fix(gateway): preserve custom_providers extra_body on reused-agent turns

[IbrahimKhan12]

Why

Gateway cached-agent turns replace request_overrides, dropping durable custom_providers[].extra_body values initialized on the agent. Custom-provider requests therefore behave differently between CLI and messaging paths, silently removing fields such as reasoning_effort.

What

Reapply matching custom-provider extra_body defaults after active-runtime restoration at the start of each agent turn. Explicit turn-level top-level and nested values retain precedence, and defaults are applied only to the active provider runtime so cooldown-active fallbacks do not inherit primary-provider settings.

Add regression coverage for a reused gateway agent across fast and normal turns, merge precedence, restored primary runtimes, and cooldown-active fallbacks.

Testing

• scripts/run_tests.sh tests/agent/test_custom_provider_extra_body.py tests/agent/test_turn_context.py tests/gateway/test_fast_command.py -q
• python -m ruff check agent/agent_init.py agent/turn_context.py tests/agent/test_custom_provider_extra_body.py tests/agent/test_turn_context.py tests/gateway/test_fast_command.py
• git diff --check origin/main..HEAD

References

• Fixes [Bug]: custom_providers[].extra_body silently dropped on gateway/messaging paths (CLI works) #54922

[alt-glitch]

This was generated by AI during triage.

Related competing fixes for #54922: #30224 preserves existing overrides then clears known transient keys; #52432 tracks/removes the previous turn payload; this PR recomputes durable extra_body from config each turn. These are different state-management contracts, so maintainer selection is needed.

[teknium1]

Thanks for isolating the durable-versus-transient override contract. The defect is still present on current main: gateway/run.py:22480 replaces the cached agent's init-time request_overrides, while agent/agent_init.py:2225-2226 establishes the config-derived custom-provider extra_body that must persist. The stateless recomputation in this PR matches the issue discussion's warning against accumulating /fast state.

Problems

• The new tests cover the helper and ChatCompletionsTransport, but do not drive the actual cached-agent reuse assignment at gateway/run.py:22480. That leaves the load-bearing gateway wiring untested.

Suggested changes

• Add a gateway-level reuse test covering custom-provider extra_body, a fast turn, and the following non-fast turn. Assert the durable body remains and the transient service tier does not leak.

Automated hermes-sweeper review.

[IbrahimKhan12]

ack will fix

[IbrahimKhan12]

Done, revised the PR.

[GottZ]

This was generated by AI during triage.

Summary

Two open PRs address #54922 through different state-management contracts: #30224 merges existing cached-agent overrides with each turn, while #68946 reconstructs durable custom-provider extra_body from configuration after runtime restoration and now exercises fast-to-normal cached-agent reuse.

Related pull requests

• fix(gateway): preserve custom provider extra_body across turns #30224 duplicate — (+91/-1) — n/a: Preserves existing extra_body at the cached-agent refresh site while clearing known transient keys, but the contributor keep_open review notes that runtime request_overrides are still omitted from gateway routing and that the tests cover only the merge helper. Despite that keep_open review, fix(gateway): preserve custom provider extra_body across turns #30224 is superseded for this issue by fix(gateway): preserve custom_providers extra_body on reused-agent turns #68946's configuration-derived reconstruction and gateway reuse coverage.
• fix(gateway): preserve custom_providers extra_body on reused-agent turns #68946 best fix — (+126/-7) — n/a: Reapplies matching custom-provider extra_body from durable configuration after restoring the active runtime, preserves explicit override precedence, and avoids applying primary-provider defaults to an active fallback. The maintainer-bot keep_open review requested a cached-agent fast-then-normal regression; the revised diff now adds that gateway-level sequence and asserts durable extra_body persistence without transient service-tier leakage, consistent with the complex's recorded best-fix designation.

Duplicates

#30224 and #68946 are competing fixes for the same cached-agent overwrite behind #54922; #30224 can be treated as superseded by #68946.

Suggested consolidation

Keep #68946 open with a salvage path: retain its stateless configuration-derived reconstruction, active-runtime scoping, precedence handling, and newly added cached-agent fast-to-normal regression, then obtain review of the revision. Close #30224 as a duplicate of #68946; despite #30224's keep_open review, its diff leaves the review-identified runtime request_overrides routing omission unresolved and tests only its helper.

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
    I54922(["issue #54922 (open)"])
    subgraph Dup30224 ["PRs duplicating each other"]
        P30224["PR #30224 (open)"]
        P68946["PR #68946 (open)"]
    end
    P68946 -->|best fix| I54922
    class I54922 open
    class P30224 open
    class P68946 open
    class P68946 best
    class P68946 target
    click I54922 "https://github.com/NousResearch/hermes-agent/issues/54922"
    click P30224 "https://github.com/NousResearch/hermes-agent/pull/30224"
    click P68946 "https://github.com/NousResearch/hermes-agent/pull/68946"

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 (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 13 kB of PR diffs, 14 kB of issue/PR text, 5 kB of discussion (6 comments), 3 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.