fix(api_server): trust full transcripts via _transcript_mode to stop repair-induced duplication

[LiangYang666]

What does this PR do?

Fixes a storage-side duplication bug in /v1/responses when previous_response_id chaining is used together with repair_message_sequence. The stored conversation_history doubles on every chained turn because the repaired transcript no longer matches the exact prefix expected by _response_messages_turn_start_index.

Root cause

AIAgent.run_conversation returns the authoritative full transcript in result["messages"]. Before each LLM call, repair_message_sequence_with_cursor may merge adjacent user/assistant messages in place. The merged transcript no longer shares an exact prefix with the loaded conversation_history, so _response_messages_turn_start_index returns 0 and _build_response_conversation_history falls through to:

full_history = prior
full_history.append(current_user)
full_history.extend(agent_messages)

Because agent_messages already contains the full transcript, this concatenation doubles the stored history every chained turn.

Fix

• In _run_agent: when result["messages"] is a list, set result["_transcript_mode"] = "full".
• In _build_response_conversation_history: after the exact-prefix heuristic fails, check result.get("_transcript_mode") == "full" and return agent_messages verbatim.

The existing _compressed flag and exact-prefix detection are kept as fallbacks for mocks, legacy adapters, and delta-shaped returns.

Related Issue

Fixes #68257

Complements #69306 (compression-triggered prefix misses via _compressed) and relates to #70695 (tolerant semantic turn-start detection).

Type of Change

• 🐛 Bug fix (non-breaking change which fixes an issue)

Changes Made

• gateway/platforms/api_server.py
  • _run_agent annotates result["_transcript_mode"] = "full" when the agent returns a message list.
  • _build_response_conversation_history trusts the explicit mode and returns the full transcript directly.
• tests/gateway/test_api_server.py
  • TestRepairedTranscriptDuplication: two regression tests covering a repaired transcript stored once and chained turns without exponential growth.
  • TestRunAgentTranscriptMode: unit test verifying _run_agent sets _transcript_mode="full".

How to Test

python -m pytest tests/gateway/test_api_server.py::TestRepairedTranscriptDuplication tests/gateway/test_api_server.py::TestRunAgentTranscriptMode -q

[teknium1]

Thanks for isolating the repaired-transcript storage path. The explicit marker matches the current agent contract: agent/turn_finalizer.py:573-578 returns the authoritative messages list, while gateway/platforms/api_server.py:5571-5574 currently prepends history after an exact-prefix miss.

Problems

• The same prefix miss still affects response output. gateway/platforms/api_server.py:5144-5152 obtains its output offset from _response_messages_turn_start_index; that helper returns 0 on a repaired prefix mismatch at gateway/platforms/api_server.py:5592-5599. _extract_output_items then scans the full transcript, so historical assistant/tool items can still be emitted as current-turn output. This is the shared-helper concern tracked by fix(api_server): robust Responses turn-start detection #70695.

Suggested changes

• Add an endpoint regression using adjacent user input items through gateway/platforms/api_server.py:4961-4966, with the mock applying the real repair helper before returning the transcript.
• Cover the resulting output boundary as well as stored history, or coordinate that portion with fix(api_server): robust Responses turn-start detection #70695.

This is an automated hermes-sweeper review.

[LiangYang666]

@teknium1 thanks for the review.

I've pushed a follow-up commit that addresses the output-side boundary:

• Extended _response_messages_turn_start_index to reverse-anchor on the last user message in the authoritative full transcript when _transcript_mode="full" is set and the exact-prefix comparison fails. This gives _extract_output_items the correct current-turn offset even after repair_message_sequence merges adjacent user messages.
• Added test_repaired_adjacent_user_input_output_boundary, an endpoint regression test that feeds adjacent user items through the input[] parser (as suggested) with a prior history containing tool calls. It asserts that the output array contains only the current turn's assistant message while the stored conversation_history remains the repaired full transcript without duplication.

Please take another look when you have a chance.

[GottZ]

This was generated by AI during triage.

Summary

Three PRs reference #68257. #68282 and part of #70695 address the reported but currently unverified leading-system transcript shape, while #75289 addresses a concrete repair-induced prefix mismatch by marking authoritative full transcripts and handling their output boundary.

Related pull requests

• fix(api): keep Responses history system-free 🩷 #68282 best fix — (+109/-9) — author action: rebase onto main, or split out the part that can merge. The diff strips leading system messages and avoids re-appending a user-prefixed suffix, but the current agent path cited in the maintainer-bot keep_open review does not produce that private-system transcript shape; the author should identify a current producer or re-scope the tests and fix to verified behavior.
• fix(api_server): robust Responses turn-start detection #70695 best fix — (+367/-8) — author action: rebase onto main, or split out the part that can merge. Its semantic prefix matching and output-boundary coverage target broader mismatch cases, but its unconditional removal of leading system messages can discard valid client-provided history; consistent with the maintainer-bot keep_open review, the salvageable mismatch/output work should be separated from the unverified system-stripping path.
• fix(api_server): trust full transcripts via _transcript_mode to stop repair-induced duplication #75289 best fix — (+250/-0) — keep open with a salvage path. The diff marks the normal AIAgent result as an authoritative full transcript, stores it verbatim after repair-induced prefix drift, and reverse-anchors output on the last user message; the follow-up adds output and chaining coverage, but the maintainer-bot keep_open review's requested end-to-end regression should still invoke the real repair helper rather than supplying an already-repaired mocked transcript.

Duplicates

#68282 and the leading-system handling in #70695 substantially duplicate the same unverified system-prefix mitigation. #75289 overlaps their history-doubling symptom but addresses a distinct, current repair-induced prefix mismatch rather than that claimed producer shape.

Suggested consolidation

Keep #75289 open with a salvage path: retain the explicit authoritative-transcript marker and storage/output-boundary fix, then strengthen the endpoint regression so it exercises the real repair helper. For #68282, request author action to identify a current leading-system producer or re-scope the patch; for #70695, request author action to split out the independently verifiable semantic mismatch and output-boundary work while removing role-only system stripping. Do not treat #75289 as resolving the unverified leading-system claim, and do not close #68282 or #70695 as exact duplicates of 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
    I68257(["issue #68257 (closed)"])
    P75289["PR #75289 (open)"]
    P75289 -->|best fix| I68257
    class I68257 closed
    class P75289 open
    class P75289 best
    class P75289 target
    click I68257 "https://github.com/NousResearch/hermes-agent/issues/68257"
    click P75289 "https://github.com/NousResearch/hermes-agent/pull/75289"

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 3 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 42 kB of PR diffs, 16 kB of issue/PR text, 1 kB of discussion (3 comments), 7 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.