* feat: add anthropic-vertex provider for Claude via GCP Vertex AI
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: sallyom <somalley@redhat.com>
* docs: add anthropic-vertex provider guide
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: sallyom <somalley@redhat.com>
* Agents: validate Anthropic Vertex project env
* Changelog: format update for Vertex entry
* Providers: rename Anthropic Vertex to Google Vertex Claude
* Providers: remove Vertex Claude provider path
* Models: normalize Vercel Claude shorthand refs
* Onboarding: default Vercel model to Claude shorthand
* Changelog: add @vincentkoc credit for #23985
* Onboarding: keep canonical Vercel default model ref
* Tests: expand Vercel model normalization coverage
---------
Signed-off-by: sallyom <somalley@redhat.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
Proxy providers returning Chinese error messages (e.g. Chinese LLM
gateways) use patterns like '上下文过长' or '上下文超出' that are not
matched by the existing English-only patterns in isContextOverflowError.
This prevents auto-compaction from triggering, leaving the session stuck.
Add the most common Chinese proxy patterns:
- 上下文过长 (context too long)
- 上下文超出 (context exceeded)
- 上下文长度超 (context length exceeds)
- 超出最大上下文 (exceeds maximum context)
- 请压缩上下文 (please compress context)
Chinese characters are unaffected by toLowerCase() so check the
original message directly.
Closes#22849
* Telegram: soft-fail reactions and fallback to inbound message id
* Telegram: soft-fail missing reaction message id
* Update CHANGELOG.md
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix: cancel compaction instead of truncating history when summarization fails
When the compaction safeguard cannot generate a summary (no model, no API
key, or LLM error), it previously returned a "Summary unavailable" fallback
string and still truncated history. This caused irreversible data loss -
older messages were discarded even though no meaningful summary was produced.
Now returns `{ cancel: true }` in all three failure paths so the framework
aborts compaction entirely and preserves the full conversation history.
Fixes#10332
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: use deterministic timestamps in compaction safeguard tests
Replace Date.now() with fixed timestamp (0) in test data to prevent
nondeterministic behavior in snapshot-based or order-dependent tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Changelog: note compaction cancellation safeguard fix
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix(compaction): pass model through runtime to fix ctx.model undefined
Fixes#3479
Root cause: extensionRunner.initialize() is never called in compact.ts workflow,
leaving ctx.model undefined. Compaction safeguard checks ctx.model and returns
fallback summary immediately without attempting LLM summarization.
Changes:
1. Pass model through compaction safeguard runtime registry (same pattern as maxHistoryShare)
2. Fall back to runtime.model when ctx.model is undefined
3. Add once-per-session warning when both models are missing (prevents log spam)
4. Add regression test for runtime.model fallback
This follows the established runtime registry pattern rather than attempting to call
extensionRunner.initialize() (which is SDK-internal and not meant for direct access).
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
* test: add comprehensive tests for compaction-safeguard model fallback
Add integration tests to verify the model fallback behavior:
- Test runtime.model fallback when ctx.model is undefined (compact.ts workflow)
- Test fallback summary when both ctx.model and runtime.model are undefined
- Test contextWindowTokens runtime storage/retrieval
- Test combined runtime values (maxHistoryShare + contextWindowTokens + model)
These tests verify the fix for issue #3479 where compaction fails due to
ctx.model being undefined in the compact.ts workflow. The runtime registry
pattern allows model to be passed when extensionRunner.initialize() is not
called, ensuring summarization works in all code paths.
Related: PR #17864
* fix(test): adapt compaction-safeguard tests to upstream type changes
- Add baseUrl to Model mock objects (now required by Model<Api>)
- Add explicit Model<Api> annotation to prevent provider string widening
- Cast modelRegistry mock through unknown (ModelRegistry expanded)
- Use non-null assertion for compactionHandler (TypeScript strict)
- Type compaction result explicitly
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Compaction: add changelog credit for model fallback fix
* Update CHANGELOG.md
* Update CHANGELOG.md
---------
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix: detect additional context overflow error patterns to prevent leak to user
Fixes#9951
The error 'input length and max_tokens exceed context limit: 170636 +
34048 > 200000' was not caught by isContextOverflowError() and leaked
to users via formatAssistantErrorText()'s invalidRequest fallback.
Add three new patterns to isContextOverflowError():
- 'exceed context limit' (direct match)
- 'exceeds the model\'s maximum context'
- max_tokens/input length + exceed + context (compound match)
These are now rewritten to the friendly context overflow message.
* Overflow: add regression tests and changelog credits
* Update CHANGELOG.md
* Update pi-embedded-helpers.isbillingerrormessage.test.ts
---------
Co-authored-by: echoVic <AkiraVic@outlook.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* feat: add Gemini (Google Search grounding) as web_search provider
Add Gemini as a fourth web search provider alongside Brave, Perplexity,
and Grok. Uses Gemini's built-in Google Search grounding tool to return
search results with citations.
- Add runGeminiSearch() with Google Search grounding via tools API
- Resolve Gemini's grounding redirect URLs to direct URLs via parallel
HEAD requests (5s timeout, graceful fallback)
- Add Gemini config block (apiKey, model) with env var fallback
- Default model: gemini-2.5-flash (fast, cheap, grounding-capable)
- Strip API key from error messages for security
- Add config validation tests for Gemini provider
- Update docs/tools/web.md with Gemini provider documentation
Closes#13074
* feat: auto-detect search provider from available API keys
When no explicit provider is configured, resolveSearchProvider now
checks for available API keys in priority order (Brave → Gemini →
Perplexity → Grok) and selects the first provider with a valid key.
- Add auto-detection logic using existing resolve*ApiKey functions
- Export resolveSearchProvider via __testing_provider for tests
- Add 8 tests covering auto-detection, priority order, and explicit override
- Update docs/tools/web.md with auto-detection documentation
* fix: merge __testing exports, downgrade auto-detect log to debug
* fix: use defaultRuntime.log instead of .debug (not in RuntimeEnv type)
* fix: mark gemini apiKey as sensitive in zod schema
* fix: address Greptile review — add externalContent to Gemini payload, add Gemini/Grok entries to schema labels/help, remove dead schema-fields.ts
* fix(web-search): add JSON parse guard for Gemini API responses
Addresses Greptile review comment: add try/catch to handle non-JSON
responses from Gemini API gracefully, preventing runtime errors on
malformed responses.
Note: FIELD_HELP entries for gemini.apiKey and gemini.model were
already present in schema.help.ts, and gemini.apiKey was already
marked as sensitive in zod-schema.agent-runtime.ts (both fixed in
earlier commits).
* fix: use structured readResponseText result in Gemini error path
readResponseText returns { text, truncated, bytesRead }, not a string.
The Gemini error handler was using the result object directly, which
would always be truthy and never fall through to res.statusText.
Align with Perplexity/xAI/Brave error patterns.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* style: fix import order and formatting after rebase onto main
* Web search: send Gemini API key via header
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
When thinking is set (e.g. thinking=low), the model produces internal
thinking blocks. The reasoning auto-default (based on model capability)
was formatting these blocks as "Reasoning:" text and delivering them to
WhatsApp/Telegram, leaking internal content to users.
Skip auto-enabling reasoning when thinkLevel is already set — the two
features serve the same purpose and enabling both causes the model's
internal thinking to be exposed as visible chat messages.
Users who explicitly set /reasoning on still get reasoning output.
Closes#24290
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: treat HTTP 502/503/504 as failover-eligible (timeout reason)
When a model API returns 502 Bad Gateway, 503 Service Unavailable, or
504 Gateway Timeout, the error object carries the status code directly.
resolveFailoverReasonFromError() only checked 402/429/401/403/408/400,
so 5xx server errors fell through to message-based classification which
requires the status code to appear at the start of the error message.
Many API SDKs (Google, Anthropic) set err.status = 503 without prefixing
the message with '503', so the message classifier never matched and
failover never triggered — the run retried the same broken model.
Add 502/503/504 to the status-code branch, returning 'timeout' (matching
the existing behavior of isTransientHttpError in the message classifier).
Fixes#20999
* Changelog: add failover 502/503/504 note with credits
* Failover: classify HTTP 504 as transient in message parser
* Changelog: credit taw0002 and vincentkoc for failover fix
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
