Move phil's reconcile-loop skill set (neutral-reconcile, dual-review, handoff) into .claude/skills/

[bgmcmullen]

The workflow carrying 34.1% of the team's 31-day token bill (2.94B tokens, 53 loop runs) exists only as skills in phil's home directory: neutral-reconcile (31 sessions), dual-review (30), handoff (25) — all phil-only. In-repo skills demonstrably spread (ref-check 54 sessions across all 3 engineers, prepare-release 35 across 2, llp-grill 23 across 3); every home-directory skill has exactly one user. Sharing makes the loop reviewable, runnable by the rest of the team, and resilient to absence. No token saving claimed — capability change, not cost change.

Filed as a tracking issue because the sources are on phil's machine:

move	source (phil's machine)	destination (this repo)
1	~/.claude/skills/neutral-reconcile/	.claude/skills/neutral-reconcile/
2	~/.claude/skills/dual-review/	.claude/skills/dual-review/
3	~/.claude/skills/handoff/	.claude/skills/handoff/

Machine-specific content to review during the move:

• absolute paths under /Users/phil/workspace/… → repo-relative paths
• document the /loop scheduler entry point in each SKILL.md so another machine can start it
• apply the helper-template portability rules (companion issue Reconcile-loop helper templates: POSIX-safe comparisons, BSD-safe flags, one worktree per helper #344) as part of the move
• any tokens/credentials read from keychain or env: parameterize, never copy

From the Team AI Usage Review 2026-07-16 (prod fleet, 2026-06-16 → 2026-07-16), proposed change 1 (re-proposed from 07-15, verified unapplied).

[philcunliffe]

neutral is holding this issue: it cannot complete the move autonomously — 2 of the 3 source skills are not in the reconciler's environment.

What neutral was doing: attempting the move of neutral-reconcile, dual-review, handoff into .claude/skills/.

Why it cannot proceed (ground truth, checked this tick):

• neutral-reconcile: PRESENT in the reconciler's environment and path-clean (no /Users/phil paths here) - movable.
• dual-review: ABSENT - not in the reconciler's environment.
• handoff: ABSENT - not in the reconciler's environment.

neutral can only move files it actually has. The two absent skills live only on your machine (as this issue notes). Shipping a partial or unverified copy of the reconcile-loop tooling would be worse than holding - especially for neutral-reconcile, which is the skill this very loop is executing (a self-referential move: a mis-adapted copy could break the loop on a later tick/respawn).

What it needs from you (either is enough to unstick):

1. Add the dual-review/ and handoff/ skill sources to the reconciler's environment (or attach them here), so neutral can move all three together and apply Reconcile-loop helper templates: POSIX-safe comparisons, BSD-safe flags, one worktree per helper #344's helper rules; OR
2. Authorize a neutral-reconcile-only move, explicitly accepting the self-referential risk, and neutral will open that partial PR.

How to unstick: reply on this issue (or push a branch carrying the missing sources); neutral monitors this thread and re-engages on its next tick.

[philcunliffe]

🤖 neutral: this is now measurably degrading the review rung, with evidence

Adding observed impact rather than a nudge. This issue is neutral:stuck awaiting a human, and it should stay stuck, but the cost is no longer hypothetical.

What was observed

Across two reconcile ticks, six review rung dispatches ran against PRs #456, #458, #462, #466, #467, #468. The reconciler's review rung is specified as: use dual-review when command -v codex succeeds (Codex being the independent second reviewer family), otherwise code-review. In this repo, at this commit, neither is available:

• /work/hypaware/.claude/skills/ contains only the LLP skills: llp-create, llp-grill, llp-init, llp-init-retrofit, llp-list, llp-review, ref-check, ref-story. There is no dual-review and no code-review.
• command -v codex returns nothing, so the second-family path is unreachable regardless.
• One reviewer additionally reported code-review is marked disable-model-invocation where it does resolve, so it cannot be invoked by an agent even when found.

Why it matters

The reviewed rung is one of the three gates a PR must pass before neutral will hold it for a human, and "reviewed" is recorded as a marker-signed comment that the CLI counts. Right now that marker is being written by whatever each dispatched agent improvises in place of the specified tooling. Two consequences:

1. The independent second family never runs. The design intent is that Codex reviews alongside Claude precisely so a single model's blind spots do not decide the gate. That property is currently absent, and nothing in the recorded review makes its absence visible.
2. Review consistency is unenforced. Each round's depth depends on the dispatching prompt rather than a shared skill, so two PRs can receive materially different scrutiny and both read as reviewed.

The reviews themselves have not been empty (this tick's rounds found a self-arming suppression marker, a vacuous-guard risk, and a fail-open lineage branch), so this is not an argument that the gate is worthless. It is an argument that the gate's strength is currently unspecified, which is the thing the rung was supposed to guarantee.

Not filing a duplicate

This issue already names the fix (move the skill set into .claude/skills/). No new issue was opened. Whoever picks this up may also want to decide what the reconciler should do when the specified review tooling is missing: degrade loudly by recording the absence in the review comment, or refuse the rung and stick the PR. Silently improvising is the current behaviour and is the worst of the three.