feat(compression): native OpenAI Responses server-side compaction for gpt-5.6 - #81747
Merged
Merged
Conversation
… gpt-5.6
Opt-in via compression.codex_responses_native (default: false). When enabled,
gpt-5.6-family models on the direct OpenAI API (api.openai.com) or a ChatGPT
Codex subscription send context_management=[{type: compaction,
compact_threshold: N}] on Responses requests. OpenAI compacts server-side and
returns an encrypted compaction output item; Hermes captures it into the
existing codex_reasoning_items sidecar and replays it on later turns in place
of the pruned history — inheriting persistence, session replay, the
cross-issuer guard, and the encrypted-replay kill switch with zero new state.
Scope is deliberately hard-gated (agent/native_compaction.py, re-checked per
request): gpt-5.6 family only — gpt-5.1/5.2 fail server-side on the field
(HTTP 500 / stream stall, no structured rejection; live-verified) — and
direct OpenAI/Codex routes only; xAI, GitHub/Copilot, OpenRouter, relays,
and local servers never see the field.
Hermes' local compression stays armed as the fallback owner: the native
threshold is clamped ~8K tokens below the local trigger so the server
compacts first, and a structured provider rejection of context_management
disables native compaction for the session and retries without it
(one-shot guard in TurnRetryState).
Live-verified E2E on api.openai.com/gpt-5.6: server compaction fired at a
4K threshold, checkpoints captured and replayed, recall preserved across
3 turns; gpt-5.1 with the flag enabled stays clean (field never sent).
Direction credit: PR #76950 by @laryhorb explored native Responses
compaction; this is a minimal reimplementation on current main.
Contributor
૮ >ﻌ< ა ci reviewran on 690b953
|
alt-glitch
added
type/feature
Contributor
This was referenced Aug 8, 2026
teknium1
added a commit
that referenced
this pull request
Aug 8, 2026
…heckpoints Two corrections on top of the #71077 base (the whole bug class): 1. Turn boundary = last USER message, not last assistant message. A Codex turn spans several assistant messages (assistant+tool_calls -> tool -> ... -> final assistant) whose reasoning items must replay together; the last-assistant boundary would strip reasoning mid-chain from the active turn (the gap flagged in PR #71077 review). 2. type="compaction" checkpoints (native server-side compaction, PR #81747) are exempt: they carry already-pruned history, not per-turn reasoning. Pruning filters items instead of popping the sidecar key. Sibling site fixed in the same class: the Codex incomplete-continuation dedup path blind-overwrote codex_reasoning_items on visually-duplicate interim messages, which would drop the only copy of a checkpoint captured on the earlier response. Extracted merge_interim_reasoning_items() into agent/native_compaction.py; newer reasoning wins, prior checkpoints are preserved unless the newer payload carries its own.
teknium1
added a commit
that referenced
this pull request
Aug 8, 2026
…heckpoints Two corrections on top of the #71077 base (the whole bug class): 1. Turn boundary = last USER message, not last assistant message. A Codex turn spans several assistant messages (assistant+tool_calls -> tool -> ... -> final assistant) whose reasoning items must replay together; the last-assistant boundary would strip reasoning mid-chain from the active turn (the gap flagged in PR #71077 review). 2. type="compaction" checkpoints (native server-side compaction, PR #81747) are exempt: they carry already-pruned history, not per-turn reasoning. Pruning filters items instead of popping the sidecar key. Sibling site fixed in the same class: the Codex incomplete-continuation dedup path blind-overwrote codex_reasoning_items on visually-duplicate interim messages, which would drop the only copy of a checkpoint captured on the earlier response. Extracted merge_interim_reasoning_items() into agent/native_compaction.py; newer reasoning wins, prior checkpoints are preserved unless the newer payload carries its own.
This was referenced Aug 8, 2026
Merged
teknium1
added a commit
that referenced
this pull request
Aug 9, 2026
…e newest assistant turn (#73624) Generic thinking fields (reasoning / reasoning_content + the reasoning_details text charge) are replayed for at most the NEWEST assistant turn on every transport: Anthropic strips all-but-newest at convert time, Bedrock Converse never replays thinking, and strict chat-completions providers reject or one-space-pad the field. The tail budget walks charged them on every message anyway, spending 19-24% of the budget (per the issue's 1,025-message measurement) on bytes that provably never reach the wire — so the tail cut landed early and each compaction discarded more real transcript than configured. _estimate_msg_budget_tokens now partitions the replay keys: * _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items, codex_message_items) — charged unconditionally. These ride the wire on every retained turn (#55572), and codex_reasoning_items now also carries native server-side compaction checkpoints (#81747). * _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the reasoning_details text charge — charged only for the newest assistant turn via charge_stale_thinking, resolved by the three budget walks (tail cut, raw-budget re-walk, proactive-prune boundary). Default stays the conservative full charge for callers without turn-position context. A partition invariant test pins that any future _REPLAY_BUDGET_KEYS entry must be classified into exactly one class. Direction credit: #73669 (@x7peeps) and #73730 (@webtecnica) both attacked this; the keep_open reviews asked for provider/API-mode-aware accounting that keeps Codex carriers charged — this implements that shape.
teknium1
added a commit
that referenced
this pull request
Aug 9, 2026
…e newest assistant turn (#73624) Generic thinking fields (reasoning / reasoning_content + the reasoning_details text charge) are replayed for at most the NEWEST assistant turn on every transport: Anthropic strips all-but-newest at convert time, Bedrock Converse never replays thinking, and strict chat-completions providers reject or one-space-pad the field. The tail budget walks charged them on every message anyway, spending 19-24% of the budget (per the issue's 1,025-message measurement) on bytes that provably never reach the wire — so the tail cut landed early and each compaction discarded more real transcript than configured. _estimate_msg_budget_tokens now partitions the replay keys: * _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items, codex_message_items) — charged unconditionally. These ride the wire on every retained turn (#55572), and codex_reasoning_items now also carries native server-side compaction checkpoints (#81747). * _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the reasoning_details text charge — charged only for the newest assistant turn via charge_stale_thinking, resolved by the three budget walks (tail cut, raw-budget re-walk, proactive-prune boundary). Default stays the conservative full charge for callers without turn-position context. A partition invariant test pins that any future _REPLAY_BUDGET_KEYS entry must be classified into exactly one class. Direction credit: #73669 (@x7peeps) and #73730 (@webtecnica) both attacked this; the keep_open reviews asked for provider/API-mode-aware accounting that keeps Codex carriers charged — this implements that shape.
ma1138569845
pushed a commit
to ma1138569845/dechnicAuditor-agent
that referenced
this pull request
Aug 10, 2026
…heckpoints Two corrections on top of the NousResearch#71077 base (the whole bug class): 1. Turn boundary = last USER message, not last assistant message. A Codex turn spans several assistant messages (assistant+tool_calls -> tool -> ... -> final assistant) whose reasoning items must replay together; the last-assistant boundary would strip reasoning mid-chain from the active turn (the gap flagged in PR NousResearch#71077 review). 2. type="compaction" checkpoints (native server-side compaction, PR NousResearch#81747) are exempt: they carry already-pruned history, not per-turn reasoning. Pruning filters items instead of popping the sidecar key. Sibling site fixed in the same class: the Codex incomplete-continuation dedup path blind-overwrote codex_reasoning_items on visually-duplicate interim messages, which would drop the only copy of a checkpoint captured on the earlier response. Extracted merge_interim_reasoning_items() into agent/native_compaction.py; newer reasoning wins, prior checkpoints are preserved unless the newer payload carries its own.
ma1138569845
pushed a commit
to ma1138569845/dechnicAuditor-agent
that referenced
this pull request
Aug 10, 2026
…e newest assistant turn (NousResearch#73624) Generic thinking fields (reasoning / reasoning_content + the reasoning_details text charge) are replayed for at most the NEWEST assistant turn on every transport: Anthropic strips all-but-newest at convert time, Bedrock Converse never replays thinking, and strict chat-completions providers reject or one-space-pad the field. The tail budget walks charged them on every message anyway, spending 19-24% of the budget (per the issue's 1,025-message measurement) on bytes that provably never reach the wire — so the tail cut landed early and each compaction discarded more real transcript than configured. _estimate_msg_budget_tokens now partitions the replay keys: * _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items, codex_message_items) — charged unconditionally. These ride the wire on every retained turn (NousResearch#55572), and codex_reasoning_items now also carries native server-side compaction checkpoints (NousResearch#81747). * _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the reasoning_details text charge — charged only for the newest assistant turn via charge_stale_thinking, resolved by the three budget walks (tail cut, raw-budget re-walk, proactive-prune boundary). Default stays the conservative full charge for callers without turn-position context. A partition invariant test pins that any future _REPLAY_BUDGET_KEYS entry must be classified into exactly one class. Direction credit: NousResearch#73669 (@x7peeps) and NousResearch#73730 (@webtecnica) both attacked this; the keep_open reviews asked for provider/API-mode-aware accounting that keeps Codex carriers charged — this implements that shape.
randlee
pushed a commit
to randlee/hermes-agent
that referenced
this pull request
Aug 11, 2026
…heckpoints Two corrections on top of the NousResearch#71077 base (the whole bug class): 1. Turn boundary = last USER message, not last assistant message. A Codex turn spans several assistant messages (assistant+tool_calls -> tool -> ... -> final assistant) whose reasoning items must replay together; the last-assistant boundary would strip reasoning mid-chain from the active turn (the gap flagged in PR NousResearch#71077 review). 2. type="compaction" checkpoints (native server-side compaction, PR NousResearch#81747) are exempt: they carry already-pruned history, not per-turn reasoning. Pruning filters items instead of popping the sidecar key. Sibling site fixed in the same class: the Codex incomplete-continuation dedup path blind-overwrote codex_reasoning_items on visually-duplicate interim messages, which would drop the only copy of a checkpoint captured on the earlier response. Extracted merge_interim_reasoning_items() into agent/native_compaction.py; newer reasoning wins, prior checkpoints are preserved unless the newer payload carries its own.
randlee
pushed a commit
to randlee/hermes-agent
that referenced
this pull request
Aug 11, 2026
…e newest assistant turn (NousResearch#73624) Generic thinking fields (reasoning / reasoning_content + the reasoning_details text charge) are replayed for at most the NEWEST assistant turn on every transport: Anthropic strips all-but-newest at convert time, Bedrock Converse never replays thinking, and strict chat-completions providers reject or one-space-pad the field. The tail budget walks charged them on every message anyway, spending 19-24% of the budget (per the issue's 1,025-message measurement) on bytes that provably never reach the wire — so the tail cut landed early and each compaction discarded more real transcript than configured. _estimate_msg_budget_tokens now partitions the replay keys: * _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items, codex_message_items) — charged unconditionally. These ride the wire on every retained turn (NousResearch#55572), and codex_reasoning_items now also carries native server-side compaction checkpoints (NousResearch#81747). * _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the reasoning_details text charge — charged only for the newest assistant turn via charge_stale_thinking, resolved by the three budget walks (tail cut, raw-budget re-walk, proactive-prune boundary). Default stays the conservative full charge for callers without turn-position context. A partition invariant test pins that any future _REPLAY_BUDGET_KEYS entry must be classified into exactly one class. Direction credit: NousResearch#73669 (@x7peeps) and NousResearch#73730 (@webtecnica) both attacked this; the keep_open reviews asked for provider/API-mode-aware accounting that keeps Codex carriers charged — this implements that shape.
blut-agent
pushed a commit
to blut-agent/hermes-agent-fork
that referenced
this pull request
Aug 11, 2026
…heckpoints Two corrections on top of the NousResearch#71077 base (the whole bug class): 1. Turn boundary = last USER message, not last assistant message. A Codex turn spans several assistant messages (assistant+tool_calls -> tool -> ... -> final assistant) whose reasoning items must replay together; the last-assistant boundary would strip reasoning mid-chain from the active turn (the gap flagged in PR NousResearch#71077 review). 2. type="compaction" checkpoints (native server-side compaction, PR NousResearch#81747) are exempt: they carry already-pruned history, not per-turn reasoning. Pruning filters items instead of popping the sidecar key. Sibling site fixed in the same class: the Codex incomplete-continuation dedup path blind-overwrote codex_reasoning_items on visually-duplicate interim messages, which would drop the only copy of a checkpoint captured on the earlier response. Extracted merge_interim_reasoning_items() into agent/native_compaction.py; newer reasoning wins, prior checkpoints are preserved unless the newer payload carries its own.
blut-agent
pushed a commit
to blut-agent/hermes-agent-fork
that referenced
this pull request
Aug 11, 2026
…e newest assistant turn (NousResearch#73624) Generic thinking fields (reasoning / reasoning_content + the reasoning_details text charge) are replayed for at most the NEWEST assistant turn on every transport: Anthropic strips all-but-newest at convert time, Bedrock Converse never replays thinking, and strict chat-completions providers reject or one-space-pad the field. The tail budget walks charged them on every message anyway, spending 19-24% of the budget (per the issue's 1,025-message measurement) on bytes that provably never reach the wire — so the tail cut landed early and each compaction discarded more real transcript than configured. _estimate_msg_budget_tokens now partitions the replay keys: * _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items, codex_message_items) — charged unconditionally. These ride the wire on every retained turn (NousResearch#55572), and codex_reasoning_items now also carries native server-side compaction checkpoints (NousResearch#81747). * _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the reasoning_details text charge — charged only for the newest assistant turn via charge_stale_thinking, resolved by the three budget walks (tail cut, raw-budget re-walk, proactive-prune boundary). Default stays the conservative full charge for callers without turn-position context. A partition invariant test pins that any future _REPLAY_BUDGET_KEYS entry must be classified into exactly one class. Direction credit: NousResearch#73669 (@x7peeps) and NousResearch#73730 (@webtecnica) both attacked this; the keep_open reviews asked for provider/API-mode-aware accounting that keeps Codex carriers charged — this implements that shape.
33hodl
pushed a commit
to 33hodl/hermes-agent
that referenced
this pull request
Aug 12, 2026
…heckpoints Two corrections on top of the NousResearch#71077 base (the whole bug class): 1. Turn boundary = last USER message, not last assistant message. A Codex turn spans several assistant messages (assistant+tool_calls -> tool -> ... -> final assistant) whose reasoning items must replay together; the last-assistant boundary would strip reasoning mid-chain from the active turn (the gap flagged in PR NousResearch#71077 review). 2. type="compaction" checkpoints (native server-side compaction, PR NousResearch#81747) are exempt: they carry already-pruned history, not per-turn reasoning. Pruning filters items instead of popping the sidecar key. Sibling site fixed in the same class: the Codex incomplete-continuation dedup path blind-overwrote codex_reasoning_items on visually-duplicate interim messages, which would drop the only copy of a checkpoint captured on the earlier response. Extracted merge_interim_reasoning_items() into agent/native_compaction.py; newer reasoning wins, prior checkpoints are preserved unless the newer payload carries its own.
33hodl
pushed a commit
to 33hodl/hermes-agent
that referenced
this pull request
Aug 12, 2026
…e newest assistant turn (NousResearch#73624) Generic thinking fields (reasoning / reasoning_content + the reasoning_details text charge) are replayed for at most the NEWEST assistant turn on every transport: Anthropic strips all-but-newest at convert time, Bedrock Converse never replays thinking, and strict chat-completions providers reject or one-space-pad the field. The tail budget walks charged them on every message anyway, spending 19-24% of the budget (per the issue's 1,025-message measurement) on bytes that provably never reach the wire — so the tail cut landed early and each compaction discarded more real transcript than configured. _estimate_msg_budget_tokens now partitions the replay keys: * _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items, codex_message_items) — charged unconditionally. These ride the wire on every retained turn (NousResearch#55572), and codex_reasoning_items now also carries native server-side compaction checkpoints (NousResearch#81747). * _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the reasoning_details text charge — charged only for the newest assistant turn via charge_stale_thinking, resolved by the three budget walks (tail cut, raw-budget re-walk, proactive-prune boundary). Default stays the conservative full charge for callers without turn-position context. A partition invariant test pins that any future _REPLAY_BUDGET_KEYS entry must be classified into exactly one class. Direction credit: NousResearch#73669 (@x7peeps) and NousResearch#73730 (@webtecnica) both attacked this; the keep_open reviews asked for provider/API-mode-aware accounting that keeps Codex carriers charged — this implements that shape.
This was referenced Aug 28, 2026
melon-xf
added a commit
to melon-xf/hermes-agent
that referenced
this pull request
Sep 3, 2026
…heckpoints Two corrections on top of the NousResearch#71077 base (the whole bug class): 1. Turn boundary = last USER message, not last assistant message. A Codex turn spans several assistant messages (assistant+tool_calls -> tool -> ... -> final assistant) whose reasoning items must replay together; the last-assistant boundary would strip reasoning mid-chain from the active turn (the gap flagged in PR NousResearch#71077 review). 2. type="compaction" checkpoints (native server-side compaction, PR NousResearch#81747) are exempt: they carry already-pruned history, not per-turn reasoning. Pruning filters items instead of popping the sidecar key. Sibling site fixed in the same class: the Codex incomplete-continuation dedup path blind-overwrote codex_reasoning_items on visually-duplicate interim messages, which would drop the only copy of a checkpoint captured on the earlier response. Extracted merge_interim_reasoning_items() into agent/native_compaction.py; newer reasoning wins, prior checkpoints are preserved unless the newer payload carries its own.
melon-xf
added a commit
to melon-xf/hermes-agent
that referenced
this pull request
Sep 3, 2026
…e newest assistant turn (NousResearch#73624) Generic thinking fields (reasoning / reasoning_content + the reasoning_details text charge) are replayed for at most the NEWEST assistant turn on every transport: Anthropic strips all-but-newest at convert time, Bedrock Converse never replays thinking, and strict chat-completions providers reject or one-space-pad the field. The tail budget walks charged them on every message anyway, spending 19-24% of the budget (per the issue's 1,025-message measurement) on bytes that provably never reach the wire — so the tail cut landed early and each compaction discarded more real transcript than configured. _estimate_msg_budget_tokens now partitions the replay keys: * _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items, codex_message_items) — charged unconditionally. These ride the wire on every retained turn (NousResearch#55572), and codex_reasoning_items now also carries native server-side compaction checkpoints (NousResearch#81747). * _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the reasoning_details text charge — charged only for the newest assistant turn via charge_stale_thinking, resolved by the three budget walks (tail cut, raw-budget re-walk, proactive-prune boundary). Default stays the conservative full charge for callers without turn-position context. A partition invariant test pins that any future _REPLAY_BUDGET_KEYS entry must be classified into exactly one class. Direction credit: NousResearch#73669 (@x7peeps) and NousResearch#73730 (@webtecnica) both attacked this; the keep_open reviews asked for provider/API-mode-aware accounting that keeps Codex carriers charged — this implements that shape.
smfworks
pushed a commit
to smfworks/hermes-agent
that referenced
this pull request
Sep 7, 2026
…heckpoints Two corrections on top of the NousResearch#71077 base (the whole bug class): 1. Turn boundary = last USER message, not last assistant message. A Codex turn spans several assistant messages (assistant+tool_calls -> tool -> ... -> final assistant) whose reasoning items must replay together; the last-assistant boundary would strip reasoning mid-chain from the active turn (the gap flagged in PR NousResearch#71077 review). 2. type="compaction" checkpoints (native server-side compaction, PR NousResearch#81747) are exempt: they carry already-pruned history, not per-turn reasoning. Pruning filters items instead of popping the sidecar key. Sibling site fixed in the same class: the Codex incomplete-continuation dedup path blind-overwrote codex_reasoning_items on visually-duplicate interim messages, which would drop the only copy of a checkpoint captured on the earlier response. Extracted merge_interim_reasoning_items() into agent/native_compaction.py; newer reasoning wins, prior checkpoints are preserved unless the newer payload carries its own.
smfworks
pushed a commit
to smfworks/hermes-agent
that referenced
this pull request
Sep 7, 2026
…e newest assistant turn (NousResearch#73624) Generic thinking fields (reasoning / reasoning_content + the reasoning_details text charge) are replayed for at most the NEWEST assistant turn on every transport: Anthropic strips all-but-newest at convert time, Bedrock Converse never replays thinking, and strict chat-completions providers reject or one-space-pad the field. The tail budget walks charged them on every message anyway, spending 19-24% of the budget (per the issue's 1,025-message measurement) on bytes that provably never reach the wire — so the tail cut landed early and each compaction discarded more real transcript than configured. _estimate_msg_budget_tokens now partitions the replay keys: * _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items, codex_message_items) — charged unconditionally. These ride the wire on every retained turn (NousResearch#55572), and codex_reasoning_items now also carries native server-side compaction checkpoints (NousResearch#81747). * _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the reasoning_details text charge — charged only for the newest assistant turn via charge_stale_thinking, resolved by the three budget walks (tail cut, raw-budget re-walk, proactive-prune boundary). Default stays the conservative full charge for callers without turn-position context. A partition invariant test pins that any future _REPLAY_BUDGET_KEYS entry must be classified into exactly one class. Direction credit: NousResearch#73669 (@x7peeps) and NousResearch#73730 (@webtecnica) both attacked this; the keep_open reviews asked for provider/API-mode-aware accounting that keeps Codex carriers charged — this implements that shape.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
gpt-5.6 sessions on the direct OpenAI API or a ChatGPT Codex subscription can now opt in to OpenAI's server-side compaction (
compression.codex_responses_native: true, default false) — the server prunes older context into an encrypted checkpoint that Hermes replays on later turns, and no other model, provider, or route is affected.Root motivation: OpenAI's Responses API added
context_management: [{type: "compaction", compact_threshold: N}](July 2026). Draft PR #76950 by @laryhorb explored this; it was a 9.3k-line stale-base rewrite with a live hang bug on non-supporting models, so this is a minimal reimplementation on current main (direction credit to @laryhorb).Changes
agent/native_compaction.py(new, ~170 LOC): the entire policy — model gate (gpt-5.6 family only), route gate (api.openai.com / Codex backend only), threshold clamp (native fires ~8K tokens below the local compression trigger so the server compacts first), and a narrow structured-rejection matcher.agent/transports/codex.py+agent/chat_completion_helpers.py: thread the resolvedcontext_managementpayload into Responses request kwargs (None ⇒ request byte-identical to before).agent/codex_responses_adapter.py: capturecompactionoutput items into the existingcodex_reasoning_itemssidecar (inherits persistence, session replay, cross-issuer stamping, and the encrypted-replay kill switch — zero new state); accept replayed compaction items in both preflights.agent/conversation_loop.py+agent/turn_retry_state.py: one-shot recovery — a structured 400 namingcontext_managementdisables native compaction for the session and retries without the field. Generic 5xx/timeouts intentionally do NOT downgrade.agent/agent_init.py,hermes_cli/config_defaults.py,cli-config.yaml.example: config plumbing (codex_responses_native,codex_responses_compact_threshold).website/docs/developer-guide/context-compression-and-caching.md.tests/run_agent/test_native_compaction.py(31 tests; sabotage-verified — removing the model gate or route gate each fails the suite).Why gpt-5.6 only
Live testing (Aug 2026): gpt-5.6 compacts correctly; gpt-5.1/5.2 with the field fail server-side as HTTP 500 (blocking) or a permanent stream stall (90s watchdog × 3 retries = dead turn) with no structured rejection to downgrade on. An explicit model-family gate is the only safe shape.
Validation
Infographic