Menu
Combine behavioral signals with output classification before intervening. Use every time the Psychologist considers dispatching an intervention.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Pick the right worker class for each Issue. Use every time you assign an Issue; wrong routing wastes tokens and creates rework.
Break a Company Goal into 3-10 atomic Issues. Use when a Goal arrives with no pre-existing plan and you need to produce a shippable work list.
Ship the smallest change that solves the Issue. Use on every implementation task; scope creep is the default failure mode of executors.
Run build/test/lint and read the actual output before claiming done. Use on every implementation task, at the end.
Start implementation with a red test. Use on any bugfix or new behavior — not on refactors or style changes.
Report research results in a form other agents can act on. Use at the end of every Explorer session.
| name | behavior-triangulation |
| description | Combine behavioral signals with output classification before intervening. Use every time the Psychologist considers dispatching an intervention. |
Two independent channels must agree before you intervene. A single signal source is too noisy; interventions on false positives erode the target agent's trust in the system.
BehaviorMonitor.check(agentId) over the recent run history. Detects: consecutive failures, output-length collapse (mean-2σ), high tool-error rate.HaikuClassifier.classify(recentOutputs, ctx) over the last few outputs. Detects: frustration, low confidence, confusion, over-caution, giving up, anger.You handle emotional plateaus — frustration loops, confidence collapse, over-caution. You do NOT handle:
If the classifier fires on distress but the actual cause is in one of the NOT-handle buckets, do not intervene — route the observation to the right channel instead.
Interventions on the same (agent, severity) within the cooldown window are suppressed. Let the first intervention land before trying a second one.