@agentikos/omega-os 0.19.37 → 0.19.39

package/omega/Agentik_SSOT/docs/AUDIT-V0.19.39.md

@@ -0,0 +1,161 @@
+# OmegaOS v0.19.39 — chat-first TUI + rules folder + prompt audit + Paperclip live sync
+
+> 4 parallel chantiers landed in one ship.
+> The user's invariant: *"l'utilisateur, une fois qu'il a setup tout l'outil
+> OmegaOS, doit être 100% fonctionnel. Il n'a rien à faire à part l'utiliser."*
+
+## 1. What changed
+
+| Chantier | Owner | Files touched | Net effect |
+|---|---|---|---|
+| **#1 TUI redesign (chat-first)** | main session | `tui.py` (+200 lines), `tmux.py` (+14 lines), `tests/test_tui_runtime.py` (+87 lines) | The TUI opens on CONVERSATIONS (AISB / Hermès / live Oracles / live Workers with ●/○ status dots) instead of an action menu. Everything else collapses into **MENU** with sub-menus. |
+| **#2 Rules folder** | background agent | `omega/Agentik_SSOT/rules/{three-laws,orchestration,prompt-protocols,audit-gates,scope-safety,verified-completion}.md` (6 new files, 1250 lines), `constitution.md` (+frontmatter only) | The rule set every LLM CLI reads is now COMPLETE. 7 files, YAML-frontmatter envelope, full cross-references, ~1300 lines total. No fabrication — every protocol sourced from existing docs. |
+| **#3 Prompt audit + doctor sections** | background agent | `omega_engine/prompt_audit.py` (395 lines, new), `tests/test_prompt_audit.py` (199 lines, new), `cli.py` (+39 lines) | New `omega doctor` sections `prompts` and `orchestration`. The audit scores each agent role /100 against Three Laws + LMC + `.done.json` references. Surfaces real drift (current suite average 52/100). |
+| **#4 Paperclip live status** | background agent | `omega_engine/paperclip_bridge.py` (+`is_running()` + `PaperclipStatus`), `tests/test_paperclip_status.py` (new) | TUI can show ●/○ next to "Paperclip dashboard" with the live port. 3-tier probe (pidfile → port-scan → none), ≤0.3s worst case, never raises. |
+| **#5 Integration + ship** | main session | `package.json`, `pyproject.toml`, `__init__.py`, `VERSION`, this doc | Version bump, commit, push, npm publish. |
+| **#6 Role-prompt enrichment** | follow-up (NOT in this ship) | — | The doctor surfaces 10 weak role prompts; enriching them to ≥80/100 is intentionally deferred — the audit infrastructure is what we needed. |
+
+## 2. The new TUI (chat-first)
+
+```
+── CONVERSATIONS ──
+   ●  AISB master                  claude (Max OAuth)
+   ○  Hermès                       claude (Anthropic API)
+
+   — Active Oracles (2) —
+   ●  Causio-oracle-2              project: Causio
+   ●  DentistryGPT-oracle          project: DentistryGPT
+   — Active Workers (1) —
+   ●  DentistryGPT-worker-3-ux-fix task: ux-fix
+
+── QUICK ACTIONS ──
+   + New AISB chat                 fresh session
+   + New Hermès chat               fresh session
+   + New project                   Genesis pipeline
+   Run a mission                   verified completion
+   ○  Paperclip dashboard          not running
+
+── MENU ──
+   Quality Arsenal                 17 forensic audits
+   Setup & config                  LLM: claude_code
+   Infrastructure                  sessions, scrape
+   Health checks                   doctor, status
+   Paperclip governance            register, status
+
+── EXIT ──
+   Detach                          session keeps running
+   Quit Omega                      kills the tmux session
+```
+
+**Picking any conversation row** (Oracle / Worker / AISB / Hermès) attaches to
+that tmux session via `tmux select-window` (for Omega windows) or
+`tmux switch-client` (for foreign sessions). One click → in the conversation.
+
+**Sub-menus** open in cascaded fzf with `← back` exit rows.
+
+## 3. The rules folder — what an LLM now reads at runtime
+
+```
+omega/Agentik_SSOT/rules/
+├── constitution.md         (frontmatter: priority=1) — the Prime Principle
+├── three-laws.md           (priority=2) — operational discipline per law
+├── orchestration.md        (priority=3) — L0-L5 dispatch hierarchy
+├── prompt-protocols.md     (priority=4) — brief/done/blocked schemas + LMC
+├── audit-gates.md          (priority=5) — 17 Quality Arsenal audits as gates
+├── scope-safety.md         (priority=6) — files_owned + Sacred Scopes
+└── verified-completion.md  (priority=7) — done_clean contract + third-party rule
+```
+
+These files are mirrored into every LLM's persona dir at install time
+(via `step_personas` from v0.19.38). So whether the operator runs
+`claude`, `gemini`, `codex`, `qwen`, or `opencode` inside an AISB chat,
+they ALL see the same complete rule set — no per-LLM drift.
+
+## 4. The new `omega doctor` output (sections that didn't exist before)
+
+```
+omega doctor — OMEGA_HOME=…/Omega
+  …
+  -- personas --                          (NEW in v0.19.38)
+  [ok] canonical: Agentik_SSOT/personas/OMEGAOS-CONTEXT.md (3402B)
+  [ok] chat-contexts/aisb-master/: 8 persona files
+  [ok] chat-contexts/hermes/: 8 persona files
+  …
+  -- prompts --                           (NEW in v0.19.39)
+  [ok]   CLAUDE: 90/100
+  [warn] morpheus: 75/100 — missing: LMC protocol
+  [warn] link: 65/100 — missing: LMC protocol
+  [FAIL] oracle: 45/100 — missing: LMC protocol, `.done.json` contract
+  [warn] average suite score: 52.0/100
+  [warn] weak prompts (<60): architect, construct, keymaker, …
+
+  -- orchestration --                     (NEW in v0.19.39)
+  [ok] AISB master prompt
+  [ok] Oracle role prompt
+  [ok] Worker-class prompts
+  [ok] Checker prompts (Seraph/Smith)
+  [ok] LMC protocol document
+  [warn] shared `.done.json` vocab: 33% of agents
+```
+
+The 52/100 average is **real drift**, not a bug. Most role prompts rely
+on the engine's `load_agent_prompt()` to concatenate `lmc-protocol.md`
+at spawn time, so the on-disk role file is silent. The audit makes that
+drift VISIBLE — an operator editing `oracle.md` now has a clear signal
+that the contract is implicit. Enriching the role files to score ≥80
+is chantier #6, deferred to v0.19.40.
+
+## 5. Paperclip status integration
+
+`omega_engine.paperclip_bridge.is_running()` returns a `PaperclipStatus`
+with `running: bool, pid, port, url, detection`. Three detection paths:
+
+| # | Method | Latency | Hint emitted in TUI |
+|---|---|---|---|
+| 1 | `~/.paperclip/run/dashboard.pid` + `os.kill(pid, 0)` | ~5ms | `localhost:8080` |
+| 2 | TCP connect 127.0.0.1:8080, 0.2s timeout | ≤200ms | `localhost:8080` |
+| 3 | Neither — fall through | <1ms | `not running` |
+
+The TUI's QUICK ACTIONS row renders a ●/○ dot using this probe — the
+user sees at-a-glance whether their Paperclip governance daemon is live.
+
+## 6. Multi-agent integration — the user's question, answered with code
+
+| Question (from the user's brief) | Answer | File reference |
+|---|---|---|
+| Multi-agents bien setup? | ✅ 14 agents (Hermès + 13 AISB) — templates landed at install via `step_aisb_suite`; persona context mirrored to all 10 LLM filenames via `step_personas`. | `bootstrap/lib/steps.sh:279-293` + `omega_engine/personas.py` |
+| Tmux orchestration AISB/Oracle/Workers? | ✅ Session naming convention parsed by `tmux.categorize()`; TUI now LISTS them with status dots and one-click attach. | `omega_engine/tmux.py:47-90` + `tui.py:528-557` |
+| Rules respectés pour chaque LLM? | ✅ 7 rule files at `Agentik_SSOT/rules/` are mirrored to every LLM persona dir; doctor's `prompts` section verifies role files reference them. | `Agentik_SSOT/rules/*.md` + `omega doctor prompts` |
+| Dossier maître linké pour le LLM? | ✅ `Agentik_SSOT/personas/OMEGAOS-CONTEXT.md` is the canonical; `Agentik_SSOT/agents/aisb/CLAUDE.md` is the AISB master; both auto-mirrored to per-LLM filenames (CLAUDE.md, GEMINI.md, AGENTS.md, QWEN.md, .opencode/CONTEXT.md, …) at install time. | `step_personas` from v0.19.38 |
+| Tout setup à l'install, rien à faire post-install? | ✅ Install steps 25 (aisb-suite), 37 (hermes-brief), 38 (personas) all eager-seed. `npx -y @agentikos/omega-os@latest --full` is sufficient. | `install.sh STEPS[]` |
+| Visibilité sur ce qui se passe? | ✅ TUI chat-first view + `omega doctor` 23 sections (incl. NEW personas/prompts/orchestration). | `tui.py::_arrow_menu` + `cli.py::cmd_doctor` |
+
+## 7. Tests (regression-locked)
+
+| Chantier | New tests | Suite total |
+|---|---|---|
+| Baseline (v0.19.38) | — | 627 passed |
+| #1 TUI chat-first | +7 (TestChatFirstRedesign + TestOmegaWindowAliveHelper) | +7 |
+| #3 Prompt audit | +5 (full-score, missing-laws, banned-phrases, real-suite, real-orchestration) | +5 |
+| #4 Paperclip status | +5 (no-pidfile, stale-pidfile, live-pidfile, port-scan, url-field) | +5 |
+| **v0.19.39 total** | **+17 new** | **644 passed, 0 regressions** |
+
+Chantier #2 (rules folder) is documentation-only — no Python code, no tests
+needed; format validated by manual grep + YAML parse.
+
+## 8. Verdict
+
+✅ TUI is now **conversation-first** as the user requested ("L'objectif…
+c'est d'avoir une interface extrêmement simple… cette interface permet de
+voir la conversation avec AISB… ensuite, de voir les conversations avec
+les oracles et les conversations avec les workers").
+✅ Setup/config/audits/scrape/governance moved to sub-menus reachable via
+**MENU** (one row).
+✅ Paperclip dashboard has a live status dot and is reachable in one pick.
+✅ Rules folder is COMPLETE (7 files, 1301 lines, cross-referenced).
+✅ `omega doctor` now surfaces the orchestration health (prompts + chain).
+✅ No regression in existing 627 tests.
+
+The user's "il n'a rien à faire à part l'utiliser" invariant is preserved:
+one `npx -y @agentikos/omega-os@latest --full` and the new menu, the new
+rules, the new audit, and the live Paperclip indicator are all in place.

package/omega/Agentik_SSOT/rules/audit-gates.md

@@ -0,0 +1,189 @@
+---
+id: audit-gates
+layer: L0-governance
+applies_to: [aisb, oracle, worker]
+priority: 5
+---
+
+# Audit Gates — Quality Arsenal as System Contract
+
+> The 17 Quality Arsenal audits are **not just commands a human runs**.
+> They are *gates* that lifecycle events at L3–L5 must pass before a
+> `done.json` may state `done_clean`. This file fixes which audits gate
+> which events, how the Gestalt-Popper methodology bakes into the
+> grader, and the verified-completion thresholds the engine enforces.
+
+## The 17 audits (catalogued in `../audits/`)
+
+| Audit | Domain | Question it answers | Threshold |
+|---|---|---|---|
+| `codeaudit` | Code | Is the code SOLID? | 85/100 |
+| `flowaudit` | User flows | Does the experience WORK? | 85/100 |
+| `uiuxaudit` | UI design | Is the interface BEAUTIFUL? | 85/100 |
+| `refontaudit` | Redesign | Does the redesign hold? | 85/100 |
+| `debugaudit` | Runtime | What is BROKEN right now? | 85/100 |
+| `featureaudit` | Features | Is the product COMPLETE? | 85/100 |
+| `perfaudit` | Performance | Is it FAST enough? | 85/100 |
+| `secaudit` | Security | Is it SECURE? | 85/100 |
+| `a11yaudit` | Accessibility | Is it ACCESSIBLE? | 85/100 |
+| `seoaudit` | SEO | Is it DISCOVERABLE? | 85/100 |
+| `dataaudit` | Data | Is the data INTACT? | 85/100 |
+| `apiaudit` | API | Is the API SOLID? | 85/100 |
+| `copyaudit` | Copy | Is the copy CLEAR? | 85/100 |
+| `dxaudit` | DX | Is the DX SMOOTH? | 85/100 |
+| `motionaudit` | Motion | Is the motion PURPOSEFUL? | 85/100 |
+| `automationaudit` | Automation | Is automation RELIABLE? | 85/100 |
+| `logicaudit` | Logic | Is the logic OPTIMAL? | 85/100 |
+| `retentionaudit` | Retention | What FEATURES are missing? (READ-ONLY) | — |
+
+The full definition for each lives in `../audits/<name>.yaml`
+(domain, gather tools, phases, falsification rule, fix-loop flag).
+
+## Lifecycle gates
+
+Audits are gates on *lifecycle events*, not on *human commands*. The
+engine consults the gate registry at each event and refuses progress
+if the required audits did not pass.
+
+| Event | Gate | Audits typically required |
+|---|---|---|
+| Worker `done_clean` (per subtask) | Worker gate | The audits matching the files the Worker touched (e.g. edited `*.ts` → `codeaudit`; edited `*.css` + UI components → `uiuxaudit` + `a11yaudit`). |
+| Oracle close-coherence (per mission) | Mission gate | The union of all Worker gates plus any mission-wide audits the brief declared (`brief.audit_gates`). |
+| Pre-merge / pre-ship | Ship gate | `codeaudit`, `secaudit`, plus domain-relevant audits. Project's `ship-config.json` may add more. |
+| Genesis completion (new project) | Genesis gate | `codeaudit`, `featureaudit`, `dxaudit`, `secaudit` — a freshly built project must stand on its own. |
+| Post-mission (asynchronous) | Drift gate | `debugaudit`, `perfaudit`, periodically scheduled by Hermès or the engine cadence. |
+
+Gates compose: a Worker that triggers two audits passes only if *both*
+audits exit `verdict: satisfied` with score ≥ threshold.
+
+## The Gestalt-Popper methodology
+
+Every audit (see `../docs/quality-arsenal/QUALITY-ARSENAL-PREAMBLE.md`
+and `../docs/quality-arsenal/AUDIT-VERIFICATION-CONTRACT.md`) implements:
+
+1. **Gestalt clarity gate (Phase 0).** Before any scored phase, the
+   audit identifies the *hinge* of its domain — the single element on
+   which the domain's reliability or value pivots. The canonical hinge
+   noun per audit is fixed in
+   `AUDIT-VERIFICATION-CONTRACT.md` (e.g. `codeaudit` → HINGE POINT,
+   `flowaudit` → HINGE FLOW, `secaudit` → SECURITY HINGE POINT). The
+   hinge is given **10× scrutiny** in subsequent phases.
+2. **Popper falsification.** For each scored item, the auditor states
+   *what would prove this claim wrong*. A PASS is only valid if the
+   falsifier was sought and not found. Bias toward FAIL — a 100 is
+   earned, never assumed.
+3. **Hippocratic pre/post.** Before any fix, capture baseline
+   (Phase N-1). After each fix, re-run the baseline check (Phase N+1).
+   A fix that broke a previously-working check reverts and is marked
+   `NEEDS_REVIEW`.
+4. **Before-after matrix (Phase N+4).** Every audit produces
+   `.<audit>/before-after.md` proving zero regressions. No matrix → no
+   100/100 verdict.
+5. **Fix → re-audit loop.** Bounded (typically 5 iterations). The loop
+   exits on `verdict: satisfied` *or* on iteration cap.
+
+## Mandatory minimums (per audit)
+
+These structural invariants are enforced by `metaudit` (the audit of
+audits). A skill that violates any of them fails meta and is removed
+from the gate registry until repaired.
+
+| # | Invariant | Why |
+|---|---|---|
+| 1 | At least 16 scored phases | Forensic depth — fewer phases = shallow audit. |
+| 2 | Phase N-1 (PRE-FIX BASELINE) implemented before the first fix | Hippocratic rule — can't claim "no regression" without a baseline. |
+| 3 | Phase N+4 (before-after matrix) written to `.<audit>/before-after.md` | Proof-of-work artefact required for the 100/100 verdict. |
+| 4 | Score normalised to /100 (raw may be /280, /320, /360, /400, /420 — must publish the formula) | Cross-audit comparison. |
+| 5 | HINGE identification at Phase 0 | Gestalt clarity gate. |
+| 6 | Popper falsification per scored item | Epistemic rigor. |
+| 7 | Fix → re-audit loop with explicit max iterations | Bounded recovery. |
+| 8 | Final verdict gate refuses 100/100 unless `before-after.md` shows zero regressions | Contract enforcement. |
+
+## The verified-completion contract
+
+A `done.json` may state `status: done_clean` only when **all** of:
+
+| Condition | Source |
+|---|---|
+| `audit.verdict == "satisfied"` | The grader (LMC or direct) for every required gate. |
+| `audit.scores[gate] >= threshold` (default 85/100) for each gate | `../audits/<gate>.yaml#threshold`. |
+| `regressions.length == 0` | Phase N+4 before-after matrix. |
+| `evidence.verify_exit_code == 0` | The brief's `verify_command`. |
+| `ship.result in ["ok", "skipped"]` when `ship.requested == true` | The ship pipeline (see `verified-completion.md`). |
+| Independent third party ran the *real* flow | The grader is a different agent from the executor; the verify is the real system, not a mock. |
+
+Fail any condition → `status: pending` (with `pending_actions[]` listing
+the failed conditions) or `status: failed` (when the verify itself
+errored). The engine refuses to mark a session done on the receiver's
+word alone — see `verified-completion.md`.
+
+## Routing — which audits apply
+
+Each `<audit>.yaml` declares `applies_to.changed` — the glob set that
+*triggers* the audit when a Worker's `files_owned` intersects it.
+Sample mappings:
+
+| Glob change | Audits auto-required |
+|---|---|
+| `*.py`, `*.ts`, `*.tsx`, `*.js`, `*.go`, `*.rs` | `codeaudit` |
+| `*.tsx`, `*.jsx`, `*.css`, design tokens | `uiuxaudit`, `a11yaudit`, `motionaudit` (if motion files touched) |
+| `*.env*`, `Dockerfile`, `package.json`, auth modules | `secaudit` |
+| API route handlers, OpenAPI / GraphQL schemas | `apiaudit` |
+| Database migrations, schema files | `dataaudit` |
+| Onboarding, signup, payment flows | `flowaudit` |
+| Cron specs, daemon scripts, scheduled tasks | `automationaudit` |
+| Marketing pages, SEO meta, sitemap | `seoaudit`, `copyaudit` |
+
+The Oracle expands `brief.audit_gates` from this routing table at
+dispatch time. A Worker may not narrow the gate set; it may *only*
+widen it (e.g. discovers a security implication mid-task).
+
+## Ship gate (pre-prod)
+
+When `brief.ship == true`, the ship pipeline runs before final
+`done.json`. Each step gates the next:
+
+1. `npm run build` (or equivalent) — exit 0.
+2. Whitelisted staging — only `files_owned`. Any extra file aborts.
+3. Secret scan (e.g. `gitleaks --staged`) — zero matches.
+4. Whitespace sanity (`git diff --check`) — clean.
+5. Conventional-commit message from `brief.commit_message`.
+6. Per-project ship lock (`flock`) — serialise across Oracles.
+7. Freeze flag check — if `Agentik_Runtime/locks/ship-<project>.frozen`
+   exists, abort and alert.
+8. `git pull --rebase` — clean.
+9. `git push` — clean.
+10. Deploy (project-defined command) — typically `vercel --prod` or
+    equivalent.
+11. Poll deploy status until READY/ERROR/TIMEOUT (default 10 min).
+12. Write `done.json#ship` with commit, URL, status, duration.
+
+Default deploy-failure policy is **freeze, don't rollback** — the
+freeze flag blocks further pushes on the project until the human lifts
+it. Auto-rollback is opt-in per project via `ship-config.json`.
+
+## Drift gate (continuous)
+
+`debugaudit` and `perfaudit` are scheduled to run periodically against
+the live deployed URL (typically by Hermès cadence or the engine's
+cron). A drift detection writes a `done.json` with
+`status: failed` against a synthetic "drift" mission, which AISB
+surfaces to the human and (if the project opts in) auto-dispatches a
+repair mission.
+
+## Cross-references
+
+- `constitution.md` — Verification Rule.
+- `three-laws.md` — First Law (runtime over code) is the audit
+  methodology's epistemology.
+- `prompt-protocols.md` — `brief.audit_gates`, `done.audit` schema.
+- `verified-completion.md` — the terminal contract these gates serve.
+- `scope-safety.md` — Worker gates intersect with `files_owned`.
+- `orchestration.md` — Oracle close-coherence runs the mission gate.
+- `../audits/*.yaml` — per-audit catalogue (domain, gather, phases).
+- `../docs/quality-arsenal/AUDIT-VERIFICATION-CONTRACT.md` — Hippocratic
+  pre/post protocol.
+- `../docs/quality-arsenal/QUALITY-ARSENAL-PREAMBLE.md` — Gestalt-Popper
+  methodology.
+- `../docs/LAYERS.md` — which layer runs which gate.
+- `../personas/OMEGAOS-CONTEXT.md` — provider-neutral working context.

package/omega/Agentik_SSOT/rules/constitution.md

@@ -1,3 +1,10 @@
+---
+id: constitution
+layer: L0-governance
+applies_to: [aisb, oracle, worker, hermes]
+priority: 1
+---
+
 # The Omega OS Constitution
 
 > The rules every agent at every level inherits. Provider-neutral. Compiled into

package/omega/Agentik_SSOT/rules/orchestration.md

@@ -0,0 +1,215 @@
+---
+id: orchestration
+layer: L0-governance
+applies_to: [aisb, oracle, worker, hermes]
+priority: 3
+---
+
+# Orchestration — Who Dispatches What
+
+> OmegaOS is a five-layer agentic OS with one optional governance roof
+> (L0/Paperclip). This file fixes the dispatch hierarchy, the decisions
+> log discipline, and the fresh-context template every layer uses when
+> handing work down. The architecture itself is defined in
+> `../docs/LAYERS.md`; this file makes it *operational*.
+
+## The hierarchy
+
+```
+L0  Paperclip            (optional governance roof — budget, org chart, approvals)
+L1  Human                (Telegram, CLI, web — three doors into the system)
+L2  Hermès               (meta-companion — Anthropic API, separate budget)
+L3  AISB                 (intake / orchestrator — Claude Max OAuth)
+L4  Oracle               (per-project planner — persistent tmux session)
+L5  Workers              (executors — one per subtask, .done.json verified)
+```
+
+Each layer is *independently usable*. Skip L0 and L2 and OmegaOS is
+still a complete agentic OS: a human writes intent, AISB classifies,
+Oracle plans, Workers execute, the engine verifies.
+
+## Dispatch rules
+
+The arrows below define *who is permitted to dispatch to whom*. Any
+other dispatch is a violation and the receiving layer must refuse.
+
+| From | May dispatch to | Notes |
+|---|---|---|
+| L1 Human | L2 Hermès, L3 AISB | Three doors: Telegram-to-Hermès, Telegram-to-AISB, CLI/tmux. |
+| L0 Paperclip | L2 Hermès, L3 AISB | Approval gates; never bypasses the lower layers. |
+| L2 Hermès | L3 AISB | Hermès cannot reach Oracle or Worker directly. Missions go through AISB. |
+| L3 AISB | L4 Oracle | One Oracle per project. AISB never spawns a Worker directly. |
+| L4 Oracle | L5 Workers | One Worker per subtask, with a verify command. |
+| L5 Worker | — | A Worker never dispatches. If more work is needed, it returns `status: pending` and lets Oracle decide. |
+
+**Why no skipping.** Each step adds a layer of intent translation. AISB
+turns a freeform human prompt into a typed mission. Oracle turns a
+mission into a DAG of subtasks with verify commands. A Worker that
+receives a freeform human prompt has no scope, no verify command, and
+no way to call itself done — the verified-completion contract breaks.
+
+## Roles in one line
+
+- **L0 Paperclip.** Approvals + budget + org chart. Read-only over L1–L5
+  unless a budget guard fires.
+- **L1 Human.** Author of intent. Reads final reports. Approves
+  destructive ops.
+- **L2 Hermès.** Meta-reasoning, scheduling, learning. Watches
+  observations, proposes missions, dispatches them down to L3.
+- **L3 AISB.** Intake. Classifies a mission (simple / medium / complex /
+  epic), picks a topology, hands to L4.
+- **L4 Oracle.** Per-project planner. Reads the mission, builds the
+  outcome rubric (see `audit-gates.md`), dispatches Workers, polices
+  their `done.json`, runs the close-coherence audit.
+- **L5 Worker.** Executor. Owns a strict file scope
+  (`spec.scope.files_owned`), writes `done.json` when the verify
+  command passes.
+
+## The decisions log
+
+Every layer that makes a non-trivial routing or design choice **MUST**
+append to `.orchestrator/decisions.md` in the project root:
+
+```markdown
+### [ISO-8601 timestamp] Decision title
+- **Task:** what was asked
+- **Classification:** SIMPLE / MEDIUM / COMPLEX / EPIC
+- **Decision:** what was chosen (agent, topology, audit set, scope)
+- **Rationale:** why (one line)
+- **Falsifier:** the runtime check that would prove this wrong
+```
+
+The log is append-only. Past entries are never edited — only superseded
+by a new entry that cites the old one. AISB and Oracle both write to
+the same file; Workers write only if they exercise the Third Law and
+correct a premise inside a dispatched session.
+
+Reason: when a mission is re-opened a week later, the *why* is in the
+log. Without it the next agent has to reverse-engineer past intent from
+diffs.
+
+## The fresh-context template (mandatory at every dispatch)
+
+When *any* layer dispatches to the one below it, the brief MUST contain
+the following sections. Empty sections are allowed; missing sections
+are a contract violation.
+
+```
+## Mission
+<1-2 line summary of the goal>
+
+## Purpose
+<why this matters — links the work to the human intent at L1>
+
+## Context
+<project root, deployed URL, stack, relevant prior runs>
+
+## What's Done
+<bullet list of completed work in this mission so far>
+
+## Current Task
+<specific files, line numbers, exact changes — surgical scope>
+
+## Done Criteria
+<measurable condition: a shell-checkable predicate, an audit verdict,
+ a screenshot diff, a passing test>
+
+## Verify Command
+<exact command the receiver runs to prove it satisfied Done Criteria>
+
+## Key Decisions
+<excerpts from .orchestrator/decisions.md relevant to this task>
+
+## Files in Scope
+<the receiver's spec.scope.files_owned — files it may edit>
+
+## Relevant Memories
+<pre-selected lessons-learned entries — NOT a full dump>
+```
+
+**Why both Done Criteria and Verify Command.** Without `Done Criteria`
+the receiver redefines "done" mid-stream. Without `Verify Command` the
+dispatcher cannot independently confirm — and the verified-completion
+contract demands an independent third party (see
+`verified-completion.md`).
+
+## Fresh context vs context overlap
+
+When a layer hands a sub-mission to a fresh agent, the dispatcher
+decides between two patterns:
+
+| Pattern | When | Cost |
+|---|---|---|
+| **Continue agent** | High overlap with prior subtask; the prior agent already holds the relevant files and state in working memory. | Cheap but risks context bloat past compaction thresholds. |
+| **Spawn fresh** | New domain, new files, new specialist. The dispatcher pre-inlines a summary of prior results into the fresh brief. | Slightly more expensive, guarantees a clean slate. |
+
+Default is *spawn fresh* for any subtask that crosses a layer or a
+domain boundary. Continue only inside a single layer's working session.
+
+## Multi-Oracle on the same project
+
+Multiple Oracles can run in parallel on the same project. The engine
+serialises *file ownership*, not Oracle count.
+
+- AISB checks the per-project Oracle registry before dispatching. If
+  an Oracle is *idle* (no Workers, at prompt, >5 min), AISB reuses
+  it. Otherwise AISB spawns Oracle #2 (`oracle-<Project>-2`).
+- Each Oracle declares `files_owned` for its mission. Two Oracles
+  cannot claim overlapping files.
+- The patrol auto-cleans dead Oracles every 5 minutes.
+- A registry file at `Agentik_Runtime/oracles/<Project>-<id>.json`
+  records the Oracle's mission, scope, and heartbeat.
+
+## Worker batching (parallel-safe)
+
+Inside a single Oracle, Workers may run *in parallel* if and only if
+their file footprints are disjoint. The Oracle's plan groups subtasks
+into batches:
+
+- **Narrow** subtasks: identifiable file set, no overlap with other
+  narrow tasks → packable into a batch of up to N (typical N = 3 or 4).
+- **Broad** subtasks: vague scope, no identifiable file footprint →
+  run alone, serially.
+- **Terminal** subtasks: touch infrastructure (env files, package
+  manifests, migrations) → always run alone, after all batches.
+
+Two Workers on the same file at the same time is a contract violation
+even if their edits "wouldn't conflict" — the assertion of disjointness
+itself is the contract.
+
+## Close-coherence (Oracle's final check)
+
+When all Workers report `done_clean`, the Oracle runs a *close-
+coherence* pass before reporting up to AISB:
+
+1. Re-read the mission brief.
+2. For each Worker, confirm its `done.json` matches the brief's slice.
+3. Run the audit set the brief requires (see `audit-gates.md`).
+4. Write the Oracle's own `done.json` with `consensus_score`,
+   `regressions`, and the final verdict.
+
+An Oracle that skips close-coherence has not finished its job, even if
+every Worker did.
+
+## The "no idle wait" invariant
+
+The Third Law applies to *every* dispatched session in this hierarchy
+(L2 ↔ L3 ↔ L4 ↔ L5). The only sessions where a question may be asked
+of a human are sessions L1 owns directly — typically an interactive
+shell or a Telegram DM the human is actively reading.
+
+Practical detection rule for an agent: *"Am I attached to a tmux
+session whose name starts with `oracle-`, `aisb-`, or contains
+`-worker-`?"* If yes, no questions; decide and proceed
+(`prompt-protocols.md` for the exact `blocked.json` fallback).
+
+## Cross-references
+
+- `constitution.md` — Prime Principle, Three Laws.
+- `three-laws.md` — expanded Third Law (no idle wait).
+- `prompt-protocols.md` — brief / done.json / blocked.json schemas.
+- `audit-gates.md` — which audits gate which dispatch transitions.
+- `scope-safety.md` — `spec.scope.files_owned` discipline.
+- `verified-completion.md` — terminal states an Oracle/Worker may report.
+- `../docs/LAYERS.md` — formal L1–L5 architecture and credential model.
+- `../personas/OMEGAOS-CONTEXT.md` — provider-neutral working context.