Activation

A playbook that opens only if the agent asks for it governs nothing. In the target, it opens according to rules, and once open, it is the law.

Declaring activation

playbooks:
  - name: group-lunch
    goal: "the group's meal is ordered and confirmed"
    activation:
      when: "'agent-entreprise' in profiles"        # a deterministic filter
      intent: "order for several people"             # ranked by the model
      exclusive_skills: []                           # skills allowed ONLY in this playbook

In what order

1. Eligibility first. Playbooks the caller is not allowed to open are removed before any ranking. A forbidden journey is never offered.
2. Deterministic triggers. start_playbook, an automation event, auto_trigger, a true when without intent: the playbook opens, with no model.
3. The model ranks the rest, and the score falls into a band:

score	mode	effect
≥ enforce	ENFORCE	the playbook opens and is the law
≥ suggest	SUGGEST	several candidates with their score, and a question if needed ("are you ordering for a group?")
below	OBSERVE	nothing opens; we observe what the agent does, and report it to the admin

Two candidates above enforce, within margin of each other, fall back to SUGGEST.

The thresholds are set on the station:

station:
  guidance: {enforce: 0.85, suggest: 0.5, margin: 0.05, max_mode: enforce}

max_mode caps the mode. A new station starts in observe: the model computes, logs, and does not act, until it is calibrated on its traffic.

seek_guidance — the explicit door

Each station exposes a seek_guidance tool, next to the selection operations and start_playbook. The agent describes what it wants to do; it receives candidates — skills and playbooks — with their score and band. A candidate is described by the config: a skill by its cognitive.objective, a playbook by its goal. Write them to be understood.

The first write — the implicit door

The agent does not have to call seek_guidance. Then Core calls the same function at the first write made with no playbook open. A write from an exclusive skill is refused with the reason ("open the group order first"), or the playbook opens according to the bands.

Why ENFORCE by a model is safe

Opening a playbook grants no right: it imposes an order and adds guards. The only exception — exclusive_skills — is the reason eligibility comes before ranking.

Leaving a playbook

Only by an explicit abandon (by a profile allowed to do so) or a replacement by another. Both are logged. If the intent seems to change along the way, the model can suggest another journey, never impose it: a replacement loses the reserved actions.

OBSERVE reports patterns

"14 times this week, refund was called outside a journey after read-ticket." Patterns, not conversations. This is the material for the suggested playbooks in the console.