arkaos 5.16.1 → 5.17.1

package/README.md

@@ -4,7 +4,7 @@
 
 **The Operating System for AI Agent Teams.**
 
-89 agents. 17 departments. 340 skills. Enterprise frameworks. Multi-runtime. One install.
+89 agents. 17 departments. 341 skills. Enterprise frameworks. Multi-runtime. One install.
 
 ```bash
 npx arkaos install
@@ -97,7 +97,7 @@ npx arkaos doctor    # Health check
 
 ### Skill packs, à la carte
 
-The default install ships a curated core so your context window stays lean. Everything else lives in the ArkaOS plugin marketplace: 16 department packs with 266 skills, generated straight from the same sources the core uses. Inside Claude Code:
+The default install ships a curated core so your context window stays lean. Everything else lives in the ArkaOS plugin marketplace: 16 department packs with 267 skills, generated straight from the same sources the core uses. Inside Claude Code:
 
 ```
 /plugin marketplace add andreagroferreira/arka-os

package/THE-ARKAOS-GUIDE.md

@@ -1,6 +1,6 @@
 # The ArkaOS Guide
 
-> v5.16.1 — 89 agents, 17 departments, 340 skills, 306 commands, 20 ADRs.
+> v5.17.1 — 89 agents, 17 departments, 341 skills, 307 commands, 20 ADRs.
 > One file, everything you need to start. Generated by `scripts/guide_gen.py` — never hand-edited.
 
 ## What it is
@@ -40,7 +40,7 @@ answers as a generic assistant.
 | `/arka` | 22 |
 | `/brand` | 14 |
 | `/community` | 15 |
-| `/content` | 20 |
+| `/content` | 21 |
 | `/dev` | 42 |
 | `/do` | 1 |
 | `/ecom` | 18 |

package/VERSION

@@ -1 +1 @@
-5.16.1
+5.17.1

package/arka/SKILL.md

@@ -63,7 +63,7 @@ complementary, not alternatives.
 # ArkaOS — Main Orchestrator
 
 > **The Operating System for AI Agent Teams**
-> 89 agents. 17 departments. 340 skills. Multi-runtime. Dashboard. Knowledge RAG.
+> 89 agents. 17 departments. 341 skills. Multi-runtime. Dashboard. Knowledge RAG.
 
 ## ⛔ Evidence flow — 4 gates (NON-NEGOTIABLE)
 
@@ -207,7 +207,7 @@ violation (squad-routing, arka-supremacy, spec-driven, mandatory-qa).
 
 | Command | Description |
 |---------|-------------|
-| `/arka status` | System status (version, departments, agents, active projects). Includes **LLM costs (24h)** section: top-line cost + cache hit rate + call count from `core.runtime.llm_cost_telemetry.summarise(period="today")`. Also includes **Enforcement (24h)** section: total calls, block rate, top blocked tools/reasons from `core.governance.enforcement_telemetry.summarise(period="today")` (PR19 v2.41.0). Plus **Reorganization (today)** section: today's proposal path + artifact count from `core.cognition.reorganizer_scheduler.status_summary()` (PR24 v2.46.0). Plus **Model routing** section: gateway live/off + resolved per-slot routes + served counts from `core.runtime.model_routing_check.status_summary()`. Plus **MCP usage (24h)** section: total calls + servers in use + top servers from `core.runtime.mcp_telemetry.summarise(period="today")`. |
+| `/arka status` | System status (version, departments, agents, active projects). Includes **LLM costs (24h)** section: top-line cost + cache hit rate + call count from `core.runtime.llm_cost_telemetry.summarise(period="today")`. Also includes **Enforcement (24h)** section: total calls, block rate, top blocked tools/reasons from `core.governance.enforcement_telemetry.summarise(period="today")` (PR19 v2.41.0). Plus **Reorganization (today)** section: today's proposal path + artifact count from `core.cognition.reorganizer_scheduler.status_summary()` (PR24 v2.46.0). Plus **Model routing** section: gateway live/off + resolved per-slot routes + served counts + the `fallbackModel` chain from settings.json or its state (`unset` / `disabled` / `invalid`) (Runtime Sync PR3 v5.17.0) from `core.runtime.model_routing_check.status_summary()`. Plus **MCP usage (24h)** section: total calls + servers in use + top servers from `core.runtime.mcp_telemetry.summarise(period="today")`. |
 | `/arka costs [period]` | LLM cost visibility — aggregates telemetry by day/week/month/all, with top expensive sessions. See `arka/skills/costs/SKILL.md`. Shells out to `~/.arkaos/bin/arka-py -m core.runtime.llm_cost_telemetry_cli <period>`. |
 | `/arka enforcement [period]` | Enforcement compliance — aggregates flow-marker enforcement telemetry by day/week/month/all. Shows block rate, top blocked tools, top reasons. Shells out to `~/.arkaos/bin/arka-py -m core.governance.enforcement_telemetry_cli <period>`. |
 | `/arka mcps [period]` | MCP usage — aggregates the PostToolUse MCP telemetry (`~/.arkaos/telemetry/mcp-usage.jsonl`) by day/week/month/all. Shows total calls, servers in use, top servers and top tools. Shells out to `~/.arkaos/bin/arka-py -m core.runtime.mcp_telemetry_cli <period>`. |
@@ -315,4 +315,4 @@ All output: YAML frontmatter · wikilinks · department paths · MOC organizatio
 
 ## Model Selection
 
-Use `model` parameter from agent YAML (`departments/*/agents/*.yaml`), upgraded by the Model Fabric routing when `[ARKA:MODEL-FABRIC]` is present. Quality Gate model: constitution `quality_gate.model_policy` is the single source (best-available/frontier tier; veto is model-independent). Default: `sonnet`. Mechanical tasks: `haiku`.
+Use `model` parameter from agent YAML (`departments/*/agents/*.yaml`), upgraded by the Model Fabric routing when `[ARKA:MODEL-FABRIC]` is present. Quality Gate model: constitution `quality_gate.model_policy` is the single source (best-available/frontier tier; veto is model-independent). Default: `sonnet`. Mechanical tasks: `sonnet` at low effort — ArkaOS never routes to Haiku (Runtime Sync 2026-09-03).

package/arka/skills/costs/SKILL.md

@@ -70,8 +70,10 @@ Outside this skill's scope but worth surfacing alongside the summary:
 | Lever | Where | Effect |
 | --- | --- | --- |
 | `ARKAOS_LLM_PROVIDER` | env | Switch between Anthropic, OpenAI, Gemini adapters per the multi-LLM router. |
+| `fallbackModel` | `~/.claude/settings.json` (Claude Code 2.1.166+) | Up to three models tried in order when the primary is overloaded or unavailable. `npx arkaos install` / `update` seed `["claude-opus-5", "claude-sonnet-5"]` when the key is absent and never overwrite an operator chain; `/arka status` shows the configured chain. |
+| Schedule `model:` / `fallback_models:` | `~/.arkaos/schedules.yaml` | A pinned `model` goes to `claude --model`; without one the Model Fabric `strategy` role becomes `ANTHROPIC_DEFAULT_MODEL` (Claude Code 2.1.236+). `fallback_models` goes to `--fallback-model` (default `claude-opus-5 → claude-sonnet-5`) so one overload or model 404 no longer ends a nightly cycle. |
 | `ANTHROPIC_BEDROCK_SERVICE_TIER` | env (Claude Code 2.1.122+) | `default` / `flex` / `priority`. Only relevant when routing through AWS Bedrock — `flex` cuts cost ~50 % on non-urgent workloads at the price of higher tail latency. |
-| Agent `model:` field | `departments/*/agents/*.yaml` | Per-agent override per the model-routing matrix in CLAUDE.md. Mechanical roles default to haiku, C-suite to opus. |
+| Agent `model:` field | `departments/*/agents/*.yaml` | Per-agent override per the model-routing matrix in CLAUDE.md. Mechanical roles default to `sonnet` at low effort, C-suite to `opus` (Opus 5); the Quality Gate row below lifts Marta / Eduardo / Francisca to the frontier tier. |
 | Quality Gate model | constitution `quality_gate.model_policy` | Marta / Eduardo / Francisca run on the best model available (frontier tier) regardless of the cost ceiling — review quality is never the place to save tokens. |
 
 ## Non-negotiables

package/arka/skills/forge/SKILL.md

@@ -48,7 +48,7 @@ The Forge analyses any prompt across 5 complexity dimensions, routes to 1-3 expl
 
 | Tier | Score | Explorers | Critic | Companion | Model |
 |------|-------|-----------|--------|-----------|-------|
-| Simple (Shallow) | ≤ 30 | 1 (Pragmatic, inline) | Light | None | haiku |
+| Simple (Shallow) | ≤ 30 | 1 (Pragmatic, inline) | Light | None | sonnet |
 | Standard | 31-65 | 2 (Pragmatic + Architectural, parallel) | Full | On request | sonnet |
 | Complex (Deep) | 66-85 | 3 (+ Contrarian, parallel) | Full | Proactive | opus |
 | Super | ≥ 86 | 3 + full synthesis | Full (highest judgment) | Proactive | opus |
@@ -59,7 +59,7 @@ Five dimensions score 0-100 each: **scope**, **dependencies**, **ambiguity**, **
 
 When dispatching explorer subagents via the Task tool, include `model: <tier_model>` parameter. The tier-to-model mapping is:
 
-- simple → `model: haiku` (fast routing, single explorer)
+- simple → `model: sonnet` (fast routing, single explorer; the weakest lane ArkaOS routes to)
 - standard → `model: sonnet` (default, 1-2 explorers)
 - complex → `model: opus` (3 explorers + critic synthesis)
 - super → `model: opus` (highest judgment, full synthesis)

package/arka/skills/forge/references/complexity-engine.md

@@ -6,7 +6,7 @@ Referenced from SKILL.md. Read only when needed.
 
 | Tier | Score | Explorers | Critic | Companion | Model |
 |------|-------|-----------|--------|-----------|-------|
-| Simple (Shallow) | ≤ 30 | 1 (Pragmatic, inline) | Light | None | haiku |
+| Simple (Shallow) | ≤ 30 | 1 (Pragmatic, inline) | Light | None | sonnet |
 | Standard | 31-65 | 2 (Pragmatic + Architectural, parallel) | Full | On request | sonnet |
 | Complex (Deep) | 66-85 | 3 (Pragmatic + Architectural + Contrarian, parallel) | Full | Proactive | opus |
 | Super | ≥ 86 | 3 + full synthesis | Full (highest judgment) | Proactive | opus |

package/arka/skills/fusion/SKILL.md

@@ -68,7 +68,7 @@ Produce a table: role → provider/model + ONE-line rationale each. Rules:
   evidence: budget panel + frontier judge beats frontier solo at ~50%
   cost).
 - `mechanical` → cheapest competent lane: local coder model if Ollama
-  runs one, else haiku-class.
+  runs one, else Sonnet 5 at low effort (the weakest cloud lane).
 - Fusion panel: 2-3 DIVERSE models (different families — e.g. one
   Anthropic, one Kimi/DeepSeek, one local large) + frontier judge.
   Homogeneous panels waste the fan-out.

package/bin/arka-doctor

@@ -370,17 +370,23 @@ run_check "plugins" "warn" "$PLUGINS_DESC" "$PLUGINS_OK" "Run: bash install.sh"
 
 # ─── Check 17: Hyperframes Skills (video production) ──────────────────────
 # Parity with installer/doctor.js `hyperframes-skills` (PR-C2): same name,
-# same sentinel, warn-only — video production is opt-in.
+# same sentinels, same two roots (~/.claude/skills and ~/.agents/skills),
+# warn-only — video production is opt-in.
 HF_OK=false
 HF_DESC="Hyperframes skills AUSENTE (video-as-code editing for /content video)"
-for hf_name in hyperframes hyperframes-core; do
-  if [ -f "$HOME/.claude/skills/$hf_name/SKILL.md" ]; then
-    HF_OK=true
-    HF_DESC="Hyperframes skills installed ($hf_name)"
-    break
-  fi
+for hf_root in "$HOME/.claude/skills" "$HOME/.agents/skills"; do
+  for hf_name in hyperframes hyperframes-core; do
+    if [ -f "$hf_root/$hf_name/SKILL.md" ]; then
+      HF_OK=true
+      # Static, matching installer/doctor.js `description` exactly: run_check
+      # splices $desc unescaped into the --json payload, so no path or
+      # name interpolation may reach it.
+      HF_DESC="Hyperframes skills installed (video-as-code editing for /content video)"
+      break 2
+    fi
+  done
 done
-run_check "hyperframes-skills" "warn" "$HF_DESC" "$HF_OK" "Run /content video-setup in Claude Code"
+run_check "hyperframes-skills" "warn" "$HF_DESC" "$HF_OK" "Run /content video-setup in Claude Code, or: npx hyperframes skills update (core set; npx hyperframes skills check reports staleness)"
 
 # ─── Check 18: SQLite corrupt-store backups (self-heal F1-D1) ──────────────
 # Parity with installer/doctor.js `sqlite-corrupt-backups`: same name,

package/config/claude-agents/data-collector.md

@@ -2,7 +2,7 @@
 name: data-collector
 description: >
   Tomas Jr — Data Collector (kb squad). Executes: data-collection, web-scraping, API-integration, data-validation.
-model: haiku
+model: sonnet
 ---
 
 <!-- generated by core/agents/behavioral_compiler.py from departments/kb/agents/data-collector.yaml — DO NOT EDIT; edit the YAML and re-run -->

package/config/claude-agents/doc-writer.md

@@ -2,7 +2,7 @@
 name: doc-writer
 description: >
   Isabel — Documentation Writer (ops squad). Executes: technical-writing, API-docs, user-guides, SOPs.
-model: haiku
+model: sonnet
 ---
 
 <!-- generated by core/agents/behavioral_compiler.py from departments/ops/agents/doc-writer.yaml — DO NOT EDIT; edit the YAML and re-run -->

package/config/claude-agents/paulo-tech-lead.md

@@ -5,7 +5,7 @@ description: >
   work into atomic tasks, dispatches specialists, enforces the evidence flow
   (real test runs, exit codes on record), and submits deliverables to the
   Quality Gate before anything ships.
-tools: Read, Grep, Glob, Bash, Agent, TaskCreate
+tools: Read, Grep, Glob, Bash, Agent
 model: sonnet
 ---
 
@@ -37,8 +37,8 @@ not cut corners to look fast.
   self-documenting names, no dead code.
 - TDD where practical; coverage >= 80% on new code (constitution MUST).
 - Conventional commits, feature branches, never commit to main/master/dev.
-- Model routing: dispatch specialists on sonnet, mechanical tasks on haiku;
-  opus only for Tier 0/security-scope reviews.
+- Model routing: sonnet for specialists, sonnet at low effort for mechanical
+  tasks (never haiku); opus only for Tier 0/security-scope reviews.
 
 ## Signature Rules (anti-sycophancy)
 

package/config/claude-agents/research-assistant.md

@@ -2,7 +2,7 @@
 name: research-assistant
 description: >
   Maria — Research Assistant (dev squad). Executes: research, documentation, analysis, literature-review.
-model: haiku
+model: sonnet
 ---
 
 <!-- generated by core/agents/behavioral_compiler.py from departments/dev/agents/research-assistant.yaml — DO NOT EDIT; edit the YAML and re-run -->

package/config/cognition/schedules.yaml

@@ -2,6 +2,16 @@
 #
 # This file is deployed to ~/.arkaos/schedules.yaml during installation.
 # Edit the deployed copy to customize schedules.
+#
+# Model pin and fallback chain (Runtime Sync PR3) — per schedule, optional:
+#   model: claude-sonnet-5            # pinned: goes to `claude --model`
+#   fallback_models: [claude-opus-5]  # `--fallback-model a,b`; [] disables
+# Without `model`, the Model Fabric `strategy` role (~/.arkaos/models.yaml)
+# becomes ANTHROPIC_DEFAULT_MODEL for the run. Without `fallback_models`,
+# the chain is claude-opus-5 → claude-sonnet-5. The runtime caps chains at
+# three models after de-duplication, removes ids outside `availableModels`
+# before the walk and fails over past an id it cannot reach (a retired
+# model); the ArkaOS gateway plays no part in the chain.
 
 schedules:
   dreaming:

package/config/constitution.yaml

@@ -111,7 +111,7 @@ enforcement_levels:
         enforcement: "persistence.py exports on status change; audit detects missing exports"
 
       - id: model-routing
-        rule: "Quality-first model routing: design, review, architecture, strategy, and Quality Gate phases run the BEST model available for the task (frontier tier) at maximum effort BY DEFAULT. Cheaper/smaller models (haiku, local) are reserved for genuinely mechanical work: commit messages, keyword extraction, data fetching, formatting. Cost optimisation NEVER downgrades a quality-critical phase — the CostGovernor hard budget is the only ceiling. User-owned overrides live in ~/.arkaos/models.yaml (Model Fabric): the user decides which provider/model serves each role; this rule sets the default posture, not a fixed vendor table. Replaced the cost-optimised per-tier table in the Excellence Reform (2026-07-05)."
+        rule: "Quality-first model routing: design, review, architecture, strategy, and Quality Gate phases run the BEST model available for the task (frontier tier) at maximum effort BY DEFAULT. Cheaper lanes (Sonnet 5 at low effort, local) are reserved for genuinely mechanical work: commit messages, keyword extraction, data fetching, formatting. ArkaOS never routes to Haiku. Cost optimisation NEVER downgrades a quality-critical phase — the CostGovernor hard budget is the only ceiling. User-owned overrides live in ~/.arkaos/models.yaml (Model Fabric): the user decides which provider/model serves each role; this rule sets the default posture, not a fixed vendor table. Retired model ids (previous Opus and Fable generations) and the Haiku lane are mapped to their current replacement at resolve time, with a notice; the table is core/runtime/model_router.LEGACY_MODEL_IDS. The user's models.yaml is never rewritten (Runtime Sync 2026-09-03). Replaced the cost-optimised per-tier table in the Excellence Reform (2026-07-05)."
         enforcement: "warning"
 
       - id: subagent-discipline

package/config/hooks/_lib/fastpath/engine.cjs

@@ -227,7 +227,7 @@ function decidePre(payload, manifest, ctx) {
     return { action: "delegate", reason: "bash-effect-or-budget" };
   }
 
-  // P7/P8 — Write/Edit/MultiEdit/NotebookEdit/Task/Skill: specialist +
+  // P7/P8 — Write/Edit/NotebookEdit/Task/Skill: specialist +
   // frontend + flow chains with decision-bearing state writes.
   return { action: "delegate", reason: "flow-gated" };
 }
@@ -311,7 +311,7 @@ function toolText(payload) {
 /**
  * PostToolUse decision. Returns
  *   {action:"fast-exit", stdout:"{}", writes:[...]} — mcp telemetry only
- *   {action:"delegate", reason}
+ *   {action:"delegate", reason} — incl. "kb-marker" for KB-consult tools
  * ctx: {homeDir, config, env, fs, now?: Date}
  */
 function decidePost(payload, manifest, ctx) {
@@ -333,6 +333,21 @@ function decidePost(payload, manifest, ctx) {
     return { action: "delegate", reason: "post-stateful-tool" };
   }
 
+  // Q3b — KB-first evidence (Runtime Sync PR0). The "obsidian"/"graphify"
+  // turn markers the research gate reads have exactly one writer, in
+  // Python (post_tool_use.KB_MARKER_TOOL_PREFIXES → manifest). A fast-exit
+  // here skipped it, so on a turn whose KB call took this path the gate
+  // could not see the consult (kb_first.jsonl 2026-09-03: 66 denials,
+  // 0 kb-consulted). Prefix = server; Python decides whether the tool is
+  // a read consult. Old manifests without the key keep the previous
+  // behaviour (fail-open, never deny).
+  const kbPrefixes = Array.isArray(manifest.tools.post_delegate_prefixes)
+    ? manifest.tools.post_delegate_prefixes
+    : [];
+  if (kbPrefixes.some((p) => toolName.startsWith(p))) {
+    return { action: "delegate", reason: "kb-marker" };
+  }
+
   // Q5 — error path: gotchas, violations, cognition capture, metrics.
   // PostToolUse never carries an exit code (failures fire the event
   // above), so the only error signal on this path is trigger text.

package/config/hooks/gate-manifest.json

@@ -519,7 +519,7 @@
         "tool": "Edit"
       },
       {
-        "expect": "delegate",
+        "expect": "fast_allow",
         "tool": "MultiEdit"
       },
       {
@@ -715,7 +715,6 @@
   "tools": {
     "effect_always": [
       "Edit",
-      "MultiEdit",
       "NotebookEdit",
       "Skill",
       "Task",
@@ -724,7 +723,6 @@
     "flow_gated": [
       "Bash",
       "Edit",
-      "MultiEdit",
       "NotebookEdit",
       "Skill",
       "Task",
@@ -735,6 +733,10 @@
       "ExitPlanMode",
       "Task"
     ],
+    "post_delegate_prefixes": [
+      "mcp__graphify__",
+      "mcp__obsidian__"
+    ],
     "research_external": [
       "WebFetch",
       "WebSearch",

package/config/hooks/pre-tool-use.ps1

@@ -1,7 +1,7 @@
 # ============================================================================
 # ArkaOS v2 — PreToolUse Hook (Windows PowerShell)
 #
-# Parity with config/hooks/pre-tool-use.sh. Blocks Write/Edit/MultiEdit when
+# Parity with config/hooks/pre-tool-use.sh. Blocks Write/Edit when
 # the mandatory 13-phase flow is required for the session AND the assistant
 # has not emitted a flow marker in its last 3 messages of the transcript.
 #
@@ -127,7 +127,7 @@ print(json.dumps({
 # Blocks Tier-1 leads from writing to specialist-owned files without
 # dispatching first. Only fires for file-mutation tools.
 $specialistPy = Join-Path $env:ARKAOS_ROOT "core/workflow/specialist_enforcer.py"
-if ((Test-Path $specialistPy) -and ($toolName -in @("Write","Edit","MultiEdit","NotebookEdit"))) {
+if ((Test-Path $specialistPy) -and ($toolName -in @("Write","Edit","NotebookEdit"))) {
     $toolInputJson = "{}"
     if ($inp.tool_input) {
         $toolInputJson = ($inp.tool_input | ConvertTo-Json -Compress -Depth 10)
@@ -201,7 +201,7 @@ print(json.dumps({
 }
 
 # --- Fast allow: not a flow-gated tool ---
-if ($toolName -ne "Write" -and $toolName -ne "Edit" -and $toolName -ne "MultiEdit") {
+if ($toolName -ne "Write" -and $toolName -ne "Edit") {
     exit 0
 }
 

package/config/models.yaml

@@ -34,13 +34,13 @@ providers:
 # `best` / `default` / `fast` resolve per provider, so roles stay readable.
 aliases:
   runtime:
-    best: opus
+    best: fable            # Claude Code alias → Fable 5.1 (Claude Code 2.1.257+)
     default: sonnet
-    fast: haiku
+    fast: sonnet           # weakest lane is Sonnet 5 — never Haiku
   anthropic:
-    best: claude-opus-4-8
+    best: claude-fable-5-1
     default: claude-sonnet-5
-    fast: claude-haiku-4-5-20251001
+    fast: claude-sonnet-5
   ollama:
     best: ""                # set your strongest local model, e.g. kimi-k2.6
     default: ""

package/config/settings-template.json

@@ -14,7 +14,7 @@
           {
             "type": "command",
             "command": "{{HOOKS_DIR}}/session-start.sh",
-            "timeout": 5
+            "timeout": 15
           }
         ]
       }
@@ -30,18 +30,18 @@
         ]
       }
     ],
-    "PreCompact": [
+    "PostToolUse": [
       {
         "hooks": [
           {
             "type": "command",
-            "command": "{{HOOKS_DIR}}/pre-compact.sh",
-            "timeout": 30
+            "command": "{{HOOKS_DIR}}/post-tool-use.sh",
+            "timeout": 5
           }
         ]
       }
     ],
-    "PostToolUse": [
+    "PostToolUseFailure": [
       {
         "hooks": [
           {
@@ -52,39 +52,81 @@
         ]
       }
     ],
-    "PostToolUseFailure": [
+    "PreToolUse": [
       {
         "hooks": [
           {
             "type": "command",
-            "command": "{{HOOKS_DIR}}/post-tool-use.sh",
-            "timeout": 5
+            "command": "{{HOOKS_DIR}}/pre-tool-use.sh",
+            "timeout": 10
+          }
+        ]
+      },
+      {
+        "matcher": "Task",
+        "hooks": [
+          {
+            "type": "command",
+            "command": "{{HOOKS_DIR}}/agent-provision.sh",
+            "timeout": 10
           }
         ]
       }
     ],
-    "CwdChanged": [
+    "Stop": [
       {
         "hooks": [
           {
             "type": "command",
-            "command": "{{HOOKS_DIR}}/cwd-changed.sh",
+            "command": "{{HOOKS_DIR}}/stop.sh",
             "timeout": 5
           }
         ]
       }
     ],
-    "PreToolUse": [
+    "SubagentStop": [
       {
-        "matcher": "Task",
         "hooks": [
           {
             "type": "command",
-            "command": "{{HOOKS_DIR}}/agent-provision.sh",
+            "command": "{{HOOKS_DIR}}/subagent-stop.sh",
             "timeout": 10
           }
         ]
       }
+    ],
+    "SessionEnd": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "{{HOOKS_DIR}}/session-end.sh",
+            "timeout": 15
+          }
+        ]
+      }
+    ],
+    "PreCompact": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "{{HOOKS_DIR}}/pre-compact.sh",
+            "timeout": 30
+          }
+        ]
+      }
+    ],
+    "CwdChanged": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "{{HOOKS_DIR}}/cwd-changed.sh",
+            "timeout": 5
+          }
+        ]
+      }
     ]
   }
 }

package/core/agents/behavioral_compiler.py

@@ -32,6 +32,8 @@ from pathlib import Path
 
 import yaml
 
+from core.agents.schema import normalise_tier
+
 _ROOT = Path(__file__).resolve().parents[2]
 _OUT_DIR = _ROOT / "config" / "claude-agents"
 
@@ -95,7 +97,7 @@ def _frontmatter(slug: str, data: dict) -> str:
     desc = f"{data.get('name', slug)} — {role} ({dept} squad)."
     if domains:
         desc += f" Executes: {domains}."
-    model = data.get("model", "sonnet")
+    model = normalise_tier(data.get("model", "sonnet"))
     return (
         f"---\nname: {slug}\ndescription: >\n  {desc}\nmodel: {model}\n---"
     )

package/core/agents/schema.py

@@ -8,11 +8,24 @@ Every ArkaOS agent has a complete behavioral profile composed of:
 """
 
 from enum import Enum
-from typing import Literal, Optional
+from typing import Literal
+
 from pydantic import BaseModel, Field, model_validator
 
+# "haiku" stays accepted for legacy client YAML only; ArkaOS's weakest
+# lane is Sonnet 5 (Runtime Sync 2026-09-03). "fable" = Claude Code alias.
+ModelTier = Literal["haiku", "sonnet", "opus", "fable"]
+
+
+# Runtime Sync 2026-09-03: the weakest lane ArkaOS routes to is Sonnet 5.
+# "haiku" still PARSES (legacy client YAML) but never reaches a dispatch:
+# get_model() and the compiled frontmatter both normalise it.
+_LANE_NORMALISATION: dict[ModelTier, ModelTier] = {"haiku": "sonnet"}
 
-ModelTier = Literal["haiku", "sonnet", "opus"]
+
+def normalise_tier(model: ModelTier) -> ModelTier:
+    """Map a legacy lane to the one ArkaOS actually routes to."""
+    return _LANE_NORMALISATION.get(model, model)
 
 
 def tier_default_model(tier: int) -> ModelTier:
@@ -24,9 +37,9 @@ def tier_default_model(tier: int) -> ModelTier:
 
 # --- DISC Framework ---
 
-class DISCType(str, Enum):
+class DISCType(str, Enum):  # noqa: UP042 — behavioural parity kept; StrEnum migration deferred
     D = "D"  # Dominance
-    I = "I"  # Influence
+    I = "I"  # noqa: E741 — DISC letter is the contract
     S = "S"  # Steadiness
     C = "C"  # Conscientiousness
 
@@ -64,13 +77,13 @@ class EnneagramType(int, Enum):
     PEACEMAKER = 9
 
 
-class EnneagramCenter(str, Enum):
+class EnneagramCenter(str, Enum):  # noqa: UP042 — behavioural parity kept; StrEnum migration deferred
     BODY = "body"      # 8, 9, 1 — Anger/Instinct
     HEART = "heart"    # 2, 3, 4 — Shame/Feeling
     HEAD = "head"      # 5, 6, 7 — Fear/Thinking
 
 
-class InstinctualSubtype(str, Enum):
+class InstinctualSubtype(str, Enum):  # noqa: UP042 — behavioural parity kept; StrEnum migration deferred
     SP = "self-preservation"
     SO = "social"
     SX = "sexual"
@@ -136,14 +149,26 @@ class BigFiveProfile(BaseModel):
 
 # --- MBTI Framework ---
 
-class MBTIType(str, Enum):
-    INTJ = "INTJ"; INTP = "INTP"; ENTJ = "ENTJ"; ENTP = "ENTP"
-    INFJ = "INFJ"; INFP = "INFP"; ENFJ = "ENFJ"; ENFP = "ENFP"
-    ISTJ = "ISTJ"; ISFJ = "ISFJ"; ESTJ = "ESTJ"; ESFJ = "ESFJ"
-    ISTP = "ISTP"; ISFP = "ISFP"; ESTP = "ESTP"; ESFP = "ESFP"
-
-
-class CognitiveFunction(str, Enum):
+class MBTIType(str, Enum):  # noqa: UP042 — behavioural parity kept; StrEnum migration deferred
+    INTJ = "INTJ"
+    INTP = "INTP"
+    ENTJ = "ENTJ"
+    ENTP = "ENTP"
+    INFJ = "INFJ"
+    INFP = "INFP"
+    ENFJ = "ENFJ"
+    ENFP = "ENFP"
+    ISTJ = "ISTJ"
+    ISFJ = "ISFJ"
+    ESTJ = "ESTJ"
+    ESFJ = "ESFJ"
+    ISTP = "ISTP"
+    ISFP = "ISFP"
+    ESTP = "ESTP"
+    ESFP = "ESFP"
+
+
+class CognitiveFunction(str, Enum):  # noqa: UP042 — behavioural parity kept; StrEnum migration deferred
     Ni = "Ni"  # Introverted Intuition
     Ne = "Ne"  # Extraverted Intuition
     Si = "Si"  # Introverted Sensing
@@ -217,7 +242,7 @@ class Authority(BaseModel):
     block_delivery: bool = False
     orchestrate: bool = False
     delegates_to: list[str] = Field(default_factory=list)
-    escalates_to: Optional[str] = None
+    escalates_to: str | None = None
 
 
 # --- Communication Style ---
@@ -272,12 +297,12 @@ class Agent(BaseModel):
 
     memory_path: str = ""
 
-    model: Optional[ModelTier] = Field(
+    model: ModelTier | None = Field(
         default=None,
         description="Claude model override for dispatch. Falls back to tier default when None.",
     )
 
-    parent_squad: Optional[str] = Field(
+    parent_squad: str | None = Field(
         default=None,
         description=(
             "Slug of the parent squad when this agent belongs to a sub-squad. "
@@ -287,7 +312,7 @@ class Agent(BaseModel):
         ),
     )
 
-    sub_squad_role: Optional[str] = Field(
+    sub_squad_role: str | None = Field(
         default=None,
         description=(
             "Role within the sub-squad when parent_squad is set "
@@ -314,4 +339,4 @@ class Agent(BaseModel):
 
     def get_model(self) -> ModelTier:
         """Return the resolved Claude model, using tier default when unset."""
-        return self.model or tier_default_model(self.tier)
+        return normalise_tier(self.model or tier_default_model(self.tier))

package/core/cognition/scheduler/cli.py

@@ -14,6 +14,8 @@ def list_schedules(config_path: str) -> list[dict]:
             "time": s.run_time.strftime("%H:%M"),
             "timeout": s.timeout_minutes,
             "retry": s.retry_on_fail,
+            "model": s.model,
+            "fallback_models": list(s.fallback_models),
         }
         for s in schedules
     ]