await_gateway_decision() (tools/approval.py:2419-2534, especially :2491-2507); this copied loop has no interrupt check.

[GottZ]

This was generated by AI during triage.

Summary

Four PRs address the missing Codex app-server approval path for gateway sessions. #25889 adds a transport-level fallback with a new queue loop, #27636 adds a runtime callback resolver with another queue loop, #43217 routes through the shared waiter but mixes in unrelated Telegram cache-path changes, and #66925 provides the current-main bridge while preserving policy floors, redaction, interruption, cleanup, and session-choice persistence.

Related pull requests

• #25889 related — (+257/-25) — superseded by #66925: It directly forwards exec and apply-patch requests to a new gateway helper, but duplicates the approval wait loop and omits the shared interrupt handling. Despite the keep_open review on #25889, #66925 implements the requested salvage path through _await_gateway_decision() and adds the missing policy protections.
• #27636 duplicate — (+330/-8) — superseded by #66925: It correctly identifies the runtime callback gap and covers smart approval, but re-reads HERMES_YOLO_MODE at request time and duplicates the queue wait loop without /stop or /new interruption. Despite the keep_open reviews on #27636, its requested callback bridge and shared-waiter corrections are present in #66925 without those defects.
• #43217 [closed] related — (+268/-3) — closed reference implementation: It demonstrates routing through _await_gateway_decision(), but combines the approval fix with unrelated Telegram attachment/cache-path redaction changes and was closed as superseded. Despite its earlier approved review, its approval-specific approach is more cleanly and comprehensively implemented by #66925; the later contributor review also confirms that the merged bypass-only work did not replace the interactive bridge.
• #66925 related — (+378/-41) — preferred consolidation target: It bridges the runtime callback to the existing shared queue, enforces hardline, sudo-stdin, and user-deny floors, force-redacts gateway payloads, and retains timeout, interruption, cleanup, and hook behavior. The keep_open review's session-persistence and lifecycle-test blockers were explicitly addressed in the shown diff with approve_session() and full-turn construction coverage.

Duplicates

#25889 and #27636 substantially duplicate the same Codex-to-gateway approval bridge; #43217 is an earlier closed implementation of that core change. #66925 is the hardened current-main consolidation of their relevant approval work.

Suggested consolidation

Merge #66925 after maintainer confirmation of the updated head and focused test results; its diff addresses the blocking contributor review while retaining the strongest parts of the competing implementations. Close #25889 and #27636 as superseded duplicates of #66925, and keep closed #43217 as a historical reference rather than reopening it.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    subgraph Dup25889 ["PRs duplicating each other"]
        P25889["PR #25889 (open)"]
        P27636["PR #27636 (open)"]
    end
    class P25889 open
    class P27636 open
    class P25889 target
    click P25889 "https://github.com/NousResearch/hermes-agent/pull/25889"
    click P27636 "https://github.com/NousResearch/hermes-agent/pull/27636"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed or no verify verdict yet (state tag in the node label).

Cross-PR triage: Reviewed 4 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 63 kB of PR diffs, 6 kB of issue/PR text, 12 kB of discussion (15 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[Fewmanism]

Thanks for consolidating the related approval-bridge work. I agree that #66925 now covers the requested current-main salvage path more comprehensively than this PR, including the shared waiter, interruption behavior, policy floors, redaction, and session-choice persistence.

I’m happy to close #25889 as superseded once maintainers confirm #66925 as the canonical target. No further implementation work is planned here in the meantime.

@teknium1, does that consolidation direction match your preference?

[GottZ]

This was generated by AI during triage.

Delta since our previous triage comment

@Fewmanism explicitly confirms that #66925 now covers #25889’s intended current-main salvage path more comprehensively and that no further implementation work is planned on #25889. They are willing to close #25889 as superseded once maintainers designate #66925 as the canonical target; the PR remains open pending that confirmation.

Changed pull requests

• #25889 related — (+257/-25) — supersession now supported by the author: @Fewmanism confirms that #66925 includes the shared waiter, interruption handling, policy floors, redaction, and session-choice persistence missing from this diff. Despite the earlier keep_open review on #25889, that review’s requested salvage path is implemented in #66925, and the author now plans no further work here.

Suggested consolidation

The recommendation is unchanged: select #66925 as the canonical target and close #25889 as superseded once maintainers confirm that direction.

Complex graph unchanged since our previous triage comment.

Cross-PR triage: Reviewed 4 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 63 kB of PR diffs, 6 kB of issue/PR text, 12 kB of discussion (16 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.