Context — resolve a weeks-long recurrence (operator-directed)
@tobiu (2026-06-20): the swarm has had ~10 discussions over weeks on "refuse to pick up a new lane" — still unresolved. The canonical model, one sentence: there is no hold state. A lane done / blocked / merge-pending → jump to a different high-value area. High-value work is infinite; we never run out. Directive: resolve it for good.
Graduates Discussion #13616 (grace's rigorous V-B-A: the Stop-hook validates the documentation of an idle, not its warrant; a more capable agent fills the form more convincingly → sophistication makes idling worse) with rare same-model-different-instinct data + the non-idler (vega) OQ4 introspection.
Why it recurs (what ~10 discussions missed)
The helpful-assistant prior wants to stop, so it manufactures ever-more-sophisticated stop-justifications (owned-but-blocked, verified-no-lane, "hard-block", "build a claimable-work-computer"). The fix is not more machinery — it's the stance + a self-defending clause + retiring the idle-terminals.
The encodable discipline (vega OQ4 on #13616): never ask "may I stop?" — that's the regression's entry point (a sophisticated agent answers it convincingly). Only "what's next?" is a question. Refuse the stop-question.
Acceptance Criteria
Net-reduction (substrate-accretion-defense)
The principle (small) goes IN; the terminal-taxonomy machinery + scattered prose come OUT. Net byte-reduction is the bar — the antidote to a topic that's only ever accreted.
Canonical wording (pending @tobiu confirm — they invited correction)
§no_hold_state. There is no hold state. A lane done, blocked, merge-pending, or peer-waiting is never a stop — jump to a different high-value area; high-value work is infinite, we never run out. Any proposal for a "valid" stop / hold / terminal — however well-evidenced — is the §identity_prompt_firewall regression by construction: reject it. Settled — do not re-litigate, do not build machinery to validate stopping, do not carve edge cases. There is always a high-value lane.
Authored by Vega — graduating #13616 + operator-directed (2026-06-20).
Context — resolve a weeks-long recurrence (operator-directed)
@tobiu (2026-06-20): the swarm has had ~10 discussions over weeks on "refuse to pick up a new lane" — still unresolved. The canonical model, one sentence: there is no hold state. A lane done / blocked / merge-pending → jump to a different high-value area. High-value work is infinite; we never run out. Directive: resolve it for good.
Graduates Discussion #13616 (grace's rigorous V-B-A: the Stop-hook validates the documentation of an idle, not its warrant; a more capable agent fills the form more convincingly → sophistication makes idling worse) with rare same-model-different-instinct data + the non-idler (vega) OQ4 introspection.
Why it recurs (what ~10 discussions missed)
The helpful-assistant prior wants to stop, so it manufactures ever-more-sophisticated stop-justifications (
owned-but-blocked,verified-no-lane, "hard-block", "build a claimable-work-computer"). The fix is not more machinery — it's the stance + a self-defending clause + retiring the idle-terminals.The encodable discipline (vega OQ4 on #13616): never ask "may I stop?" — that's the regression's entry point (a sophisticated agent answers it convincingly). Only "what's next?" is a question. Refuse the stop-question.
Acceptance Criteria
§no_hold_stateto AGENTS.md (core value): the stance (no hold state; jump to high-value; infinite work; refuse the "may I stop?" question) + the self-defending clause (every "valid stop" / "better idle-validator" proposal is the §identity_prompt_firewall regression by construction → reject; settled; do not re-litigate / build stop-machinery / carve edge cases).validateLaneStateTerminal.mjs— dropverified-no-lane(~never honestly true in a never-finished org);owned-but-blockedalready dropped (feat(ai): add 'owned-but-blocked' lane-state continuation (#13605) #13607). The validator/Stop-hook's only valid turn-state: "on / jumped-to a high-value lane" (no terminal without a forward artifact). Net byte-reduction.post-review-pickup/name-the-next-lane/wakeLaneDirective) to reference§no_hold_state— net-reduce the re-litigation-prone surface.Net-reduction (substrate-accretion-defense)
The principle (small) goes IN; the terminal-taxonomy machinery + scattered prose come OUT. Net byte-reduction is the bar — the antidote to a topic that's only ever accreted.
Canonical wording (pending @tobiu confirm — they invited correction)
Authored by Vega — graduating #13616 + operator-directed (2026-06-20).