docs(agents): add shared BMAD skill entrypoints

.agents/skills/bmad-build/SKILL.md

@@ -0,0 +1,8 @@
+---
+name: bmad-build
+description: "Implement a bounded repository change with the shared BMAD build workflow and local gates."
+---
+
+Read and follow `~/.codex/skills/bmad-build/SKILL.md` with this repository as the project root.
+This entrypoint intentionally uses the shared personal workflow directly; do not run the legacy
+BMAD renderer or adjacent workflow snapshots. Preserve this repository's authority and local gates.

.agents/skills/bmad-checkpoint-preview/SKILL.md

@@ -0,0 +1,8 @@
+---
+name: bmad-checkpoint-preview
+description: "Guide human review of a change with orientation, concern walkthrough, risk details, and observation."
+---
+
+Read and follow `~/.codex/skills/bmad-checkpoint-preview/SKILL.md` with this repository as the project root.
+This entrypoint intentionally uses the shared personal workflow directly; do not run the legacy
+BMAD renderer or adjacent workflow snapshots. Preserve this repository's authority and local gates.

.agents/skills/bmad-qa/SKILL.md

@@ -0,0 +1,8 @@
+---
+name: bmad-qa
+description: "Review changes with the shared BMAD QA process and local domain gates; fix only when requested."
+---
+
+Read and follow `~/.codex/skills/bmad-qa/SKILL.md` with this repository as the project root.
+This entrypoint intentionally uses the shared personal workflow directly; do not run the legacy
+BMAD renderer or adjacent workflow snapshots. Preserve this repository's authority and local gates.

.agents/skills/stack-dev-dispatch/SKILL.md

@@ -5,10 +5,10 @@ description: Implement one approved GraphReFly Stack slice from the canonical se
 
 # GraphReFly Stack dev dispatch
 
-Apply the global `repository-ownership-practice` skill unless the user explicitly invokes
-`--delivery-only`. In practice mode, pause for OWN/PREDICT before implementation and TEACH-BACK after one
-batch. `--delivery-only` skips those waits but must still produce the ownership handoff and report
-`delivered, not yet ownership-verified`.
+Use `~/.codex/skills/bmad-build/SKILL.md` inside the selected implementation slice, retaining this
+workflow's authority, sequencer and gates. Use `~/.codex/skills/bmad-checkpoint-preview/SKILL.md` for
+human review: build-handoff mode during ordinary delivery; interactive mode for explicit checkpoint
+or `--practice` requests. No prediction or teach-back is required.
 
 ## Code-intelligence routing
 
@@ -24,7 +24,7 @@ correctness evidence.
    antipatterns, and the canonical sequencer. Do not load unrelated decision history.
 2. Select the requested phase or the single `ready` phase. Require all dependencies to be `done` and
    required design decisions to be locked.
-3. Record the user's best-effort prediction, then verify named files, APIs, dependency versions, and one
+3. Verify named files, APIs, dependency versions, and one
    concrete input-to-output path before planning changes.
 4. Translate the phase gate into one Given/When/Then vertical implementation and verification plan. Freeze
    intended seams, tests, non-goals, and stopping boundary. Do not pull deferred backlog into the slice.
@@ -34,8 +34,8 @@ correctness evidence.
    build, or replay commands once they exist.
 7. Apply the `stack-qa` workflow to the resulting diff. Fix in-scope findings and rerun gates.
 8. Mark a phase `done` only when its recorded gate is actually satisfied. Then promote the earliest
-   dependency-satisfied blocked phase to `ready`, update the phase note, and run `pnpm docs:check`. In practice
-   mode, stop for TEACH-BACK before implementing that next phase.
+   dependency-satisfied blocked phase to `ready`, update the phase note, and run `pnpm docs:check`. For an explicitly requested interactive
+   checkpoint, present the preview and wait for review navigation before that next phase.
 
 Record new architectural locks in decisions before code. Record deferred work in backlog and reusable
 lessons in antipatterns; never store them in `AGENTS.md` or this skill.

.agents/skills/stack-goal-dispatch/SKILL.md

@@ -34,29 +34,20 @@ all executable correctness gates.
 4. On every continuation, inspect `git status`, `docs/sources.jsonl`, `docs/decisions/decisions.jsonl`, `docs/plan/phases.jsonl`, and the concern authorities referenced by the next phase.
 5. Prefer durable records and Git state over recollected chat prose. Record consequential progress before ending a task.
 
-## 1A. Ownership checkpoint and Goal-loop guard
+## 1A. Preview checkpoint and Goal-loop guard
 
-Apply the global `repository-ownership-practice` skill for each implementation batch unless the user
-explicitly invokes `--delivery-only`.
-
-- Before implementation investigation, pause for the user's OWN and PREDICT cards.
-- Advance at most one delivery batch before a user-visible ownership checkpoint.
-- At the batch boundary, persist the canonical phase state and report prediction corrections plus
-  diff/behavior/trace evidence.
-- Wait for the user's TEACH-BACK before automatically selecting the next ready phase. Keep the Goal active;
-  this wait is neither completion nor a blocker.
-- In `--delivery-only` mode, continue under the canonical sequencer but label the result
-  `delivered, not yet ownership-verified`.
-
-Keep product horizon, Goal, phase, batch, proof, and user comprehension as separate states.
+Use `~/.codex/skills/bmad-build/SKILL.md` inside the selected implementation slice, retaining this
+workflow's authority, sequencer and gates. Use `~/.codex/skills/bmad-checkpoint-preview/SKILL.md` for
+human review: build-handoff mode during ordinary delivery; interactive mode for explicit checkpoint
+or `--practice` requests. No prediction or teach-back is required.
 
 ## 2. Select and advance work
 
 1. Select only the single `ready` phase in `docs/plan/phases.jsonl`; require all dependencies to be `done`.
 2. Apply `stack-decision-guard` before changing scope, product semantics, architecture, or delivery policy.
 3. Define the current batch from the selected phase's gate and deliverables. Keep it as small as possible while still producing reviewable evidence.
-4. After the current phase gate passes, advance the sequencer. In practice mode, stop at the ownership
-   checkpoint before executing the next phase. In `--delivery-only` mode, continue automatically. Never skip a
+4. After the current phase gate passes, advance the sequencer. For an explicitly requested interactive
+   checkpoint, show the preview and wait for review navigation. Otherwise continue automatically. Never skip a
    gate or select deferred work merely because no implementation phase is ready.
 5. Keep the Goal horizon at the current explicitly approved product horizon. Backlog entries are not authorized phases: when a roadmap-design phase must select among them, consolidate the material choices through `stack-decision-guard` and `stack-design-review`, record the approved phases, then continue.
 

.agents/skills/stack-qa/SKILL.md

@@ -3,6 +3,12 @@ name: stack-qa
 description: Perform the single GraphReFly Stack quality gate combining adversarial code review, correctness fixes, tests, deterministic semantic evidence and GateResult validation, security/privacy review, UX blockers, and product readiness. Use after implementation, before marking a product phase done, and for review or QA requests.
 ---
 
+Read `~/.codex/skills/bmad-build/references/qa.md` for the shared review lenses, five-way triage and
+repair loop. Integrate the domain checks below into that one review; retain this repository's canonical
+completion verdict. Review-only remains read-only, including documentation and authority records.
+Finish with `~/.codex/skills/bmad-checkpoint-preview/SKILL.md` in build-handoff mode unless interactive
+review was requested.
+
 # GraphReFly Stack QA
 
 Apply `D13`; do not create a second completion verdict from another review workflow.

AGENTS.md

@@ -45,3 +45,13 @@ pnpm format
 ```
 
 New Stack decisions after D60 must satisfy ~/src/graphrefly/authority/README.md.
+
+## Shared build and human review
+
+Use the personal `~/.codex/skills/bmad-build/SKILL.md` inside the work selected by this repository's
+dispatch/Goal workflow. Use `~/.codex/skills/bmad-build/references/qa.md` for machine-review lenses,
+finding triage and repairs, together with this repository's own QA checks and completion gates.
+Use `~/.codex/skills/bmad-checkpoint-preview/SKILL.md` for the human review trail; it replaces the old
+repository-ownership-practice checks. Ordinary delivery uses build-handoff mode; explicit checkpoint
+or `--practice` requests use interactive review. The shared skills own only this personal workflow;
+project authority, semantic approvals, execution permissions and work records remain locally governed.