Agent console guide

Console guide for Weblab agents

Use this guide when the user asks you to follow it for console work. It explains tool usage; it does not grant permission to configure, restart or replace a lab.

Read, inspect, send, read

1. Read get_lab_state at the start of each turn. Use returned internal node IDs, not display names such as SW3. Identify the vendor before choosing commands. In compact mode, follow next_node_offset and next_link_offset until both are null; hold an exhausted offset at its corresponding total. All pages must have the same revision. A node missing from one page is not absent from the lab.
2. Call get_console_output. For an initial check, omit cursor to get recent output. Inspect the end for the current prompt, unfinished input, pager, login request or confirmation. Old boot logs do not establish current status.
3. If capture_end is limit, keep reading with each returned cursor before typing. If gap is true, inspect the replacement history and read again.
4. Before every send_console_command or send_console_input, make a fresh get_console_output call in this conversation and use its exact revision and cursor. Do not shorten, invent or reuse older values. The Agent requires this read even if the previous send returned output and a newer cursor.
5. Prefer send_console_command for one ordinary command, wait for output, and read subsequent bytes with the returned cursor. Verify the result and return to the appropriate prompt before the next command. A successful send or a timeout does not establish command completion.

Submit commands without encoding Enter

Use send_console_command with plain command text, for example command: "show clock". The tool appends exactly one real carriage return. Do not add Enter escapes or line endings. It rejects empty commands, control characters, literal \r/\n, and commands over 4095 UTF-8 bytes without typing. Send one command per call and verify the resulting prompt before continuing.

This helper uses the same permissions, approvals, revision/cursor checks and human locks as raw input. It does not clear pending input, enter privileged mode, log in, handle pagers or retry automatically. Establish an empty command line first; never append a new command to an unfinished line. Use the raw tool for bare Enter, control keys, pager input or deliberately literal escape text.

The vendor examples below describe the bytes for the raw tool. With the command helper, pass the command itself, such as enable or terminal length 0, without the trailing Enter escape.

Raw input: Enter and control keys are bytes

send_console_input types exactly what its input contains. It does not append Enter. For example, in JSON tool arguments:

{"input": "show clock\r"}

This JSON escape becomes a carriage-return byte (Enter). "show clock" alone leaves an unsubmitted command. "show clock\\r" types a literal backslash and r; it does not press Enter. When calling from JavaScript, use input: "show clock\r". The same distinction applies with any model/provider: in JavaScript source, input: "show clock\\r" produces printable backslash-and-r characters. Do not double-escape Enter when writing a JavaScript string passed to a tool.

Intended key	JSON value for input
Enter	"\r"
Ctrl+C	"\u0003"
Ctrl+U	"\u0015"
Tab	"\t"
Space (one pager page)	" "

Literal "^C" is two printable characters, not Ctrl+C. Control-key behavior depends on the guest CLI. Do not send Ctrl+D or exit as a generic cleanup step.

If a command is echoed but there is no result, inspect the pending line. Do not type the whole command again: that can concatenate commands. Submit an intact authorized command with Enter, or cancel/clear a malformed line using a key supported by that CLI, then read the resulting prompt before retrying. Never submit an unfamiliar line left by a human or another session.

Obtaining a prompt after boot

• A running process is not proof of a ready CLI. Read the latest console first.
• If it says Press RETURN to get started, send one Enter, then read again. If necessary, repeat once after inspecting the new output. Do not blindly send multiple Enters: they may submit pending commands or answer confirmations.
• With no prompt and no new output, inspect a larger recent window for a login, pager or startup question. An idle read alone does not prove the node is booting.
• If login is requested, use only credentials provided for that device. Do not guess passwords. Stop at an unknown setup/confirmation prompt and report it.
• Warnings about PKI, NTP or licensing do not by themselves prove the console is unavailable. Establish the actual prompt and command response.

Cisco IOS / IOSv / IOL

1. SW3> is user EXEC. For privileged inspection such as show running-config, send enable\r and verify SW3#. If it requests a password, obtain the authorized credential; do not try guessed passwords or command abbreviations to work around the missing privilege.
2. At the EXEC prompt, send terminal length 0\r as a separate command. Read its response, check for errors, and verify the prompt has returned.
3. Then send show running-config\r, or a focused operational command matching the question. For interface checks, examples include show interfaces Ethernet5/0 switchport\r and show running-config interface Ethernet5/0\r on images supporting that syntax. Use the interface name from the current lab; do not copy this example blindly.
4. If output still ends in --More--, pagination is active. Read fresh state and send a single space without Enter for each page. Alternatively, use the device's supported pager quit key, verify the EXEC prompt, then disable paging and rerun the show command. Do not type terminal length 0 into a pager.
5. After reconnect/login, recheck paging instead of assuming an earlier terminal setting still applies. A rejected tool call never applied that setting.

Do not search configuration for an exercise label such as task 3; inspect the interfaces, VLANs and protocol settings required by the task. Do not treat (config)# as an EXEC prompt or leave a user's configuration context silently.

Junos

• Distinguish the operating-system shell, operational CLI and configuration CLI. If in the shell and CLI access is intended, enter cli\r and verify the prompt.
• At the operational CLI prompt, use set cli screen-length 0\r to disable pagination for the session, then inspect its response before the next command.
• show configuration\r inspects configuration from operational mode. Choose operational protocol/interface commands appropriate to the question.
• Do not enter configuration mode, commit, clean storage or power off as part of ordinary inspection. Shutdown requires separate authorization and the Junos guest shutdown procedure.

Complete output versus a recent excerpt

Device pagination and Weblab capture limits are independent. Disabling pagination does not remove max_bytes or the capture deadline.

Ollama's compact mode caps each capture at 1024 raw bytes, reported as capture_max_bytes, even if you request a larger budget. Continue with the cursor instead of repeatedly increasing max_bytes. OpenAI keeps the ordinary budget. Any chars truncated marker from the surrounding client is another incomplete result; do not treat missing text or missing nodes as evidence of absence.

• For a full configuration, use a reasonable capture budget (for example 65536 bytes), then append subsequent cursor-based reads until the command ends and the CLI prompt returns. Continue after a timeout if completion is unverified.
• Do not switch to a new cursorless latest read while collecting a command: it may omit its beginning. omitted_bytes > 0 explicitly marks an excerpt.
• To inspect older retained history, omit the cursor and set latest=false, then follow cursors. History is bounded; expired output may require rerunning an authorized read-only command from a verified prompt.
• Report incomplete output as incomplete. Never conclude that configuration is absent from a truncated excerpt, an unanswered command, or a topology's original startup_config snippet.

Error recovery and ownership

expected_cursor must be the exact opaque cursor from the latest console read for that node, never a visible CLI prompt such as SW4#. Missing-read, cursor mismatch and read-revision errors mean no input was sent; they are not requests for approval or an authorization code. Call get_console_output again, inspect the output, then use its exact cursor and revision for the next input request. Normal Settings permissions, approvals and input locks still apply.

If input is rejected for a stale cursor, no command was sent. Read from the last cursor, drain remaining output and inspect the new prompt before sending again. If input_status is unknown, delivery may have occurred: inspect output before any retry. Respect human input locks and stop when another operator takes over. YOLO changes approval handling; it does not bypass these checks.

In your final answer distinguish what you read, what you executed, what the device actually returned, and what remains unverified. Never claim a successful configuration change solely because a tool accepted input.