Repository navigation
feat(nodes): add gemini vision node - #672
Conversation
|
No description provided. |
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
📝 WalkthroughWalkthroughAdded a new Gemini Vision node: package implements IGlobal lifecycle and validation, an IInstance handling image and document lanes, a Chat wrapper calling Google GenAI, service registration and requirements, and documentation. The node validates API keys, initializes a Chat, processes images (base64), and emits text responses. Changes
Sequence Diagram(s)sequenceDiagram
participant Instance as IInstance
participant Global as IGlobal/Chat
participant Gemini as Google GenAI
participant Config as services.json
Note over Config: Node configured (apikey, model, prompts)
Instance->>Global: request Chat (beginGlobal initializes Chat)
Instance->>Instance: receive image (AVI chunks / Doc)
Instance->>Global: chat(question with data:image/... base64 + prompt)
Global->>Gemini: models.generate_content / request
Gemini-->>Global: response (text) or error
Global-->>Instance: Answer (text) or formatted error
Instance->>Downstream: emit Doc(type=Text) / preventDefault original
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Possibly related PRs
Suggested reviewers
Poem
🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@nodes/src/nodes/llm_vision_gemini/IInstance.py`:
- Around line 38-40: The current single-string cache _cached_answer should
become a mapping keyed by a stable frame identity (e.g., frame number,
timestamp, or a hash of the image bytes) so cached replies are only reused for
the exact same frame; change _cached_answer: str to something like
_cached_answers: Dict[str, str], compute a stable frame_key inside
writeDocuments (or the method that generates the answer) and store
_cached_answers[frame_key] = answer, then in writeImage (or the consumer that
reuses the cache) only read and use _cached_answers[frame_key] when the computed
key matches exactly; also update the other usages of _cached_answer in the same
class (the other cached fields/sections around the write/create methods) to use
this keyed cache and add a small eviction/cleanup policy to avoid unbounded
growth.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: ASSERTIVE
Plan: Pro
Run ID: bc14fa30-5535-4c2a-9d2f-28b54e4932ce
📒 Files selected for processing (6)
nodes/src/nodes/llm_vision_gemini/IGlobal.pynodes/src/nodes/llm_vision_gemini/IInstance.pynodes/src/nodes/llm_vision_gemini/__init__.pynodes/src/nodes/llm_vision_gemini/gemini_vision.pynodes/src/nodes/llm_vision_gemini/requirements.txtnodes/src/nodes/llm_vision_gemini/services.json
There was a problem hiding this comment.
♻️ Duplicate comments (1)
nodes/src/nodes/llm_vision_gemini/IInstance.py (1)
38-40:⚠️ Potential issue | 🟠 MajorKey the cache by frame identity instead of using one shared answer slot.
_cached_answeris still “last document wins”. Clearing it on Line 46 makes reuse depend on lane ordering, and overwriting it on Line 98 lets the nextAVI_ACTION.ENDpick up a different frame’s answer if documents arrive late or in batches. This breaks the dual-lane dedupe guarantee and can emit the wrong analysis downstream.Suggested direction
- _cached_answer: str = None + _cached_answers: dict[str, str] ... - self._cached_answer = None + # Do not clear cached answers here; evict only on exact frame match. ... - if self._cached_answer is not None: - self.instance.writeText(self._cached_answer) - self._cached_answer = None + frame_key = self._frame_key(bytes(self.image_data)) + cached_answer = self._cached_answers.pop(frame_key, None) + if cached_answer is not None: + self.instance.writeText(cached_answer) else: ... ... - self._cached_answer = answer_text + self._cached_answers[self._frame_key(decoded_doc_bytes)] = answer_textYou'd also need a per-instance initializer for
_cached_answersand a shared_frame_key(...)helper, e.g. a hash of the raw image bytes.Also applies to: 44-55, 97-99
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@nodes/src/nodes/llm_vision_gemini/IInstance.py` around lines 38 - 40, The cache currently uses a single slot _cached_answer which causes last-write-wins bugs across lanes; change to a per-frame cache (e.g., _cached_answers dict) initialized in the instance initializer and use a helper _frame_key(...) to compute a unique key (hash of raw image bytes or similar) for each frame; update writeDocuments to store answers into _cached_answers[_frame_key(...)] instead of _cached_answer, update writeImage and the AVI_ACTION.END handling to read and clear only that frame's entry, and remove/stop using the global _cached_answer to preserve dual-lane dedupe guarantees.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Duplicate comments:
In `@nodes/src/nodes/llm_vision_gemini/IInstance.py`:
- Around line 38-40: The cache currently uses a single slot _cached_answer which
causes last-write-wins bugs across lanes; change to a per-frame cache (e.g.,
_cached_answers dict) initialized in the instance initializer and use a helper
_frame_key(...) to compute a unique key (hash of raw image bytes or similar) for
each frame; update writeDocuments to store answers into
_cached_answers[_frame_key(...)] instead of _cached_answer, update writeImage
and the AVI_ACTION.END handling to read and clear only that frame's entry, and
remove/stop using the global _cached_answer to preserve dual-lane dedupe
guarantees.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: ASSERTIVE
Plan: Pro
Run ID: 8065ad0f-56d6-4509-add5-3fb03417ba02
📒 Files selected for processing (1)
nodes/src/nodes/llm_vision_gemini/IInstance.py
|
Do you need to add the documentation markdown to this nodes folder? |
There was a problem hiding this comment.
Actionable comments posted: 3
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@nodes/src/nodes/llm_vision_gemini/IInstance.py`:
- Around line 52-69: The image-lane path in IInstance currently calls Gemini but
does not populate the shared cache, so duplicate suppression only works if
writeDocuments runs first; fix by storing the result into the same cache used by
writeDocuments after the image-lane Gemini call (set self._cached_answer to the
returned text or the same object used by writeDocuments immediately after answer
= self.IGlobal._chat.chat(question)), and keep the existing behavior of
instance.writeText(answer.getText()) and clearing the cache when the other path
consumes it; apply the same symmetric caching change to the analogous block
around lines 92-104 so either lane can fill/reuse _cached_answer.
- Around line 51-74: In the AVI_ACTION.END branch of IInstance (the block
handling end-of-frame logic), skip calling Gemini when there is no image payload
by checking that self.image_data is non-empty before building the data URL and
creating Question; if image_data is empty or zero-length, avoid the base64
encode/Question/chat call and simply clear image_data and return
self.preventDefault() (same cleanup path as when using _cached_answer). This
prevents writeImage()/Gemini inference on empty frames—add the check right
before importing Question/base64 and constructing image_base64/image_data_url.
In `@nodes/src/nodes/llm_vision_gemini/README.md`:
- Line 15: The README statement claiming "All profiles support a 1M token
context window" is inaccurate; update the sentence and any duplicate claims
(notably at lines 49-52) to reflect the actual context-window sizes shown in the
profile table (e.g., "Gemini 3.1 Flash Image Preview" has a 131,072 token
limit). Specifically, replace the blanket 1M-token claim with a precise
description that some profiles support up to 1M tokens while others (like Gemini
3.1 Flash Image Preview) use smaller windows, and ensure the table entries and
summary text (mentions of "Gemini 3.1 Flash Image Preview" and the general
context-window claim) are consistent.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: ASSERTIVE
Plan: Pro
Run ID: 0a806af1-d807-46d0-bebd-a51824c6b800
📒 Files selected for processing (2)
nodes/src/nodes/llm_vision_gemini/IInstance.pynodes/src/nodes/llm_vision_gemini/README.md
There was a problem hiding this comment.
Actionable comments posted: 3
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@nodes/src/nodes/llm_vision_gemini/IInstance.py`:
- Around line 42-78: The writeImage method currently handles
AVI_ACTION.BEGIN/WRITE/END but lets unknown action values fall through without
calling preventDefault(); update writeImage (method name writeImage) to include
an explicit final fallback branch for any unrecognized action (e.g., else or
final return) that calls self.preventDefault() to ensure consistent behavior for
unknown AVI_ACTION values; keep existing branches and logging intact and only
add the deterministic fallback after the existing if/elif/elif block so unknown
actions are handled.
- Around line 48-50: The WRITE branch for AVI_ACTION.WRITE in IInstance.py can
raise when no frame has been opened (self.image_data is None); update the
handling so it first checks that self.image_data is initialized (e.g., not None
or a bytearray) before doing self.image_data += buffer, and if it is None return
self.preventDefault() (or otherwise initialize self.image_data only when a prior
BEGIN/AVI_ACTION.BEGIN handler ran); ensure you reference AVI_ACTION.WRITE,
self.image_data, and preventDefault() when applying the guard so WRITE never
performs None += buffer.
- Around line 96-111: The single-slot cache self._cached_answer causes answers
to be misassigned across multiple docs; change the cache to be per-document
(e.g., replace self._cached_answer with a dict like self._cached_answers) keyed
by a unique doc identifier (use doc.metadata.chunkId or generate an 'unknown'
fallback) and update the read/write sites: when looping over for doc in
documents, check self._cached_answers.get(key) instead of self._cached_answer,
store answer_text into self._cached_answers[key] after calling
IGlobal._chat.chat(question), and ensure writeImage reads and clears the same
per-doc entry; also clean up entries after use to avoid memory growth.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: ASSERTIVE
Plan: Pro
Run ID: bf4fca37-5647-422d-a9dc-8b0885247369
📒 Files selected for processing (2)
nodes/src/nodes/llm_vision_gemini/IInstance.pynodes/src/nodes/llm_vision_gemini/README.md
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@nodes/src/nodes/llm_vision_gemini/IInstance.py`:
- Around line 123-125: The comment next to self._cached_answer and
self._cache_from_image is misleading: writeImage only checks self._cached_answer
(not _cache_from_image), so update the comment in IInstance near the assignment
of _cached_answer/_cache_from_image to accurately reflect behavior—explain that
we cache the answer for writeImage to reuse regardless of origin, and remove or
rephrase the claim "writeImage checks for this"; reference the symbols
_cached_answer, _cache_from_image, and the writeImage method to guide the
change.
- Around line 129-130: In the writeImage method of IInstance
(nodes/src/nodes/llm_vision_gemini/IInstance.py) change the bare call to
self.preventDefault() to return self.preventDefault() so the return value is
propagated to the framework consistently with other preventDefault() usages in
writeImage; update the single occurrence in writeImage to use a leading return.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: ASSERTIVE
Plan: Pro
Run ID: e826500a-d9b6-49ea-a285-005ed27ff44a
📒 Files selected for processing (1)
nodes/src/nodes/llm_vision_gemini/IInstance.py
Rod-Christensen
left a comment
There was a problem hiding this comment.
Clean, well-structured Gemini Vision node — follows the existing vision node architecture correctly. Good error handling and API key validation. Approved.
PRs #672 (gemini-vision) and #677 (openai-vision) merged with `ruff format --check` failures on three files, leaving develop CI red and gating the post-CI nightly publish workflow_run. Pure whitespace changes — line wraps long tuple unpacks and except-clause type tuples per ruff's default style — no semantic edits. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* add gemini vision node * clear cache to prevent bleed * error handling + README * fix non-symmetric cache + empty frame guard * bug fixes + CR comments * add openai vision node * update README + update models * fix: use chunk reference in _processFullTables to resolve chunkId scoping bug In the second phase of _processFullTables, the code incorrectly used `doc.metadata.chunkId` when initializing a new table entry. `doc` was a stale variable from the outer `for docKey, doc in documents.items()` loop, not the current chunk being processed. Additionally, Doc() was called with invalid constructor params (objectId, chunk) that don't exist on the Pydantic model. Fix mirrors the pattern used in _processFullDocuments: reset page_content on the chunk itself and store it directly, so the subsequent append builds the correct concatenated table content. Resolves #776 * fix: address CodeRabbit review - score assignment and first-chunk aliasing Two follow-up fixes from code review: 1. Line 157: tableDocs[tableKey] = doc.score replaced the Doc object with a raw float, crashing Phase 2 with AttributeError when page_content was accessed. Fixed to tableDocs[tableKey].score = doc.score. 2. Lines 184-189: tableDocs[tableKey] and chunk aliased the same object, so clearing chunk.page_content then appending chunk.page_content accumulated empty string, silently dropping the first chunk's text. Fixed by saving original_content before clearing, restoring it after assignment, and continuing to skip the duplicate append below. * fix: guard against falsy-value traps in psycopg2, chromadb, weaviate mocks Three test mocks mishandled falsy values, causing mock tests to pass while real-service tests would fail: psycopg2: bool is a subclass of int in Python, so isinstance(False, int) returns True. SQL WHERE params like isDeleted=False were incorrectly matched as LIMIT values. Fixed by adding not isinstance(p, bool) guards in _handle_semantic_search and _extract_limit. chromadb: 'if limit and ...' treated limit=0 as falsy, returning all rows instead of zero rows. Fixed with explicit 'limit is not None' check. weaviate: sort key 'distance if distance else 1.0' collapsed 0.0 (exact match, best possible distance) to 1.0, ranking perfect matches last. Fixed with 'distance if distance is not None else 1.0'. Resolves #748 * fix(client-python): break drain cycle causing RecursionError on disconnect ## Summary <!-- Describe your changes in 1-3 bullet points --> - Fix `RecursionError: maximum recursion depth exceeded` in `_GatheringFuture.cancel` during `/task/service` WebSocket teardown by breaking the cycle between an in-flight message-handler task and the disconnect-drain gather it waits on. - Fix C — `dap_client.request()`'s `ConnectionError` handler now uses fire-and-forget `self._transport.disconnect()` instead of `await self.disconnect()`. Develop's `disconnect()` is sync-returns-`Task`, matching the documented carve-out at `transport_websocket.py:516-517`. - Fix A — `_drain_message_tasks` walks each candidate's `_fut_waiter` chain and excludes tasks that transitively await the drainer itself, so the gather can never close back on itself even if a future caller awaits disconnect from a handler context. ## Type <!-- What kind of change is this? (feature, fix, refactor, docs, chore, etc.) --> fix ## Testing - [x] Tests added or updated - [x] Tested locally - [ ] `./builder test` passes `packages/client-python/tests/test_drain_cycle_regression.py` adds two regression tests, one pinning each fix: - `test_drain_cycle_does_not_recurse` — builds the smallest task graph that exhibits the cycle (a handler awaiting `transport.disconnect()` while it is itself in `_message_tasks`); fails on pre-fix code via deadlock + recursion, passes after Fix A. - `test_request_does_not_await_self_disconnect_on_connection_error` — asserts `dap_client.request()` does not call `self.disconnect()` on `ConnectionError`; fails pre-fix, passes after Fix C. The pytest teardown after a pre-fix run reproduces the exact `tasks.py:721` recursion as the production trace, confirming the unit test covers the production bug rather than a similar-looking variant. `./builder test` reports one unrelated failure on `test_node_cases[llm_openai:services:test2:openai-4o-mini]` — a live OpenAI call (`avoidMocks: true`, `requires: ['ROCKETRIDE_OPENAI_KEY']`) where the LLM produced no answer for `"What is 5*5?"`. Pre-existing flakiness in the real-API path, unrelated to these changes (transport-layer fixes don't touch LLM provider calls). ## Checklist - [x] Commit messages follow [conventional commits](https://www.conventionalcommits.org/) - [x] No secrets or credentials included - [ ] Wiki updated (if applicable) - [ ] Breaking changes documented (if applicable) ## Linked Issue <!-- REQUIRED: Every PR must be linked to an issue. Use one of: --> <!-- Fixes #123 / Closes #123 / Resolves #123 --> Fixes #791 * chore(deps)(deps): update requests requirement in /nodes/src/nodes Updates the requirements on [requests](https://github.com/psf/requests) to permit the latest version. - [Release notes](https://github.com/psf/requests/releases) - [Changelog](https://github.com/psf/requests/blob/main/HISTORY.md) - [Commits](psf/requests@v2.34.0...v2.34.2) --- updated-dependencies: - dependency-name: requests dependency-version: 2.34.2 dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com> * feat(shell-ui): apply saved theme on init to prevent tan loading flash - index.html: inline script sets --rr-bg-default and body background synchronously before React renders, so the page starts dark for users who have a dark preference saved in localStorage (rr:home:theme). - createShellConfig onInit: reads rr:home:theme then rr:theme to apply the user's saved theme instead of always defaulting to rocketride-light, eliminating the tan background during the 1.5s OAuth code-exchange loading window. Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com> * feat(shell-ui): apply saved theme on init to prevent tan loading flash (#878) - index.html: inline script sets --rr-bg-default and body background synchronously before React renders, so the page starts dark for users who have a dark preference saved in localStorage (rr:home:theme). - createShellConfig onInit: reads rr:home:theme then rr:theme to apply the user's saved theme instead of always defaulting to rocketride-light, eliminating the tan background during the 1.5s OAuth code-exchange loading window. Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com> * chore(deps): pin cryptography==46.0.7 in text_output (#884) Previously unpinned, leaving the resolved version subject to PyPI latest. Three open Dependabot alerts all clear at 46.0.7: - GHSA-r6ph-v2qm-q3c2 (high) subgroup attack on SECT curves - GHSA-p423-j2cm-9vmq (medium) buffer overflow on non-contiguous buffers - GHSA-m959-cc7f-wv43 (low) incomplete DNS name constraint enforcement Exact pin matches the sibling node db_mysql/requirements.txt (also pins ==46.0.7) for reproducible builds and consistent dependency policy across nodes. Future bumps will arrive via Dependabot PRs. * chore(deps): patch minimatch ReDoS across all four major lines via overrides (#887) Resolves 6 high-severity Dependabot alerts on root pnpm-lock.yaml. (3 additional alerts on packages/shared-ui/pnpm-lock.yaml clear when #886 lands, which deletes that dead sub-lockfile.) Adds four bounded pnpm.overrides entries — one per major — so each transitive caller stays on the API surface it expects, just patched: minimatch@<3.1.4 → >=3.1.4 <4 was 3.1.2, fix 3.1.4* minimatch@5 → >=5.1.8 <6 was 5.1.6, fix 5.1.8 minimatch@9 → >=9.0.7 <10 was 9.0.5, fix 9.0.7 minimatch@10 → >=10.2.3 <11 was 10.1.1, fix 10.2.3 * 3.1.4 covers GHSA-23c5-xmqv-rm74 in the 3.x line (Dependabot only flagged 3.1.3 for GHSA-7r86-cg39-jmmj; the tighter bound pre-empts the second advisory too). Post-install resolves to minimatch@3.1.5, 5.1.9, 9.0.9, 10.2.5. Advisories cleared (all high): GHSA-3ppc-4f35-3m26 minimatch ReDoS: repeated wildcards with non-matching literal in pattern (5.x) GHSA-23c5-xmqv-rm74 minimatch ReDoS: nested *() extglobs generate catastrophically backtracking regex (3.x, 5.x) GHSA-7r86-cg39-jmmj minimatch ReDoS: matchOne() combinatorial backtracking via multiple non-adjacent GLOBSTAR segments (3.x, 5.x, 9.x, 10.x) A single high pin (e.g. minimatch>=10.2.3) was considered and rejected: minimatch had breaking API changes across majors (default vs named exports between 3→5→7); forcing all callers to 10.x risks breaking deps that still rely on 3.x/5.x semantics. The bounded approach matches the existing repo convention (handlebars: >=4.7.9 <5; protobufjs: >=7.5.5 <8). * chore(deps): bump dompurify to 3.4.x in shared-ui, drop dead sub-lockfile (#886) Resolves all 16 open Dependabot alerts on dompurify (8 unique GHSAs, all medium severity) ranging in fix versions 3.3.2 → 3.4.0. Changes: - packages/shared-ui/package.json: peerDependency dompurify ~3.2.5 → ~3.4.0 - package.json: add "dompurify": ">=3.4.0" to root pnpm.overrides, matching the existing security-bump convention (cross-spawn, glob, handlebars, js-yaml, postcss, protobufjs, qs). - packages/shared-ui/pnpm-lock.yaml: deleted. This file was a dead artifact — last (and only) touched in the initial commit (8197744), zero references in CI, build scripts, or publish workflows. Its sole effect was duplicating dompurify alerts since Dependabot scans every pnpm-lock.yaml independently. - pnpm-lock.yaml: regenerated. dompurify resolves to 3.4.3 (latest 3.4.x available); no 3.2.x/3.3.x copies remain. Advisories cleared (all medium): GHSA-cj63-jhhr-wcxv GHSA-cjmm-f4jc-qw8r GHSA-h8r8-wccr-v5f2 GHSA-v2wj-7wpq-c8vv GHSA-39q2-94rc-95cp GHSA-crv5-9vww-q3g8 GHSA-h7mw-gpvr-xq4m GHSA-v9jr-rg53-9pgp Call site (packages/shared-ui/src/components/canvas/util/helpers.tsx:228) uses DOMPurify.sanitize(input, { ADD_ATTR: [...] }) — stable across the 3.x line; no API breakage. * docs(agents): add observability & tracing integration guide (#716) * docs(agents): add observability & tracing integration guide Documents how external services consume runtime logs, lifecycle events, and pipeline traces over the WebSocket DAP channel so integrators have a single source of truth on what the server actually exposes (no OTel, Prometheus, or audit-log DB). Wires the new doc into the .rocketride/docs sync list and lists it in every agent stub (Claude Code, CLAUDE.md, AGENTS.md, Cursor, Windsurf, Copilot) so installed workspaces surface it alongside the existing README/quickstart/API references. Co-Authored-By: Claude <noreply@anthropic.com> * docs(agents): tag observability doc websocket fence with language Adds the `text` language identifier to the WebSocket URL fenced code block so markdownlint MD040 stops flagging it, per CodeRabbit review on PR #716. Co-Authored-By: Claude <noreply@anthropic.com> * docs(agents): remove HTTP transport from observability guide Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Dmitrii Karataev <dmitry.karataev@gmail.com> * fix(cli): implement signal handlers for graceful shutdown (RR-655) (#891) This is from an already approved PR Rebased onto current develop, dropping changes already merged separately (redaction regex, CSS comment fix, PageStatus UI, upload math fix). Only the signal handler work remains: - SIGINT/SIGTERM handlers with proper exit codes (130/143) - Double-signal force exit, 5s cleanup timeout - Idempotent cleanupClient via dedup promise - isCancelled guards on command actions to prevent exit races - run()/main() shutdown awareness via awaitShutdown() Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore(deps): patch undici + lodash family ReDoS/RCE via overrides (#888) Resolves 16 open Dependabot alerts (6 undici, 5 lodash, 5 lodash-es) in a single PR. Affected manifests: root pnpm-lock.yaml (8 alerts) and packages/shared-ui/pnpm-lock.yaml (8 alerts, all clear once #886 lands deleting that dead sub-lockfile). Changes: - package.json: add three bounded entries to pnpm.overrides "lodash": ">=4.18.0 <5" was 4.17.23 "lodash-es": ">=4.18.0 <5" was 4.17.23 "undici": ">=7.24.0 <8" was 7.20.0 - packages/shared-ui/package.json: bump lodash peerDep ~4.17.21 → ~4.18.0 (declaration alignment; shared-ui has zero direct lodash imports in src/, so this is purely a doc-correctness fix). Post-install resolves to lodash@4.18.1, lodash-es@4.18.1, undici@7.25.0. Advisories cleared: undici (all in 7.0.0..<7.24.0): GHSA-f269-vfmq-vjvj (high) WebSocket 64-bit length parser overflow / crash GHSA-v9p9-hfj2-hcw8 (high) WebSocket unhandled exception via server_max_win GHSA-vrm6-8vpv-qv8q (high) WebSocket permessage-deflate unbounded memory GHSA-2mjp-6q6p-2qxm (medium) HTTP request/response smuggling GHSA-phc3-fgpg-7m6h (medium) DeduplicationHandler unbounded memory GHSA-4992-7rv2-5pvq (medium) CRLF injection via `upgrade` option lodash + lodash-es (<= 4.17.23): GHSA-r5fr-rjxr-66jc (high) Code injection via _.template imports key names GHSA-f23m-r3pf-42rh (medium) Prototype pollution via _.unset / array path bypass Note: GHSA-xxjr-mmjv-4gpg (medium, fix=4.17.23) was already covered by the prior dependabot-bumped 4.17.23 and will auto-close when Dependabot rescans. Bounded ranges match the existing repo convention (handlebars: ">=4.7.9 <5", protobufjs: ">=7.5.5 <8") and avoid pulling in unreleased lodash 5.x or undici 8.x majors that would carry breaking changes. * chore(deps): bump shared-ui devDependencies - @rollup/plugin-commonjs 28.0.6 → 28.0.9 - @rollup/plugin-replace 6.0.2 → 6.0.3 - @rollup/plugin-typescript 12.1.4 → 12.3.0 - @types/react 18.2.21 → 18.3.28 - @types/react-dom 18.2.7 → 18.3.7 - rollup 4.59.0 → 4.60.4 - rollup-plugin-dts 6.2.3 → 6.4.1 Replaces Dependabot PR #881 which failed due to stale pnpm-lock.yaml. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat(database): add QuestionType.EXECUTE for direct SQL/Cypher execution Add a new QuestionType.EXECUTE value plus a `client.database.query()` namespace on both the Python and TypeScript SDKs so trusted callers can issue raw SQL or Cypher against database pipelines without paying the LLM-translation tax. db_postgres / db_mysql / db_neo4j skip both the LLM call and the is_sql_safe / _is_cypher_safe gating when the question arrives with type=EXECUTE. End-to-end verified against local Postgres + Neo4j with the matching SDK on both languages (8/8 cases pass; nodes contract suite 733/733). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(database): gate EXECUTE path, bound results, validate SDK inputs Addresses CodeRabbit review on PR #782: - Gate QuestionType.EXECUTE behind opt-in allow_execute config (default off) on both SQL (DatabaseGlobalBase) and Neo4j (db_neo4j.IGlobal) backends so arbitrary write/admin SQL/Cypher cannot run without explicit operator consent. - Bound _executeRawQuery SELECT results with fetchmany(max_execute_rows+1) to avoid memory blowups on large reads; normalize negative rowcount to 0. - Wrap EXECUTE output blocks (SQL + Neo4j) in try/except so formatting and lane writes can't escape writeQuestions; serialize with default=str and sanitize SQL rows before json.dumps to handle non-serializable DB values. - Validate non-empty token/sql in client.database.query() for both Python and TypeScript SDKs; raise ValueError / throw Error instead of dispatching. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(database): bound Neo4j EXECUTE rows, parse allow_execute strictly Addresses second CodeRabbit review on PR #782: - _run_query_raw now caps result collection at max_execute_rows (default 25000), mirroring the SQL path, so a large EXECUTE MATCH...RETURN cannot exhaust the worker. Overflow raises ValueError, caught at the IInstance handler. - allow_execute config now parses strings explicitly: only '1', 'true', 'yes', 'on' (case-insensitive) enable EXECUTE. Plain bool() would have let 'false' or '0' weaken the opt-in gate. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * refactor(database): hoist max_execute_rows default to a single constant per module Replace the three repeated 25000 literals in each of db_global_base.py and db_neo4j/IGlobal.py with a module-level DEFAULT_MAX_EXECUTE_ROWS constant. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * refactor(database): reference DEFAULT_MAX_EXECUTE_ROWS in get_data limits Replace the remaining 25000 literals in the get_data tool description and _clamp_limit / inline limit clamp with the existing DEFAULT_MAX_EXECUTE_ROWS constant. Single source of truth across the EXECUTE path and the non-EXECUTE get_data path within each module. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(database): add QuestionType.DIALECT for engine discovery Lets SDK callers ask a database pipeline which engine it's connected to so applications can branch on dialect-specific behavior (SQL syntax, type coercion) or assert they aren't pointed at the wrong kind of database. - New QuestionType.DIALECT in both Python and TypeScript client schemas. - DatabaseInstanceBase: new abstract _db_dialect() method + DIALECT branch in writeQuestions that emits {'dialect': '...'} on the answers lane. No gate (the value is metadata, not data). - db_postgres / db_mysql implement _db_dialect() -> 'postgres' / 'mysql'. - db_neo4j handles DIALECT in writeQuestions and emits 'neo4j'. - Python SDK: client.database.dialect(token=...) returns DatabaseDialect enum. - TypeScript SDK: client.database.dialect({token}) returns DatabaseDialect enum; re-export from client/index.ts so callers can import from the package root. - services.json: each db node now exposes a UI-toggleable allow_execute boolean (default false) so operators can opt-in to the EXECUTE path from the Dropper UI instead of hand-editing pipe JSON. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(client-python): clean pending DAP request on send failure (#736) Co-authored-by: Dmitrii Karataev <dmitry.karataev@gmail.com> * chore: removed obsolote input section from 72 service configuration files. Input array in services*.jsons were duplicate of lane. Closes #629 (#728) Co-authored-by: Dmitrii Karataev <dmitry.karataev@gmail.com> * chore: bump minor version on all release packages Bump all package versions to the next minor release (patch reset to 0): - rocketride-server: 3.1.2 → 3.2.0 - chat-ui, dropper-ui: 2.0.1 → 2.1.0 - All other packages: 1.0.x → 1.1.0 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * style(nodes): ruff format llm_vision_gemini + llm_vision_openai PRs #672 (gemini-vision) and #677 (openai-vision) merged with `ruff format --check` failures on three files, leaving develop CI red and gating the post-CI nightly publish workflow_run. Pure whitespace changes — line wraps long tuple unpacks and except-clause type tuples per ruff's default style — no semantic edits. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(nodes): remove llm_vertex, fix llm_gemini key handling, migrate vision nodes to LLMBase - Remove deprecated llm_vertex node (Vertex AI now supported via llm_gemini) - Fix llm_gemini API key validation: drop strict startsWith('AI') check, add sub-key fallback for old pipe files where sub-key is profile.split('_',1)[1] - Fix llm_gemini validateConfig to early-return when apikey is empty (secure fields are not decrypted at validate time) - Migrate llm_vision_gemini and llm_vision_openai IInstance from deprecated IInstanceGenericLLM to ai.common.llm_base.LLMBase - Remove startsWith('AI') key format check from gemini_vision.py - Remove vertex icon from UI icon map and asset bundle - Hide modelSource field in services.common.llm.json (was readonly) - Bump client-typescript SDK version to 1.1.0 - Update test mocks to remove langchain_google_vertexai mock Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(task): add backward-compat fallback for token in DAP request root get_task_token() now checks arguments.token first, then falls back to request.token so older clients and debug command token injection still work. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix(ui): prioritize URL auth param over sessionStorage token The URL param carries the freshly-minted key for the current task and must not be shadowed by a stale sessionStorage value from a previous task on the same origin. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix(ui): remove redundant auth check in dropper-ui to match chat-ui Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: ryan-t-christensen <ryan.christensen@rocketride.ai> Co-authored-by: koushik sreevatsa <koushiksreevatsa100@gmail.com> Co-authored-by: Alexandru Sclearuc <asclearuc@gmail.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: Anand Ray <anandray@users.noreply.github.com> Co-authored-by: Joshua Phillips <J.dspears@yahoo.com> Co-authored-by: Dmitrii Karataev <dmitry.karataev@gmail.com> Co-authored-by: Dylan Savage <dylanmsavage@hotmail.com> Co-authored-by: wdwd720 <mihirmodi808@gmail.com> Co-authored-by: Nechar KC <126724860+SpeedRelativity@users.noreply.github.com> Co-authored-by: Dmitrii Kataraev <dmitrii.kataraev@rocketride.ai>
…sion-gemini feat(nodes): add gemini vision node
PRs rocketride-org#672 (gemini-vision) and rocketride-org#677 (openai-vision) merged with `ruff format --check` failures on three files, leaving develop CI red and gating the post-CI nightly publish workflow_run. Pure whitespace changes — line wraps long tuple unpacks and except-clause type tuples per ruff's default style — no semantic edits. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Summary
Type
feature
Testing
./builder testpassesChecklist
Linked Issue
Fixes #
Summary by CodeRabbit
New Features
Documentation