Skip to content
Back to skills

Bitrix24 Agent Neversight

ASecurity

Design, implement, debug, and harden integrations between AI agents and Bitrix24 REST API (webhooks, OAuth 2.0, scopes, events, batch, limits, and REST 3.0). Use when asked to connect AI assistants/agents to Bitrix24, automate CRM/tasks/chats, process Bitrix24 events, choose an auth model, or resolve Bitrix24 API errors and performance issues.

  • 57 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 17, 2026
businesspythongobashrailsapiperformancedocumentation

Works with

  • cli
  • api

Security analysis

A100/100

Scanned October 17, 2026

npx -y skills add diegosouzapw/awesome-omni-skill --skill bitrix24-agent-neversight --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Bitrix24 Agent Neversight?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Bitrix24 Agent Neversight
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/diegosouzapw-bitrix24-agent-neversight/badge)](https://www.skillsdirectory.com/skills/diegosouzapw-bitrix24-agent-neversight)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: bitrix24-agent
description: Design, implement, debug, and harden integrations between AI agents and Bitrix24 REST API (webhooks, OAuth 2.0, scopes, events, batch, limits, and REST 3.0). Use when asked to connect AI assistants/agents to Bitrix24, automate CRM/tasks/chats, process Bitrix24 events, choose an auth model, or resolve Bitrix24 API errors and performance issues.
---

# Bitrix24 Agent (Lean + Reliable)

Use this skill to deliver correct Bitrix24 integrations with low token usage and production-safe defaults.

## Quick Start

Use this flow unless the user asks for a different one:

1. Pick intent + one minimal pack (`core` by default).
2. Run a read probe first.
3. For writes, use plan then execute with confirmation.

Read probe:

```bash
python3 skills/bitrix24-agent/scripts/bitrix24_client.py user.current --params '{}'
```

Safer write flow:

```bash
python3 skills/bitrix24-agent/scripts/bitrix24_client.py crm.lead.add \
  --params '{"fields":{"TITLE":"Plan demo"}}' \
  --packs core \
  --plan-only

python3 skills/bitrix24-agent/scripts/bitrix24_client.py \
  --execute-plan <plan_id> \
  --confirm-write
```

## Runtime Prerequisites

Required environment:

- `B24_DOMAIN`
- `B24_AUTH_MODE` = `webhook` or `oauth`

Webhook mode:

- `B24_WEBHOOK_USER_ID`
- `B24_WEBHOOK_CODE`

OAuth mode:

- `B24_ACCESS_TOKEN`
- `B24_REFRESH_TOKEN`
- `B24_CLIENT_ID` and `B24_CLIENT_SECRET` (for `--auto-refresh`)

Useful safety/reliability flags:

- `B24_REQUIRE_PLAN=1` for mandatory plan->execute on write/destructive calls
- `B24_PACKS=core,...` for default pack set
- `B24_RATE_LIMITER=file` with `B24_RATE_LIMITER_RATE` and `B24_RATE_LIMITER_BURST`

## Default Mode: Lean

Apply these limits unless the user asks for deep detail:

- Load at most 2 reference files before first actionable step.
- Start from `references/packs.md`.
- Then open only one target file: `references/catalog-<pack>.md`.
- Open `references/chains-<pack>.md` only if the user needs workflow steps.
- Open `references/bitrix24.md` only for auth architecture, limits, event reliability, or unknown errors.

Response limits:

- Use concise output (goal + next action + one command).
- Do not retell documentation.
- Do not dump large JSON unless requested.
- Return only delta if guidance was already given.

## Routing Workflow

1. Determine intent:
- method call
- troubleshooting
- architecture decision
- event/reliability setup

2. Normalize product vocabulary:

- "collabs", "workgroups", "projects", "social network groups" -> `collab` (and `boards` for scrum).
- "Copilot", "CoPilot", "BitrixGPT", "AI prompts" -> `platform` (`ai.*`).
- "open lines", "contact center connectors", "line connectors" -> `comms` (`imopenlines.*`, `imconnector.*`).
- "feed", "live feed", "news feed" -> `collab` (`log.*`).
- "sites", "landing pages", "landing" -> `sites` (`landing.*`).
- "booking", "calendar", "work time", "time tracking" -> `services` (`booking.*`, `calendar.*`, `timeman.*`).
- "orders", "payments", "catalog", "products" -> `commerce` (`sale.*`, `catalog.*`).
- "consents", "consent", "e-signature", "sign" -> `compliance` (`userconsent.*`, `sign.*`).

3. Choose auth quickly:

- one portal/internal integration: webhook
- app or multi-portal lifecycle: OAuth

4. Select minimal packs:

- default `core`
- add only required packs: `comms`, `automation`, `collab`, `content`, `boards`, `commerce`, `services`, `platform`, `sites`, `compliance`, `diagnostics`

## Execution Flow (Safe by Default)

Command template:

```bash
python3 skills/bitrix24-agent/scripts/bitrix24_client.py <method> \
  --params '<json>' \
  --packs core
```

Guardrails to enforce:

- allowlist via packs and `--method-allowlist`
- write gate with `--confirm-write`
- destructive gate with `--confirm-destructive`
- optional two-phase write with `--plan-only` and `--execute-plan`
- idempotency for writes (auto or `--idempotency-key`)
- audit trail unless `--no-audit` is explicitly needed

## Reliability and Performance

Pagination and sync safety:

- Never stop after first `*.list` page.
- Keep deterministic ordering and persist checkpoints after successful page persistence.

Batch rules:

- Maximum 50 commands per `batch`.
- No nested `batch`.
- Split oversized batches and parse per-command errors.

Limits and retries:

- Treat `QUERY_LIMIT_EXCEEDED` and `5xx` as transient.
- Use exponential backoff with jitter (client default).
- Use shared rate limiter keyed by portal in multi-worker setups.

Events:

- Online events are not guaranteed delivery.
- For no-loss pipelines, use offline flow:
  - `event.offline.get(clear=0)`
  - process idempotently with retry budget
  - `event.offline.error` for failed items
  - `event.offline.clear` only for successful/DLQ'ed items
- Use `scripts/offline_sync_worker.py` as baseline.

## Error Handling

Fast mapping:

| Error code | Typical cause | Immediate action |
|---|---|---|
| `WRONG_AUTH_TYPE` | method called with wrong auth model | switch webhook/OAuth model for this method |
| `insufficient_scope` | missing scope | add scope and reinstall/reissue auth |
| `expired_token` | OAuth token expired | refresh token (`--auto-refresh` or external refresh flow) |
| `QUERY_LIMIT_EXCEEDED` | burst above portal budget | backoff, queue, tune limiter, reduce concurrency |
| `ERROR_BATCH_LENGTH_EXCEEDED` | batch payload too large | split batch |
| `ERROR_BATCH_METHOD_NOT_ALLOWED` | unsupported method in batch | call directly |

Escalate to deep reference (`references/bitrix24.md`) on:

- unknown auth/permission behavior
- recurring limit failures
- offline event loss concerns
- OAuth refresh race or tenant isolation issues

## Quality Guardrails

- Never expose webhook/OAuth secrets.
- Enforce least-privilege scopes and tenant isolation.
- Keep writes idempotent where possible.
- Validate `application_token` in event handlers.
- Prefer REST v3 where compatible; fallback to v2 where needed.

## Reference Loading Map

1. `references/packs.md` for pack and loading strategy.
2. `references/catalog-<pack>.md` for method shortlist.
3. `references/chains-<pack>.md` for implementation chains.
4. `references/bitrix24.md` for protocol-level troubleshooting and architecture decisions.

Useful search shortcuts:

```bash
rg -n "^# Catalog|^# Chains" references/catalog-*.md references/chains-*.md
rg -n "WRONG_AUTH_TYPE|insufficient_scope|QUERY_LIMIT_EXCEEDED|expired_token" references/bitrix24.md
rg -n "offline|event\\.bind|event\\.offline|application_token" references/bitrix24.md
```

## Scripts

- `scripts/bitrix24_client.py`: method calls, packs, allowlist, confirmations, plans, idempotency, audit, rate limiting, retries.
- `scripts/offline_sync_worker.py`: offline queue polling, bounded retries, DLQ handling, safe clear flow, graceful shutdown.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.