* feat(middleware): inject dynamic context via DynamicContextMiddleware
Move memory and current date out of the system prompt and into a
dedicated <system-reminder> HumanMessage injected once per session
(frozen-snapshot pattern) via a new DynamicContextMiddleware.
This keeps the system prompt byte-exact across all users and sessions,
enabling maximum Anthropic/Bedrock prefix-cache reuse.
Key design decisions:
- ID-swap technique: reminder takes the first HumanMessage's ID
(replacing it in-place via add_messages), original content gets a
derived `{id}__user` ID (appended after). Preserves correct ordering.
- hide_from_ui: True on reminder messages so frontend filters them out.
- Midnight crossing: date-update reminder injected before the current
turn's HumanMessage when the conversation spans midnight.
- INFO-level logging for production diagnostics.
Also adds prompt-caching breakpoint budget enforcement tests and
updates ClaudeChatModel docs to reference the new pattern.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(token-usage): log input/output token detail breakdown in middleware
Extend the LLM token usage log line to include input_token_details and
output_token_details (cache_creation, cache_read, reasoning, audio, etc.)
when present. Adds tests covering Anthropic cache detail logging from
both usage_metadata and response_metadata.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: fix nginx
* fix(middleware): always inject date; gate memory on injection_enabled
Date injection is now unconditional — it is part of the static system
prompt replacement and should always be present. Memory injection
remains gated by `memory.injection_enabled` in the app config.
Previously the entire DynamicContextMiddleware was skipped when
injection_enabled was False, which also suppressed the date.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(lint): format files and correct test assertions for token usage middleware
- ruff format dynamic_context_middleware.py and test_claude_provider_prompt_caching.py
- Remove unused pytest import from test_dynamic_context_middleware.py
- Fix two tests that asserted response_metadata fallback logic that
doesn't exist: replace with tests that match actual middleware behavior
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(middleware): address Copilot review comments on DynamicContextMiddleware
- Use additional_kwargs flag for reminder detection instead of content
substring matching, so user messages containing '<system-reminder>'
are not mistakenly treated as injected reminders
- Generate stable UUID when original HumanMessage.id is None to prevent
ambiguous 'None__user' derived IDs and message collisions
- Downgrade per-turn no-op log to DEBUG; keep actual injection events at INFO
- Add two new tests: missing-id UUID fallback and user-text false-positive
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: cap prompt caching breakpoints at 4 to prevent API 400 errors (fixes#2448)
The previous _apply_prompt_caching() attached cache_control to every text
block in the system prompt, every content block in the last N messages, and
the last tool definition. In multi-turn conversations with structured content
blocks this easily exceeded the 4-breakpoint hard limit enforced by both the
Anthropic API and AWS Bedrock, producing a 400 Bad Request (or a silent
"No generations found in stream" when streaming).
Fix: collect all candidate blocks in document order, then apply cache_control
only to the last MAX_CACHE_BREAKPOINTS (4) of them. Later breakpoints cover a
larger prefix and therefore yield better cache hit rates, making this the
optimal placement strategy as well as the safe one.
Adds 13 unit tests covering the budget cap, edge cases, and correct
last-candidate placement.
* docs: clarify _apply_prompt_caching docstring includes tool definitions
Per Copilot review: the implementation also caches the last tool definition
(see the candidates list at lines 202-205), so the docstring summary should
explicitly mention tools alongside system and recent messages.
* Fix the lint error
* style: fix ruff format check for test_claude_provider_prompt_caching.py
Add the missing blank line before the 'Edge cases' section comment so
that ruff format --check passes in CI.
---------
Co-authored-by: octo-patch <octo-patch@github.com>
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
* fix(oauth): inject billing header for non-Haiku model access
The Anthropic Messages API requires a billing identification block
in the system prompt when using Claude Code OAuth tokens (sk-ant-oat*)
to access non-Haiku models (Opus, Sonnet). Without it, the API returns
a generic 400 "Error" with no actionable detail.
This was discovered by intercepting Claude Code CLI requests — the CLI
injects an `x-anthropic-billing-header` text block as the first system
prompt entry on every request. Third-party consumers of the same OAuth
tokens must do the same.
Changes:
- Add `_apply_oauth_billing()` to `ClaudeChatModel` that prepends the
billing header block to the system prompt when `_is_oauth` is True
- Add `metadata.user_id` with device/session identifiers (required by
the API alongside the billing header)
- Called from `_get_request_payload()` before prompt caching runs
Verified with Claude Max OAuth tokens against all three model tiers:
- claude-opus-4-6: 200 OK
- claude-sonnet-4-6: 200 OK
- claude-haiku-4-5-20251001: 200 OK (was already working)
Closes#1245
* fix(oauth): address review feedback on billing header injection
- Make OAUTH_BILLING_HEADER configurable via ANTHROPIC_BILLING_HEADER env var
- Normalize billing block to always be first in system list (strip + reinsert)
- Guard metadata with isinstance check for non-dict values
- Replace os.uname() with socket.gethostname() for Windows compat
- Fix docstrings to say "all OAuth requests" instead of "non-Haiku"
- Move inline imports to module level (fixes ruff I001)
- Add 9 unit tests for _apply_oauth_billing
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add Claude Code OAuth and Codex CLI providers
Port of bytedance/deer-flow#1136 from @solanian's feat/cli-oauth-providers branch.\n\nCarries the feature forward on top of current main without the original CLA-blocked commit metadata, while preserving attribution in the commit message for review.
* fix: harden CLI credential loading
Align Codex auth loading with the current ~/.codex/auth.json shape, make Docker credential mounts directory-based to avoid broken file binds on hosts without exported credential files, and add focused loader tests.
* refactor: tighten codex auth typing
Replace the temporary Any return type in CodexChatModel._load_codex_auth with the concrete CodexCliCredential type after the credential loader was stabilized.
* fix: load Claude Code OAuth from Keychain
Match Claude Code's macOS storage strategy more closely by checking the Keychain-backed credentials store before falling back to ~/.claude/.credentials.json. Keep explicit file overrides and add focused tests for the Keychain path.
* fix: require explicit Claude OAuth handoff
* style: format thread hooks reasoning request
* docs: document CLI-backed auth providers
* fix: address provider review feedback
* fix: harden provider edge cases
* Fix deferred tools, Codex message normalization, and local sandbox paths
* chore: narrow PR scope to OAuth providers
* chore: remove unrelated frontend changes
* chore: reapply OAuth branch frontend scope cleanup
* fix: preserve upload guards with reasoning effort wiring
---------
Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
DanielWalnut