fix(bedrock): align Claude context-window table with Anthropic docs

[patrick-muller]

What does this PR do?

The BEDROCK_CONTEXT_LENGTHS table in agent/bedrock_adapter.py reported
stale values for recent Anthropic Claude models on Bedrock:

• No entry for claude-opus-4-7 — falls through via substring match to
  the generic claude-opus-4 entry at 200K, i.e. 5x too small.
• claude-opus-4-6 reported 200K, but is 1M generally available.
• claude-sonnet-4-6 reported 200K, but is 1M generally available.

This table feeds get_bedrock_context_length() → get_model_context_length(),
which drives the displayed context window in /usage, the compression
threshold (compression.threshold * context_length), and subagent
context budgets. A 5x-too-small value causes premature context
compression and under-utilizes the context window users are paying for.

This PR is metadata-only — no API routing, transport behavior, or
beta-header handling is changed. Claude-on-Bedrock continues to route
through AnthropicBedrock SDK via api_mode: anthropic_messages exactly
as before.

Values per Anthropic's official models comparison

Source: https://platform.claude.com/docs/en/about-claude/models/overview
(the "Latest models comparison" table, column: "Context window"):

Bedrock model ID	Before	After	Anthropic docs
anthropic.claude-opus-4-7	(missing)	1_000_000	1M tokens
anthropic.claude-opus-4-6	200_000	1_000_000	1M tokens
anthropic.claude-sonnet-4-6	200_000	1_000_000	1M tokens
anthropic.claude-sonnet-4-5	200_000	200_000 (unchanged)	200K (1M beta retired Apr 30 2026)
anthropic.claude-haiku-4-5	200_000	200_000 (unchanged)	200k tokens

Why Sonnet 4.5 and Haiku 4.5 remain at 200K — and claude-opus-4 /
claude-sonnet-4 / claude-3-* also remain at 200K:

• Sonnet 4.5: per the April 30, 2026 release note,

  "We've retired the 1M token context window beta (context-1m-2025-08-07)
  for Claude Sonnet 4.5 and Claude Sonnet 4. The beta header now has no
  effect on these models, and requests exceeding the standard 200k-token
  context window return an error."

• Haiku 4.5: listed as 200k tokens in the Anthropic models comparison.
• Sonnet 4 / Opus 4 / Claude 3: never had 1M support.

Related Issue

No existing issue — discovered while running us.anthropic.claude-opus-4-7
on Bedrock and seeing a 200K window displayed instead of 1M.

Related (but distinct) bugs in the same resolution chain:

• get_model_context_length step 2 short-circuits to 128K default; resolve_display_context_length skips config_context_length #15563 — separate custom-endpoint short-circuit in
  get_model_context_length() step 2
• Add config flag to disable context-1m-2025-08-07 beta header for Anthropic Enterprise OAuth subscriptions #21557 — inverse concern: some Anthropic Enterprise OAuth accounts
  cannot use the 1M beta

Fixes #

Prior art

A previous PR — #16686 by @mmcclean-aws (opened April 27, 2026) — had
the same core insight and the same three source lines (add
claude-opus-4-7, bump claude-opus-4-6 and claude-sonnet-4-6 to 1M).
This PR is a superset of that work:

• Identical source diff to fix(bedrock): correct context lengths for Claude 4.x models to 1M tokens #16686 on those three entries.
• Adds tests — 3 updated assertions (opus-4-6, sonnet-4-6 versioned,
  inference-profile) + 3 new guard tests (opus-4-7 → 1M,
  sonnet-4-5 → 200K, haiku-4-5 → 200K) with docstring citations to
  Anthropic's sources.
• Adds inline Anthropic-doc citations in the BEDROCK_CONTEXT_LENGTHS
  header comment so the next contributor who touches this table has
  an upstream source of truth.
• Leaves claude-sonnet-4-5 and claude-haiku-4-5 at 200K explicitly
  (with commentary + guard tests), citing the April 30 2026 release
  note that retired the 1M beta for Sonnet 4.5. fix(bedrock): correct context lengths for Claude 4.x models to 1M tokens #16686 left those
  lines untouched with no commentary; my worry was that a future
  contributor might "helpfully" bump them alongside the others without
  knowing Anthropic has actually capped them at 200K now.

Happy for maintainers to close #16686 in favor of this PR, or to
incorporate @mmcclean-aws's commit with authorship preserved and layer
the tests/docs on top — whatever's easier for review.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• agent/bedrock_adapter.py (BEDROCK_CONTEXT_LENGTHS):
  • Add anthropic.claude-opus-4-7: 1_000_000
  • Change anthropic.claude-opus-4-6: 200_000 → 1_000_000
  • Change anthropic.claude-sonnet-4-6: 200_000 → 1_000_000
  • Update header comment with links to Anthropic's models overview and
    the April 30 2026 release note, so future readers have an upstream
    source of truth when this table needs another update.
• tests/agent/test_bedrock_adapter.py (TestBedrockContextLength):
  • Update test_claude_opus_4_6, test_claude_sonnet_versioned, and
    test_inference_profile_resolves to assert the new 1M values.
  • Add test_claude_opus_4_7 — guards the new entry + doc-links it.
  • Add test_claude_sonnet_4_5_is_200k — guards against regressions
    that would re-bump it, citing the retirement release note.
  • Add test_claude_haiku_4_5_is_200k — guards against regressions
    that would bump it to 1M.

agent/model_metadata.py's non-Bedrock DEFAULT_CONTEXT_LENGTHS table
already lists claude-opus-4-7, claude-opus-4-6, and
claude-sonnet-4-6 at 1M for the chat-completions / Anthropic direct
paths. This PR brings the Bedrock-specific fallback table into alignment.

How to Test

1. Run the focused test module:

   scripts/run_tests.sh tests/agent/test_bedrock_adapter.py tests/agent/test_bedrock_1m_context.py tests/agent/test_bedrock_integration.py -q
   

   Expected: 178 passed.

2. Manual verification with any Claude-on-Bedrock model:

   hermes --model us.anthropic.claude-opus-4-7 --provider bedrock
   > /usage
   

   Context window should now show ~1,000,000 (was 200,000).

3. Verify older Claude models (Sonnet 4 / Opus 4 / 3.x) still display 200K
   and that Haiku 4.5 / Sonnet 4.5 still display 200K — those are
   Anthropic's current context windows for those models, not stale values.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(bedrock):, test(bedrock):)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix (no unrelated commits)
• I've run scripts/run_tests.sh tests/agent/test_bedrock_adapter.py tests/agent/test_bedrock_1m_context.py tests/agent/test_bedrock_integration.py -q — 178 passed
• I've added tests for my changes (3 new tests doc-linking Anthropic's published context windows)
• I've tested on my platform: macOS 26.3 (Apple Silicon), Python 3.11, against us.anthropic.claude-opus-4-7 on Bedrock us-east-1

Documentation & Housekeeping

• I've updated relevant documentation — header comment in
  BEDROCK_CONTEXT_LENGTHS now links to Anthropic's models overview and
  the April 30 2026 release note
• N/A — no config keys added or changed
• N/A — no architecture or workflow changes
• N/A — no file I/O, process management, or cross-platform code touched
• N/A — no tool descriptions or schemas changed

Screenshots / Logs

Net diff against main (single file, +10/-3):

 BEDROCK_CONTEXT_LENGTHS: Dict[str, int] = {
-    # Anthropic Claude models on Bedrock
-    "anthropic.claude-opus-4-6":     200_000,
-    "anthropic.claude-sonnet-4-6":   200_000,
+    # Anthropic Claude models on Bedrock.
+    # Context windows per Anthropic's official models comparison
+    # (https://platform.claude.com/docs/en/about-claude/models/overview).
+    # Opus 4.7 / Opus 4.6 / Sonnet 4.6 have 1M generally available
+    # (no beta header required as of April 2026). Sonnet 4.5 and Sonnet 4
+    # had their `context-1m-2025-08-07` beta retired on April 30, 2026,
+    # so they are standard 200K; Haiku 4.5 is 200K.
+    "anthropic.claude-opus-4-7":     1_000_000,
+    "anthropic.claude-opus-4-6":     1_000_000,
+    "anthropic.claude-sonnet-4-6":   1_000_000,
     "anthropic.claude-sonnet-4-5":   200_000,
     "anthropic.claude-haiku-4-5":    200_000,
     "anthropic.claude-opus-4":       200_000,

[alt-glitch]

Supersedes #16686 — both fix BEDROCK_CONTEXT_LENGTHS for Claude 4.x to 1M, but this PR is more comprehensive (also updates sonnet-4-5 and haiku-4-5, adds opus-4-7 entry).

[patrick-muller]

Thanks for the spot! Quick clarification so the comparison is accurate:

#16686 also adds the claude-opus-4-7 entry — that's not actually a difference between the two PRs. And we don't really change sonnet-4-5 / haiku-4-5 either; both PRs leave those entries at 200_000. The source diff on those two lines is zero in both.

The actual delta vs #16686:

• Tests. fix(bedrock): correct context lengths for Claude 4.x models to 1M tokens #16686 has no tests. This PR updates 3 existing assertions (opus-4-6, sonnet-4-6-versioned, inference-profile) and adds 3 new guard tests: test_claude_opus_4_7 (→ 1M), test_claude_sonnet_4_5_is_200k, test_claude_haiku_4_5_is_200k, each with Anthropic-doc citations in the docstring.
• Anthropic-doc citations in the BEDROCK_CONTEXT_LENGTHS header comment pointing at the models overview + the April 30 2026 release note, so the next contributor has an upstream source of truth.
• Explicit "stay at 200K" decision for sonnet-4-5 and haiku-4-5. fix(bedrock): correct context lengths for Claude 4.x models to 1M tokens #16686 leaves those lines alone without commentary — my concern is that someone reading this table in 3 months might "helpfully" bump them alongside the others without knowing the 1M beta was retired for Sonnet 4.5 on Apr 30 2026 and that Haiku 4.5 is genuinely a 200K model. The guard tests + inline comment pin those values so that mistake can't silently land.

I've added a Prior Art section to the PR body crediting @mmcclean-aws. Happy for maintainers to take either path — close #16686 in favor of this one, or land #16686 first and let me layer the tests/doc-citations on top. Whatever's easier for review.

[mmcclean-aws]

Happy to close #16686 and use this PR as it has the tests missing from my PR

[teknium1]

Thanks for correcting a live Bedrock context-resolution defect. Current main still returns the stale static values from agent/bedrock_adapter.py:1300-1301 through the provider path at agent/model_metadata.py:2189-2196.

Problems

• The table remains incomplete for current 1M Bedrock models. Anthropic's current overview lists Bedrock Opus 4.8 and Sonnet 5 as 1M. With this PR, Opus 4.8 still matches the existing 200K anthropic.claude-opus-4 fallback (agent/bedrock_adapter.py:1304, :1338), while Sonnet 5 falls back to 128K (:1325).
• The table update leaves the explanatory claim that Bedrock Claude is capped at 200K in agent/model_metadata.py:2180-2184, which would no longer describe the updated behavior.
• The newly added model/value assertions are catalog snapshots. The repository rubric asks for resolution invariants rather than tests that fail whenever provider metadata changes.

Suggested changes

• Include current 1M Bedrock families and test longest-match/regional-ID resolution.
• Update the stale model_metadata.py comment.
• Keep a behavior-focused lookup invariant instead of adding per-model snapshot guards.

Automated hermes-sweeper review.

[teknium1]

Please extend this same table update to the current 1M Bedrock families: Anthropic's current overview lists Bedrock Opus 4.8 and Sonnet 5 at 1M. Otherwise Opus 4.8 still matches the existing 200K anthropic.claude-opus-4 fallback and Sonnet 5 falls through to the 128K default.

[alt-glitch]

This was generated by AI during triage.

Supersedes #16686 (now closed, not merged) — same core BEDROCK_CONTEXT_LENGTHS bump to 1M for Claude 4.x, but this PR adds the guard tests + Anthropic-doc citations #16686 lacked, and #16686's author agreed to close in its favor. This is the earliest-open canonical of the Bedrock 1M-context family (#54901/#54914/#54918 dedupe here). Related, not a duplicate, since the superseded PR is already closed.

[teknium1]

Merged via PR #67977 — all three of your commits were cherry-picked onto current main with your authorship preserved (rebase merge). You were the earliest submission in this cluster (May 11) and your doc citations (GA dates, the Sonnet 4.5 beta retirement) anchored the whole table. Opus 4.8 rows from a later PR (#54918) were added on top. Thanks!

#67977