feat(nodes): add tool_pipedrive CRM tool node - #1687
Open
joshuadarron wants to merge 14 commits into
Open
Conversation
Exposes the Pipedrive REST API v1 to agents: 256 tools across 24 resource
groups (deals, persons, organizations, activities, pipelines, stages, notes,
leads, products, fields, files, users, roles, permission sets, teams, goals,
filters, webhooks, subscriptions, mailbox, call logs, projects, misc), plus a
generic `request` tool that reaches any remaining v1 endpoint through the same
auth, rate-limit and read-only layer.
Full coverage is more tools than an LLM can choose between reliably, so
`pipedrive.toolGroups` controls which groups are published. The default eight
core CRM groups publish 108 tools; `["all"]` publishes everything. Filtering
happens in `IInstance._collect_tool_methods()`, so a gated tool is invisible to
`tool.query` and refused by `tool.invoke` alike.
Client details specific to Pipedrive:
- `{"success", "data", "additional_data"}` envelope is unwrapped; `success:
false` on an HTTP 200 is raised as an error.
- `x-ratelimit-reset` is the seconds remaining in the window, not an epoch
timestamp as in the GitHub client this otherwise mirrors.
- Offset pagination (`start`/`limit`, max 500) everywhere except projects,
which use cursor pagination.
- Personal API tokens go in the `api_token` query parameter; JWT, long or
`Bearer `-prefixed values go in an Authorization header.
- Custom fields are grouped under `custom_fields` on read and written through
the `extra` passthrough that every create and update tool accepts.
Tests: 81 unit tests that need no credentials or network, plus an env-gated
live suite behind PIPEDRIVE_API_TOKEN, with record-creating tests additionally
behind PIPEDRIVE_ALLOW_WRITES.
Also corrects the stale node/service counts in docs/README-nodes.md
(88 -> 118 directories, 118 -> 148 services).
Closes #1675
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Contributor
|
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:
📝 WalkthroughWalkthroughThe PR adds a Pipedrive CRM tool node, changes CrewAI planning and tool handling, normalizes typed AI content blocks, adds in-memory live event sessions, and improves test import isolation. ChangesPipedrive CRM tool node
CrewAI manager and tool bridge
Typed content blocks and live events
Estimated code review effort: 5 (Critical) | ~120 minutes Possibly related issues
Possibly related PRs
Suggested reviewers: 🚥 Pre-merge checks | ✅ 3 | ❌ 2❌ Failed checks (2 warnings)
✅ Passed checks (3 passed)
✨ Finishing Touches 💡 1📝 Generate docstrings 💡
🧪 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 |
Contributor
🤖 Internal: Discord sync markerAuto-managed by the Discord notification workflow. Stores the linked Discord message ID and forum thread ID. Do not edit or delete. |
…icon The first icon was a green uppercase "P" next to an unrelated dark dot. The actual Pipedrive mark is a solid green disc carrying a white lowercase "p" whose counter shows the disc through it. Redrawn as plain geometry (disc r=128, stem 36x150, bowl r=56, counter r=27) so it stays crisp at 16px with no gradients, clip paths or embedded raster. Uses the default nonzero fill rule with the counter arc wound opposite to the bowl. Under evenodd the stem and bowl would XOR where they overlap and punch a green notch out of their junction; nonzero merges them solid while the opposing winding still renders the counter as a hole. Keeps the viewBox (SVGR runs with dimensions:false, so an icon without one computes to zero width) and keeps two distinct colours, which makes the auto-currentcolor svgo plugin treat it as a multicolour brand mark and pass the authored colours through untinted. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…work The previous version approximated the mark with circles. The real letterform differs in three ways that are visible even at 16px: the stem carries a spur at the top left, the bowl is an ellipse rather than a circle, and the counter is an ellipse tangent to the stem's right edge. Geometry was measured off the supplied 272x272 artwork rather than eyeballed — connected-component analysis to isolate the counter, then a least-squares fit of the bowl ellipse to the glyph's right-hand edge (max residual 2px). Rasterising the traced primitives and diffing against the source leaves 1.47% of opaque pixels differing, all of it a sub-pixel outline along the shape boundary. Still a two-colour multicolour mark, so the auto-currentcolor svgo plugin passes the authored colours through, and still nonzero winding with the counter wound opposite the bowl so the stem and bowl merge solid while the counter stays open. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… agent Full coverage is 256 tools. An operator can already reach that by selecting groups, and nothing told them the agent's tool choice degrades long before it. Adds a guard rail that counts published tools rather than groups, because group sizes are lopsided: deals is 28 tools and permission_sets is 3, so a group count bounds nothing useful. The threshold is 120, just under the 128-function cap some providers impose on a single request and above the 107 the default groups publish. Crossing it warns and carries on. The tools are still published, in both validateConfig (so it shows in the config panel while editing) and beginGlobal (so it shows at pipeline start). Silently dropping tools an operator asked for surfaces later as "the agent won't do X", which is worse than a noisy panel. The warning is deliberately not raised from _collect_tool_methods, which runs on every tool.query and tool.invoke. "all" is exempt: it is already an explicit opt-in to the whole surface, and scripted callers that never route through an LLM legitimately want it. Listing every group by name is not exempt. Group sizes are derived from the __pipedrive_group__ stamps rather than a hand-maintained table, so a new tool cannot drift out of sync; a test asserts every group in ALL_GROUPS has a count. Another test pins the advisory semantics so the warning is not later "fixed" into a silent truncation. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Anthropic with extended thinking returns `content` as a list of typed blocks, not a string. Every non-streaming path handed that list straight to callers, who stringified it — so the answer arrived as the blocks' repr, carrying the raw reasoning and its base64 signatures. For agents it was worse than cosmetic: the ReAct parser reads output back with a greedy `Action Input: (.*)`, so the serialized wrapper's trailing `", "type": "text"}]` landed inside the tool arguments and every declared field then reported as missing. Extract the block vocabulary that only the streaming path knew into `ai.common.utils.flatten_content_blocks` and route chat's invoke, the stream fallback, and the agent's `extract_text` through it. The streaming path now shares that vocabulary instead of duplicating it, so the paths cannot drift; the fallback also forwards recovered reasoning to `on_reasoning_chunk` rather than dropping it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three defects made the CrewAI nodes fail on runs that should have worked. Tool arguments were dropped. `_ToolInput` declared a single `input` field, and CrewAI's argument filter intersects the model's emitted keys with the schema's property names — so every real parameter was discarded before validation, which then reported all required fields missing. Generate a typed model per tool instead: JSON Schema types map to real annotations (so pydantic coerces "123" -> 123), optional fields accept an explicit null, and paramless tools get an empty schema. The JSON schema is no longer appended to the description either; CrewAI renders args_schema into the prompt itself, and the two copies disagreed. `HostInvokeLLM` never declared `supports_function_calling()`. CrewAI defines it on `LLM` but not on `BaseLLM`, and several call sites invoke it unguarded — converter.py, tool_usage.py, reasoning_handler, and task_evaluator. The AttributeError was swallowed by converter's broad except and re-reported as "Failed to convert text into a Pydantic model". Declare it False: the host channel is text-in / text-out and cannot honour native tool calls, so every call belongs on the ReAct path. Pin crewai <2 for the same reason — 1.15.x moves the BaseLLM contract. Hierarchical crews then died in planning. `Crew(planning=True)` runs a task with `output_pydantic` set and drives it with the host wrapper, which cannot promise JSON; unparseable plan text dead-ends the converter and raises "Failed to get the Planning output". Make planning opt-in via a new **Crew Planning** field, off by default. Also clear `Task.tools` after construction: CrewAI back-fills them from the task's agent, and a hierarchical crew resolves the executing agent's toolset for the manager — which let the manager work a delegate's tools itself instead of delegating. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Pipedrive retired the v1 search routes: `/persons/search` and its siblings answer 404 "Unknown method ." and `/itemSearch` answers a plain 404, so every search tool the node published was dead. Non-search v1 endpoints still route normally, so only the search calls move — a blanket swap would break the many endpoints with no v2 equivalent. Adds a v2 base URL (company-domain aware, sharing v1's normalisation) and a v2 envelope wrapper: v2 replaced offset pagination with an opaque `next_cursor`, which `paginated` would have missed entirely and reported as a single page. Search tool schemas now advertise `cursor` instead of `start`, and `item_search_by_field` follows v2's renames — `field_type` -> `entity_type` with the "Field" suffix dropped from each value, and `exact_match` replaced by the three-way `match` mode. Also turns Tool groups into a multi-select dropdown with per-group tool counts, so the field no longer requires typing group names from memory. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
PR #1661 rewired every accumulating pane — Trace, Errors, Status, Tokens, Flow, Log, Analyze — to read through a DVR session obtained from the host's `openEventStream`, a prop no host in this repo binds. Hosts that have not bridged it across their transport (the VS Code webview is one) passed a null session, `useTaskEvents.ingestLive` dropped every event, and the panes rendered their empty states mid-run with no error anywhere. Fall back to an in-memory store over the live events the host already delivers. It satisfies the same session contract `deliver` expects — a pure append, each event handed over exactly once, oldest first, across repeated `play` calls — and is keyed per stream identity so the several views `SourceSection` opens all share one buffer. Replay of recorded runs still requires the real stream. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
github-actions
Bot
added
module:ai
nodes/test is a package whose subdirectories are named after node packages (text_output/, response/, image_cleanup/, telegram/, tool_git/, ...). Two places put that directory on sys.path so they could bare-import a test-root module: conftest (for _sys_modules_guard) and cloud_tts/test_cloud_tts (for test_contracts). Once nodes/test sits ahead of src/nodes, `import text_output` resolves to the *test* package, whose __path__ holds no node module — so `from text_output.instance import Instance` dies with ModuleNotFoundError. Whether it bites was pure collection order. The affected modules bootstrap with "insert src/nodes only if absent", which never re-prioritises: previously nothing seeded src/nodes before cloud_tts (c) inserted nodes/test, so image_cleanup (i) was the first to insert src/nodes and landed it in front. The new agent_crewai tests seed src/nodes during collection of agent_crewai (a) instead, so cloud_tts now stacks nodes/test on top of it and the guard short-circuits in all four later modules — four collection errors, none of them in code this branch touches. Import both test-root modules package-relative and drop the sys.path inserts, so nodes/test never shadows a node package. cloud_tts gains the __init__.py the rest of the node-named test dirs already have, which is what makes the relative import resolvable. The four bootstraps now move src/nodes to the front unconditionally rather than skipping when it is already somewhere on the path, and the guard suite pins both invariants: nodes/test stays off sys.path, and a node package imported by name resolves under src/nodes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Collaborator
Author
|
@coderabbitai review |
Contributor
✅ Action performedReview finished.
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 19
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
packages/shared-ui/src/modules/project/ProjectView.tsx (1)
486-511: 🩺 Stability & Availability | 🟡 Minor | ⚡ Quick winClose the live-edge store when the project identity changes.
SourceSectioncloses each session it opens (paneandfetchChapterEventswalker), butProjectViewkeepsliveStoresRef.currentacross renders and only creates newLiveEventStores; it never closes the store whensource.id,runKind, orprojectIdchanges. Track the current identity (at leastprojectId) and close the old store before re-creating it, otherwiseLiveEventStore.viewscan retain detached sessions and backstop trim events that can no longer be read.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/shared-ui/src/modules/project/ProjectView.tsx` around lines 486 - 511, Update the liveStore callback and liveStoresRef lifecycle so stores are closed when their identity changes, including at least projectId and the source.id/runKind key. Before replacing or discarding a store in liveStoresRef.current, call its close method, and ensure the active identity is tracked so unchanged keys continue reusing their existing store without closing it.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@nodes/src/nodes/agent_crewai/crewai_base.py`:
- Around line 461-466: Update the docstring of _build_crew_tools to remove the
claim that each tool’s JSON Schema is embedded in its description, and document
that the schema is provided through CrewAI’s args_schema instead. Keep the rest
of the method documentation unchanged.
- Around line 363-370: Update the nested _annotation_for function used by
_build_crew_tools to inspect prop.get('type') before _JSON_TYPE_MAP lookup,
returning Any whenever the type value is not a string (including JSON Schema
arrays such as ["string", "null"]); retain the existing mapped annotation
behavior for string types.
In `@nodes/src/nodes/tool_pipedrive/pipedrive_client.py`:
- Around line 175-182: The company domain normalization in the relevant
Pipedrive client helper must parse and validate exactly one Pipedrive subdomain
before returning it. Reject fragments, queries, credentials, ports, URL
delimiters, and any label containing characters outside lowercase letters,
digits, or hyphens; return None for invalid input, while preserving valid bare
subdomains and supported Pipedrive URLs.
In `@nodes/src/nodes/tool_pipedrive/requirements.txt`:
- Line 1: Pin the requests dependency in requirements.txt to the
constraints-backed safe version requests==2.34.2, matching the workspace
constraint and preventing resolution to older versions.
In `@nodes/src/nodes/tool_pipedrive/tools/call_logs.py`:
- Around line 121-124: URL-encode every model-supplied string path segment with
urllib.parse.quote(value, safe='') before interpolation, preferably through a
shared helper in _base.py. Apply this to log_id in call_log_delete,
call_log_get, and call_log_recording_add; lead UUIDs in lead_get and
lead_delete; label UUIDs in lead_label_delete and lead_label_update; and set_id
in permission_set_get and permission_set_assignments_list across the specified
call_logs.py, leads.py, and roles.py sites.
In `@nodes/src/nodes/tool_pipedrive/tools/deals.py`:
- Around line 65-99: The write-key tuples contain fields unavailable through
their corresponding schemas; make each key/schema pair consistent. In
nodes/src/nodes/tool_pipedrive/tools/deals.py:65-99, organizations.py:63-72, and
persons.py:62-72, either expose add_time in _DEAL_WRITE_PROPS, _ORG_WRITE_PROPS,
and _PERSON_WRITE_PROPS or remove it from the matching write-key tuple. In
nodes/src/nodes/tool_pipedrive/tools/leads.py:52-65, either expose origin_id in
_LEAD_WRITE_PROPS or remove it from _LEAD_WRITE_KEYS.
- Around line 576-585: Add a shared _delete_bulk helper in _base.py that calls
_require_write() before validating ids, validates the required non-empty list,
builds the CSV parameters, applies optional extra passthrough, and performs the
DELETE request. Mark ids as required and delegate deal_delete_bulk in
nodes/src/nodes/tool_pipedrive/tools/deals.py#L576-L585, activity_delete_bulk
and activity_type_delete_bulk in
nodes/src/nodes/tool_pipedrive/tools/activities.py#L176-L183 and `#L253-L260`,
organization_delete_bulk in
nodes/src/nodes/tool_pipedrive/tools/organizations.py#L179-L186,
person_delete_bulk in nodes/src/nodes/tool_pipedrive/tools/persons.py#L208-L215,
and stage_delete_bulk in
nodes/src/nodes/tool_pipedrive/tools/pipelines.py#L280-L287 to the helper,
preserving the deals extra passthrough.
- Around line 558-566: Remove the deal_followers_users_list tool and its
associated registration, leaving deal_followers_list as the sole tool for the
/deals/{deal_id}/followers endpoint; do not apply clean_user to follower rows.
In `@nodes/src/nodes/tool_pipedrive/tools/fields.py`:
- Around line 102-144: The declared EXTRA fields are omitted from request
bodies. In nodes/src/nodes/tool_pipedrive/tools/fields.py lines 102-144, update
field_create and field_update; in nodes/src/nodes/tool_pipedrive/tools/misc.py
lines 127-164, update channel_message_receive. After each body_from call and
before self._write, merge args.get('extra') or {} into the body.
In `@nodes/src/nodes/tool_pipedrive/tools/files.py`:
- Around line 63-70: Update file_list in
nodes/src/nodes/tool_pipedrive/tools/files.py (lines 63-70) to use a
Files-specific schema or validation that caps limit at 100 while preserving the
existing pagination behavior. Update nodes/src/nodes/tool_pipedrive/README.md
(lines 77-82) to qualify the general 1–500 limit statement and document the
Files endpoint’s 100-item maximum.
In `@nodes/src/nodes/tool_pipedrive/tools/products.py`:
- Around line 307-312: Replace the response cleaner argument dict with the
already imported passthrough cleaner in product_variation_update at
nodes/src/nodes/tool_pipedrive/tools/products.py:307-312,
project_plan_activity_update at
nodes/src/nodes/tool_pipedrive/tools/projects.py:234-239, and
project_plan_task_update at
nodes/src/nodes/tool_pipedrive/tools/projects.py:256-261. Preserve the existing
PUT requests and bodies.
In `@nodes/src/nodes/tool_pipedrive/tools/projects.py`:
- Around line 55-58: Replace the local _CURSOR_PROPS and _cursor_list
cursor-pagination implementation with the existing PAGING_V2(),
paging_params_v2(), and paginated_v2 helpers already used by the search tools.
Update the affected project-listing flow to preserve the shared clamping, cursor
trimming, count semantics, and paginated envelope behavior.
In `@nodes/test/tool_pipedrive/test_pipedrive.py`:
- Around line 198-203: Update the domain parameter list in
test_company_domain_forms so the case-normalization case passes an uppercase
value directly, rather than calling lower() before base_url_for receives it.
Keep the existing expected normalized URL and other domain-form cases unchanged.
- Around line 388-402: Reorder parametrization and patch injection for all four
tests in nodes/test/tool_pipedrive/test_pipedrive.py: lines 388-402
(test_malformed_retry_after_falls_back_to_backoff), 575-597
(test_writes_are_blocked), 870-886 (test_search_tools_target_v2), and 910-924
(test_non_search_tools_stay_on_v1). Make pytest.mark.parametrize the outermost
decorator and update signatures respectively to (self, bad_value, mock_request,
mock_sleep), (self, tool, args, mock_request), (self, tool, args, path,
mock_request), and (self, tool, args, path, mock_request).
In `@nodes/test/tool_pipedrive/test_tools.py`:
- Around line 422-433: Update test_create_and_delete to resolve a valid deal
field from the /dealFields endpoint before constructing conditions, replacing
the hard-coded field_id '1'. If no suitable field exists, skip the test as the
prerequisite is unavailable; otherwise use the resolved field identifier in the
filter body.
- Around line 413-415: Remove the vacuous base64.b64encode assertion in the
download test, since the preceding downloaded == payload assertion already
verifies the raw response; if validating encoded output is required, compare it
against the expected base64 value returned by file_download instead of asserting
truthiness.
- Around line 270-277: Update test_pipeline_movement_statistics to assert that
pipelines is non-empty before indexing pipelines[0], matching the guard in
test_pipeline_conversion_statistics. Keep the existing movement_statistics
request and assertions unchanged after the guard.
- Around line 27-28: Update the imports in the live test to use the node package
path, matching test_pipedrive.py: import PipedriveAPIError, base_url_for,
base_url_v2_for, call, and call_envelope from tool_pipedrive.pipedrive_client,
and remove the sys.path insertion and bare pipedrive_client import.
In `@packages/shared-ui/src/modules/project/hooks/liveEventSession.ts`:
- Around line 158-170: Update TaskEventSession.getStatus() to scan only events
at or before the current session watermark established by seek(anchor), rather
than the entire buffer tail. Preserve the existing status-event filtering and
null fallback, ensuring useTaskEvents receives the status snapshot corresponding
to the replay position.
---
Outside diff comments:
In `@packages/shared-ui/src/modules/project/ProjectView.tsx`:
- Around line 486-511: Update the liveStore callback and liveStoresRef lifecycle
so stores are closed when their identity changes, including at least projectId
and the source.id/runKind key. Before replacing or discarding a store in
liveStoresRef.current, call its close method, and ensure the active identity is
tracked so unchanged keys continue reusing their existing store without closing
it.
🪄 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 Plus
Run ID: 5f6cbec3-08b8-4c0e-b062-f1d70c93a4ca
⛔ Files ignored due to path filters (1)
nodes/src/nodes/tool_pipedrive/pipedrive.svgis excluded by!**/*.svg
📒 Files selected for processing (67)
CHANGELOG.mddocs/README-nodes.mdnodes/src/nodes/agent_crewai/README.mdnodes/src/nodes/agent_crewai/crewai_base.pynodes/src/nodes/agent_crewai/crewai_manager/IGlobal.pynodes/src/nodes/agent_crewai/crewai_manager/README.mdnodes/src/nodes/agent_crewai/crewai_manager/manager.pynodes/src/nodes/agent_crewai/requirements.txtnodes/src/nodes/agent_crewai/services.manager.jsonnodes/src/nodes/tool_pipedrive/IGlobal.pynodes/src/nodes/tool_pipedrive/IInstance.pynodes/src/nodes/tool_pipedrive/README.mdnodes/src/nodes/tool_pipedrive/__init__.pynodes/src/nodes/tool_pipedrive/pipedrive_client.pynodes/src/nodes/tool_pipedrive/requirements.txtnodes/src/nodes/tool_pipedrive/services.jsonnodes/src/nodes/tool_pipedrive/tool_groups.pynodes/src/nodes/tool_pipedrive/tools/__init__.pynodes/src/nodes/tool_pipedrive/tools/_base.pynodes/src/nodes/tool_pipedrive/tools/activities.pynodes/src/nodes/tool_pipedrive/tools/call_logs.pynodes/src/nodes/tool_pipedrive/tools/deals.pynodes/src/nodes/tool_pipedrive/tools/fields.pynodes/src/nodes/tool_pipedrive/tools/files.pynodes/src/nodes/tool_pipedrive/tools/filters.pynodes/src/nodes/tool_pipedrive/tools/goals.pynodes/src/nodes/tool_pipedrive/tools/leads.pynodes/src/nodes/tool_pipedrive/tools/mailbox.pynodes/src/nodes/tool_pipedrive/tools/misc.pynodes/src/nodes/tool_pipedrive/tools/notes.pynodes/src/nodes/tool_pipedrive/tools/organizations.pynodes/src/nodes/tool_pipedrive/tools/persons.pynodes/src/nodes/tool_pipedrive/tools/pipelines.pynodes/src/nodes/tool_pipedrive/tools/products.pynodes/src/nodes/tool_pipedrive/tools/projects.pynodes/src/nodes/tool_pipedrive/tools/roles.pynodes/src/nodes/tool_pipedrive/tools/search.pynodes/src/nodes/tool_pipedrive/tools/subscriptions.pynodes/src/nodes/tool_pipedrive/tools/teams.pynodes/src/nodes/tool_pipedrive/tools/users.pynodes/src/nodes/tool_pipedrive/tools/webhooks.pynodes/test/agent_crewai/test_llm_contract.pynodes/test/agent_crewai/test_manager_planning.pynodes/test/agent_crewai/test_manager_tool_scoping.pynodes/test/agent_crewai/test_stop_words.pynodes/test/agent_crewai/test_tool_schema.pynodes/test/cloud_tts/__init__.pynodes/test/cloud_tts/test_cloud_tts.pynodes/test/conftest.pynodes/test/image_cleanup/test_provenance.pynodes/test/local_text_output/test_instance.pynodes/test/response/test_media_metadata.pynodes/test/test_sys_modules_guard.pynodes/test/text_output/test_instance.pynodes/test/tool_pipedrive/__init__.pynodes/test/tool_pipedrive/test_pipedrive.pynodes/test/tool_pipedrive/test_tools.pypackages/ai/src/ai/common/agent/_internal/utils.pypackages/ai/src/ai/common/chat.pypackages/ai/src/ai/common/utils/__init__.pypackages/ai/src/ai/common/utils/content_blocks.pypackages/ai/tests/ai/common/agent/__init__.pypackages/ai/tests/ai/common/agent/test_extract_text.pypackages/ai/tests/ai/common/utils/test_content_blocks.pypackages/shared-ui/src/modules/project/ProjectView.tsxpackages/shared-ui/src/modules/project/hooks/liveEventSession.test.tspackages/shared-ui/src/modules/project/hooks/liveEventSession.ts
…e PR Security - pipedrive_client._company_host accepted anything up to the first slash, so a configured domain of "evil.example#" built https://evil.example#.pipedrive.com and Requests sent it — with the api_token or Authorization header attached — to evil.example. Cut at the first path/query/fragment delimiter, refuse userinfo and ports, lower-case, and require a single DNS label; anything else falls back to the generic host. - Free-form ids (lead uuids, permission-set ids, channel ids, goal ids, note comment ids) were f-stringed into request paths unencoded, so a value carrying "/", "..", "?" or "#" retargeted the call at another endpoint — for the DELETE tools, a destructive one. Added _base.path_segment() and applied it at all 20 string-id sites. - Pin requests>=2.34.2 to match constraints.lock, as tool_cognee/tool_github/ tool_n8n already do. Correctness - crewai_base._annotation_for passed prop['type'] straight to dict.get. JSON Schema spells a nullable field as ["string", "null"], and an unhashable key raises TypeError instead of returning the default — one such property aborted the whole _build_crew_tools loop. Non-str types now fall back to Any, which is what the docstring already promised. - deal_followers_users_list GET the same /deals/{id}/followers as deal_followers_list and only differed by running follower rows through clean_user, emitting a user-shaped object built from mismatched fields. Dropped it; resolving followers to users needs a /users/{id} call per follower. - add_time (deals, organizations, persons) and origin_id (leads) were in the write-key tuples but not in the matching schema props, so body_from could never populate them. Added the properties. - product_variation_update and both project-plan updates passed the builtin dict as the response cleaner: dict(None) raises TypeError on an empty 204 body. Switched to passthrough, as every other write already uses. - liveEventSession.getStatus scanned from the live edge, contradicting the documented "as of the session position" and the caller, which seeks to an anchor first — a repositioned view seeded a replayed run with the current status. Scan back from the watermark. Maintainability - Eight bulk deletes each repeated the ids guard, the CSV build and a trailing _require_write(), and left ids optional so a missing argument surfaced as a hand-rolled ValueError. Added _base._delete_bulk (write gate first), marked ids required in every schema, and delegated all eight. - projects._cursor_list re-implemented paging_params_v2 and the v2 envelope shaping; it now uses the shared helpers, so the cursor contract has one implementation instead of two that can drift. - file_list advertised the shared 1-500 range, but /files documents a maximum limit of 100. PAGING/paging_params/_list take a max_limit override and file_list passes 100. - _build_crew_tools' docstring still claimed the JSON Schema is embedded in the tool description, which stopped being true when args_schema took over. - Tests: 'ACME'.lower() evaluated to 'acme' before the call, so the case path was never exercised; test_tools imported pipedrive_client bare, giving the module a second identity in a run that also collects test_pipedrive; a vacuous b64encode assert, a missing empty-pipelines guard, and a hard-coded field_id of 1 that fails rather than skips on a fresh sandbox. Two findings were not applied. `extra` is not dropped by fields/misc — body_from merges args['extra'] by default (tools/_base.py). And the @patch/parametrize signatures are already correct: mock injects positionally innermost-first while pytest passes parametrize values by keyword, so the suggested reorder would bind the mocks to the parametrized names. New tests cover the rejected domains, path-segment encoding, non-str schema types, the write-gate ordering, and the watermark-bounded status. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Collaborator
Author
|
Addressed the CodeRabbit review in 7b3cc53. Security
Correctness
Maintainability
Dropping the duplicate followers tool moved the surface from 256 to 255 tools; Not applied — 2 findings
New tests cover the rejected domains, path-segment encoding, non- |
joshuadarron
marked this pull request as ready for review
July 30, 2026 01:41
joshuadarron
requested review from
Rod-Christensen,
jmaionchi and
stepmikhaylov
as code owners
July 30, 2026 01:41
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (3)
nodes/src/nodes/tool_pipedrive/services.json (1)
59-89: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick winMake tool counts and the warning guard account for the raw request consistently.
The listed groups sum to 254, and the default eight sum to 106. Since
requestis separately published whenallowRawRequestis enabled, the documented default of 108 cannot be reached from these counts; moreover, the guard currently omits that extra published tool.
nodes/src/nodes/tool_pipedrive/services.json#L59-L89: Correct the displayed per-group, default, and “All groups” counts after deriving them from registered tools; keep raw-request counting separate from group selection.nodes/src/nodes/tool_pipedrive/README.md#L38-L50: Align the 255/108 claims with the corrected configuration counts and clarify whetherrequestis included.nodes/src/nodes/tool_pipedrive/tool_groups.py#L29-L35: Update the guard contract and implementation soallowRawRequest=Truecontributes one tool before applying the 120-tool warning.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@nodes/src/nodes/tool_pipedrive/services.json` around lines 59 - 89, Derive the Pipedrive group, default, and all-groups tool counts from the registered tools, then update nodes/src/nodes/tool_pipedrive/services.json lines 59-89 accordingly while keeping the raw request separate from group counts. Align the documented totals and explicitly state whether request is included in nodes/src/nodes/tool_pipedrive/README.md lines 38-50. Update the warning guard contract and implementation in nodes/src/nodes/tool_pipedrive/tool_groups.py lines 29-35 so allowRawRequest=True adds one tool before applying the 120-tool threshold.nodes/src/nodes/agent_crewai/crewai_base.py (2)
297-315: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winAdd a PEP 257 docstring to the new override.
supports_function_callingis a new public method innodes/**/*.py, but currently has only inline comments. Add a short triple-double-quoted docstring while retaining the detailed rationale if useful.As per path instructions, Python pipeline nodes require PEP 257 docstrings; the repository convention uses triple-double-quoted docstrings.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@nodes/src/nodes/agent_crewai/crewai_base.py` around lines 297 - 315, Add a concise triple-double-quoted PEP 257 docstring immediately inside the supports_function_calling method, describing that it reports whether native function calling is supported. Retain the existing detailed inline rationale and return behavior unchanged.Sources: Path instructions, Learnings
390-396: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick winGuard malformed
requiredlists before building the model.
set(required)raisesTypeErrorfor malformed schemas such asrequired: [[]]. This occurs before thecreate_modelfallback and aborts the entire tool-construction loop instead of isolating the invalid descriptor.Proposed fix
- required_keys = set(required) if isinstance(required, (list, tuple, set)) else set() + required_keys = ( + {item for item in required if isinstance(item, str)} + if isinstance(required, (list, tuple, set)) + else set() + )Based on the PR objective that malformed descriptors must not prevent valid tools from being included, malformed required entries need the same safe fallback.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@nodes/src/nodes/agent_crewai/crewai_base.py` around lines 390 - 396, Update the required-key handling near input_schema and required_keys to validate that every required entry is a string before constructing the set; treat malformed values such as nested lists as an invalid descriptor and return _NoArgsInput, allowing the surrounding tool-construction loop to continue processing valid tools.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@nodes/src/nodes/tool_pipedrive/tools/_base.py`:
- Around line 211-214: Update the delete parameter construction in the
bulk-delete method to apply the validated `extra` dictionary before assigning
the CSV value derived from `ids`; ensure the final `params['ids']` always uses
the validated input and cannot be overridden by `args[extra_key]`.
- Around line 281-290: Update path_segment to reject literal "." and ".." values
before URL encoding, while continuing to quote all other values with safe='' so
reserved characters remain encoded. Ensure callers receive a clear validation
error for these dot-segment IDs and preserve existing behavior for valid
identifiers.
---
Outside diff comments:
In `@nodes/src/nodes/agent_crewai/crewai_base.py`:
- Around line 297-315: Add a concise triple-double-quoted PEP 257 docstring
immediately inside the supports_function_calling method, describing that it
reports whether native function calling is supported. Retain the existing
detailed inline rationale and return behavior unchanged.
- Around line 390-396: Update the required-key handling near input_schema and
required_keys to validate that every required entry is a string before
constructing the set; treat malformed values such as nested lists as an invalid
descriptor and return _NoArgsInput, allowing the surrounding tool-construction
loop to continue processing valid tools.
In `@nodes/src/nodes/tool_pipedrive/services.json`:
- Around line 59-89: Derive the Pipedrive group, default, and all-groups tool
counts from the registered tools, then update
nodes/src/nodes/tool_pipedrive/services.json lines 59-89 accordingly while
keeping the raw request separate from group counts. Align the documented totals
and explicitly state whether request is included in
nodes/src/nodes/tool_pipedrive/README.md lines 38-50. Update the warning guard
contract and implementation in nodes/src/nodes/tool_pipedrive/tool_groups.py
lines 29-35 so allowRawRequest=True adds one tool before applying the 120-tool
threshold.
🪄 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 Plus
Run ID: 25ca362e-f6fd-4add-a0b0-c0c997c7680c
📒 Files selected for processing (28)
nodes/src/nodes/agent_crewai/crewai_base.pynodes/src/nodes/tool_pipedrive/README.mdnodes/src/nodes/tool_pipedrive/pipedrive_client.pynodes/src/nodes/tool_pipedrive/requirements.txtnodes/src/nodes/tool_pipedrive/services.jsonnodes/src/nodes/tool_pipedrive/tool_groups.pynodes/src/nodes/tool_pipedrive/tools/_base.pynodes/src/nodes/tool_pipedrive/tools/activities.pynodes/src/nodes/tool_pipedrive/tools/call_logs.pynodes/src/nodes/tool_pipedrive/tools/deals.pynodes/src/nodes/tool_pipedrive/tools/fields.pynodes/src/nodes/tool_pipedrive/tools/files.pynodes/src/nodes/tool_pipedrive/tools/filters.pynodes/src/nodes/tool_pipedrive/tools/goals.pynodes/src/nodes/tool_pipedrive/tools/leads.pynodes/src/nodes/tool_pipedrive/tools/misc.pynodes/src/nodes/tool_pipedrive/tools/notes.pynodes/src/nodes/tool_pipedrive/tools/organizations.pynodes/src/nodes/tool_pipedrive/tools/persons.pynodes/src/nodes/tool_pipedrive/tools/pipelines.pynodes/src/nodes/tool_pipedrive/tools/products.pynodes/src/nodes/tool_pipedrive/tools/projects.pynodes/src/nodes/tool_pipedrive/tools/roles.pynodes/test/agent_crewai/test_tool_schema.pynodes/test/tool_pipedrive/test_pipedrive.pynodes/test/tool_pipedrive/test_tools.pypackages/shared-ui/src/modules/project/hooks/liveEventSession.test.tspackages/shared-ui/src/modules/project/hooks/liveEventSession.ts
Two follow-ups from the second CodeRabbit pass, both on helpers added in the
previous commit.
_delete_bulk merged the caller's `extra` pass-through after assigning the
validated `ids`, so `deal_delete_bulk({'ids': [1], 'extra': {'ids': '2,3'}})`
deleted 2 and 3 instead of 1. Merge `extra` first and let the validated CSV win.
The hand-rolled version this replaced had the same ordering, so the flaw predates
the helper.
path_segment relied on quote() alone, but "." and ".." contain nothing quote
escapes — `/notes/{id}/comments/..` survived encoding intact and addressed the
parent resource. Reject those two segments instead of encoding them.
Tests cover both: a dot-segment id raises before any request is made, and a
pass-through "ids" no longer displaces the validated list while other extra
parameters still reach the query string.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
asclearuc
requested changes
Jul 30, 2026
asclearuc
left a comment
asclearuc
left a comment
Collaborator
There was a problem hiding this comment.
Thanks @joshuadarron — this is a large, careful node: 256 tools across 24 resource groups, a clean read-only/write gate, offset + cursor pagination, and a strong credential-free unit suite. You have already fixed every Major/Critical CodeRabbit finding (SSRF on companyDomain, path-segment encoding, the shared bulk-delete helper, the dict-vs-passthrough cleaners, the array-type schema fix), and two of its Major findings were withdrawn as false positives.
Requesting changes for one security defect: the personal API token can leak into error text and logs on a transport-level failure — see the inline comment on pipedrive_client.py:292. Blocking alongside it: the tool_* checklist requires a test proving the token is absent from the raised error even when the underlying requests exception text contains it, and TestErrors in nodes/test/tool_pipedrive/test_pipedrive.py has none yet — please add one with the fix.
Non-blocking, your call: this PR also bundles unrelated agent_crewai / packages/ai changes (the _chat return-contract change in chat.py affects every chat caller and has no direct test) and a shared-ui live-event-session hook — consider splitting them or calling them out in the description so those wider-blast-radius changes get focused review. Please also confirm nodes/src/nodes/constraints.lock was regenerated for the new node (it has no tool_pipedrive references while tool_github is listed; runtime resolves fine, so low risk). One inline nit on notes.py:56.
| timeout=timeout, | ||
| ) | ||
| except requests.RequestException as exc: | ||
| raise ValueError(f'Pipedrive request failed: {exc}') from exc |
Collaborator
There was a problem hiding this comment.
MUST fix — the CRM API token can leak into error messages and logs.
On the personal-token path, _auth() returns the credential as an api_token query parameter, so requests bakes it into the request URL: https://<host>.pipedrive.com/api/v1/deals?api_token=<SECRET>. When the request fails at transport level (DNS failure, connection refused, TLS error, timeout), requests raises a RequestException whose text contains that full URL. This line stringifies it:
raise ValueError(f'Pipedrive request failed: {exc}') from exc{exc} copies the token verbatim into the ValueError the agent framework surfaces and logs. The project tool_* checklist requires error messages to be built only from structured fields, never from str(exc).
Build the message from non-secret parts (exception type only), and drop the raw {exc}. Note that from exc chains the token-bearing exception, so a printed traceback can still re-expose it — suppress the chain with from None (below), or if you want to keep the cause detail, scrub the token, e.g. str(exc).replace(token, '***').
Suggested change
| raise ValueError(f'Pipedrive request failed: {exc}') from exc | |
| raise ValueError(f'Pipedrive request failed: {type(exc).__name__}') from None |
Please also add the secret-leak regression test the checklist calls for: assert the token does not appear in the raised error's message even when the mocked requests call raises an exception whose text contains the token.
Collaborator
Author
There was a problem hiding this comment.
Fixed in cdc66fa.
except requests.RequestException as exc:
raise ValueError(
f'Pipedrive request failed: {type(exc).__name__} on {method.upper()} {path}'
) from NoneOne deviation from your suggestion: I kept method and path in the message. Both are secret-free — path is the API path with no query string, which is where the credential lives — and without them the message can't tell you which call died. Say the word if you'd rather have the type alone.
Verified this was the only leak site in the client:
- line 308 stringifies
resp.text[:200]— the response body, not the URL. _raise_pipedrive_errorbuilds from parsederror/error_info/codefields.- the tenacity wrapper uses
reraise=True, so noRetryErrorrepr is introduced.
Three tests added to TestErrors, and all three fail against the previous code:
test_api_token_is_not_in_the_transport_error_or_its_traceback— mocksrequests.requestto raise aConnectionErrorwhose text embeds...?api_token=<TOKEN>, then asserts the token is in neitherstr(exc)nortraceback.format_exception(...). The traceback half is what pinsfrom None— you were right that the chained cause re-exposes it, and scrubbing the message alone wouldn't have been enough.test_transport_failure_names_the_failure_and_the_call— guards against the fix regressing into a bare string.- the existing
test_transport_failure_raises_value_error.
Red check before/after:
# against the old `f'...: {exc}' from exc`
FAILED test_transport_failure_names_the_failure_and_the_call
FAILED test_api_token_is_not_in_the_transport_error_or_its_traceback
E AssertionError: assert 'pipedrive-p...-placeholder' not in 'Pipedrive r...lutionError)'
# after
3 passed
Reused the existing PERSONAL_TOKEN placeholder rather than a token-shaped literal, so gitleaks stays quiet.
| 'lead_id', | ||
| 'project_id', | ||
| 'user_id', | ||
| 'add_time', |
Collaborator
There was a problem hiding this comment.
nit — add_time is unreachable for notes.
add_time is listed in _NOTE_WRITE_KEYS (line 56) but not in _NOTE_WRITE_PROPS (lines 63-74). note_create / note_update publish **_NOTE_WRITE_PROPS as the input schema, and body_from() only copies keys present in args; the framework validates args against that published schema. A key that is not a published property can never be populated, so an agent cannot set add_time on a note (it only works if smuggled through extra).
CodeRabbit raised this same class for deals and organizations, and you fixed those by adding add_time to their *_WRITE_PROPS; notes was missed. Either add 'add_time': STR(...) to _NOTE_WRITE_PROPS, or drop it from _NOTE_WRITE_KEYS.
Collaborator
Author
There was a problem hiding this comment.
Fixed in cdc66fa — added add_time to _NOTE_WRITE_PROPS with the same wording as deals/organizations/persons.
Good catch that notes was missed. I scanned every *_WRITE_KEYS / *_WRITE_PROPS pair across tools/*.py after fixing it, and notes was the last mismatch:
$ python - <<'PY' # keys present in *_WRITE_KEYS but absent from the matching *_WRITE_PROPS
...
PY
no mismatches
The personal-token path puts the credential in the `api_token` query parameter,
so requests bakes it into the URL. A transport-level failure (DNS, refused
connection, TLS, timeout) raises a RequestException whose text carries that URL,
and `raise ValueError(f'Pipedrive request failed: {exc}') from exc` copied the
token verbatim into the error the agent framework surfaces and logs. The chained
cause held the same URL, so scrubbing the message alone would still have leaked
it through any printed traceback.
Build the message from the exception type plus method and path — all secret-free,
and the type still distinguishes ConnectTimeout from SSLError from ConnectionError
— and raise `from None`. Verified this was the only such site: the JSON-decode
path stringifies `resp.text[:200]` rather than the URL, `_raise_pipedrive_error`
builds from parsed response fields, and the tenacity wrapper uses `reraise=True`
so no RetryError repr is introduced.
Three tests pin it, and all three fail against the previous code: the token
appears in neither the message nor the formatted traceback (that half is what
pins `from None`), and the message still names the failure and the call.
Also:
- notes' `add_time` was in _NOTE_WRITE_KEYS but not _NOTE_WRITE_PROPS, so
body_from could never populate it — the same class fixed for deals,
organizations, persons and leads in 7b3cc53. A scan of every
*_WRITE_KEYS/*_WRITE_PROPS pair confirms notes was the last one.
- ChatBase._chat had no direct test. flatten_content_blocks and extract_text are
covered, but nothing pinned the wiring that routes the provider's content
through them, which is the contract every non-streaming chat caller depends on.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Collaborator
Author
|
@asclearuc — both blocking items are fixed in cdc66fa, details in the two inline threads. Answers to the three non-blocking notes: 1. Bundled Fair point. I've rewritten the PR description with an "Also in this PR — changes outside the node" table naming each one, its blast radius, why it rides along, and where its tests are. Not splitting them: the CrewAI schema fix and the content-block flattening are what make the Pipedrive tools usable from an agent at all — without them every tool call arrives with 2.
3. Expected, and nothing to regenerate. Per # * push to develop / weekly / manual -> regenerate + auto-commit the lock.
# * pull_request touching a requirements.txt -> compile only, so a cross-node
# version conflict fails HERE (at CI) instead of bricking a user's first run.That gate ("Compile universal node constraints") is green on this PR, and the lock picks the node up when this merges. Local verification on cdc66fa: 2408 passed in the full non-dynamic node suite (the one failure is Re-requesting your review. |
…pipedrive # Conflicts: # nodes/src/nodes/agent_crewai/services.manager.json
Contributor
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
nodes/src/nodes/agent_crewai/README.md (1)
10-10: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick winDocument Crew Planning in the Manager configuration table.
The README now refers to the Crew Planning field, but the hand-written Manager field table omits it. Add a row describing its advanced-mode visibility, default
false, and JSON-output requirement. Update the hand-written section rather than the generated block.As per path instructions, generated README content must not be hand-edited.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@nodes/src/nodes/agent_crewai/README.md` at line 10, Update the hand-written CrewAI Manager configuration table in README.md to add a Crew Planning row, documenting that it is visible in advanced mode, defaults to false, and requires JSON output; do not modify the generated content block.Source: Path instructions
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Outside diff comments:
In `@nodes/src/nodes/agent_crewai/README.md`:
- Line 10: Update the hand-written CrewAI Manager configuration table in
README.md to add a Crew Planning row, documenting that it is visible in advanced
mode, defaults to false, and requires JSON output; do not modify the generated
content block.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: ASSERTIVE
Plan: Pro Plus
Run ID: 9e165474-28f8-43be-8a26-b384ed26cd75
📒 Files selected for processing (3)
nodes/src/nodes/agent_crewai/README.mdnodes/src/nodes/agent_crewai/crewai_manager/manager.pynodes/src/nodes/agent_crewai/services.manager.json
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Exposes the Pipedrive REST API v1 to agents: 255 tools across 24 resource groups (deals, persons, organizations, activities, pipelines, stages, notes, leads, products, fields, files, users, roles, permission sets, teams, goals, filters, webhooks, subscriptions, mailbox, call logs, projects, misc), plus a generic
requesttool that reaches any remaining v1 endpoint through the same auth, rate-limit and read-only layer.Full coverage is more tools than an LLM can choose between reliably, so
pipedrive.toolGroupscontrols which groups are published. The default eight core CRM groups publish 108 tools;["all"]publishes everything. Filtering happens inIInstance._collect_tool_methods(), so a gated tool is invisible totool.queryand refused bytool.invokealike.Client details specific to Pipedrive:
{"success", "data", "additional_data"}envelope is unwrapped;success: falseon an HTTP 200 is raised as an error.x-ratelimit-resetis the seconds remaining in the window, not an epoch timestamp as in the GitHub client this otherwise mirrors.start/limit, max 500) everywhere except projects (cursor) and/files(max 100).api_tokenquery parameter; JWT, long orBearer-prefixed values go in an Authorization header. Because the token rides in the query string, transport-level errors are built from the exception type plus method and path, neverstr(exc), and the cause chain is suppressed so a traceback cannot re-expose the URL.custom_fieldson read and written through theextrapassthrough that every create and update tool accepts.Also corrects the stale node/service counts in docs/README-nodes.md (88 -> 118 directories, 118 -> 148 services).
Also in this PR — changes outside the node
These are not Pipedrive code and carry a wider blast radius. Flagged here so they get
focused review rather than being read as part of the node.
packages/ai(chat.py,utils/content_blocks.py,agent/_internal/utils.py)ChatBase._chatreturns flattened visible text instead of the provider's rawcontent. Anthropic with extended thinking returns a list of typed blocks, so every caller was stringifying the blocks' repr — the raw reasoning and its base64 signatures landed in the answer.Action Input: (.*)swallowed the serialized wrapper's trailing", "type": "text"}]into the tool arguments, so every declared field reportedField required.tests/ai/common/utils/test_content_blocks.py,tests/ai/common/agent/test_extract_text.py, andtests/ai/common/test_chat_return_contract.py(the_chatreturn contract itself)nodes/src/nodes/agent_crewaiargs_schema;supports_stop_wordsguarded; planning made opt-in; hierarchical manager tool scoping.nodes/test/agent_crewai/—test_tool_schema.py,test_llm_contract.py,test_manager_planning.py,test_manager_tool_scoping.pypackages/shared-ui(liveEventSession.ts)openEventStream.liveEventSession.test.ts(18 tests)nodes/test/conftest.py,cloud_tts, four node test bootstrapsnodes/testreachingsys.path, where its node-named subdirectories shadow the real node packages.develop-adjacent code.nodes/test/test_sys_modules_guard.pyTesting
PIPEDRIVE_API_TOKEN(record-creating tests additionally behindPIPEDRIVE_ALLOW_WRITES)packages/ai689 passed./builder testpasses — CI green on Ubuntu, Windows and macOSChecklist
Linked Issue
Closes #1675
Summary by CodeRabbit
Summary
New Features
Bug Fixes
Documentation