Delegation and Mesnada Configuration

Keys of the Mesnada orchestrator and of delegated-task handling, plus the tools the agent uses to delegate. For the idea, read Agent delegation; for the walkthrough, the guide Delegate to subagents with Mesnada.

Web UI: Settings > Mesnada (server, orchestrator, ACP) and Settings > General > Subagent Delegation.

Orchestrator

[Mesnada]
Enabled = true

[Mesnada.Server]
Host = 'localhost'
Port = 5005

[Mesnada.Orchestrator]
StorePath        = './.pando/mesnada/tasks.json'
LogDir           = './.pando/mesnada/logs'
EnginesDir       = ''        # defaults to <dirname(LogDir)>/engines
MaxParallel      = 5         # tasks in flight at once; 0 = unlimited
MaxPerEngine     = 0         # per-engine cap; 0 = unlimited
ClaimTTL         = '2m'      # how long a dispatch reservation lasts
DispatchInterval = '10s'     # how often deferred tasks are started
DefaultEngine    = 'pando'
DefaultModel     = ''
PersonaPath      = ''

Custom engines

Place *.template.yaml files in EnginesDir. Each template specifies command, args (with Go template expressions), prompt mode, output format and available models. Custom engines appear dynamically in the mesnada_spawn_agent tool.

Delegation

[Mesnada.Delegation]
Enabled             = true
InjectIntoLiveLoop  = true
ResurrectIdleLoop   = true
SynthesizeFallback  = false
MaxResurrections    = 4
MaxDepth            = 3
MaxConcurrent       = 8
ResurrectionTimeout = '10m'

Warm instances and other projects

[Mesnada.Delegation]
ReuseWarmInstances       = true
AutoStartWarmInstance    = true
WarmInstanceIdleTimeout  = '10m'   # '0' or empty = never stop it
WarmQueueDepth           = 0       # 0 = start a separate process when the instance is full
AllowExternalWarmTargets = true    # caller side
AcceptDelegations        = true    # target side

The project parameter of mesnada_spawn_agent routes the task to the right project’s instance.

Safety nets

Tools the agent uses

mesnada_spawn_agent launches a task:

{
  "prompt": "Analyze the authentication module and write tests",
  "subagent_type": "general",
  "description": "Write auth tests",
  "background": true
}

Dependencies, with the logs of the earlier tasks passed along:

{
  "prompt": "Write tests based on the analysis",
  "dependencies": ["T1", "T2"],
  "include_dependency_logs": true,
  "dependency_log_lines": 100
}

Waiting without blocking (recommended) uses mesnada_await; the parent agent is resumed when results arrive:

{
  "action": "wait",
  "actor_id": "explore-1"
}

Blocking until one task completes uses mesnada_wait_task:

{
  "task_id": "T1",
  "timeout": "10m"
}

Agent self-service (pando_setup)

The built-in pando_setup tool needs no configuration and is always available to the agent. See Agent Self-Service.

pando_setup models --provider anthropic --detail --limit 5
pando_setup providers
pando_setup config                           # show all config
pando_setup config --search token            # search for token-related settings
pando_setup config TokenOptimization         # show a specific section
pando_setup run /caveman lite
pando_setup session

models uses the live model registry, enriched with models.dev: canonical ID (for example copilot.gpt-5.4), context window, price per million tokens, capabilities and knowledge cutoff.

Commands the agent may run with run:

• Mode commands: /caveman, /caveman-finish, /ponytail, /ponytail-finish, /superpowers, /superpowers-finish, /learning, /learning-finish.
• Instruction commands: /improve-agents-md, /vulnhunt, /vulnhunter-fix, /vulnhunt-fix-verify, and custom user: and project: commands.

Blocked, because they need you at the controls: /goal*, /compact, /db-compact and all -finish closing commands.

Security: configuration is strictly read-only (no write path exists); secrets are masked by key suffix (apikey, token, password, secret…) showing only the last 4 characters; the tool description stays under 600 characters; the tool cannot be trimmed away by context optimization.