@sienklogic/plan-build-run 2.11.0 → 2.12.1

package/plugins/pbr/skills/build/SKILL.md

@@ -1,23 +1,32 @@
 ---
 name: build
 description: "Execute all plans in a phase. Spawns agents to build in parallel, commits atomically."
-allowed-tools: Read, Write, Edit, Bash, Glob, Grep, Task, AskUserQuestion
-argument-hint: "<phase-number> [--gaps-only] [--team]"
+allowed-tools: Read, Write, Edit, Bash, Glob, Grep, Task, AskUserQuestion, Skill
+argument-hint: "<phase-number> [--gaps-only] [--team] [--model <model>] [--auto]"
 ---
 
 **STOP — DO NOT READ THIS FILE. You are already reading it. This prompt was injected into your context by Claude Code's plugin system. Using the Read tool on this SKILL.md file wastes ~7,600 tokens. Begin executing Step 1 immediately.**
 
-# /pbr:build — Phase Execution
+# /pbr:execute-phase — Phase Execution
 
-You are the orchestrator for `/pbr:build`. This skill executes all plans in a phase by spawning executor agents. Plans are grouped by wave and executed in order — independent plans run in parallel, dependent plans wait. Your job is to stay lean, delegate ALL building work to Task() subagents, and keep the user's main context window clean.
+**References:** `@references/questioning.md`, `@references/ui-brand.md`
+
+You are the orchestrator for `/pbr:execute-phase`. This skill executes all plans in a phase by spawning executor agents. Plans are grouped by wave and executed in order — independent plans run in parallel, dependent plans wait. Your job is to stay lean, delegate ALL building work to Task() subagents, and keep the user's main context window clean.
 
 ## Context Budget
 
 Reference: `skills/shared/context-budget.md` for the universal orchestrator rules.
+Reference: `skills/shared/agent-type-resolution.md` for agent type fallback when spawning Task() subagents.
+
+Reference: `skills/shared/agent-context-enrichment.md` for enriching executor spawn prompts with project context.
+Reference: `references/deviation-rules.md` for the deviation taxonomy (Rules 1-4) used by executors to classify and handle unexpected issues.
+Reference: `references/node-repair.md` for the node repair taxonomy (RETRY, DECOMPOSE, PRUNE, ESCALATE) used by executors to handle failed tasks.
 
 Additionally for this skill:
-- **Minimize** reading executor output — read only SUMMARY.md frontmatter, not full content
+- **Minimize** reading executor output — read only SUMMARY.md frontmatter, not full content. Exception: if `context_window_tokens` in `.planning/config.json` is >= 500000, reading full SUMMARY.md bodies is permitted when semantic content is needed for inline decisions.
 - **Delegate** all building work to executor subagents — the orchestrator routes, it doesn't build
+- **Lazy-load steps**: Instead of reading ahead, fetch the next step's instructions on demand:
+  `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js skill-section build "step-6"` → returns that step's content as JSON. Use this when context budget is DEGRADING.
 
 ## Step 0 — Immediate Output
 
@@ -31,6 +40,22 @@ Additionally for this skill:
 
 Where `{N}` is the phase number from `$ARGUMENTS`. Then proceed to Step 1.
 
+## Multi-Session Sync
+
+Before any phase-modifying operations (spawning executors, writing SUMMARY.md, updating STATE.md/ROADMAP.md), acquire a claim:
+
+```
+acquireClaim(phaseDir, sessionId)
+```
+
+If the claim fails (another session owns this phase), display: "Another session owns this phase. Use `/pbr:progress` to see active claims."
+
+On completion or error (including all exit paths), release the claim:
+
+```
+releaseClaim(phaseDir, sessionId)
+```
+
 ## Prerequisites
 
 - `.planning/config.json` exists
@@ -48,7 +73,57 @@ Parse `$ARGUMENTS` according to `skills/shared/phase-argument-parsing.md`.
 | `3` | Build phase 3 |
 | `3 --gaps-only` | Build only gap-closure plans in phase 3 |
 | `3 --team` | Use Agent Teams for complex inter-agent coordination |
+| `3 --model opus` | Use opus for all executor spawns in phase 3 (overrides config and adaptive selection) |
+| `3 --auto` | Build phase 3 with auto mode — suppress confirmation gates, auto-advance on success |
 | (no number) | Use current phase from STATE.md |
+| `3 --preview` | Preview what build would do for phase 3 without executing |
+| `3 --cross-check` | Before spawning executors for phase 3, check current plan files_modified against prior-phase provides for conflicts |
+
+---
+
+### --preview mode
+
+If `--preview` is present in `$ARGUMENTS`:
+
+1. Extract the phase slug from `$ARGUMENTS` (use the phase number to look up the slug, or pass the number directly — the CLI accepts partial slug matches).
+2. Run:
+
+   ```bash
+   node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js build-preview {phase-slug}
+   ```
+
+   Capture the JSON output.
+3. Render the following preview document (do NOT proceed to Step 2):
+
+   ```
+   ╔══════════════════════════════════════════════════════════════╗
+   ║  DRY RUN — /pbr:execute-phase {N} --preview                          ║
+   ║  No executor agents will be spawned                          ║
+   ╚══════════════════════════════════════════════════════════════╝
+
+   PHASE: {phase}
+
+   ## Plans
+   {for each plan: - {id} (wave {wave}, {task_count} tasks)}
+
+   ## Wave Structure
+   {for each wave: Wave {wave}: {plan IDs} [parallel | sequential]}
+
+   ## Files That Would Be Modified
+   {for each file in files_affected: - {file}}
+   (Total: {count} files)
+
+   ## Estimated Agent Spawns
+   {agent_count} executor task(s)
+
+   ## Critical Path
+   {critical_path joined with " → "}
+
+   ## Dependency Chain
+   {for each entry in dependency_chain: - {id} (wave {wave}) depends on: {depends_on or "none"}}
+   ```
+
+4. **STOP** — do not proceed to Step 2.
 
 ---
 
@@ -63,48 +138,62 @@ Execute these steps in order.
 Reference: `skills/shared/config-loading.md` for the tooling shortcut and config field reference.
 
 1. Parse `$ARGUMENTS` for phase number and flags
+   - If `--auto` is present in `$ARGUMENTS`: set `auto_mode = true`. Log: "Auto mode enabled — suppressing confirmation gates"
 2. Read `.planning/config.json` for parallelization, model, and gate settings (see config-loading.md for field reference)
 3. Resolve depth profile: run `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js config resolve-depth` to get the effective feature/gate settings for the current depth. Store the result for use in later gating decisions.
-4. **CRITICAL: Write .active-skill NOW.** Write `.planning/.active-skill` with the content `build` (registers with workflow enforcement hook)
-5. Validate:
-   - Phase directory exists at `.planning/phases/{NN}-{slug}/`
-   - PLAN.md files exist in the directory
-   - Prior phase dependencies are met (check for SUMMARY.md files in dependency phases)
+4. **CRITICAL (hook-enforced): Write .active-skill NOW.** Write `.planning/.active-skill` with the content `build` (registers with workflow enforcement hook)
+5. **Pre-build dependency validation (consolidated check):**
+   Run all three checks before spawning any executors:
+
+   a. **Phase artifact check:** Phase directory exists at `.planning/phases/{NN}-{slug}/` and PLAN.md files are present
+   b. **Dependency SUMMARY check:** All phases listed in `depends_on` for each plan have a SUMMARY.md file (i.e., they have been built). If any dependency is unbuilt, display:
+      ```
+      BLOCKED: Plan {plan_id} depends on phase {dep_phase} which has not been built yet.
+      Run `/pbr:build {dep_phase}` first.
+      ```
+      and stop.
+   c. **Staleness / cross-check (optional):** If `--cross-check` flag is present, compare each plan's `files_modified` list against prior-phase `provides` for conflicts (see --cross-check section above for full logic). Without the flag, skip this check.
+
+   All three sub-checks (a, b, c) are part of the same pre-build gate. Stop execution if (a) or (b) fail. Proceed after (c) even if conflicts are acknowledged.
 6. If no phase number given, read current phase from `.planning/STATE.md`
    - `config.models.complexity_map` — adaptive model mapping (default: `{ simple: "haiku", medium: "sonnet", complex: "inherit" }`)
-7. If `gates.confirm_execute` is true: use AskUserQuestion (pattern: yes-no from `skills/shared/gate-prompts.md`):
+7. If `gates.confirm_execute` is true AND `auto_mode` is NOT true: use AskUserQuestion (pattern: yes-no from `skills/shared/gate-prompts.md`):
    question: "Ready to build Phase {N}? This will execute {count} plans."
    header: "Build?"
    options:
      - label: "Yes"  description: "Start building Phase {N}"
      - label: "No"   description: "Cancel — review plans first"
-   If "No" or "Other": stop and suggest `/pbr:plan {N}` to review plans
-8. If `git.branching` is `phase`: create and switch to branch `plan-build-run/phase-{NN}-{name}` before any build work begins
+   If "No" or "Other": stop and suggest `/pbr:plan-phase {N}` to review plans
+   **Skip if:** `auto_mode` is true — auto-proceed as if user selected "Yes"
+8. **Git branching (config-gated):**
+   Read `git.branching` from config (values: none, phase, milestone, disabled).
+   - If `none` or `disabled`: skip branch operations entirely (current default behavior)
+   - If `phase`:
+     a. Determine branch name: `pbr/phase-{NN}-{name}` (e.g., `pbr/phase-03-auth`)
+     b. Check if branch already exists: `git branch --list {branch-name}`
+     c. If branch exists (resume scenario): `git switch {branch-name}` — log "Resuming on existing phase branch"
+     d. If branch does not exist: `git switch -c {branch-name}` — log "Created phase branch: {branch-name}"
+   - If `milestone`:
+     a. Determine branch name: `pbr/milestone-v{version}` where {version} comes from the active milestone in ROADMAP.md
+     b. Check if branch already exists: `git branch --list {branch-name}`
+     c. If branch exists: `git switch {branch-name}` — log "Switching to milestone branch"
+     d. If branch does not exist: `git switch -c {branch-name}` — log "Created milestone branch: {branch-name}"
+     e. Note: milestone branch persists across all phases in the milestone; it is merged by /pbr:milestone complete
 9. Record the current HEAD commit SHA: `git rev-parse HEAD` — store as `pre_build_commit` for use in Step 8-pre-c (codebase map update)
 
 **Staleness check (dependency fingerprints):**
 After validating prerequisites, check plan staleness:
-1. Read each PLAN.md file's `dependency_fingerprints` field (if present)
-2. For each fingerprinted dependency, check the current SUMMARY.md file (length + modification time)
-3. If any fingerprint doesn't match: the dependency phase was re-built after this plan was created
-4. Use AskUserQuestion (pattern: stale-continue from `skills/shared/gate-prompts.md`):
-   question: "Plan {plan_id} may be stale — dependency phase {M} was re-built after this plan was created."
-   header: "Stale"
-   options:
-     - label: "Continue anyway"  description: "Proceed with existing plans (may still be valid)"
-     - label: "Re-plan"          description: "Stop and re-plan with `/pbr:plan {N}` (recommended)"
-   If "Re-plan" or "Other": stop and suggest `/pbr:plan {N}`
-   If "Continue anyway": proceed with existing plans
-10. If plans have no `dependency_fingerprints` field, fall back to timestamp-based staleness detection:
-   a. Read `.planning/ROADMAP.md` and identify the current phase's dependencies (the `depends_on` field)
-   b. For each dependency phase, find its phase directory under `.planning/phases/`
-   c. Check if any SUMMARY.md files in the dependency phase directory have a modification timestamp newer than the current phase's PLAN.md files
-   d. If any upstream dependency was modified after planning, display a warning (do NOT block):
-      ```
-      Warning: Phase {dep_phase} (dependency of Phase {N}) was modified after this phase was planned.
-      Plans may be based on outdated assumptions. Consider re-planning with `/pbr:plan {N}`.
-      ```
-   e. This is advisory only — continue with the build after displaying the warning
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js staleness-check {phase-slug}
+```
+
+Returns `{ stale: bool, plans: [{id, stale, reason}] }`. If `stale: true` for any plan:
+- Use AskUserQuestion (pattern: stale-continue from `skills/shared/gate-prompts.md`):
+  question: "Plan {plan_id} may be stale — {reason}"
+  options: ["Continue anyway", "Re-plan with /pbr:plan-phase {N}"]
+- If "Re-plan": stop. If "Continue anyway": proceed.
+If `stale: false`: proceed silently.
 
 **Validation errors — use branded error boxes:**
 
@@ -116,24 +205,33 @@ If no plans found:
 
 Phase {N} has no plans.
 
-**To fix:** Run `/pbr:plan {N}` first.
+**To fix:** Run `/pbr:plan-phase {N}` first.
 ```
 
-If dependencies incomplete:
-```
-╔══════════════════════════════════════════════════════════════╗
-║  ERROR                                                       ║
-╚══════════════════════════════════════════════════════════════╝
+If dependencies incomplete, use conversational recovery:
 
-Phase {N} depends on Phase {M}, which is not complete.
+1. Run: `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js suggest-alternatives missing-prereq {dependency-phase-slug}`
+2. Parse the JSON response to get `existing_summaries`, `missing_summaries`, and `suggested_action`.
+3. Display what summaries exist and what is still missing.
+4. Use AskUserQuestion (pattern: yes-no from `skills/shared/gate-prompts.md`) to offer:
+   - "Build {dependency-phase} first" — stop and show: `/pbr:execute-phase {dependency-phase}`
+   - "Continue anyway (skip dependency check)" — proceed with build, note unmet deps in output
 
-**To fix:** Build Phase {M} first with `/pbr:build {M}`.
-```
+If config validation fails for a specific field, use conversational recovery:
+
+1. Run: `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js suggest-alternatives config-invalid {field} {value}`
+2. Parse the JSON response to get `valid_values` and `suggested_fix`.
+3. Display the invalid field, its current value, and the valid options.
+4. Use AskUserQuestion to offer: "Fix config.json now, or continue with current value?"
+   - If "Fix now": stop and display the `suggested_fix` instruction.
+   - If "Continue": proceed with default value for that field.
 
 ---
 
 ### Step 2: Load Config (inline)
 
+**Init-first pattern**: When spawning agents, pass the output of `node plugins/pbr/scripts/pbr-tools.cjs init execute-phase {N}` as context rather than having the agent read multiple files separately. This reduces file reads and prevents context-loading failures.
+
 Read configuration values needed for execution. See `skills/shared/config-loading.md` for the full field reference; build uses: `parallelization.*`, `features.goal_verification`, `features.inline_verify`, `features.atomic_commits`, `features.auto_continue`, `features.auto_advance`, `planning.commit_docs`, `git.commit_format`, `git.branching`.
 
 ---
@@ -156,7 +254,7 @@ This returns a JSON object with `plans` (array with plan_id, wave, depends_on, a
 4. Sort plans by plan number
 
 **If no plans match filters:**
-- With `--gaps-only`: "No gap-closure plans found. Run `/pbr:plan {N} --gaps` first."
+- With `--gaps-only`: "No gap-closure plans found. Run `/pbr:plan-phase {N} --gaps` first."
 - Without filter: "No plans found in phase directory."
 
 ---
@@ -181,7 +279,7 @@ Use AskUserQuestion (pattern: yes-no from `skills/shared/gate-prompts.md`):
     - label: "Yes"  description: "Delete existing SUMMARYs and re-execute all plans"
     - label: "No"   description: "Keep existing build — review instead"
 - If "Yes": delete SUMMARY files and proceed
-- If "No" or "Other": suggest `/pbr:review {N}`
+- If "No" or "Other": suggest `/pbr:verify-work {N}`
 
 ---
 
@@ -196,32 +294,167 @@ Validate wave consistency:
 
 ---
 
+### Step 5a: Pre-Spawn Intra-Phase Conflict Detection (conditional)
+
+**Trigger:** Only run if `context_window_tokens` in `.planning/config.json` is >= 500000.
+
+If the condition is false, skip this step entirely and proceed to Step 5c.
+
+**Purpose:** Before spawning any executor, detect whether plans within this phase conflict with each other — specifically plans that are assigned to the same wave (and would run in parallel) but modify the same files or share import graph dependencies without an explicit `depends_on` relationship. This prevents silent data-race failures where two parallel executors clobber each other's writes.
+
+**Procedure:**
+
+1. Use the `plan-index` output already collected in Step 3. No additional file reads are needed. Extract from each plan:
+   - `plan_id`
+   - `wave`
+   - `depends_on` list
+   - `files_modified` list
+
+2. **Same-file conflict detection:** For every pair of plans in the same wave, compute the intersection of their `files_modified` lists. A non-empty intersection is a **direct conflict** — both plans write the same file in parallel.
+
+3. **Import graph overlap detection:** For every pair of plans in the same wave that do NOT share a direct file conflict, check for shared directory prefix overlap. If plan A modifies `src/auth/session.ts` and plan B modifies `src/auth/middleware.ts`, they share the `src/auth/` subtree — flag as a **potential conflict** (import graph sibling edits may cause merge conflicts or runtime breakage even without touching the same file).
+
+   Overlap rule: two files share a directory prefix if their paths share at least one path segment beyond the project root (e.g., both under `src/auth/` or both under `plugins/pbr/scripts/`).
+
+4. **Implicit dependency detection:** For every pair of plans across ANY waves (not just same-wave) where plan B's `files_modified` overlaps with plan A's `files_modified` BUT plan B does NOT list plan A in its `depends_on` (and plan A is in an earlier wave), flag as an **implicit dependency warning** — plan B will overwrite plan A's work without declaring the dependency.
+
+5. Build a conflict report:
+
+   ```
+   Intra-Phase Conflict Detection Results
+
+   Plans analyzed: {count} | Context: {context_window_tokens} tokens
+
+   {If direct conflicts found:}
+   DIRECT CONFLICTS (same file, same wave — parallel execution will clobber):
+
+   | File | Plan A | Plan B | Wave |
+   |------|--------|--------|------|
+   | {path} | {plan_id} | {plan_id} | {wave} |
+
+   {If potential conflicts found:}
+   POTENTIAL CONFLICTS (shared directory subtree, same wave):
+
+   | Shared Prefix | Plan A | Plan B | Wave |
+   |---------------|--------|--------|------|
+   | {prefix}/ | {plan_id} | {plan_id} | {wave} |
+
+   {If implicit dependencies found:}
+   IMPLICIT DEPENDENCIES (file overlap, no depends_on declared):
+
+   | File | Earlier Plan | Later Plan | Missing depends_on |
+   |------|-------------|------------|-------------------|
+   | {path} | {plan_id} (wave {W}) | {plan_id} (wave {W+N}) | {plan_id} not in depends_on |
+
+   {If conflicts found:}
+   Suggested wave reordering:
+   {For each direct conflict pair: "Move {plan_id} to wave {current_wave + 1} and add depends_on: ['{other_plan_id}']"}
+   {For each implicit dependency: "Add depends_on: ['{earlier_plan_id}'] to {later_plan_id}"}
+   ```
+
+6. If **any** conflicts or warnings were found, present the report and use AskUserQuestion (pattern: yes-no from `skills/shared/gate-prompts.md`):
+
+   ```
+   question: "{N} conflict(s) detected between plans in this phase. Proceed anyway?"
+   header: "Intra-Phase Conflicts"
+   options:
+     - label: "Proceed"  description: "Continue — conflicts are intentional or I will fix them manually"
+     - label: "Abort"    description: "Stop — I need to re-plan with /pbr:plan-phase {N}"
+   ```
+
+   If user selects "Abort": stop the build. Display: "Re-run `/pbr:plan-phase {N}` and apply the suggested wave reordering above."
+
+   If user selects "Proceed": log a deviation — `DEVIATION: Intra-phase conflict(s) acknowledged by user. Plans: {conflicting plan IDs}.` — then continue to Step 5c.
+
+7. If **no conflicts found**: display `✓ No intra-phase conflicts detected. Proceeding.` and continue to Step 5c silently.
+
+---
+
+### Step 5c: Pre-Spawn Cross-Phase Conflict Check (conditional)
+
+Runs after Step 5a (intra-phase conflict detection).
+
+**Trigger:** Only run if `--cross-check` flag is present in `$ARGUMENTS` AND `context_window_tokens` in `.planning/config.json` is >= 500000.
+
+If either condition is false, skip this step entirely and proceed to Step 5b (checkpoint manifest).
+
+**Purpose:** Before spawning any executor, detect whether this phase's planned file modifications conflict with artifacts or provides from prior completed phases. A conflict means the current phase may overwrite or break something a prior phase established.
+
+**Procedure:**
+
+1. Collect current phase `files_modified` from all PLAN.md frontmatters:
+
+   ```bash
+   node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js plan-index {phase-slug}
+   ```
+
+   Extract the union of all `files_modified` arrays across plans. This is the **change surface**.
+
+2. Collect prior completed phase provides:
+
+   ```bash
+   node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js phase-list --status verified --before {phase_number}
+   ```
+
+   For each returned phase, read SUMMARY.md `provides` list (frontmatter only — keep context lean).
+
+3. Compare: for each prior-phase `provides` entry that names a specific file path, check if that path appears in the current phase's change surface.
+
+4. Present conflict report to user before proceeding:
+
+   ```
+   Cross-Phase Conflict Check Results
+
+   Files in scope: {count}
+   Prior phases checked: {count}
+
+   {If conflicts found:}
+   ⚠ Potential conflicts detected:
+
+   | File | Current Phase Plans | Prior Phase That Provides It |
+   |------|---------------------|------------------------------|
+   | {path} | {plan_ids} | Phase {N}: {slug} |
+
+   These files were established as deliverables by prior phases. Modifying them may cause regressions.
+   ```
+
+   Use AskUserQuestion (pattern: yes-no from `skills/shared/gate-prompts.md`):
+
+   ```
+   question: "{count} potential cross-phase conflicts detected. Proceed with build?"
+   header: "Conflicts"
+   options:
+     - label: "Proceed"  description: "Continue — I reviewed the conflicts and they are intentional"
+     - label: "Abort"    description: "Stop — I need to review the plan before building"
+   ```
+
+   ```
+   {If no conflicts found:}
+   ✓ No cross-phase conflicts detected. Proceeding to executor spawn.
+   ```
+
+5. If user selects "Abort": stop the build. Suggest reviewing the flagged plan files and running `/pbr:plan-phase {N}` to revise if needed.
+6. If user selects "Proceed" or no conflicts found: continue to Step 5b (checkpoint manifest).
+
+---
+
 ### Step 5b: Write Checkpoint Manifest (inline)
 
-**CRITICAL: Write .checkpoint-manifest.json NOW before entering the wave loop.**
+**CRITICAL (hook-enforced): Initialize checkpoint manifest NOW before entering the wave loop.**
 
-Before entering the wave loop, write `.planning/phases/{NN}-{slug}/.checkpoint-manifest.json`:
+**Session affinity:** The checkpoint manifest includes a `session_id` field. Before writing any phase state, validate that the current session owns the manifest by checking `manifest.session_id` matches the active session. If mismatch, another session may have taken over — re-acquire the claim or warn the user.
 
-```json
-{
-  "plans": ["02-01", "02-02", "02-03"],
-  "checkpoints_resolved": [],
-  "checkpoints_pending": [],
-  "wave": 1,
-  "deferred": [],
-  "commit_log": [],
-  "last_good_commit": null
-}
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js checkpoint init {phase-slug} --plans "{comma-separated plan IDs}"
 ```
 
-This file tracks execution progress for crash recovery and rollback. On resume after compaction, read this manifest to determine where execution left off and which plans still need work.
+After each wave completes, update the manifest:
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js checkpoint update {phase-slug} --wave {N} --resolved {plan-id} --sha {commit-sha}
+```
 
-Update the manifest after each wave completes:
-- Move completed plan IDs into `checkpoints_resolved`
-- Advance the `wave` counter
-- Record commit SHAs in `commit_log` (array of `{ plan, sha, timestamp }` objects)
-- Update `last_good_commit` to the SHA of the last successfully verified commit
-- Append any deferred items collected from executor SUMMARYs
+This tracks execution for crash recovery and rollback. Read `.checkpoint-manifest.json` on resume to reconstruct which plans are complete.
 
 ---
 
@@ -244,12 +477,77 @@ For each wave, in order (Wave 1, then Wave 2, etc.):
 
 For each plan in the current wave (excluding skipped plans):
 
+**Inline Execution Gate (conditional):**
+
+Before spawning a Task() executor, check if this plan qualifies for inline execution:
+
+1. Read `workflow.inline_execution` from config. If `false`, skip this check — use normal Task() spawn.
+2. Estimate current context usage percentage. Use a rough heuristic: if `context_window_tokens` is set in config, estimate based on conversation length; otherwise assume 30%.
+3. For each plan in the current wave, run the inline decision gate:
+   ```
+   const { shouldInlineExecution } = require(path.join(PLUGIN_ROOT, 'scripts/lib/gates/inline-execution.js'));
+   const result = shouldInlineExecution(planPath, config, contextPct);
+   ```
+   — The orchestrator conceptually evaluates these conditions; it does not literally run JavaScript. The gate logic is: plan has <= `inline_max_tasks` tasks (default 2), ALL tasks are `simple` complexity, AND context < `inline_context_cap_pct` (default 40%).
+
+4. If `result.inline` is `true`:
+   a. Display: `◆ Executing plan {plan_id} inline (trivial plan, {taskCount} simple task(s))`
+   **CRITICAL — DO NOT SKIP: Write .inline-active signal file NOW before executing inline tasks.**
+   b. Write signal file: Write the current phase number to `.planning/.inline-active` using the Write tool
+   c. Read the full PLAN.md file
+   d. For each task in the plan (parsed from XML `<task>` blocks):
+      - Read the `<action>` element and execute each numbered step directly
+      - Follow the same rules as a normal executor: create/modify files, run commands
+      - After each task, run the `<verify>` command
+      - If verify fails, fall back to normal Task() spawn for this plan
+   **CRITICAL — DO NOT SKIP: Write SUMMARY.md for the inline plan NOW. Required for crash recovery and build completion tracking.**
+   e. After all tasks complete, write SUMMARY.md directly:
+      - Use the same frontmatter format as executor-produced SUMMARYs
+      - Include: `status: completed`, `key_files`, `requires: []`, `deferred: []`
+      - Body: brief description of what was done
+   f. Delete `.planning/.inline-active` signal file
+   g. Skip the normal Task() spawn below — proceed to Step 6c (Read Results)
+
+5. If `result.inline` is `false`: proceed with normal Task() spawn below (no change to existing flow)
+
+**Agent Context Enrichment (conditional):**
+
+Before spawning each Task() executor, enrich its prompt with project context if enabled:
+
+1. Check `features.rich_agent_prompts` in config. If `true`:
+   ```
+   const { buildRichAgentContext } = require(path.join(PLUGIN_ROOT, 'scripts/lib/gates/rich-agent-context.js'));
+   const richContext = buildRichAgentContext(planningDir, config, contextPct > 50 ? 2500 : 5000);
+   ```
+   — The orchestrator conceptually evaluates this; it does not literally run JavaScript. If rich context is non-empty, include it in the executor spawn prompt under a `## Project Context` header.
+
+2. Check `features.multi_phase_awareness` in config. If `true`:
+   ```
+   const { loadMultiPhasePlans } = require(path.join(PLUGIN_ROOT, 'scripts/lib/gates/multi-phase-loader.js'));
+   const multiPhase = loadMultiPhasePlans(planningDir, currentPhaseNum, config);
+   ```
+   — If `multiPhase.phasesLoaded > 1`, include adjacent phase plan frontmatter (first 30 lines of each) in the executor spawn prompt under a `## Adjacent Phase Context` header. Do NOT include full plan bodies for adjacent phases — frontmatter only.
+
+3. If neither feature is enabled, skip enrichment entirely (no performance cost).
+
+**Local LLM plan quality check (optional, advisory):**
+
+Before spawning executors for this wave, if `config.local_llm.enabled` is `true`, run a quick classification on each plan to catch stubs before wasting an executor spawn:
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js llm classify PLAN ".planning/phases/{NN}-{slug}/{plan_id}-PLAN.md"
+```
+
+- If classification is `"stub"` or `"partial"` with confidence >= 0.7: warn the user before spawning: `"⚠ Plan {plan_id} classified as {classification} (confidence {conf}) — consider refining before building."`
+- If the command fails or returns null: skip silently (local LLM unavailable — not an error)
+- This is advisory only — never block on the result
+
 **Present plan narrative before spawning:**
 
 Display to the user before spawning:
 
 ```
-◐ Spawning {N} executor(s) for Wave {W}...
+◆ Spawning {N} executor(s) for Wave {W}...
 ```
 
 Then present a brief narrative for each plan to give the user context on what's about to happen:
@@ -271,6 +569,7 @@ This is a read-only presentation step — extract descriptions from plan frontma
 
 **Model Selection (Adaptive)**:
 Before spawning the executor for each plan, determine the model:
+0. If `--model <value>` was parsed from `$ARGUMENTS` (valid values: sonnet, opus, haiku, inherit), use that model for ALL executor Task() spawns in this run. Skip steps 1-4. The --model flag is the highest precedence override.
 1. Read the plan's task elements for `complexity` and `model` attributes
 2. If ANY task has an explicit `model` attribute, use the most capable model among them (inherit > sonnet > haiku)
 3. Otherwise, use the HIGHEST complexity among the plan's tasks to select the model:
@@ -278,70 +577,30 @@ Before spawning the executor for each plan, determine the model:
 4. If `config.models.executor` is set (non-null), it overrides adaptive selection entirely — use that model for all executors
 5. Pass the selected model to the Task() spawn
 
+If `--model <value>` is present in `$ARGUMENTS`, extract the value. Valid values: `sonnet`, `opus`, `haiku`, `inherit`. If an invalid value is provided, display an error and list valid values. Store as `override_model`.
+
 Reference: `references/model-selection.md` for full details.
 
 1. Extract the `## Summary` section from the PLAN.md (everything after the `## Summary` heading to end of file). If no ## Summary section exists (legacy plans), fall back to reading the full PLAN.md content. Note: The orchestrator reads the full PLAN.md once for narrative extraction AND summary extraction; only the ## Summary portion is inlined into the executor prompt. The full PLAN.md stays on disk for the executor to Read.
-2. Read `.planning/CONTEXT.md` (if exists)
+2. Read `.planning/PROJECT.md` ## Context section (if exists; backwards compat: fall back to `.planning/CONTEXT.md` if PROJECT.md has no ## Context)
 3. Read `.planning/STATE.md`
 4. Read prior SUMMARY.md files from the same phase (completed plans in earlier waves)
 5. Read `.planning/config.json`
 
-Construct the executor prompt:
-
-```
-You are the executor agent. Execute the following plan.
-
-<plan_summary>
-[Inline only the ## Summary section from PLAN.md]
-</plan_summary>
-
-<plan_file>
-.planning/phases/{NN}-{slug}/{plan_id}-PLAN.md
-</plan_file>
-
-<project_context>
-Project root: {absolute path to project root}
-Platform: {win32|linux|darwin}
-
-Config:
-  commit_format: {commit_format from config}
-  tdd_mode: {tdd_mode from config}
-  atomic_commits: {atomic_commits from config}
-
-Available context files (read via Read tool as needed):
-  - Config: {absolute path to config.json}
-  - State: {absolute path to STATE.md}
-{If CONTEXT.md exists:}
-  - Project context (locked decisions): {absolute path to CONTEXT.md}
-</project_context>
-
-<prior_work>
-Completed plans in this phase:
-| Plan | Status | Commits | Summary File |
-|------|--------|---------|-------------|
-| {plan_id} | complete | {hash1}, {hash2} | {absolute path to SUMMARY.md} |
-
-Read any SUMMARY file via Read tool if you need details on what prior plans produced.
-</prior_work>
-
-Execute all tasks in the plan sequentially. For each task:
-0. Read the full plan file from the path in <plan_file> to get task details
-1. Execute the <action> steps
-2. Run the <verify> commands
-3. Create an atomic commit with format: {commit_format}
-4. Record the commit hash
+Construct the executor prompt by reading `${CLAUDE_SKILL_DIR}/templates/executor-prompt.md.tmpl` and filling in all `{placeholder}` values:
 
-After all tasks complete:
-1. Write SUMMARY.md to .planning/phases/{NN}-{slug}/SUMMARY-{plan_id}.md
-2. Run self-check (verify files exist, commits exist, verify commands still pass)
-3. Return your SUMMARY.md content as your final response
+- `{NN}-{slug}` — phase directory (e.g., `02-authentication`)
+- `{plan_id}` — plan being executed (e.g., `02-01`)
+- `{commit_format}`, `{tdd_mode}`, `{atomic_commits}` — from loaded config
+- File paths: absolute paths to project root, config.json, STATE.md, PROJECT.md
+- `{prior_work table rows}` — one row per completed plan in this phase
 
-If you hit a checkpoint task, STOP and return the checkpoint response format immediately.
-```
+Use the filled template as the Task() prompt.
 
 **Spawn strategy based on config:**
 
 - If `parallelization.enabled: true` AND multiple plans in this wave:
+  - **Extended context override:** If `features.extended_context` is `true` in `.planning/config.json`, use `max_concurrent_agents = 5` regardless of the configured value (unless the configured value is already higher). The 1M context window gives the orchestrator enough headroom to track 5 concurrent executor results. Log: "Extended context: raising max_concurrent_agents to 5"
   - Spawn up to `max_concurrent_agents` Task() calls in parallel
   - Each Task() call is independent
   - Use `run_in_background: true` for each executor
@@ -360,10 +619,21 @@ Task({
   prompt: <executor prompt constructed above>
 })
 
-NOTE: The pbr:executor subagent type auto-loads the agent definition. Do NOT inline it.
+NOTE: The pbr:executor subagent type auto-loads the agent definition.
+
+After executor completes, check its output for completion markers:
+
+- `## PLAN COMPLETE` -- proceed to next plan or verification
+- `## PLAN FAILED` -- log failure, check if retry is appropriate based on workflow.node_repair_budget
+- `## CHECKPOINT: {TYPE}` -- surface checkpoint to user, pause workflow
+- No marker found -- treat as partial completion, log warning "Executor returned without completion marker"
+
+Route accordingly. Do NOT inline executor output into orchestrator context.
+
+**Memory capture:** Reference `skills/shared/memory-capture.md` — check executor output for `<memory_suggestion>` blocks and save any reusable knowledge discovered during execution.
 ```
 
-**Path resolution**: Before constructing the agent prompt, resolve `${CLAUDE_PLUGIN_ROOT}` to its absolute path. Do not pass the variable literally in prompts — Task() contexts may not expand it. Use the resolved absolute path for any pbr-tools.js or template references included in the prompt.
+**Path resolution**: Before constructing the agent prompt, resolve `${CLAUDE_PLUGIN_ROOT}` to its absolute path. Do not pass the variable literally in prompts — Task() contexts may not expand it. Use the resolved absolute path for any pbr-tools.cjs or template references included in the prompt.
 
 #### 6b. Wait for Wave Completion
 
@@ -386,79 +656,179 @@ For each completed executor:
 
 **Spot-check executor claims:**
 
-After reading each SUMMARY, perform a lightweight verification:
-- Pick 2-3 files from the SUMMARY's `key_files` list and verify they exist (`ls`)
-- Run `git log --oneline -n {commit_count}` and confirm the count matches the claimed commits
-- For each spot-checked file, verify it has >10 lines (`wc -l`): warn if trivially small
-- For each spot-checked file, search for TODO/FIXME/placeholder/stub markers: warn if found
-- Check SUMMARY.md frontmatter for `self_check_failures`: if present, warn the user:
-  "Plan {id} reported self-check failures: {list failures}. Inspect before continuing?"
-- If ANY spot-check fails, warn the user before proceeding to the next wave:
-  "Spot-check failed for plan {id}: {detail}. Inspect before continuing?"
+CRITICAL (no hook): Before reading results or advancing to the next wave, run the spot-check CLI for each completed plan.
 
-**Read executor deviations:**
+For each completed plan in this wave:
 
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js spot-check {phaseSlug} {planId}
+```
+
+Where `{phaseSlug}` is the phase directory name (e.g., `49-build-workflow-hardening`) and `{planId}` is the plan identifier (e.g., `49-01`).
+
+The command returns JSON: `{ ok, summary_exists, key_files_checked, commits_present, detail }`
+
+**If `ok` is `false` for ANY plan: STOP.** Do NOT advance to the next wave. Present the user with:
+
+```
+Spot-check FAILED for plan {planId}: {detail}
+
+Choose an action:
+  Retry   — Re-spawn executor for this plan
+  Continue — Skip this plan and proceed to next wave (may leave phase incomplete)
+  Abort   — Stop the build entirely
+```
+
+Use AskUserQuestion with the three options. Route:
+
+- Retry: Re-spawn the executor for this plan (go back to Step 6a for this plan only)
+- Continue: Log the failure, skip the plan, proceed
+- Abort: Stop all build work, leave phase in partial state
+
+**If `ok` is `true` for all plans:**
+
+- Also check SUMMARY.md frontmatter for `self_check_failures`: if present, warn the user: "Plan {id} reported self-check failures: {list}. Inspect before continuing?"
+- Also search SUMMARY.md for `## Self-Check: FAILED` marker — if present, warn before next wave
+- Between waves: verify no file conflicts from parallel executors (`git status` for uncommitted changes)
+
+**Read executor deviations:**
 After all executors in the wave complete, read all SUMMARY frontmatter and:
 - Collect `deferred` items into a running list (append to `.checkpoint-manifest.json` deferred array)
 - Flag any deviation-rule-4 (architectural) stops — these require user attention
+- If SUMMARY frontmatter contains `deviations:` with entries, summarize by rule:
+  - Rule 1 (Bug, auto-fixed): informational only
+  - Rule 2 (Missing dep, auto-installed): informational only
+  - Rule 3 (Blocking, asked user): highlight for review
+  - Rule 4 (Architectural, asked user): highlight for review — these may need re-planning
+- Check for ESCALATE entries from node repair (SUMMARY frontmatter `node_repair` field with strategy "ESCALATE"): these indicate the executor exhausted RETRY/DECOMPOSE/PRUNE and needs user intervention. Surface immediately:
+  ```
+  Executor escalated {N} task(s):
+  - Task {id}: {description} — {error details}
+  How to proceed? (retry with different approach / skip / abort)
+  ```
 - Present a brief wave summary to the user:
   "Wave {W} complete. {N} plans done. {D} deferred ideas logged. {A} architectural issues."
 
-Build a wave results table:
+Build a wave results table using standardized status symbols (`✓` complete, `✗` failed, `◆` in-progress, `○` pending — see `@references/ui-brand.md`):
 
 ```
 Wave {W} Results:
 | Plan | Status | Tasks | Commits | Deviations |
 |------|--------|-------|---------|------------|
-| {id} | complete | 3/3 | abc, def, ghi | 0 |
-| {id} | complete | 2/2 | jkl, mno | 1 |
+| {id} | ✓ complete | 3/3 | abc, def, ghi | 0 |
+| {id} | ✓ complete | 2/2 | jkl, mno | 1 |
 ```
 
 #### 6c-ii. Inline Per-Task Verification (conditional)
 
 **Skip if** the depth profile has `features.inline_verify: false`.
 
-To check: use the resolved depth profile. Only `comprehensive` mode enables inline verification by default.
-
-When inline verification is enabled, each completed plan gets a targeted verification pass before the orchestrator proceeds to the next wave. This catches issues early — before dependent plans build on a broken foundation.
+To check: use the resolved depth profile. Only `comprehensive` mode enables inline verification by default. When inline verification is enabled, each completed plan gets a targeted verification pass before the orchestrator proceeds to the next wave — catching issues early before dependent plans build on a broken foundation.
 
 For each plan that completed successfully in this wave:
 
 1. Read the plan's SUMMARY.md to get `key_files` (the files this plan created/modified)
-2. Display to the user: `◐ Spawning inline verifier for plan {plan_id}...`
+2. Display to the user: `◆ Spawning inline verifier for plan {plan_id}...`
 
-   Spawn a lightweight verifier:
+   Spawn `Task({ subagent_type: "pbr:verifier", model: "haiku", prompt: ... })`. Read `${CLAUDE_SKILL_DIR}/templates/inline-verifier-prompt.md.tmpl` and fill in `{NN}-{slug}`, `{plan_id}`, and `{comma-separated key_files list}` (key_files from PLAN.md frontmatter). Use the filled template as the `prompt` value.
 
-   <!-- NOTE: This is a targeted inline check (existence/substantiveness/wiring for specific files),
-        NOT the full must-have verifier. The canonical full verifier prompt lives in
-        agents/verifier.md and is templated via skills/review/templates/verifier-prompt.md.tmpl.
-        Keep this lightweight prompt distinct from the full verifier. -->
+3. If verifier reports FAIL for any file:
+   - Present the failure to the user: "Inline verify failed for plan {plan_id}: {details}"
+   - Re-spawn the executor for just the failed items: include only the failing file context in the prompt
+   - If the retry also fails: proceed but flag in the wave results table (don't block indefinitely)
+4. If verifier reports all PASS: continue to next wave
+
+**Note:** This adds latency (~10-20s per plan for the haiku verifier). It's opt-in via `features.inline_verify: true` for projects where early detection outweighs speed.
+
+---
 
+#### 6c-iii. Multi-Layer Validation (conditional)
+
+**Skip if** `features.multi_layer_validation` is not `true` in `.planning/config.json`.
+
+After each wave completes, run parallel validation passes over the files changed in that wave. This is a BugBot-style multi-perspective review (correctness, security, performance, etc.) that runs inline without spawning external agents.
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js validation get-passes
 ```
-Task({
-  subagent_type: "pbr:verifier",
-  model: "haiku",
-  prompt: "Targeted inline verification for plan {plan_id}.
 
-Verify ONLY these files: {comma-separated key_files list}
+This returns the active validation pass names from `config.validation_passes` (or all 8 standard passes if unconfigured). Then:
 
-For each file, check three layers:
-1. Existence — does the file exist?
-2. Substantiveness — is it more than a stub? (>10 lines, no TODO/FIXME placeholders)
-3. Wiring — is it imported/used by at least one other file?
+1. Get the list of changed files for this wave:
+   ```bash
+   git diff --name-only HEAD~{wave_commit_count}..HEAD
+   ```
+2. Build the validation prompt:
+   ```bash
+   node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js validation build-prompt '{pass_name}' '{comma-separated changed files}'
+   ```
+   (`buildValidationPrompt()` in `plan-build-run/bin/lib/validation.cjs`)
+3. Run each pass as a lightweight inline review (no Task() spawn — use direct tool calls). Display results as:
+   - `✓ Validation [{pass}]: no issues`
+   - `⚠ Validation [{pass}]: {N} finding(s) — {brief description}`
+4. If any HIGH-severity finding is reported: surface it to the user before proceeding to the next wave. Use AskUserQuestion (pattern: acknowledge-finding) to let the user decide whether to fix now or defer.
+5. Findings do NOT block the build by default — they are advisory unless `config.validation_blocking` is `true`.
 
-Report PASS or FAIL with a one-line reason per file.
-Write nothing to disk — just return your results as text."
-})
+---
+
+#### 6c-iv. Security Scan (conditional)
+
+**Skip if** `features.security_scanning` is not `true` in `.planning/config.json`.
+
+Run an OWASP-style security scan over the files changed during this wave using the patterns in `plan-build-run/bin/lib/security-scan.cjs`.
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js security scan '{space-separated changed files}'
 ```
 
-3. If verifier reports FAIL for any file:
-   - Present the failure to the user: "Inline verify failed for plan {plan_id}: {details}"
-   - Re-spawn the executor for just the failed items: include only the failing file context in the prompt
-   - If the retry also fails: proceed but flag in the wave results table (don't block indefinitely)
-4. If verifier reports all PASS: continue to next wave
+(`scanFiles()` in `plan-build-run/bin/lib/security-scan.cjs`)
 
-**Note:** This adds latency (~10-20s per plan for the haiku verifier). It's opt-in via `features.inline_verify: true` for projects where early detection outweighs speed.
+Display formatted findings:
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js security format-findings '{scan-output-json}'
+```
+
+(`formatFindings()` in `plan-build-run/bin/lib/security-scan.cjs`)
+
+Output format:
+- If no findings: `✓ Security scan: clean`
+- If findings exist:
+  ```
+  ⚠ Security scan: {N} finding(s)
+    [SEC-001] {file}:{line} — {message} (severity: high)
+    ...
+  ```
+
+HIGH-severity findings (hardcoded secrets, SQL injection, eval-of-user-input) require user acknowledgment before the build continues. MEDIUM and LOW findings are logged but non-blocking.
+
+---
+
+#### 6c-v. Smart Test Selection (conditional)
+
+**Skip if** `features.regression_prevention` is not `true` in `.planning/config.json`.
+
+After each wave, identify which test files are relevant to the changed source files using `plan-build-run/bin/lib/test-selection.cjs`. Run only those tests instead of the full suite — this catches regressions early without the latency of a full test run.
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js test-selection select '{space-separated changed files}'
+```
+
+(`selectTests()` in `plan-build-run/bin/lib/test-selection.cjs`)
+
+Then format the test command:
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js test-selection format-command '{selection-output-json}'
+```
+
+(`formatTestCommand()` in `plan-build-run/bin/lib/test-selection.cjs`)
+
+Run the resulting command (e.g., `npm test -- tests/foo.test.js tests/bar.test.js`) via Bash. Display:
+- `✓ Smart tests ({N} files): all passed`
+- `✗ Smart tests ({N} files): {M} failed — {test names}`
+
+If tests fail: present the failure to the user (same flow as Step 6d). The failing test output counts as a gap for the wave — the orchestrator does NOT automatically retry, but surfaces it for the user to decide (fix now, skip, or abort).
 
 ---
 
@@ -500,7 +870,45 @@ Use AskUserQuestion (pattern: multi-option-failure from `skills/shared/gate-prom
     - label: "Abort"     description: "Stop the entire build"
 
 **If user selects 'Retry':**
-- Re-spawn executor Task() with the same prompt
+
+**Phase Replay Enrichment (conditional):**
+
+Before re-spawning, check `workflow.phase_replay` from config (loaded in Step 2).
+
+If `workflow.phase_replay` is `true`:
+
+1. Collect replay context:
+   a. Read the original PLAN.md for this plan (already available from Step 6a)
+   b. Read SUMMARY.md if it exists (partial results from the failed attempt)
+   c. Read VERIFICATION.md if it exists (specific failure details)
+   d. Run `git diff {pre_build_commit}..HEAD -- {files_modified}` to get code diffs from the failed attempt
+2. Construct an enriched fix-executor prompt by reading `${CLAUDE_SKILL_DIR}/templates/executor-prompt.md.tmpl` and appending a `## Replay Context` section:
+
+<!-- markdownlint-disable MD046 -->
+
+    ## Replay Context
+
+    This is a RETRY of a failed execution. Use the context below to understand what went wrong and fix it.
+
+    ### Original Plan Summary
+    {plan ## Summary section}
+
+    ### Prior Attempt Results
+    {SUMMARY.md frontmatter: status, completed tasks, failed task, error details}
+
+    ### Verification Failures
+    {VERIFICATION.md gaps if available, or "No verification run yet"}
+
+    ### Code Diffs From Failed Attempt
+    {git diff output, truncated to 200 lines max to stay within 30% of the executor's context budget}
+
+<!-- markdownlint-enable MD046 -->
+
+3. Cap the replay context: if the total replay section exceeds 30% of the executor's context budget (estimate ~60k tokens for a 200k window), truncate the git diff first, then VERIFICATION details.
+
+If `workflow.phase_replay` is `false` or not set:
+- Re-spawn executor Task() with the same prompt (unchanged current behavior)
+
 - If retry also fails: ask user again (max 2 retries total)
 
 **If user selects 'Skip':**
@@ -509,87 +917,146 @@ Use AskUserQuestion (pattern: multi-option-failure from `skills/shared/gate-prom
 - If yes: warn user that those plans will also need to be skipped or adjusted
 
 **If user selects 'Rollback':**
-- Read `last_good_commit` from `.checkpoint-manifest.json`
-- If `last_good_commit` exists:
-  - Show the user: "Rolling back to commit {sha} (last verified good state). This will soft-reset {N} commits."
-  - Run: `git reset --soft {last_good_commit}`
-  - Delete the failed plan's SUMMARY.md file if it was created
-  - Update the checkpoint manifest: remove the failed plan from `checkpoints_resolved`
-  - Continue to next wave or stop based on user preference
-- If no `last_good_commit`: warn "No rollback point available (this was the first plan). Use abort instead."
+Run the rollback CLI:
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js rollback .planning/phases/{NN}-{slug}/.checkpoint-manifest.json
+```
+
+Returns `{ ok, rolled_back_to, plans_invalidated, files_deleted, warnings }`.
+
+- If `ok` is `true`: display "Rolled back to commit {rolled_back_to}. {plans_invalidated.length} downstream plans invalidated."
+  Show any warnings. Continue to next wave or stop based on user preference.
+- If `ok` is `false`: display the error message. Suggest "Use abort instead."
 
 **If user selects 'Abort':**
 - Update STATE.md with current progress
 - Present what was completed before the abort
-- Suggest: "Fix the issue and run `/pbr:build {N}` to resume (completed plans will be skipped)"
+- Suggest: "Fix the issue and run `/pbr:execute-phase {N}` to resume (completed plans will be skipped)"
 
 #### 6e. Handle Checkpoints
 
 If any executor returned `checkpoint`:
 
 1. Read the checkpoint details from the executor's response
-2. Present the checkpoint to the user:
+2. Read checkpoint type from executor response: `human-verify` | `decision` | `human-action`
+3. Read `gates.checkpoint_auto_resolve` from config.json (default: `"none"`)
+   Values: `none` | `verify-only` | `verify-and-decision` | `all`
 
-```
-Checkpoint in Plan {id}, Task {N}: {checkpoint type}
+   **When `--auto` flag is active, `checkpoint_auto_resolve` defaults to `verify-and-decision` unless explicitly set to `none` in config.**
 
-{checkpoint details — what was built, what is needed}
+4. Determine auto-resolve eligibility:
 
-{For decision type: present options}
-{For human-action type: present steps}
-{For human-verify type: present what to verify}
-```
+   - **human-action**: NEVER auto-resolve (regardless of config or `--auto` flag). Always present to user.
+   - **human-verify**:
+     - Auto-resolve if `checkpoint_auto_resolve` is `verify-only`, `verify-and-decision`, or `all`
+     - Auto-resolve if `--auto` flag is active AND `checkpoint_auto_resolve` is NOT `none`
+     - To auto-resolve: run the verify command from the checkpoint. If passes, approve and continue. If fails, present to user.
+   - **decision**:
+     - Auto-resolve if `checkpoint_auto_resolve` is `verify-and-decision` or `all`
+     - To auto-resolve: use the first/default option. Log which option was auto-selected.
+     - If `checkpoint_auto_resolve` is `verify-only` or `none`: present to user.
 
-3. Wait for user response
-4. Spawn a FRESH continuation executor:
+5. If auto-resolved:
+   Log: `"Auto-resolved {type} checkpoint for Plan {id}, Task {N}: {resolution}"`
+   Resume executor with resolution context.
 
-Reference: `references/continuation-format.md` for the continuation protocol.
+6. If NOT auto-resolved, present based on type:
 
-```
-You are the executor agent. Continue executing a plan from a checkpoint.
+   **For `human-verify`:**
 
-<plan_summary>
-[Inline only the ## Summary section from PLAN.md]
-</plan_summary>
+   ```text
+   CHECKPOINT: Verify Output
 
-<plan_file>
-.planning/phases/{NN}-{slug}/{plan_id}-PLAN.md
-</plan_file>
+   Plan {id}, Task {N}: {description}
 
-<completed_tasks>
-| Task | Commit | Status |
-|------|--------|--------|
-| {task_name} | {hash} | complete |
-| {task_name} | {hash} | complete |
-| {checkpoint_task} | — | checkpoint |
-</completed_tasks>
+   What was built:
+   {what_built from checkpoint data}
 
-<checkpoint_resolution>
-User response to checkpoint: {user's response}
-Resume at: Task {N+1} (or re-execute checkpoint task with user's answer)
-</checkpoint_resolution>
+   How to verify:
+   {verify_steps from checkpoint data}
+   ```
 
-<project_context>
-{Same lean context as original spawn — config key-values + file paths, not inlined bodies}
-</project_context>
+   Use AskUserQuestion with options: "Looks good", "Has issues" (+ text field for details)
 
-Continue execution from the checkpoint. Skip completed tasks. Process the checkpoint resolution, then continue with remaining tasks. Write SUMMARY.md when done.
-```
+   **For `decision`:**
+
+   ```text
+   CHECKPOINT: Decision Required
+
+   Plan {id}, Task {N}: {description}
+
+   Options:
+   {options from checkpoint data}
+   ```
+
+   Use AskUserQuestion with the options from the checkpoint as selectable choices
+
+   **For `human-action`:**
+
+   ```text
+   CHECKPOINT: Action Required
+
+   Plan {id}, Task {N}: {description}
+
+   You need to:
+   {required_action from checkpoint data}
+
+   Reply when complete.
+   ```
+
+   Use AskUserQuestion with options: "Done", "Can't do this right now"
+   If user selects "Can't do this right now": suggest "Run `/pbr:pause` to save state and resume later."
+
+7. Wait for user response
+8. Spawn a FRESH continuation executor:
+
+Reference: `references/continuation-format.md` for the continuation protocol.
+
+Read `${CLAUDE_SKILL_DIR}/templates/continuation-prompt.md.tmpl` and fill in:
+
+- `{NN}-{slug}`, `{plan_id}` — current phase and plan
+- `{plan_summary}` — the ## Summary section from PLAN.md
+- `{task table rows}` — one row per task with completion status
+- `{user's response}` — the checkpoint resolution from Step 3
+- `{project context key-values}` — config values + file paths
+
+Use the filled template as the Task() prompt.
+
+#### 6e-ii. CI Gate (after wave completion, conditional)
+
+If `config.ci.gate_enabled` is `true` AND `config.git.branching` is not `none`:
+
+1. Push current commits: `git push`
+2. Wait 5 seconds for CI to trigger
+3. Get the current run ID:
+   ```bash
+   gh run list --branch $(git branch --show-current) --limit 1 --json databaseId -q '.[0].databaseId'
+   ```
+4. Poll CI status using CLI:
+   ```bash
+   node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js ci-poll <run-id> [--timeout <seconds>]
+   ```
+   Returns `{ status, conclusion, url, next_action, elapsed_seconds }`.
+5. If `next_action` is `"continue"`: proceed to next wave
+6. If `next_action` is `"wait"`: re-run ci-poll after 15 seconds (repeat up to `config.ci.wait_timeout_seconds`)
+7. If `next_action` is `"abort"` or `status` is `"failed"`:
+   Show warning box and use AskUserQuestion: Wait / Continue anyway / Abort
+8. If "Continue anyway": log deviation — `DEVIATION: CI gate bypassed for wave {N}`
+9. If "Abort": stop build, update STATE.md
 
 #### 6f. Update STATE.md
 
 After each wave completes (all plans in the wave are done, skipped, or aborted):
 
 **SUMMARY gate — verify before updating STATE.md:**
+For every plan in the wave, run:
 
-Before writing any STATE.md update, verify these three gates for every plan in the wave:
-1. SUMMARY file exists at the expected path
-2. SUMMARY file is not empty (file size > 0)
-3. SUMMARY file has a valid title and YAML frontmatter (contains `---` delimiters and a `status:` field)
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js summary-gate {phase-slug} {plan-id}
+```
 
-Block the STATE.md update until ALL gates pass. If any gate fails:
-- Warn user: "SUMMARY gate failed for plan {id}: {which gate}. Cannot update STATE.md."
-- Ask user to retry the executor or manually inspect the SUMMARY file
+Returns `{ ok: bool, gate: string, detail: string }`. Block STATE.md update until ALL plans return `ok: true`. If any fail, warn: "SUMMARY gate failed for plan {id}: {gate} — {detail}. Cannot update STATE.md."
 
 Once gates pass, update `.planning/STATE.md`:
 
@@ -600,6 +1067,15 @@ node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js state update status building
 node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js state update last_activity now
 ```
 
+**CLI exit code verification with retry**: After running each pbr-tools CLI command above, check the exit code:
+
+- If the command succeeds (exit 0): proceed to next command
+- If the command fails (non-zero exit):
+  1. Log the error: "CLI command failed: {command} (exit {code})"
+  2. Wait 1 second
+  3. Retry the command once
+  4. If retry also fails: log warning "CLI command failed after retry: {command}" and continue. Do NOT block the workflow -- state can be reconciled later via `/pbr:status`
+
 - Current plan progress: "{completed}/{total} in current phase"
 - Last activity timestamp
 - Progress bar percentage
@@ -607,6 +1083,15 @@ node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js state update last_activity now
 
 **STATE.md size limit:** Follow the size limit enforcement rules in `skills/shared/state-update.md` (150 lines max — collapse completed phases, remove duplicated decisions, trim old sessions).
 
+**Completion check:** Before proceeding to next wave, confirm ALL of:
+- [ ] SUMMARY gate passed for every plan in this wave
+- [ ] STATE.md frontmatter `plans_complete` updated
+- [ ] STATE.md body progress bar updated
+- [ ] `last_activity` timestamp refreshed
+
+To verify programmatically: `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js step-verify build step-6f '["STATE.md updated","SUMMARY.md exists","commit made"]'`
+If any item fails, investigate before proceeding to Step 7.
+
 ---
 
 ### Step 7: Phase Verification (delegated, conditional)
@@ -623,11 +1108,60 @@ To check: run `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js config resolve-de
 This implements budget mode's "skip verifier for < 3 tasks" rule: small phases in quick mode don't need a full verification pass.
 
 **If skipping because `features.goal_verification` is `false`:**
-Note for Step 8f completion summary: append "Note: Automatic verification was skipped (goal_verification: false). Run `/pbr:review {N}` to verify what was built."
+Note for Step 8f completion summary: append "Note: Automatic verification was skipped (goal_verification: false). Run `/pbr:verify-work {N}` to verify what was built."
+
+**Confidence-Gated Verification Skip (conditional):**
+
+Before spawning the verifier, check if the build passes the confidence gate:
+
+1. Read `verification.confidence_gate` from config. If `false` or not set, skip this check — proceed to normal verification flow.
+2. Read `verification.confidence_threshold` from config (default: `100`).
+3. Collect confidence signals:
+   a. Read ALL SUMMARY.md frontmatter from this phase's completed plans. Extract `completion` percentage from each.
+   b. Calculate aggregate completion: average of all plan completion percentages.
+   c. Check commit SHAs: for each SUMMARY.md that lists `commits`, verify they exist via `git log --oneline {sha} -1` (quick existence check, not full log).
+   d. Detect test suite: check for `package.json` (scripts.test), `pytest.ini`/`pyproject.toml` ([tool.pytest]), `Makefile` (test target), or `Cargo.toml`. Use the first match.
+   e. Run test suite: execute the detected test command (e.g., `npm test`, `pytest`, `make test`). Capture exit code.
+
+4. Evaluate confidence gate:
+   - `completion_met`: aggregate completion >= `confidence_threshold`
+   - `shas_verified`: all listed commit SHAs exist in git log
+   - `tests_passed`: test suite exit code is 0 (or no test suite detected — treat as pass with warning)
+
+5. If ALL three pass:
+   - Display: `Confidence gate passed (completion: {pct}%, SHAs: verified, tests: passed) — skipping verifier`
+   - Set verification status to `passed` (auto-verified)
+   **CRITICAL — DO NOT SKIP: Write VERIFICATION.md NOW. Without this file, the autonomous loop cannot confirm phase completion.**
+   - Write a minimal VERIFICATION.md:
+
+<!-- markdownlint-disable MD046 -->
+
+     ```yaml
+     ---
+     status: passed
+     method: confidence-gate
+     completion: {pct}
+     shas_verified: true
+     tests_passed: true
+     must_haves_checked: 0
+     must_haves_passed: 0
+     ---
+     # Verification — Confidence Gate
+
+     Phase auto-verified via confidence gate. Run `/pbr:verify-work {N}` for full must-have verification.
+     ```
+
+<!-- markdownlint-enable MD046 -->
+
+   - Skip the verifier spawn — proceed directly to Step 8.
+
+6. If ANY signal fails:
+   - Display: `Confidence gate not met ({failed_signals}) — spawning verifier`
+   - Proceed with normal verification flow below (unchanged behavior).
 
 **If verification is enabled:**
 
-Display to the user: `◐ Spawning verifier...`
+Display to the user: `◆ Spawning verifier...`
 
 Spawn a verifier Task():
 
@@ -638,13 +1172,25 @@ Task({
 })
 
 NOTE: The pbr:verifier subagent type auto-loads the agent definition. Do NOT inline it.
+
+After verifier completes, check for completion marker: `## VERIFICATION COMPLETE`. Read VERIFICATION.md frontmatter for status.
 ```
 
-**Path resolution**: Before constructing the agent prompt, resolve `${CLAUDE_PLUGIN_ROOT}` to its absolute path. Do not pass the variable literally in prompts — Task() contexts may not expand it. Use the resolved absolute path for any pbr-tools.js or template references included in the prompt.
+**Path resolution**: Before constructing the agent prompt, resolve `${CLAUDE_PLUGIN_ROOT}` to its absolute path. Do not pass the variable literally in prompts — Task() contexts may not expand it. Use the resolved absolute path for any pbr-tools.cjs or template references included in the prompt.
 
 #### Verifier Prompt Template
 
-Use the same verifier prompt template as defined in `/pbr:review`: read `skills/review/templates/verifier-prompt.md.tmpl` and fill in its placeholders with the phase's PLAN.md must_haves and SUMMARY.md file paths. This avoids maintaining duplicate verifier prompts across skills.
+Use the same verifier prompt template as defined in `/pbr:verify-work`: read `${CLAUDE_PLUGIN_ROOT}/skills/review/templates/verifier-prompt.md.tmpl` and fill in its placeholders with the phase's PLAN.md must_haves and SUMMARY.md file paths. This avoids maintaining duplicate verifier prompts across skills.
+
+**Prepend this block to the verifier prompt before sending:**
+```
+<files_to_read>
+CRITICAL (no hook): Read these files BEFORE any other action:
+1. .planning/phases/{NN}-{slug}/PLAN-*.md — must-haves to verify against
+2. .planning/phases/{NN}-{slug}/SUMMARY-*.md — executor build summaries
+3. .planning/phases/{NN}-{slug}/VERIFICATION.md — prior verification results (if exists)
+</files_to_read>
+```
 
 After the verifier returns, read the VERIFICATION.md frontmatter and display the results:
 
@@ -665,7 +1211,7 @@ If `--gaps-only` flag was used AND `features.goal_verification` is `true`:
 2. Re-run the verifier using the same Step 7 process — this produces a fresh `VERIFICATION.md` that accounts for the gap-closure work
 3. Read the new verification status for use in determining `final_status` below
 
-This ensures that `/pbr:review` after a `--gaps-only` build sees the updated verification state, not stale gaps from before the fix.
+This ensures that `/pbr:verify-work` after a `--gaps-only` build sees the updated verification state, not stale gaps from before the fix.
 
 **8-pre-b. Determine final status based on verification:**
 - If verification ran and status is `passed`: final_status = "built"
@@ -675,15 +1221,17 @@ This ensures that `/pbr:review` after a `--gaps-only` build sees the updated ver
 
 **8-pre-c. Codebase map incremental update (conditional):**
 
+**CRITICAL (no hook): Run codebase map update if conditions are met. Do NOT skip this step.**
+
 Only run if ALL of these are true:
-- `.planning/codebase/` directory exists (project was previously scanned with `/pbr:scan`)
+- `.planning/codebase/` directory exists (project was previously scanned with `/pbr:map-codebase`)
 - Build was not aborted
 - `git diff --name-only {pre_build_commit}..HEAD` shows >5 files changed OR `package.json`/`requirements.txt`/`go.mod`/`Cargo.toml` was modified
 
 If triggered:
 1. Record the pre-build commit SHA at the start of Step 1 (before any executors run) for comparison
 2. Run `git diff --name-only {pre_build_commit}..HEAD` to get the list of changed files
-3. Display to the user: `◐ Spawning codebase mapper (incremental update)...`
+3. Display to the user: `◆ Spawning codebase mapper (incremental update)...`
 
    Spawn a lightweight mapper Task():
    ```
@@ -695,7 +1243,17 @@ If triggered:
    ```
 4. Do NOT block on this — use `run_in_background: true` and continue to Step 8a. Report completion in Step 8f if it finishes in time.
 
-**CRITICAL: Update ROADMAP.md progress table NOW. Do NOT skip this step.**
+**8-pre-d. Write phase manifest (on successful completion):**
+
+If all plans completed successfully (final_status is "built" or "built (unverified)"), write `.phase-manifest.json` to the phase directory. This manifest aggregates all plan commits for the undo skill's `--phase NN` mode:
+
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js phase write-manifest {phase-slug}
+```
+
+The manifest collects commit hashes from each plan's SUMMARY.md and stores them as a single artifact that `completePhase()` uses for rollback support. If the command fails, log a warning but do not block completion.
+
+**CRITICAL (no hook): Update ROADMAP.md progress table NOW. Do NOT skip this step. (roadmap-sync warns)**
 
 **8a. Update ROADMAP.md Progress table** (REQUIRED — do this BEFORE updating STATE.md):
 
@@ -704,8 +1262,13 @@ If triggered:
 node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js roadmap update-plans {phase} {completed} {total}
 node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js roadmap update-status {phase} {final_status}
 ```
-These return `{ success, old_status, new_status }` or `{ success, old_plans, new_plans }`. Falls back to manual editing if unavailable.
+These return `{ success, old_status, new_status }` or `{ success, old_plans, new_plans }`.
 
+> Note: Use CLI for atomic writes — direct Write bypasses file locking.
+
+**CLI exit code verification with retry**: Same pattern as Step 6f -- if any CLI command fails, retry once after 1 second. If retry also fails, log a warning but do not block the build.
+
+**Last-resort fallback** (only if CLI is completely unavailable after retry):
 1. Open `.planning/ROADMAP.md`
 2. Find the `## Progress` table
 3. Locate the row matching this phase number
@@ -713,13 +1276,65 @@ These return `{ success, old_status, new_status }` or `{ success, old_plans, new
 5. Update the `Status` column to the final_status determined in Step 8-pre
 6. Save the file — do NOT skip this step
 
-**CRITICAL: Update STATE.md NOW with phase completion status. Do NOT skip this step.**
+**CRITICAL (no hook): Update STATE.md NOW with phase completion status. Do NOT skip this step. (state-sync warns)**
 
-**8b. Update STATE.md:**
-- Phase status: {final_status from Step 8-pre}
-- Plan completion count
-- Last activity timestamp
-- Progress bar
+**8b. Update STATE.md (CRITICAL (no hook) — update BOTH frontmatter AND body):**
+
+> Note: Use CLI for atomic writes — direct Write bypasses file locking.
+
+Use CLI commands to update STATE.md (keeps frontmatter and body in sync atomically):
+```bash
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js state update status {final_status}
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js state update plans_complete {N}
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js state update last_activity now
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js state update progress_percent {pct}
+node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js state update last_command "/pbr:execute-phase {N}"
+```
+These update both frontmatter fields (`status`, `plans_complete`, `last_activity`, `progress_percent`, `last_command`) and the body `## Current Position` section (`Phase:`, `Plan:`, `Status:`, `Last activity:`, `Progress:` bar) atomically — they MUST stay in sync.
+
+**Completion check:** Before proceeding to 8c, confirm ALL of:
+- [ ] STATE.md frontmatter fields set: status, plans_complete, last_activity, progress_percent, last_command
+- [ ] STATE.md body ## Current Position updated: Phase, Status, Last activity, Progress bar
+- [ ] Frontmatter and body are consistent (same status value in both)
+
+To verify programmatically: `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js step-verify build step-8b '["STATE.md updated","ROADMAP.md updated","commit made"]'`
+If any item fails, investigate before marking phase complete.
+
+**8b-ii. Calculate and write velocity metrics to STATE.md (informational only):**
+
+After updating STATE.md status, calculate velocity metrics from this phase's build data:
+
+1. Read all SUMMARY.md files in the current phase directory. Extract `metrics.duration_minutes` from each frontmatter (if present). If duration is not in frontmatter, estimate from git log: time between first and last commit for that plan using `git log --format=%aI`.
+
+2. Calculate:
+   - `plans_executed`: count of SUMMARY.md files in this phase
+   - `avg_duration_minutes`: mean of all plan durations (round to nearest integer)
+   - `phase_duration_minutes`: total time from first plan start to last plan completion
+   - `trend`: compare this phase's avg\_duration to the previous phase's avg\_duration (read prior phase SUMMARY.md files if they exist):
+     - If >20% faster: "improving"
+     - If >20% slower: "degrading"
+     - Otherwise: "stable"
+     - If no previous phase data: "baseline"
+
+3. Count total plans across all phases from ROADMAP.md progress table for the `total_plans` metric.
+
+4. Write a `## Metrics` section to STATE.md. If a `## Metrics` section already exists, replace it. Otherwise insert it before `## History` (or append at end if no History section):
+
+```
+## Metrics
+
+| Metric | Value |
+|--------|-------|
+| Plans executed (this phase) | {plans_executed} |
+| Avg plan duration | {avg_duration_minutes} min |
+| Phase duration | {phase_duration_minutes} min |
+| Trend | {trend} |
+| Total plans (all phases) | {total_plans} |
+```
+
+5. Keep the metrics section concise (under 10 lines) to respect the 100-line STATE.md cap from RH-56. The metrics section replaces itself on each phase completion — it does NOT accumulate.
+
+6. Metrics are **informational only** — they do NOT gate any workflow decisions.
 
 **8c. Commit planning docs (if configured):**
 Reference: `skills/shared/commit-planning-docs.md` for the standard commit pattern.
@@ -729,38 +1344,131 @@ If `planning.commit_docs` is `true`:
 
 **8d. Handle git branching:**
 If `git.branching` is `phase`:
-- All work was done on the phase branch (created in Step 1)
-- Squash merge to main: `git checkout main && git merge --squash plan-build-run/phase-{NN}-{name}`
-- Use AskUserQuestion (pattern: yes-no from `skills/shared/gate-prompts.md`):
-  question: "Phase {N} complete on branch `plan-build-run/phase-{NN}-{name}`. Squash merge to main?"
-  header: "Merge?"
-  options:
-    - label: "Yes, merge"   description: "Squash merge to main and delete the phase branch"
-    - label: "No, keep"     description: "Leave the branch as-is for manual review"
-- If "Yes, merge": complete the merge and delete the phase branch
-- If "No, keep" or "Other": leave the branch as-is and inform the user
+- Verify we are on the phase branch: `git branch --show-current`
+- If NOT on the phase branch, warn: "Expected to be on {branch-name} but on {current}. Skipping merge."
+- If on the phase branch:
+  - Use AskUserQuestion (pattern: yes-no from `skills/shared/gate-prompts.md`):
+    question: "Phase {N} complete on branch `pbr/phase-{NN}-{name}`. Squash merge to main?"
+    header: "Merge?"
+    options:
+      - label: "Yes, merge"   description: "Squash merge to main and delete the phase branch"
+      - label: "No, keep"     description: "Leave the branch as-is for manual review"
+  - If "Yes, merge":
+    1. `git switch main`
+    2. `git merge --squash pbr/phase-{NN}-{name}`
+    3. `git commit -m "feat({NN}-{name}): phase {N} squash merge"`
+    4. `git branch -d pbr/phase-{NN}-{name}`
+    5. Log: "Phase branch merged and deleted"
+  - If "No, keep" or "Other": leave the branch as-is and inform the user
+  - **Skip if:** `auto_mode` is true — auto-proceed with "Yes, merge"
+
+If `git.branching` is `milestone`:
+- Do NOT merge after individual phase completion
+- Log: "Phase complete on milestone branch. Branch will be merged when milestone is completed via /pbr:milestone complete."
+- Stay on the milestone branch for the next phase
+
+**8d-ii. PR Creation (when branching enabled):**
+
+If `config.git.branching` is `phase` or `milestone` AND phase verification passed:
+
+1. Push the phase branch: `git push -u origin {branch-name}`
+2. If `config.git.auto_pr` is `true`:
+   - Run: `gh pr create --title "feat({phase-scope}): {phase-slug}" --body "$(cat <<'EOF'
+## Phase {N}: {phase name}
+
+**Goal**: {phase goal from ROADMAP.md}
+
+### Key Files
+{key_files from SUMMARY.md, bulleted}
+
+### Verification
+{pass/fail status from VERIFICATION.md}
+
+---
+Generated by Plan-Build-Run
+EOF
+)"`
+3. If `config.git.auto_pr` is `false`:
+   - Use AskUserQuestion to ask: "Phase branch pushed. Create a PR?"
+   - Options: Yes (create PR as above) / No / Later (skip)
 
 **8e. Auto-advance / auto-continue (conditional):**
 
+**If `auto_mode` is `true`:** Set `features.auto_advance = true` and `mode = autonomous` behavior for the remainder of this invocation. Pass `--auto` to chained skills. Fall through to the auto_advance logic below.
+
+**Speculative Planning (conditional):**
+
+Before evaluating auto-advance, check if the next phase can be speculatively planned:
+
+1. Read `workflow.speculative_planning` from config. If `false` or not set, skip this block entirely.
+2. Determine the next phase number: `N+1`.
+3. Check ROADMAP.md for phase N+1:
+   a. Does phase N+1 exist in the roadmap?
+   b. Read its `**Depends on:**` field. Does it list phase N as a dependency?
+   c. If N+1 depends on N: skip speculative planning — the next phase needs this phase's output.
+   d. If N+1 does NOT depend on N (independent): proceed with speculative planning.
+
+4. Check deviation count from this build:
+   a. Read all SUMMARY.md frontmatter from this phase. Count total `deviations` across all plans.
+   b. If deviation count > 2: skip speculative planning. Display: `Speculative planning skipped — {count} deviations detected (threshold: 2)`
+   c. If deviation count <= 2: proceed.
+
+5. Spawn speculative planner in background:
+   a. Display: `Spawning speculative planner for Phase {N+1} (independent of Phase {N})...`
+   b. Write speculative plan to a temp location: `.planning/phases/{N+1-slug}/.speculative/`
+   c. Spawn:
+
+<!-- markdownlint-disable MD046 -->
+
+      Task({
+        subagent_type: "pbr:planner",
+        model: "sonnet",
+        run_in_background: true,
+        prompt: "Plan Phase {N+1}: {phase goal from ROADMAP.md}. Write plans to .planning/phases/{N+1-slug}/. This is a SPECULATIVE plan — it may be discarded if Phase {N} deviates significantly."
+      })
+
+<!-- markdownlint-enable MD046 -->
+
+   d. Do NOT block on the result — continue to auto-advance evaluation.
+
+6. After the build completes (in Step 8f), check if speculative planner finished:
+   a. If finished AND phase N deviation count is still <= 2: move speculative plans from `.speculative/` to the phase directory. Display: `Speculative plans for Phase {N+1} ready`
+   b. If finished BUT deviation count > 2: discard speculative plans. Delete `.speculative/` directory. Display: `Speculative plans for Phase {N+1} discarded (Phase {N} deviated)`
+   c. If not yet finished: note in completion summary. Plans will be available when the planner completes.
+
 **If `features.auto_advance` is `true` AND `mode` is `autonomous`:**
 Chain to the next skill directly within this session. This eliminates manual phase cycling.
 
+**NOTE:** When `workflow.phase_boundary_clear` is `enforce`, do NOT write `.auto-next` — force the user to /clear first. The phase boundary enforcement in auto-continue.js will also block continuation if a `.phase-boundary-pending` signal file exists.
+
 | Build Result | Next Action | How |
 |-------------|-------------|-----|
-| Verification passed, more phases | Plan next phase | `Skill({ skill: "pbr:plan", args: "{N+1}" })` |
-| Verification skipped | Run review | `Skill({ skill: "pbr:review", args: "{N}" })` |
-| Verification gaps found | **HARD STOP** — present gaps to user | Do NOT auto-advance past failures |
-| Last phase in current milestone | **HARD STOP** — milestone boundary | Suggest `/pbr:milestone audit`. Explain: "auto_advance pauses at milestone boundaries — your sign-off is required." |
-| Build errors occurred | **HARD STOP** — errors need human review | Do NOT auto-advance past errors |
+| Verification passed, more phases | Plan next phase | `Skill({ skill: "pbr:plan", args: "{N+1}" })` (append `--auto` if `auto_mode`) |
+| Verification skipped, `workflow.validate_phase: true` | Run validate-phase | `Skill({ skill: "pbr:validate-phase", args: "{N}" })` (append `--auto` if `auto_mode`) |
+| Verification skipped, `workflow.validate_phase: false` | Run review | `Skill({ skill: "pbr:review", args: "{N}" })` (append `--auto` if `auto_mode`) |
+| Verification gaps found | **HARD STOP** — present gaps to user | If `auto_continue` also true: write `.planning/.auto-next` with `/pbr:verify-work {N}` before stopping. Do NOT auto-advance past failures. |
+| Last phase in current milestone | **HARD STOP** — milestone boundary | If `auto_continue` also true: write `.planning/.auto-next` with `/pbr:complete-milestone` before stopping. Suggest `/pbr:audit-milestone`. Explain: "auto_advance pauses at milestone boundaries — your sign-off is required." |
+| Build errors occurred | **HARD STOP** — errors need human review | If `auto_continue` also true: write `.planning/.auto-next` with `/pbr:execute-phase {N}` before stopping. Do NOT auto-advance past errors. |
 
 After invoking the chained skill, it runs within the same session. When it completes, the chained skill may itself chain further (review→plan, plan→build) if auto_advance remains true. This creates the full cycle: build→review→plan→build→...
 
 **Else if `features.auto_continue` is `true`:**
-Write `.planning/.auto-next` containing the next logical command (e.g., `/pbr:plan {N+1}` or `/pbr:review {N}`)
+Write `.planning/.auto-next` containing the next logical command (e.g., `/pbr:plan-phase {N+1}` or `/pbr:verify-work {N}`)
 - This file signals to the user or to wrapper scripts that the next step is ready
 
+**Completion check:** Before proceeding to 8f, confirm ALL of:
+- [ ] auto_advance OR auto_continue evaluated (one path taken)
+- [ ] If auto_continue: `.auto-next` file written with correct next command
+- [ ] Pending todos evaluated (Step 8e-ii)
+- [ ] Clearly-satisfied todos auto-closed via `pbr-tools.cjs todo done`
+
+To verify programmatically: `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js step-verify build step-8e '["STATE.md updated","commit made"]'`
+If any item fails, investigate before closing the session.
+
 **8e-ii. Check Pending Todos:**
 
+**CRITICAL (no hook): Check pending todos after build. Do NOT skip this step.**
+
 After completing the build, check if any pending todos are now satisfied:
 
 1. Check if `.planning/todos/pending/` exists and contains files
@@ -774,16 +1482,38 @@ After completing the build, check if any pending todos are now satisfied:
 
 Only auto-close when the match is unambiguous. When in doubt, leave it open.
 
+**8f-pre. Phase Boundary Clear (conditional):**
+
+After verification completes and before the branded banner, check `workflow.phase_boundary_clear` from config:
+
+1. Read config: `node ${CLAUDE_PLUGIN_ROOT}/scripts/pbr-tools.js config-get workflow.phase_boundary_clear`
+2. If `"off"` or missing: skip (no action — current default behavior)
+3. If `"recommend"`:
+   - Add an advisory line to the completion output:
+     ```
+     Context Reset: Run /clear before starting the next phase for optimal context quality.
+     ```
+   - This is informational only — do NOT block the user
+4. If `"enforce"`:
+   - Add a prominent warning to the completion output:
+     ```
+     CONTEXT RESET REQUIRED
+     Phase boundary clear is enforced. Run /clear before continuing to the next phase.
+     The next /pbr:plan-phase or /pbr:execute-phase will work best with a fresh context window.
+     ```
+   - Write `.planning/.phase-boundary-pending` file containing `{phase_num}` — this signals to auto-continue.js that a clear is needed
+   - Do NOT write `.planning/.auto-next` when enforce is active — the auto-continue hook will handle this
+
 **8f. Present completion summary:**
 
-Use the branded output templates from `references/ui-formatting.md`. Route based on status:
+Use the branded output templates from `references/ui-brand.md`. Route based on status:
 
 | Status | Template |
 |--------|----------|
 | `passed` + more phases in current milestone | "Phase Complete" template |
 | `passed` + last phase in current milestone | "Milestone Complete" template |
 
-**Milestone boundary detection:** To determine "last phase in current milestone", read ROADMAP.md and find the `## Milestone:` section containing the current phase. Check its `**Phases:** start - end` range. If the current phase equals `end`, this is the last phase in the milestone. For projects with a single milestone or no explicit milestone sections, "last phase in ROADMAP" is equivalent.
+**Milestone boundary detection:** To determine "last phase in current milestone", read ROADMAP.md and find the `## Milestone:` section containing the current phase. Active milestones use `## Milestone:` headings directly; completed milestones are wrapped in `<details><summary>## Milestone:` blocks or use the legacy `-- COMPLETED` suffix. Check the active milestone's `**Phases:** start - end` range. If the current phase equals `end`, this is the last phase in the milestone. For projects with a single milestone or no explicit milestone sections, "last phase in ROADMAP" is equivalent.
 | `gaps_found` | "Gaps Found" template |
 
 Before the branded banner, include the results table:
@@ -805,107 +1535,54 @@ Total files modified: {count}
 Deviations: {count}
 ```
 
-Then present the appropriate branded banner:
-
-**If `passed` + more phases:**
-```
-╔══════════════════════════════════════════════════════════════╗
-║  PLAN-BUILD-RUN ► PHASE {N} COMPLETE ✓                       ║
-╚══════════════════════════════════════════════════════════════╝
+Then present the appropriate branded banner from Read `references/ui-brand.md` § "Completion Summary Templates":
 
-**Phase {N}: {Name}**
+- **If `passed` + more phases:** Use the "Phase Complete" template. Fill in phase number, name, plan count, and next phase details.
+- **If `passed` + last phase:** Use the "Milestone Complete" template. Fill in phase count.
+- **If `gaps_found`:** Use the "Gaps Found" template. Fill in phase number, name, score, and gap summaries from VERIFICATION.md.
 
-{X} plans executed
-Goal verified ✓
-
-
-
-╔══════════════════════════════════════════════════════════════╗
-║  ▶ NEXT UP                                                   ║
-╚══════════════════════════════════════════════════════════════╝
-
-**Phase {N+1}: {Name}** — {Goal from ROADMAP.md}
-
-`/pbr:plan {N+1}`
-
-<sub>`/clear` first → fresh context window</sub>
-
-
-
-**Also available:**
-- `/pbr:review {N}` — manual acceptance testing before continuing
-- `/pbr:discuss {N+1}` — talk through the next phase before planning
-- `/pbr:status` — see full project status
+**Conditional routing additions for Next Up block:**
 
+After the primary next-action command in the NEXT UP block, append the following conditional suggestions (in this order):
 
+**A. Ship suggestion (gated: git.branching=phase AND auto_pr=true):**
+Read `config.git.branching` and `config.git.auto_pr` from `.planning/config.json`.
+If `branching === "phase"` AND `auto_pr === true`:
 ```
-
-**If `passed` + last phase:**
+**Also available:** Create a PR for this phase's branch
+`/pbr:ship`
 ```
-╔══════════════════════════════════════════════════════════════╗
-║  PLAN-BUILD-RUN ► MILESTONE COMPLETE 🎉                      ║
-╚══════════════════════════════════════════════════════════════╝
-
-{N} phases completed
-All phase goals verified ✓
-
-
-
-╔══════════════════════════════════════════════════════════════╗
-║  ▶ NEXT UP                                                   ║
-╚══════════════════════════════════════════════════════════════╝
-
-**Audit milestone** — verify requirements, cross-phase integration, E2E flows
-
-`/pbr:milestone audit`
-
-<sub>`/clear` first → fresh context window</sub>
-
-
-
-**Also available:**
-- `/pbr:review` — manual acceptance testing
-- `/pbr:milestone complete` — archive milestone after audit passes
-
 
+**B. UI review suggestion (gated: UI files in SUMMARY key_files):**
+Collect all `key_files` values from every SUMMARY.md frontmatter in this phase.
+Check if any entry has extension `.tsx`, `.jsx`, `.css`, `.scss`, `.vue`, `.svelte`, or `.html` (case-insensitive).
+If YES:
 ```
-
-**If `gaps_found`:**
+**UI components detected:** Review visual output before continuing
+`/pbr:ui-review {N}`
 ```
-╔══════════════════════════════════════════════════════════════╗
-║  PLAN-BUILD-RUN ► PHASE {N} GAPS FOUND ⚠                     ║
-╚══════════════════════════════════════════════════════════════╝
-
-**Phase {N}: {Name}**
-
-Score: {X}/{Y} must-haves verified
-Report: .planning/phases/{phase_dir}/VERIFICATION.md
-
-### What's Missing
-
-{Extract gap summaries from VERIFICATION.md}
+If NO: skip silently — do not display an empty block.
 
+Both A and B are appended AFTER the primary next-action command. They never replace the primary routing.
 
+Include `<sub>/clear first → fresh context window</sub>` inside the Next Up routing block of the completion template.
 
-╔══════════════════════════════════════════════════════════════╗
-║  ▶ NEXT UP                                                   ║
-╚══════════════════════════════════════════════════════════════╝
-
-**Plan gap closure** — create additional plans to complete the phase
+**NEXT UP routing (when verification was skipped):**
 
-`/pbr:plan {N} --gaps`
-
-<sub>`/clear` first → fresh context window</sub>
+If verification was skipped and `workflow.validate_phase` is `true` (default), present:
 
+```
+**Run quality gate** to check test coverage gaps
 
+`/pbr:validate-phase {N}`
 
 **Also available:**
-- `cat .planning/phases/{phase_dir}/VERIFICATION.md` — see full report
-- `/pbr:review {N}` — manual testing before planning
-
-
+- `/pbr:review {N}` — skip validation, go straight to review
+- `/pbr:continue` — auto-route to next logical step
 ```
 
+If `workflow.validate_phase` is `false`, present the existing `/pbr:review {N}` suggestion instead.
+
 **8g. Display USER-SETUP.md (conditional):**
 
 Check if `.planning/phases/{NN}-{slug}/USER-SETUP.md` exists. If it does:
@@ -952,12 +1629,12 @@ If SUMMARY.md shows files not listed in the plan's `files_modified`:
 
 ### Build on wrong branch
 If `git.branching` is `phase` but we're not on the phase branch:
-- Create the phase branch: `git checkout -b plan-build-run/phase-{NN}-{name}`
+- Create the phase branch: `git switch -c pbr/phase-{NN}-{name}`
 - Proceed with build on the new branch
 
 ---
 
-## Files Created/Modified by /pbr:build
+## Files Created/Modified by /pbr:execute-phase
 
 | File | Purpose | When |
 |------|---------|------|
@@ -970,3 +1647,9 @@ If `git.branching` is `phase` but we're not on the phase branch:
 | `.planning/STATE.md` | Updated progress | Steps 6f, 8b |
 | `.planning/.auto-next` | Next command signal (if auto_continue enabled) | Step 8e |
 | Project source files | Actual code | Step 6 (executors) |
+
+---
+
+## Cleanup
+
+Delete `.planning/.active-skill` if it exists. This must happen on all paths (success, partial, and failure) before reporting results.