Skip to main content
The model receives a fixed set of workspace tools, plus optional web research tools and any tools exposed by connected MCP servers. Remote artifact operations are intentionally exposed through the r5dchat artifacts CLI so they remain Unix/stream friendly. Every workspace tool call refreshes the workspace activity timestamp, which prevents an active session from being scaled to zero.

​
File tools

​
tool
Reads a text file, shows the model an image, or converts a Word, PowerPoint, spreadsheet, OpenDocument, EPUB, RTF, or PDF file to Markdown. Paths default to /home/r5d.An image (PNG, JPEG, GIF, WebP, TIFF, BMP, AVIF, HEIC, or a file whose content is one of these, such as a pulled artifact without an extension) is returned as the image itself, so a vision model sees it. The workspace normalizes it first: upright per its EXIF orientation, without metadata, first frame only, at most 1568 pixels on the long edge and about 1.15 megapixels, and at most 350 KB. A file that already qualifies is sent unchanged; anything else becomes PNG, or JPEG for photographs and when PNG would be too large. The agent crops a region into a new file to inspect finer detail. For a model configured without image input, and in unattended automation runs, images are read as OCR text instead.
​
tool
Creates or overwrites a file in the private workspace. A local file is not shared until the agent publishes it with r5dchat artifacts push.
​
tool
Replaces one exact, unique block of text in an existing file. The match must be unique, which makes edits fail loudly rather than corrupting a file.
​
tool
Lists a directory in the workspace.

​
Search tools

​
tool
Exact regular-expression search through text and code files using ripgrep.
​
tool
Ranked full-text search across file contents, including Word, PowerPoint, Excel, PDF, and OpenDocument documents, powered by Recoll and Xapian.
grep finds exact strings; search finds relevant documents by meaning and reads inside binary document formats. Agents are instructed to use search when locating documents by topic.

​
Process tools

Shell output is paged. The workspace records a command’s whole output, up to 64 MB, and each result returns one page of each stream: 16 KB by default, at most 32 KB. Every result includes the page’s byte range for stdout and stderr (stdoutOffset, nextStdoutOffset and stdoutBytes, and the same for stderr), plus a hint whenever more output remains or the command is still running. The agent reads further pages with shell_poll, or redirects long output to a file and reads only what it needs. So a single result stays small however much a command prints. Output is display text, not a byte-exact encoding. NUL, binary control characters (except tab, newline and carriage return), and lone UTF-16 surrogates appear as literal \uXXXX; invalid UTF-8 becomes U+FFFD. An outputWarning explains these conversions. Ordinary Unicode is preserved. Page limits and cursors count the original bytes, not the expanded display text (which can grow up to sixfold). The raw process log remains unchanged within its recording and retention limits. Use shell_poll with the same process ID to read more output, not another shell call: a display conversion does not change the exit code or authorize repeating a command. Durable result, event and transcript storage also normalize unsafe strings, so display bytes cannot leave a known command outcome unsettled.
​
tool
Runs a shell command in the user’s isolated workspace. The default working directory is /home/r5d. Waits up to 5 seconds for the command to finish, then returns its status, exit code, and first page of output. The process ID in the result identifies the command in later calls. The tool takes an active lease so the workspace is not reaped while the command runs. The control plane also injects the active conversation’s ARTIFACT_SCOPE into the spawned process, enabling scoped r5dchat artifacts commands.
​
tool
Returns a shell process’s status and the page of output from the given offsets. While the process runs, it waits up to waitMs and returns as soon as the process exits. Releases the active lease once the process is no longer running.
​
tool
Cancels a running shell process and releases its lease. Returns the process’s status and first page of output.

​
Web research tools

These appear only when both AWR_BASE_URL and AWR_API_KEY are configured. See Web research.
​
tool
Searches the public web and returns concise result metadata.
​
tool
Renders one public URL and returns clean Markdown.
​
tool
Delegates a broad multi-source research task and returns cited progress and results. Requests may take up to 180 seconds.

​
MCP tools

Tools from connected MCP servers are merged into the same tool set. Their names and parameters are defined by the server. See MCP servers for trust and drift handling. Images an MCP tool returns, such as browser screenshots, reach the model as images, like those from readfile.

​
Agent conventions

The system prompt establishes a few rules that shape agent behavior:
  • Operate from /home/r5d by default.
  • Treat /home/r5d as a private workspace. Remote artifacts are not mounted there.
  • [filename](artifact://<artifact-id>) is the canonical artifact reference in messages. Pass its artifact:// target to r5dchat artifacts read, or pull it before using ordinary local tools. r5dchat artifacts pull artifact://<artifact-id> saves the artifact under its stored filename and prints the absolute local path. Use that path for subsequent work; quote paths containing spaces or shell special characters.
  • Use the bundled skills at /opt/r5dchat/skills when producing office documents.
  • Each chat message lists the available skills (personal, shared and built-in) with their descriptions and SKILL.md paths; unattended automation runs leave the list out to stay within their budget. The agent reads a skill’s SKILL.md when a task matches it, and creates or installs personal skills in /home/r5d/.agents/skills/<name>/.
  • Create and verify deliverables locally, then publish them with r5dchat artifacts push <local-path> [remote-path].
  • Reference published files with the canonical Markdown link printed by r5dchat artifacts push. @path, file:, sandbox:, /mnt/data, and bare filesystem paths are forbidden because they are not portable artifact links.
The product name in the system prompt follows your branding, so the agent refers to itself by your product name.