External assessment from claude-playlist-corpus, verified per tiltak against
code in this repo and against the claude-code-llm-wiki OKF bundle (bundle wins
on feature claims). No implementation — scope-guard.
Verdicts across the six tiltak:
1. Workflow port — split, not global. /trekreview is already ported (opt-in
--workflow) and measured fidelity-equivalent; /trekplan and /trekresearch are
unported, but the claimed benefit is UNMEASURED, not refuted. T2's +4.4% is
subagent tokens and states main context is uncounted; T1's Δ≈0 covers Phase 7
synthesis only. The measurement design already exists in T1-cc26. Measure
before porting.
2. Holdout — (a) and (b) hold today but are uncodified; (d) is already partly
there; (c) fail-closed is a REAL defect. computeVerdict counts only kept
findings, so a dropped or unreturned finding is arithmetically identical to
one that never existed. Codified in lib/review/coordinator-contract.mjs, so
it is testable test-first.
3. Lightweight lane — narrower than stated: --quick exists but does not bypass
the 2.2 framing gate. Already an open operator decision.
4. Definition-of-done as object — confirmed gap, but it is a Handover 1
breaking change (brief_version bump), not a cheap edit.
5. Prune — mostly already decided: synthesis-agent dormant by measurement, the
24 opus pins are an operator pin with a decision record (40d8742), and
gemini-bridge is already flagged DROP→NATIVE.
6. LSP — confirmed absent.
Bundle-gap: V2 overstates two Workflow primitives (token budgets, automatic
retry — neither documented on the mirrored page) and omits that resumability is
session-scoped, which matters for a multi-session tool.
Kjell Tore Guttormsen
9ee120dd7e