EPS Dashboard

Source: 2026-W22 weekly proposal #3. Thomas asked directly (2026-05-27): "How can we change the workflow so that if a concern is surfaced it must actually be fixed?" after #399's probe-position concern was flagged by code-review, deferred as "advisory", and then proved claim-relevant (costing a full re-run). The fix was drafted + reviewed but stranded on branch worktree-agent-a2da4775a3a1f5fbc, unmerged. Merge-readiness was assessed (read-only) for weekly-W22; it does NOT merge cleanly. Assessment follows.

MERGE-READINESS ASSESSMENT — branch worktree-agent-a2da4775a3a1f5fbc (binding-concerns protocol)

Branch exists locally only (no upstream, no live worktree). HEAD 710aa5600. Merge-base 30f0d5b68 (2026-05-27 01:31 PT). Current main HEAD was fdfbb6b70 at assessment time. Divergence: 7 commits on branch, 1056 on main (mostly task-state churn) over 4 days.

(1) COMMITS ON THE BRANCH (7, oldest->newest)

2880d3feb task_workflow: add binding-concerns persistence (concerns.jsonl)
604a44b0f task.py: add raise/address/defer/list-concerns subcommands
83db5c5d6 tests: cover binding-concerns library API + CLI roundtrip
aa15b2bd3 workflow.yaml: binding-concerns contract — CONCERNS no longer auto-passes
d5be3f371 agents + skills: wire binding-concerns into reviewer pipeline
3174ace91 verify_task_body: Lens 14 — mechanical concerns audit
710aa5600 agent-memory: workflow-improver pitfalls from binding-concerns landing

(2) DOES IT MERGE CLEANLY? — NO. 14 conflict hunks across 7 files.

Conflicting files (severity):

tests/test_workflow_yaml.py (2 hunks) — SEVERE/semantic. Branch and main make mutually exclusive assertions about the SAME invariants: branch asserts inline gate id=11 = concern_deferral_request and len(halt_criteria)==6; main asserts 7 inline gates incl. smoke_architecture and len(halt_criteria)==5. Competing contracts, not text noise.
.claude/workflow.yaml (1 hunk, root of the collision) — SEVERE. Main has ALREADY claimed inline gate id=11 (now a conditional gate) and id=6 = experiment_goal; main added smoke_architecture as gate id=10. The branch assigns id=11=concern_deferral_request and a new halt id=6=concern_unresolved — direct ID-space collision. Resolving this is a re-design of the gate/halt numbering, not a textual merge.
.claude/agents/code-reviewer.md (3 hunks) — MODERATE/semantic. Branch changes the implementer marker from 4-section to 5-section (adds (e) Concerns addressed) and verdict vocabulary to PASS/CONCERNS/BLOCKER/FAIL; main independently rewrote the SAME passages (mechanical-contract-only strip, Blocker tags: parse line, verdict table).
.claude/agents/clean-result-critic.md (3 hunks) — MODERATE, same flavor.
.claude/agents/codex-code-reviewer.md (2 hunks) — MODERATE.
.claude/agents/codex-clean-result-critic.md (2 hunks) — MODERATE.
.claude/skills/issue/SKILL.md (1 hunk) — MODERATE/semantic. Branch's binding-concerns triage (5c-pending) collides with main's PASS+CONCERNS auto-advance / mechanical-contract-strip rules at the Step 5c review gate.

Auto-merged cleanly (7 files, no conflict hunks) — including all 3 NEW-functionality core files: scripts/task.py (despite 5 main-side commits), scripts/verify_task_body.py (5 main-side commits), src/explore_persona_space/task_workflow.py (3 main-side commits), plus analyzer.md, critic.md, adversarial-planner/SKILL.md, issue/markers.md. The persistence layer + CLI mostly land cleanly; the conflicts are concentrated in the orchestration CONTRACT (gate IDs, verdict vocabulary, review-gate prose) — exactly where main churned this week.

(3) RECOMMENDATION: needs-rebase/rework — NOT a clean merge, NOT a from-scratch re-implement.

The substantive engine (concerns.jsonl persistence, task.py subcommands, library API, verify_task_body audit) auto-merges and is worth salvaging. But the contract-surface conflicts are semantic: the branch's PASS+CONCERNS-is-binding rule directly contradicts main's now-evolved PASS+CONCERNS auto-advance + mechanical-contract-only-strip rules, and the gate/halt-criteria ID space has hard collisions (main took id=11 and id=6). A rebase onto current main requires: (a) renumber the new gate + halt criterion off main's current IDs; (b) reconcile binding-concerns vs the mechanical-contract-strip anti-gate-hopping logic main added (make the two policies coherent, not just textually merged); (c) re-derive the 4->5-section marker shape against main's current reviewer-agent text. Roughly a one-session rebase-with-design-reconciliation, gated on re-running tests/test_workflow_yaml.py + workflow_lint.py --check-asks after.

Do NOT trust an -X ours/theirs auto-resolve on the contract files — it would silently pick one of two contradictory review-gate policies. workflow.yaml + the two skill/agent gate passages need a deliberate merge of intent.

(Read-only assessment; nothing was merged, checked out, or modified.)