feat(bstack): doctor + frontmatter + primitives.md + templates for P13

SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: bstack
 description: |
-  The Broomva Stack — twelve irreducible primitives (P1–P12) that turn any
+  The Broomva Stack — thirteen irreducible primitives (P1–P13) that turn any
   agent-driven workspace into a self-operating system, plus 28 curated skills
   that ship with the stack. The primitives are not optional features; they are
   the substrate. P1 captures every session as episodic memory. P2 gates
@@ -25,7 +25,7 @@ description: |
 
 # bstack — The Broomva Stack
 
-**Twelve irreducible primitives. Twenty-nine curated skills. One self-operating workspace.**
+**Thirteen irreducible primitives. Twenty-nine curated skills. One self-operating workspace.**
 
 bstack is a *portable harness metalayer* — it composes existing skills into a binding primitive contract that the agent enforces by reasoning, the doctor enforces by checking, and the bootstrap enforces by scaffolding.
 
@@ -64,6 +64,7 @@ The twelve primitives. Each closes one specific failure mode that drifts into en
 | **P10** | Worktree Hygiene Discipline | dirty-tree drift across the PR lifecycle |
 | **P11** | Empirical Feedback Loop | shipping code that compiles but doesn't work |
 | **P12** | Persistent Loop Discipline (`broomva/persist` skill) | long-horizon work decaying as the context window rots |
+| **P13** | Dream Cycle Discipline | tier-crossing consolidation corrupting upper-tier rules without replay (the *shadow dream* failure mode) |
 
 Full reference: see [references/primitives.md](references/primitives.md).
 
@@ -138,9 +139,9 @@ Report results. If any checks fail, fix them before proceeding.
 `scripts/doctor.sh`. Seven check sections:
 
 1. Governance files exist (CLAUDE.md, AGENTS.md, .control/policy.yaml)
-2. CLAUDE.md primitives table has all P1–P12 rows + correct count header
+2. CLAUDE.md primitives table has all P1–P13 rows + correct count header
 3. AGENTS.md has each primitive section (`### P1:` through `### P11:`)
-4. Reflexive Trigger Rules present for P6, P7, P10, P11, P12 (the reasoning-enforced primitives)
+4. Reflexive Trigger Rules present for P6, P7, P10, P11, P12, P13 (the reasoning-enforced primitives)
 5. `.control/policy.yaml` has required blocks (`ci_watch:`, `ci_heal:`, `auto_merge:`)
 6. `.claude/settings.json` wires the expected hook scripts (P1, P2, P8)
 7. Each primitive's mechanism is reachable on disk
@@ -214,7 +215,7 @@ This is the f₃ dynamics function at L3 of the RCS hierarchy. See [references/p
 
 ## See also
 
-- [references/primitives.md](references/primitives.md) — full P1–P12 reference with reflexive triggers
+- [references/primitives.md](references/primitives.md) — full P1–P13 reference with reflexive triggers
 - [references/skills-roster.md](references/skills-roster.md) — all 28 skills with install commands
 - [references/stack-architecture.md](references/stack-architecture.md) — layer dependency diagram
 - [references/quickstart.md](references/quickstart.md) — 5-minute install walkthrough

assets/templates/AGENTS.md.template

@@ -134,19 +134,55 @@ The agent picks the right subset, runs as parallel watchers, and **captures evid
 5. When CI or tests fail — capture full context first before attempting a fix.
 6. At session end — produce a *dogfood receipt*.
 
+### P12: Persistent Loop Discipline (Cross-Context Restart Loop)
+
+**What**: Reflexive discipline binding every agent to *restart the context window when it rots*, while preserving state in the filesystem. Long-horizon work (>1h, the METR 80%-reliability ceiling) decays inside a single conversation as the context window passes ~100K tokens.
+
+**How**: `python3 skills/persist/scripts/persist.py iterate <PROMPT.md>` substrate. Each iteration spawns a fresh agent context; state persists in PROMPT.md + git tree + state.jsonl. Validation backpressure from compilers/tests/linters, not model self-grading.
+
+**Invariant**: state lives in the filesystem. Each iteration is a fresh subprocess. Budget: default 50 iterations / 14400s wall-clock.
+
+**Reflexive Trigger Rule**: P12 is a reflex. Apply without being prompted:
+
+1. Before any work that may exceed ~1h of unsupervised agent time — write PROMPT.md, call `persist iterate`. Don't try >1h work in-context.
+2. When session token usage crosses ~100K — restart, don't continue in the rotted context.
+3. When the same fix has been attempted ≥3 times without convergence — stop in-context; spawn fresh persist loop.
+4. When orchestrating long-horizon work — default to persist + periodic checkpoints; compose with P5 worktrees.
+5. When the user says "run this in the background for an hour" — that's persist territory.
+
+### P13: Dream Cycle Discipline (Tier-Crossing Consolidation)
+
+**What**: Reflexive discipline binding every agent to apply the **5-phase dream shape** (*gather → replay → prune → consolidate → index*) for any consolidation that crosses a cadence-tier boundary. Closes the *shadow dream* corruption mode.
+
+**How**: Reasoning-enforced. Composes with P6 (`bookkeeping replay` is the canonical reference instance) and future Life primitives (askesis T1→T2, anamnesis T2→T3, when shipped).
+
+**Invariant**: any consolidation that crosses a cadence-tier boundary MUST replay against a frozen substrate before committing. Without replay, dense lower-tier signal corrupts sparse upper-tier rules.
+
+**Reflexive Trigger Rule**: P13 is a reflex. Apply without being prompted:
+
+1. Before any consolidation that promotes lower-tier signal to upper-tier rules — verify the consolidation primitive has a replay phase.
+2. For knowledge-graph promotion — use `bookkeeping replay` (not `bookkeeping run`) for substantial promotion runs.
+3. For governance changes (L3 tier) — every PR is a dream cycle: gather (PR description), replay (worktree + CI + doctor), prune (CI failures), consolidate (squash merge), index (commit history).
+4. When designing a NEW consolidation primitive — implement the 5-phase shape from day 1; don't ship shadow-dream form.
+5. When you observe a new dream instance shipping — record it for the rule-of-three counter.
+
+The morpheus crate (shared abstraction across implementations) is deferred per rule-of-three until ≥2 dream instances ship beyond P6.
+
 ---
 
-These eleven primitives compose into the full autonomous development loop:
+These thirteen primitives compose into the full autonomous development loop:
 
 ```
 User intent → Linear ticket (P3) → Agent dispatched (P5)
   → Prior context loaded (P1) [+ P8 freshness check] [+ P10 cleanup audit]
   → Safety gates active (P2)
   → P10 worktree decision → P11 validation plan
+  → IF long-horizon → P12 persist loop with PROMPT.md + budget
   → Code written + parallel watchers (P11 log-tails) → PR created (P4)
   → CI watched + heal loop (P7)
   → P11 deploy verification (preview URL, screenshots, browser session)
   → Merge → P10 post-merge cleanup via P9 janitor → Deploy
+  → P13 dream cycle for any consolidation (P6 replay first; future Life dreams compose here)
   → P11 dogfood receipt → Session captured (P1) → Knowledge bookkept (P6)
   → System improved (EGRI)
 ```

assets/templates/CLAUDE.md.template

@@ -6,7 +6,7 @@ This workspace is governed by **bstack** — eleven irreducible primitives (P1
 
 ## Bstack Core Automation Primitives
 
-Eleven irreducible building blocks that make this workspace self-operating. All are always active. Full specification in `AGENTS.md`.
+Thirteen irreducible building blocks that make this workspace self-operating. All are always active. Full specification in `AGENTS.md`.
 
 | # | Primitive | Mechanism | Invariant |
 |---|-----------|-----------|-----------|
@@ -21,6 +21,8 @@ Eleven irreducible building blocks that make this workspace self-operating. All
 | P9 | Branch + Worktree Janitor | `make janitor` → detects squash-merged branches + dead worktrees, removes safely | Default `--dry-run`; never touches protected branches |
 | P10 | Worktree Hygiene Discipline | Reflexive rule: decide worktree-or-not before first file; keep `git status` clean; auto-run P9 janitor after every merge | A clean tree is the only reliable reset point |
 | P11 | Empirical Feedback Loop | Reflexive rule: validate by *interacting* — log-tails, browser E2E, screenshots, deploy verification, multi-level test composition | Reasoning isn't validation; interaction is |
+| P12 | Persistent Loop Discipline (`broomva/persist` skill) | Reflexive rule: cross-context restart loop — state in filesystem (PROMPT.md + git tree), each iteration spawns a fresh agent context | At long-horizon work (>1h), in-context loops decay; restart fresh, backpressure from compilers/tests |
+| P13 | Dream Cycle Discipline | Reflexive rule: any consolidation that crosses a cadence-tier boundary MUST follow the 5-phase shape (gather → replay → prune → consolidate → index) | Replay against frozen substrate is the runtime form of stop-gradient; without it, dense lower-tier signal corrupts sparse upper-tier rules |
 
 > **Naming note.** Skill names are historical and do not always match primitive numbers. P6's skill repo is `broomva/bookkeeping`. P7's skill repo is `broomva/p9` (named when it was the ninth primitive; renaming would break every `npx skills add` install). Primitive numbers are sequential identifiers in the bstack itself; skills are independent npm-style packages with stable names. `bstack doctor` checks AGENTS.md compliance with this and all other primitives.
 

references/primitives.md

@@ -15,6 +15,8 @@ The eleven primitives that make a workspace self-operating. This is the canonica
 - [P9 — Branch + Worktree Janitor](#p9--branch--worktree-janitor)
 - [P10 — Worktree Hygiene Discipline](#p10--worktree-hygiene-discipline)
 - [P11 — Empirical Feedback Loop](#p11--empirical-feedback-loop)
+- [P12 — Persistent Loop Discipline](#p12--persistent-loop-discipline)
+- [P13 — Dream Cycle Discipline](#p13--dream-cycle-discipline)
 - [Cohesion narrative](#cohesion-narrative)
 - [RCS L3 stability constraint](#rcs-l3-stability-constraint)
 
@@ -202,28 +204,102 @@ Mental checklist: *Did I interact with it? Did I capture evidence? Was the evide
 
 ---
 
+## P12 — Persistent Loop Discipline
+
+**Closes**: long-horizon work decaying as the context window rots past ~100K tokens (the *"Dumb Zone"*). METR's Time Horizon 1.1 (Jan 2026) puts the **80%-reliability deployable horizon at ~1h on Opus 4.6** — a 14× reliability gap vs the 14.5h 50%-horizon. Above 1h, in-context loops fail.
+
+**Skill name note**: P12's skill repo is `broomva/persist` — non-anthropomorphized rename of the pattern Geoffrey Huntley popularized as the "Ralph loop" (Jan 2026).
+
+**How**: `python3 skills/persist/scripts/persist.py iterate <PROMPT.md>` substrate. Each iteration spawns a fresh agent context. State persists in the filesystem (PROMPT.md + git tree + state.jsonl). Validation backpressure from compilers/tests/linters, not model self-grading. Five-state machine: `SPAWNED → ITERATING (self-loop) → SUCCESS | BUDGET_EXHAUSTED | ABANDONED`. Default budget: 50 iterations / 14400s wall-clock (METR's 80%-horizon ceiling).
+
+**Invariant**: state lives in the filesystem. Each iteration starts from PROMPT.md content, not conversation history. Validation backpressure is external. Each iteration is a fresh subprocess.
+
+### P12 Reflexive Trigger Rule (binding on every agent)
+
+P12 is a reflex, not a request. Apply without being prompted:
+
+1. Before any work that may exceed ~1h of unsupervised agent time — write PROMPT.md, call `persist iterate`. Don't try >1h work in-context.
+2. When session token usage crosses ~100K — restart, don't continue in the rotted context.
+3. When the same fix has been attempted ≥3 times without convergence — stop in-context; spawn fresh persist loop.
+4. When orchestrating long-horizon work — default to persist + periodic checkpoints; compose with P5 (one persist loop per worktree) and P7 (each iteration's PR uses `p9 watch`).
+5. When the user says "run this in the background for an hour" — that's persist territory.
+
+---
+
+## P13 — Dream Cycle Discipline
+
+**Closes**: the *shadow dream* corruption mode — consolidation runs that gather + consolidate + index without the **replay** phase. Without replay, dense lower-tier signal corrupts sparse upper-tier rules. Pattern documented in `research/entities/concept/multi-tier-dreaming.md` (scored 9/9, promoted 2026-04-30).
+
+**How**: Reasoning-enforced. P13 has no dedicated substrate skill — it composes with primitives that already implement the dream shape:
+
+| Tier crossing | Implementation | Status |
+|---|---|---|
+| Knowledge graph (raw → promoted entities) | P6 with `bookkeeping replay` | **Reference instance — shipped 2026-05-06** |
+| Agent traces → plans (T0→T1) | Life autonomic compression | shipped (eager / shadow form) |
+| Trace bundles → prompt/tool diffs (T1→T2) | Life askesis | designed, not yet shipped |
+| Diffs → governance amendments (T2→T3) | Life anamnesis | proposed, not yet shipped |
+
+The 5-phase canonical shape:
+
+| Phase | Function |
+|---|---|
+| **Gather** | Collect a bounded bundle of dense lower-tier signal as a frozen, addressable artifact. |
+| **Replay** | Re-execute the bundle against a *frozen substrate* — sandbox, world model, retrieval cache. |
+| **Prune** | Reject replayed signal that fails the gate: no improvement, schema violation, regression. |
+| **Consolidate** | Commit the kept signal as a sparse, structured update to the upper-tier substrate. Atomic and versioned. |
+| **Index** | Re-validate the upper-tier resource graph: reference integrity, contradiction detection, garbage-collect orphans. |
+
+Three independent observations converge on this shape — biological REM sleep, Anthropic's `/dream` skill, Physical Intelligence's knowledge-insulation training (Driess et al. 2025). The replay phase is the runtime form of stop-gradient.
+
+**Invariant**: any agent-driven consolidation that crosses a cadence-tier boundary MUST replay against a frozen substrate before committing. If a consolidation primitive doesn't have a replay phase, it's a *shadow dream* and is unsafe — the agent's job is to either (a) use the dream-cycle form, or (b) explicitly justify why this consolidation is single-tier and doesn't need replay.
+
+**Deferred per rule-of-three**: the `morpheus` crate (shared abstraction across implementations). Extract only when ≥2 dream instances ship end-to-end beyond P6+replay. Currently at 1.
+
+### P13 Reflexive Trigger Rule (binding on every agent)
+
+P13 is a reflex, not a request. Apply without being prompted:
+
+1. Before any consolidation that promotes lower-tier signal to upper-tier rules — verify the consolidation primitive has a replay phase. If not, request the replay-extension before consolidating, or document why this case is single-tier.
+2. For knowledge-graph promotion — use `bookkeeping replay` (not `bookkeeping run`) for substantial promotion runs.
+3. For governance changes (L3 tier) — every PR in this workspace is a dream cycle: gather (PR description), replay (worktree + CI + doctor), prune (CI failures, doctor gaps), consolidate (squash merge), index (commit history).
+4. When designing a NEW consolidation primitive — implement the 5-phase shape from day 1; don't ship shadow-dream form.
+5. When you observe a new dream instance shipping — record it in `multi-tier-dreaming.md` (the rule-of-three counter for morpheus extraction).
+
+---
+
 ## Cohesion narrative
 
-P11 is the **cohesion glue** for the whole stack:
+P11, P12, and P13 are structural siblings at different scales:
+
+| Primitive | Discipline | Surface | Evidence | Scale |
+|---|---|---|---|---|
+| **P11** Empirical Feedback | "validate by interacting" | live deployed system | screenshots, logs, browser session | in-session (≤1h) |
+| **P12** Persistent Loop | "restart fresh when context rots" | filesystem (PROMPT.md + git) | state.jsonl + each iteration's evidence | cross-session (>1h) |
+| **P13** Dream Cycle | "consolidate by replaying" | frozen substrate | diff against frozen snapshot | tier-crossing |
+
+The whole stack composes:
 
-- **P4** (PR Pipeline) and **P7** (CI Watcher) catch what CI sees; **P11** catches what CI can't.
-- **P10** (Worktree Hygiene) keeps the working tree clean enough for empirical checks to be meaningful.
-- **P6** (Bookkeeping) records the validation evidence as durable context.
-- **P1** (Conversation Bridge) preserves the dogfood receipt across sessions.
-- **P8** (Skill Freshness) ensures the validation tools the agent reaches for (gstack, agent-browser, dogfood, qa) are themselves current.
-- **P9** (Janitor) ensures cleanup state is automatic so the next P10/P11 cycle starts from zero.
+- **P4** (PR Pipeline) and **P7** (CI Watcher) catch what CI sees; **P11** catches what CI can't; **P13** catches what consolidation without replay can't.
+- **P10** (Worktree Hygiene) keeps the working tree clean enough for empirical checks to be meaningful — same shape as P13's "frozen substrate" requirement at the knowledge layer.
+- **P6** (Bookkeeping) is the first concrete implementation of P13's discipline — `bookkeeping replay` is the canonical reference dream cycle.
+- **P12** (Persist) is the substrate for long-horizon work that needs P11/P13 discipline across many iterations.
+- **P1** (Conversation Bridge) preserves dogfood receipts and dream-cycle audit trails across sessions.
+- **P8** (Skill Freshness) ensures the validation/replay tools (gstack, bookkeeping, persist) are themselves current.
+- **P9** (Janitor) ensures cleanup state is automatic so the next cycle starts from zero.
 
-The eleven primitives compose into the full autonomous development loop:
+The thirteen primitives compose into the full autonomous development loop:
 
 ```
 User intent → Linear ticket (P3) → Agent dispatched (P5)
   → Prior context loaded (P1) [+ P8 freshness check] [+ P10 cleanup audit]
   → Safety gates active (P2)
   → P10 worktree decision → P11 validation plan
+  → IF long-horizon → P12 persist loop with PROMPT.md + budget
   → Code written + parallel watchers (P11 log-tails) → PR created (P4)
   → CI watched + heal loop (P7)
   → P11 deploy verification (preview URL, screenshots, browser session)
   → Merge → P10 post-merge cleanup via P9 janitor → Deploy
+  → P13 dream cycle for any consolidation (P6 replay first; future Life dreams compose here)
   → P11 dogfood receipt → Session captured (P1) → Knowledge bookkept (P6)
   → System improved (EGRI)
 ```