lsm / HyperNeo / 30719450760

83%

push

github

fix(space): graduate L3/L4/L5 coordinator decision-style guidance off reversibility (#2296)

The L4 space_chat coordinator over-escalated because its only behavioral
guidance conflated the gate-threshold autonomy number with decision style.
Four prompt defects drove it: 'uncertainty' was listed as an escalation
trigger (a coordinator is always somewhat uncertain), the Escalation
section hard-coded a question into every escalation, there was no act-first
framing at high levels, and L3/L4/L5 were byte-identical.

buildSpaceChatSystemPrompt now derives decision style from autonomy level
gated on reversibility: L3 acts on routine reversible work then reports,
L4 defaults to acting/reporting and escalates only on irreversible or
unrecoverable actions, L5 acts broadly. The Escalation section prefers
acting and reporting at L3+ (a question only if genuinely blocked).
Clarification is reworded so it can't be read as permission-seeking.

This is a coordinator prompt-wording change only — the gate/checkpoint
auto-pass mechanism (space.autonomyLevel >= requiredLevel) and the
SpaceAutonomyLevel type are unchanged.

11737 of 16822 branches covered (69.77%)

Branch coverage included in aggregate %.

100411 of 118771 relevant lines covered (84.54%)

302.49 hits per line