v1 (autonomous 5-step + reviewer swarm + /trekreview gate) produced confident
fabrications and burned context for what are usually 0-1 line edits. v2 is a
slow, dialogic, one-command-per-session gate with the operator as truth source
and every mechanical claim tool-grounded. Method, command-class predicates,
session queue, and STATE.md end-of-session handoff are documented in plan.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New phase after the baseline-audit remediation (S1-S17, 12ff2bf, complete):
a command-hardening pass that simulates each of the 29 commands and tightens
it to its stated intention (intention-fidelity + prompt-quality only — no
structural redesign, no new features, no GUI/M0). Runs over ~8 journey-grouped
Voyage sessions, gated by lint + /trekreview ALLOW before push.
Foundation laid this session (execution starts next session):
- docs/hardening/brief.md (valid; 3 locked forks: hybrid simulation,
intention-fidelity+prompt-quality, per-journey cadence; research skipped —
the 2026 bar is frozen in algorithm-signals-reference.md)
- docs/hardening/plan.md (8 sessions, Grade B+ 87)
Adversarial review: brief-reviewer PROCEED_WITH_RISKS (3 findings folded:
self-graded quality axis, SC-H deferral contradiction, no stopping rule);
plan-critic REVISE (2 blockers + 10 major — all addressed: anchored coverage
predicate kills substring false-positives, full-session cold-reviewer oracle
on concrete before/after output, per-type mechanical predicate, Failed:0 gate
not literal-74, S1 HALT circuit-breaker); scope-guardian ALIGNED.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
