Skip to content

configuration.md

Latest commit

 

History

History
86 lines (66 loc) · 3.3 KB

configuration.md

File metadata and controls

86 lines (66 loc) · 3.3 KB
Raw

configuration

afk stores tasks in SQLite.

Default path:

~/.claude/queue/tasks.sqlite

Override per command:

afk --queue /tmp/tasks.sqlite status

Override through the environment:

AFK_QUEUE=/tmp/tasks.sqlite afk tasks

Non-.sqlite paths are normalized to a sibling .sqlite database. For example, /tmp/tasks.jsonl becomes /tmp/tasks.sqlite.

The queue path is the only setting controlled by flag/env. There is no config file for it.

config files

afk loop and afk goal read operator settings from YAML files under the user config dir (~/.config/afk/ on Linux/macOS). Each file is written with defaults on first run; edit it in place. These are plain yaml.v3 files — there is no viper layer and no env-var overrides for these settings (per-run flags do override the file).

~/.config/afk/loop.yaml

Settings for the afk loop worker-driver:

command: ""                 # agent command template, e.g. "claude -p {{.Prompt}}"
                            # empty by default — fail-closed; loop errors until set
prompt_template: |          # text/template rendered against the task
  Task ID: {{.ID}}
  Status:  {{.Status}}

  {{.Body}}
task_timeout: 10m0s          # cap on a single task execution
cooldown: 5s                 # pause between ticks when no task is found
max_consecutive_failures: 3  # halt after this many back-to-back failures
lease: 30m0s                 # exclusive claim duration per task
heartbeat_interval: 2m0s     # how often the lease is extended while running
worker: ""                   # worker identity; empty derives "loop-<pid>"

~/.config/afk/goal.yaml

Settings for the afk goal workflow:

setup_command: ""            # agent command template for contract compilation
                            # empty by default — fail-closed; goal errors until set
audit_command: ""            # agent command for the independent auditor
                            # empty by default — goal audit errors until set
setup_prompt_template: ...   # prompt that compiles the objective into a contract
audit_prompt_template: ...   # prompt the auditor uses to check real artifacts
max_tokens: 0                # per-goal token budget (0 = unlimited)
max_iterations: 0            # per-goal iteration cap (0 = unlimited)
max_duration: 0s             # per-goal wall-clock cap (0 = unlimited)
token_regex: ""              # regex to parse a token count from agent output
setup_timeout: 5m0s          # cap on the setup (compile) agent call
audit_timeout: 5m0s          # cap on the audit agent call

Both setup_command and audit_command are empty by default: the workflow is fail-closed, so afk goal errors until you configure a setup command and afk goal audit errors until an audit command is configured.

A nonzero token budget requires a compilable token_regex with exactly one decimal capture group. During grouped invocations, AFK streams agent output unchanged and retains only a 1 MiB tail per stream for accounting. It uses the last parseable stdout match, then stderr; missing or overflowing usage fails closed as token-usage-unavailable. Iteration and token usage are cumulative and restart-persistent. Duration is measured from the first invocation in the current epoch; afk goal resume resets that epoch without clearing cumulative usage.