Loading ToolBench
A continuity engine for AI-assisted development that captures, synthesizes, and retrieves project context from coding sessions. Integrates with Claude Code, Codex, and other AI agents via MCP and webhooks.
Continuum has 20 tools with mostly complete schemas and descriptions, but significant gaps in parameter documentation and error handling guidance. Tool names follow verb_noun convention well (capture, ingest, synthesize-brain, chat). Descriptions average ~120 chars and explain WHAT the tool does, but lack WHEN to use it and recovery guidance. Input schemas are present with types, but many parameters lack descriptions or have minimal ones (e.g., 'The project slug' with no format hints). Output schemas are not documented, LLMs cannot plan downstream calls. Error handling is absent; no guidance on retryability, user-fixable vs fatal errors, or recovery paths. Security concerns: cli-auth tools expose pairing codes as parameters (should be server-side); settings tools accept apiKey parameters (should use secret injection). Composition is reasonable, tools chain well (list-projects → get-project → chat), but some tools combine concerns (ingest summarizes AND extracts AND persists AND triggers synthesis).
Auto-register a project from a Claude Code session. Given a cwd, finds an existing project or creates a new one with inferred name, slug, and icon. Enforces free-tier project limits.
Capture and classify a piece of context (note, decision, blocker, etc.) into a project. Optionally auto-classifies the content and assigns it to a project, or creates an inbox project if no match is found.
Chat with a project's synthesized brain. Retrieves recent updates, decisions, and chat history to ground responses in project context.
Authorize a CLI device pairing. Called by the logged-in user clicking 'Authorize' on the /cli-auth page. Mints a long-lived CLI token and binds it to the current workspace.
Poll for CLI pairing authorization status. Returns pending/authorized/expired. Single-use — deletes the pairing on successful read.
Revoke a paired CLI device. Hooks running on that machine will get 401 on their next call. Scoped to the current workspace.
Output schemas not documented. LLMs cannot infer what fields are returned by list-projects, get-project, chat, or get-chat-history. This breaks downstream tool chaining and forces agents to guess field names.
API keys and secrets exposed as tool parameters. update-settings and onboarding-provider accept apiKey as a parameter, which will be logged in agent traces and prompt history. Use server-side secret injection via environment variables or vault.
No error handling guidance. Tools lack descriptions of what errors can occur, whether they are retryable, and what the LLM should do next. E.g., delete-project should warn it is destructive and irreversible.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Begin a device-pairing session for CLI authentication. Returns a short-lived pairing code and auth URL. No authentication required — holding the code is the credential.
Create a new project in the current workspace. Enforces free-tier project limits.
Delete a project and all its associated data (updates, decisions, brain, chat messages).
Retrieve recent chat messages for a project (up to 50 messages, newest first).
Get a single project by slug, including its brain, recent threads, decisions, and updates.
Get the current workspace's AI provider settings (smart and cheap tiers), including configured providers, models, and available presets.
Ingest a coding session transcript from Claude Code or Codex. Summarizes the session, extracts decisions/blockers/next actions, persists them as updates, and triggers brain synthesis. Enforces free-tier session limits.
List all projects in the current workspace, ordered by most recently updated first. Includes brain synthesis status and focus.
Sign out the current user by deleting the session cookie.
Set up AI provider during onboarding. Configures both smart and cheap tiers with the same provider/key/model. Power users can split them later in /settings.
Manually trigger brain synthesis for a project. Bypasses the freshness gate that ingest uses, allowing on-demand re-synthesis.
Test a tier configuration by pinging it and returning ok/error plus latency. Used by the TEST_CONNECTION buttons on the Settings page.
Update a project's metadata (name, icon, description, cwd, state, identifier).
Update the workspace's AI provider settings for smart and/or cheap tiers. Allows configuring provider, baseUrl, model, and API key.
Parameter descriptions are minimal or missing. Many parameters lack format hints, constraints, or examples. E.g., 'slug' parameters lack format (lowercase, hyphens only); 'transcript' lacks guidance on expected structure or max length enforcement.
Ingest tool combines multiple concerns: summarization, extraction, persistence, and synthesis triggering. Split into separate tools (summarize_session, extract_decisions, trigger_synthesis) so agents can compose them independently.