Cognitive_workbench

Fetch standardized financial statements (income, balance sheet, cash flow, earnings, company overview) for a ticker from Alpha Vantage. Returns combined annual+quarterly JSON for analysis.

Aim the ChatterBot head to find and center a target in view — the user, a person, an object, an animal. Runs a closed visual loop (capture, judge where the target is, nudge pan/tilt, repeat) until the target is centered, or reports it could not find the target after searching. Use when the user says point at me, look at me, turn to face someone, find the cat, center on the person. For a one-off snapshot without re-aiming use camera-capture; for a manual fixed angle use head-move.

Capture a still photo from the ChatterBot head camera. The captured frame is attached to your own visual input, so you can SEE it and answer questions about what is in view — whether the user is present, whether there is a cat, what the scene looks like. The camera rides the pan/tilt head, so it shows whatever the head is currently aimed at; aim first with head-move if needed. To also show the photo to the user on screen, follow with the display tool (the observation includes a ready <img> URL). Use when the user asks what you see, to take a picture or snapshot, or to check whether something or someone is in view.

Move the ChatterBot head — aim the pan/tilt camera or play an expressive gesture. The bot is a stationary companion head; this points its gaze, it does NOT drive or navigate. Use when the user asks you to look somewhere, turn toward/away, look up/down, re-center, or nod/shake/scan. Angles are degrees 0-180 with 90 centered (pan 0=full right, 180=full left; tilt 0=down, 180=up, mounting-dependent). Give pan and/or tilt for absolute aim, OR a gesture (not both). Returns the confirmed pose once the head settles.

Return one genuine saying of Ramana Maharshi, drawn verbatim from his recorded talks, with source attribution. Use when delivering an authentic Ramana quote with attribution — not a paraphrase or a synthesized reflection. The returned text is a raw quote; add your own brief framing before presenting it.

Generate an original image from a text description, locally (Bonsai-Image 4B, ternary-quantized, on a long-lived studio server). Use when the user wants a picture, illustration, avatar, or face created from a description that does not already exist on the web. For existing photos of real things, prefer image search instead; for simple diagrams or line drawings, prefer authoring inline SVG.

Check whether text matches a natural-language condition. Returns "true" or "false". Auto-chunks long texts and short-circuits on first match.

Read recent emails from a Gmail inbox via IMAP. Returns text with subject, sender, date, and body per email. Read-only — never marks messages as read.

Run a bash/shell script — for system commands the existing tools don't cover. Trusted by default; treat with the same care as any code-execution action.

Filter a list of items to those matching a natural-language condition (e.g. "only the items that mention X"). NOT YET WIRED in chat-mode — present in src/tools/ for future enablement.

Read, search, write, and list notes in the user's Obsidian vault. Multi-action — set `action` first, then provide the args required by that action.

Search academic papers and journals. Returns paper text — full text when available via GROBID, otherwise abstracts. Use for scholarly literature, not general web content.

Look up the current stock price (price, change, volume) for a ticker symbol. End-of-day data on the free tier.

Fetch the text content of a web page or PDF from a URL. Handles HTML, PDF, Markdown, plain text — auto-detects format.

Compute the value of a math expression. Handles arithmetic, algebra, trig, calculus. Use this instead of LLM arithmetic — calculate is exact and deterministic.

Search the web for information on a topic. Returns a synthesized summary with citations. Use when you need current information or facts not in training data.

Returns API instructions (not data) for retrieving Bluesky metrics via AT Protocol. Read the returned instructions, then use fetch-text + extract to get actual data.

Browser automation via agent-browser CLI. Navigate, snapshot accessibility tree, interact by element ref.

Collect system and cognitive health metrics. Returns a structured health report with status classifications (ok/warning/critical) for hardware, OS, and cognitive subsystems.

Delegate coding, research, and code analysis tasks to Claude Code CLI. PREFERRED tool for reading, searching, or analyzing source code (including your own — use target=self). Use whenever a goal asks you to look at, analyze, or reason about code files (your src/, your codebase, your implementation, fs/src/). Ground claims about your own architecture, capabilities, or behavior in actual source code rather than generating from general knowledge.

Generate a new tool. Python tools go to staging for review. Instruction tools (no code) are created and registered immediately.

Extract bibliography/references from PDF files using GROBID and return a Collection of Notes (one per reference).

Derive content from a single Note via LLM-guided extraction, compression, or transformation. Output is grounded entirely in the input — no new information is introduced.

Extract structured metadata (title, authors, year) from paper text using LLM. Use to convert online search results metadata to JSON

Format paper citations from semantic-scholar Notes/Collections to BibTeX format

Generate new text or code content from scratch using natural language prompt via LLM. Creates content from the LLM's own knowledge — no source documents.

Generate content with awareness of the agent's current reasoning state — goal analysis, situation context, character orientation. Like generate-note but informed by internal planner state.

Create, list, update, and delete scheduled goals. Goals are persistent tasks that can run manually, automatically, or on a daily schedule.

Create, list, update, and delete tasks. Tasks are work items that go through establishment phases (specification, capability evaluation, execution) with milestone-based progress.

Post to Bluesky (AT Protocol). Body from target/value. Returns confirmation Note with post URI in metadata.

Send an email via Gmail SMTP. Body from target/value, headers as parameters. Returns confirmation Note with message-id in metadata.

Integrate content across multiple documents to produce new understanding. Use for cross-document synthesis, comparison, and reporting from Collections.

Count words in text. Use to determine length of a document, e.g. to determine if it needs to be summarized before further use

Find a substring or pattern in a text document. Returns position with surrounding context. Deterministic — use this instead of asking the LLM to find substrings.

Time heartbeat for chat-mode autonomy — fires every schedule period to trigger concern cadence checks

Global embedding-based search across all Notes; returns a Collection of matches

Mark a Note or Collection as persistent so it survives goal-completion cleanup

Create a Collection object and bind to variable

Monitors RSS feeds for new items matching configured keywords

List files and directories under scenarios/<world_name>/fs; returns one Note whose body is a text listing (like ls).