fix(bedrock): add Claude Sonnet/Opus/Haiku 5 to Bedrock context-length table by Polyhistor · Pull Request #66167 · NousResearch/hermes-agent

Polyhistor · 2026-07-17T08:14:12Z

Problem

au.anthropic.claude-sonnet-5 (and the opus-5/haiku-5 siblings) have no entry in BEDROCK_CONTEXT_LENGTHS in agent/bedrock_adapter.py. Since Bedrock's ListFoundationModels API doesn't expose context window sizes, get_bedrock_context_length() falls back to BEDROCK_DEFAULT_CONTEXT_LENGTH (128,000) for any unmatched model — silently under-reporting the real window.

AWS documents a 200,000 token default context window for Claude Sonnet 5 on Bedrock (same as the Sonnet 4.5/Opus 4 family already in the table), with a 1M beta available via header — not the 128K fallback.

Impact

ContextCompressor applies a small-context floor: any model resolved under 512K is treated as 'small' and its compaction threshold is raised to 75% of window (to avoid thrashing on tight windows). With the wrong 128K window this floor fires at 96,000 tokens; at the correct 200K window it fires at 150,000 tokens — about 36% less usable headroom before Hermes force-compacts the conversation. On long agentic sessions (e.g. heavy read_file/grep/web_search tool output) this causes materially more frequent compaction than intended.

Fix

Add anthropic.claude-opus-5, anthropic.claude-sonnet-5, anthropic.claude-haiku-5 → 200,000 to BEDROCK_CONTEXT_LENGTHS. Minimal, additive, no behavior change for existing entries.

Verification

>>> from agent.bedrock_adapter import get_bedrock_context_length
>>> get_bedrock_context_length('au.anthropic.claude-sonnet-5')
200000   # was 128000 before this change

Also confirmed downstream in ContextCompressor: with this fix (or an equivalent model.context_length: 200000 config override), threshold_tokens for this model correctly resolves to 150,000 instead of 96,000.

…h table au.anthropic.claude-sonnet-5 (and opus-5/haiku-5) had no entry in BEDROCK_CONTEXT_LENGTHS, so get_bedrock_context_length() silently fell back to BEDROCK_DEFAULT_CONTEXT_LENGTH (128_000) instead of the actual 200_000 token window AWS documents for this model family on Bedrock. Impact: ContextCompressor's small-context floor treats any resolved window under 512K as 'small' and raises the compaction threshold to 75% of window. With the wrong 128K window that floor fires at 96,000 tokens instead of 150,000 tokens at the correct 200K window -- roughly 36% less usable headroom before Hermes force-compacts the conversation, causing more frequent/premature compaction on long agentic sessions. Verified locally: get_bedrock_context_length('au.anthropic.claude-sonnet-5') -> 200000 (was 128000 before this change)

alt-glitch · 2026-07-17T08:20:17Z

This was generated by AI during triage.

Related: #24059 corrects Claude 4.x 1M Bedrock entries; this independently adds the 200K Claude 5 identifiers. Not a duplicate.

tonydwb

Code Review Summary

Verdict: Approved

Context length table update for Claude 5 family on Bedrock. Simple safe addition. No security concerns.

Reviewed by Hermes Agent

Bedrock Converse rejects text content blocks that are empty OR whitespace-only (ValidationException: "text content blocks must contain non-whitespace text"). The prior fix attempt substituted a single space (" ") for missing content -- but a lone space IS whitespace, so it was rejected by the exact same validation rule it was meant to satisfy. This caused a deterministic, unrecoverable retry-loop failure once any blank/whitespace assistant, tool, or user turn entered history (most commonly via context-compaction rewriting a turn to a blank string). Adds _safe_text()/_EMPTY_TEXT_PLACEHOLDER ("(empty)") and applies it everywhere a blank text block could reach the wire: user/assistant content conversion, tool results, the assistant-empty-turn fallback, and the first/last-message user-alternation padding. System-prompt blocks are the one exception: blank parts are dropped entirely rather than placeholder-filled, since a system prompt block should never carry meaningless placeholder text. Adds tests/agent/test_bedrock_empty_text_blocks.py (11 tests, was already present uncommitted -- codifies the exact failing history from issue NousResearch#9486 and asserts no blank block ever reaches Bedrock). Verified against the actual failed request dump from this session (27-message payload) -- replaying it through the fixed converter now produces zero blank/whitespace-only blocks.

teknium1

Thanks for identifying the missing Claude 5 Bedrock context entries. The premise is present on current main: agent/model_metadata.py:2259-2266 routes Bedrock models to get_bedrock_context_length(), while agent/bedrock_adapter.py:1307-1350 has no Claude 5 keys and returns the 128K default for unmatched IDs.

Problems

The placeholder change will fail current tests: tests/agent/test_bedrock_adapter.py:1310-1323 asserts that None, empty, and whitespace-only content produce a literal space, while this PR changes those outputs to (empty) without updating the assertions.
The new _safe_text() is not applied to supported string entries in list content. agent/bedrock_adapter.py:511-512 still emits {"text": part} directly, so [" "] can still send a whitespace-only block to Bedrock.

Suggested changes

Update the existing placeholder assertions and route list string items through _safe_text(); add a list-string whitespace regression case.

Automated hermes-sweeper review.

teknium1 · 2026-07-18T21:58:39Z

+# Bedrock's Converse API rejects any text content block whose text is empty
+# OR whitespace-only (ValidationException: "text content blocks must contain
+# non-whitespace text"). A lone space is whitespace and is rejected too — the
+# placeholder MUST itself be non-whitespace. Ref: issue #9486.


Please update tests/agent/test_bedrock_adapter.py:1310-1323 in this PR: those existing tests still assert the old literal-space placeholder and will fail after this constant changes the converter output.

Addressed in bbb233e:

Updated tests/agent/test_bedrock_adapter.py::TestEmptyTextBlockFix — the placeholder assertions were stale (still expecting the old literal-space behavior); now assert _EMPTY_TEXT_PLACEHOLDER (non-whitespace) instead.
Fixed _convert_content_to_converse(): plain-string items inside a content list (e.g. [" "]) were emitted as {"text": part}" directly, bypassing _safe_text(). Now routed through _safe_text() like every other path. Added a regression test (test_whitespace_only_list_string_item_gets_placeholder`) plus a happy-path counterpart.
Full adapter test suite: 135 passed, 10 skipped (pre-existing, unrelated botocore-version-gated skips).

…text, update stale placeholder assertions Addresses hermes-sweeper review on PR NousResearch#66167: 1. _convert_content_to_converse() still emitted {"text": part} directly for plain-string items inside a content list (as opposed to {"type": "text"} dicts), bypassing _safe_text() entirely. A whitespace-only string item (e.g. [" "]) could still reach Bedrock as a blank block. Now routed through _safe_text(). 2. tests/agent/test_bedrock_adapter.py::TestEmptyTextBlockFix asserted the pre-fix behavior (whitespace -> literal space " "), contradicting the new _safe_text()/_EMPTY_TEXT_PLACEHOLDER behavior added in 4618095. Updated assertions to expect the non-whitespace placeholder, plus added a regression test for the list-string-item case above.

Polyhistor · 2026-07-18T22:59:24Z

Addressed in bbb233e:

Updated tests/agent/test_bedrock_adapter.py::TestEmptyTextBlockFix — the placeholder assertions were stale (still expecting the old literal-space behavior); now assert _EMPTY_TEXT_PLACEHOLDER (non-whitespace) instead.
Fixed _convert_content_to_converse(): plain-string items inside a content list (e.g. [" "]) were emitted as {"text": part}" directly, bypassing _safe_text(). Now routed through _safe_text() like every other path. Added a regression test (test_whitespace_only_list_string_item_gets_placeholder`) plus a happy-path counterpart.

Full adapter test suite: 135 passed, 10 skipped (pre-existing, unrelated botocore-version-gated skips).

…text, update stale placeholder assertions Addresses hermes-sweeper review on PR #66167: 1. _convert_content_to_converse() still emitted {"text": part} directly for plain-string items inside a content list (as opposed to {"type": "text"} dicts), bypassing _safe_text() entirely. A whitespace-only string item (e.g. [" "]) could still reach Bedrock as a blank block. Now routed through _safe_text(). 2. tests/agent/test_bedrock_adapter.py::TestEmptyTextBlockFix asserted the pre-fix behavior (whitespace -> literal space " "), contradicting the new _safe_text()/_EMPTY_TEXT_PLACEHOLDER behavior added in 4618095. Updated assertions to expect the non-whitespace placeholder, plus added a regression test for the list-string-item case above.

teknium1 · 2026-07-20T10:28:39Z

Merged the fix half via PR #67978 — both of your _safe_text commits were cherry-picked onto current main with your authorship preserved (rebase merge), including the review follow-up routing list-string items through the guard. Nice catch that the lone-space placeholder was itself whitespace.

The context-table half of this PR was superseded by #67977: current main already carried anthropic.claude-sonnet-5 at 1M (from #67932), so the 200K row would have been a regression, and the opus-5/haiku-5 rows were speculative (those model IDs don't exist on Bedrock yet — happy to add them when AWS publishes the model cards). Thanks!

#67978

…text, update stale placeholder assertions Addresses hermes-sweeper review on PR NousResearch#66167: 1. _convert_content_to_converse() still emitted {"text": part} directly for plain-string items inside a content list (as opposed to {"type": "text"} dicts), bypassing _safe_text() entirely. A whitespace-only string item (e.g. [" "]) could still reach Bedrock as a blank block. Now routed through _safe_text(). 2. tests/agent/test_bedrock_adapter.py::TestEmptyTextBlockFix asserted the pre-fix behavior (whitespace -> literal space " "), contradicting the new _safe_text()/_EMPTY_TEXT_PLACEHOLDER behavior added in 4618095. Updated assertions to expect the non-whitespace placeholder, plus added a regression test for the list-string-item case above.

tonydwb reviewed Jul 17, 2026

View reviewed changes

teknium1 reviewed Jul 18, 2026

View reviewed changes

teknium1 added the sweeper:blast-moderate Sweeper blast radius: moderate — a subsystem or single platform label Jul 18, 2026

This was referenced Jul 20, 2026

fix(bedrock): 1M context for current-gen Claude on Bedrock (salvage #24059/#54918) #67977

Merged

fix(bedrock): non-whitespace placeholder for empty Converse text blocks (salvage #66167) #67978

Merged

teknium1 closed this in #67978 Jul 20, 2026

GottZ mentioned this pull request Jul 31, 2026

[Bug]: Native Bedrock Converse adapter does not forward context-1m beta — Opus 4.6/4.7 stuck at 200K (PR #16793 only patched OpenAI-compat path) #31277

Open

Navigation Menu

Uh oh!

fix(bedrock): add Claude Sonnet/Opus/Haiku 5 to Bedrock context-length table - #66167

fix(bedrock): add Claude Sonnet/Opus/Haiku 5 to Bedrock context-length table#66167
Polyhistor wants to merge 3 commits into
NousResearch:mainfrom
Polyhistor:fix/bedrock-sonnet5-context-length

Polyhistor commented Jul 17, 2026

Uh oh!

alt-glitch commented Jul 17, 2026

Uh oh!

tonydwb left a comment

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 18, 2026

Uh oh!

Polyhistor Jul 18, 2026

Uh oh!

Polyhistor commented Jul 18, 2026

Uh oh!

teknium1 commented Jul 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

Polyhistor commented Jul 17, 2026

Problem

Impact

Fix

Verification

Uh oh!

alt-glitch commented Jul 17, 2026

Uh oh!

tonydwb left a comment

Choose a reason for hiding this comment

Code Review Summary

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jul 18, 2026

Choose a reason for hiding this comment

Uh oh!

Polyhistor Jul 18, 2026

Choose a reason for hiding this comment

Uh oh!

Polyhistor commented Jul 18, 2026

Uh oh!

teknium1 commented Jul 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants