pantheon-opencode 1.6.0-beta.5 → 1.6.0-beta.7

package/src/plugins/tui/src/index.tsx

@@ -53,16 +53,7 @@ import { homedir } from 'node:os'
 import { dirname, join } from 'node:path'
 import { fileURLToPath } from 'node:url'
 import type { TuiPlugin, TuiPluginApi, TuiPluginModule } from '@opencode-ai/plugin/tui'
-import {
-  createEffect,
-  createMemo,
-  createResource,
-  createSignal,
-  For,
-  onCleanup,
-  onMount,
-  Show,
-} from 'solid-js'
+import { createMemo, createResource, createSignal, For, onCleanup, onMount, Show } from 'solid-js'
 
 // ─── Double-load guard (mirrors src/pantheon/plugin-once.ts) ───────────────
 // Opencode merges global + project plugin paths, so this module can load twice
@@ -749,12 +740,27 @@ async function loadConfig(configPath: string | undefined): Promise<UsageBarConfi
   }
 }
 
-/** Vendored usage-bar plugin init: reads config, seeds the kv cache, starts
- *  the poll loop and registers the `app_bottom` slot. Skips everything (and
+/** Vendored usage-bar plugin init: reads config, seeds the kv cache, registers
+ *  the `app_bottom` slot, then starts the poll loop. Skips everything (and
  *  registers nothing) when no provider is enabled — mirrors the standalone
  *  plugin's early return. */
 async function setupUsageBar(api: TuiPluginApi) {
+  let disposed = false
+  let clockInterval: ReturnType<typeof setInterval> | undefined
+  const pollTimeouts = new Set<ReturnType<typeof setTimeout>>()
+
+  // This optional feature owns timers independently from the sidebar. A slow
+  // config read may finish after shutdown, so register cleanup before awaiting
+  // anything and check the disposed flag at every async boundary.
+  api.lifecycle.onDispose(() => {
+    disposed = true
+    if (clockInterval !== undefined) clearInterval(clockInterval)
+    for (const timeout of pollTimeouts) clearTimeout(timeout)
+    pollTimeouts.clear()
+  })
+
   const config = await loadConfig(api.state.path?.config)
+  if (disposed) return
   if (api.state.path?.state) opencodeAuthFile = join(api.state.path.state, 'auth.json')
 
   const enabled = providers.filter((p) => config.providers[p.id].enabled)
@@ -771,28 +777,6 @@ async function setupUsageBar(api: TuiPluginApi) {
   const [byStatus, setByStatus] = createSignal<Record<string, StatusIndicator>>({})
   const [now, setNow] = createSignal(Date.now())
 
-  setInterval(() => setNow(Date.now()), 1_000)
-
-  for (const p of enabled) {
-    const cfg = config.providers[p.id]
-    const poll = async () => {
-      // Fetch usage and status concurrently; status pages are CDN-backed and
-      // never throttle, so we poll them on the same cadence as usage.
-      const statusP = config.showStatus && p.statusUrl ? fetchStatus(p.statusUrl) : null
-      const [all, status] = await Promise.all([p.fetchUsage(cfg), statusP])
-      if (all) {
-        const windows = all.filter((w) => cfg.show[w.category] && w.resetsAt > Date.now())
-        setByProvider((prev) => ({ ...prev, [p.id]: windows }))
-        api.kv.set(`usage-bar.${p.id}.windows`, windows)
-      }
-      // `null` means the status fetch failed — keep the last known indicator.
-      if (status !== null) setByStatus((prev) => ({ ...prev, [p.id]: status }))
-      // Back off when the usage fetch failed (e.g. 429 — these endpoints throttle).
-      setTimeout(poll, all ? POLL_MS : POLL_MS * 3)
-    }
-    void poll()
-  }
-
   api.slots.register({
     order: 60,
     slots: {
@@ -880,6 +864,46 @@ async function setupUsageBar(api: TuiPluginApi) {
       },
     },
   })
+
+  // Start background work only once the optional UI slot was accepted. If
+  // registration fails, the outer fail-open handler can suppress it without
+  // leaving a timer or provider request running in the background.
+  clockInterval = setInterval(() => setNow(Date.now()), 1_000)
+
+  const schedulePoll = (run: () => void, delay: number) => {
+    if (disposed) return
+    const timeout = setTimeout(() => {
+      pollTimeouts.delete(timeout)
+      if (!disposed) run()
+    }, delay)
+    pollTimeouts.add(timeout)
+  }
+
+  for (const p of enabled) {
+    const cfg = config.providers[p.id]
+    const poll = async () => {
+      if (disposed) return
+      // Fetch usage and status concurrently; status pages are CDN-backed and
+      // never throttle, so we poll them on the same cadence as usage.
+      const statusP = config.showStatus && p.statusUrl ? fetchStatus(p.statusUrl) : null
+      const [all, status] = await Promise.all([p.fetchUsage(cfg), statusP])
+      if (disposed) return
+      if (all) {
+        const windows = all.filter((w) => cfg.show[w.category] && w.resetsAt > Date.now())
+        setByProvider((prev) => ({ ...prev, [p.id]: windows }))
+        api.kv.set(`usage-bar.${p.id}.windows`, windows)
+      }
+      // `null` means the status fetch failed — keep the last known indicator.
+      if (status !== null) setByStatus((prev) => ({ ...prev, [p.id]: status }))
+      // Back off when the usage fetch failed (e.g. 429 — these endpoints throttle).
+      schedulePoll(runPoll, all ? POLL_MS : POLL_MS * 3)
+    }
+    const runPoll = () => {
+      if (disposed) return
+      void poll().catch(() => schedulePoll(runPoll, POLL_MS * 3))
+    }
+    runPoll()
+  }
 }
 
 /* ─── Delegations (real-time panel) ────────────────────────
@@ -919,9 +943,9 @@ async function setupUsageBar(api: TuiPluginApi) {
  *      (filterDelegationsToSession) drops every non-focused report — the panel
  *      lists only the focused session.
  *
- *   3. REFRESH — session.created/updated/deleted/status events and
- *      message.part.updated/removed (via liveStore.version bumps) re-fetch
- *      children; a 1s safety poll (oh-my-opencode-slim pattern) re-fetches
+ *   3. REFRESH — session.created/updated/deleted/status events and live-store
+ *      subscription notifications re-fetch children; a 1s safety poll
+ *      (oh-my-opencode-slim pattern) re-fetches
  *      children + re-reads the md so the panel updates EVEN without events.
  *
  *   4. DIAGNOSTICS — every re-fetch logs
@@ -947,8 +971,8 @@ async function setupUsageBar(api: TuiPluginApi) {
  *      panel renders EMPTY — zero rows, zero errors.
  *
  * The live tool-part lifecycle helpers below remain (tested, exported) but
- * are no longer the source of truth: they feed a version signal that merely
- * re-triggers the children fetch. Everything is fail-open: a missing
+ * are no longer the source of truth: they update the shared live store, whose
+ * subscribers re-trigger the children fetch. Everything is fail-open: a missing
  * children API or directory yields [], and any unreadable input is skipped —
  * never crash. */
 
@@ -2631,18 +2655,88 @@ function DelegationRow(props: {
 
 /* ─── Main Sidebar View ──────────────────────────────────── */
 
-/** Plugin-level live delegation store shared with the event subscriptions
- *  in `tui()`: the map of live entries + a version signal bumped on every
- *  mutation. The View subscribes to the version (in an effect) to refresh the
- *  durable child list and also reads the map as an optimistic live source. */
+/** Plugin-level live delegation store shared by event subscriptions and View.
+ *  Mutations notify subscribers so the View can refresh its durable child
+ *  list while also reading the map as an optimistic live source. */
 export type LiveDelegationStore = {
   map: Map<string, LiveDelegationEntry>
-  /** Reactive version getter — View reads it inside an effect to re-fetch. */
-  version: () => number
-  /** Bump the version after a live mutation. */
+  /** Subscribe to invalidations consumed by the sidebar's refresh path. */
+  subscribe: (listener: () => void) => () => void
+  /** Notify sidebar subscribers after a live mutation. */
   bump: () => void
 }
 
+/** Create the observable live store shared by task events and the sidebar. */
+export function createLiveDelegationStore(): LiveDelegationStore {
+  const listeners = new Set<() => void>()
+  return {
+    map: new Map<string, LiveDelegationEntry>(),
+    subscribe(listener) {
+      listeners.add(listener)
+      return () => listeners.delete(listener)
+    },
+    bump() {
+      for (const listener of listeners) listener()
+    },
+  }
+}
+
+/** Register the live task-part events that invalidate the sidebar store. */
+export function registerLiveDelegationEvents(
+  api: TuiPluginApi,
+  liveStore: LiveDelegationStore,
+): () => void {
+  const unsubLive: (() => void)[] = []
+
+  try {
+    unsubLive.push(
+      api.event.on('message.part.updated', (event: unknown) => {
+        // SDK v2 shape: properties.part (v1 drift: properties.info.part).
+        const props = ((event as { properties?: unknown } | null)?.properties ?? {}) as {
+          part?: unknown
+          info?: { part?: unknown }
+        }
+        const part = (props.part ?? props.info?.part) as DelegationToolPart | undefined
+        if (part === undefined) return
+        // beta.6: track the live tool call per child session (panel's
+        // activity line). Fire-and-forget — display-only state.
+        const sample = extractToolActivity(part as Parameters<typeof extractToolActivity>[0])
+        if (sample !== null)
+          trackToolActivity(latestToolActivity, sample.sessionID, sample.activity)
+        if (reduceDelegationToolPart(liveStore.map, part)) liveStore.bump()
+      }),
+    )
+  } catch {
+    /* events API unavailable — children API + 1s poll cover the panel */
+  }
+
+  try {
+    unsubLive.push(
+      api.event.on('message.part.removed', (event: unknown) => {
+        // SDK v2 shape: properties.partID (v1 drift: properties.part.id).
+        const props = ((event as { properties?: unknown } | null)?.properties ?? {}) as {
+          partID?: string
+          part?: { id?: string }
+        }
+        const partID = props.partID ?? props.part?.id
+        if (partID !== undefined && removeDelegationEntry(liveStore.map, partID)) liveStore.bump()
+      }),
+    )
+  } catch {
+    /* events API unavailable — children API + 1s poll cover the panel */
+  }
+
+  return () => {
+    for (const unsub of unsubLive) {
+      try {
+        unsub()
+      } catch {
+        /* ignore */
+      }
+    }
+  }
+}
+
 function View(props: {
   api: TuiPluginApi
   sessionID: string | undefined
@@ -2758,7 +2852,7 @@ function View(props: {
   // means the event bus isn't delivering, and the 1s poll is what keeps
   // the panel honest.
   let eventRefreshCount = 0
-  // In-flight guard: overlap between the poll, events and the version effect
+  // In-flight guard: overlap between the poll, events and store subscribers
   // collapses into one fetch — no duplicate work, no duplicate entries.
   let delegationsInflight: Promise<void> | null = null
   const refreshDelegations = () => {
@@ -2910,14 +3004,17 @@ function View(props: {
       /* events API not available — the 1s poll still covers it */
     }
 
-    // message.part.updated/removed → liveStore.version() bumps (tui()):
-    // subscribe here so the panel ALSO re-fetches children when tool-part
-    // events DO arrive and the updated live row is rendered immediately.
-    createEffect(() => {
-      props.liveStore.version() // subscribe to live mutations
-      eventRefreshCount += 1
-      void refreshDelegations()
-    })
+    // Task-part events invalidate the same store the panel reads, refreshing
+    // children when those events arrive and making optimistic live rows visible.
+    cleanup.push(
+      props.liveStore.subscribe(() => {
+        eventRefreshCount += 1
+        void refreshDelegations()
+      }),
+    )
+    // Fetch the initial state; future updates use the store subscription above.
+    eventRefreshCount += 1
+    void refreshDelegations()
 
     // Safety poll (OMO-slim pattern): 1s re-fetch children + md + tick the
     // elapsed counter. Guarantees the panel updates even with zero events.
@@ -2935,8 +3032,8 @@ function View(props: {
     // compacting, so after a compaction/attach the live map would start
     // empty even though the session still holds the delegation tool parts.
     // Re-seed the live map from the session's existing parts (in message
-    // order) right after the subscriptions are registered — the version
-    // bump re-triggers the children fetch. Fail-open: if the SDK access is
+    // order) right after the subscriptions are registered — the store
+    // notification re-triggers the children fetch. Fail-open: if the SDK access is
     // unavailable we skip — the children API + the 1s poll keep the panel
     // working.
     try {
@@ -3070,63 +3167,22 @@ const tui: TuiPlugin = (api, _options, _meta) => {
     .catch(() => setVersion(null))
 
   // Vendored feature (MIT): AI subscription usage gauges.
-  void setupUsageBar(api) // async init — never blocks sidebar registration
+  // Usage indicators are optional: a config, storage, or slot failure must
+  // never turn an otherwise healthy TUI startup into an unhandled rejection.
+  void setupUsageBar(api).catch(() => undefined)
 
-  // ── Live delegation store (tool-part tracker, kept as a REFRESH
-  // TRIGGER) ──
+  // ── Live delegation store (tool-part tracker + sidebar invalidation) ──
   // The Delegations panel's PRIMARY source is `api.client.session.children`
   // (delegations-sidebar pattern) — `message.part.updated` may not fire for
   // the native task tool on runtime 1.18.13 (the "(0)" symptom). The
-  // tool-part lifecycle below still tracks `task` parts and bumps a version
-  // signal the View subscribes to: when part events DO arrive, the panel
+  // tool-part lifecycle below still tracks `task` parts and invalidates the
+  // store subscription consumed by the View: when part events DO arrive, the panel
   // re-fetches children. Fail-open: if the events API is unavailable nothing
   // subscribes and the 1s safety poll keeps the panel working (never crash).
-  const [liveVersion, setLiveVersion] = createSignal(0)
-  const liveStore: LiveDelegationStore = {
-    map: new Map<string, LiveDelegationEntry>(),
-    version: liveVersion,
-    bump: () => setLiveVersion((v) => v + 1),
-  }
-  const unsubLive: (() => void)[] = []
-
-  try {
-    unsubLive.push(
-      api.event.on('message.part.updated', (event: any) => {
-        // SDK v2 shape: properties.part (v1 drift: properties.info.part).
-        const props = (event?.properties ?? {}) as { part?: unknown; info?: { part?: unknown } }
-        const part = (props.part ?? props.info?.part) as DelegationToolPart | undefined
-        if (part === undefined) return
-        // beta.6: track the live tool call per child session (panel's
-        // activity line). Fire-and-forget — display-only state.
-        const sample = extractToolActivity(part as any)
-        if (sample !== null)
-          trackToolActivity(latestToolActivity, sample.sessionID, sample.activity)
-        if (reduceDelegationToolPart(liveStore.map, part)) liveStore.bump()
-      }),
-    )
-  } catch {
-    /* events API unavailable — children API + 1s poll cover the panel */
-  }
-  try {
-    unsubLive.push(
-      api.event.on('message.part.removed', (event: any) => {
-        // SDK v2 shape: properties.partID (v1 drift: properties.part.id).
-        const props = (event?.properties ?? {}) as { partID?: string; part?: { id?: string } }
-        const partID = props.partID ?? props.part?.id
-        if (partID !== undefined && removeDelegationEntry(liveStore.map, partID)) liveStore.bump()
-      }),
-    )
-  } catch {
-    /* events API unavailable — children API + 1s poll cover the panel */
-  }
+  const liveStore = createLiveDelegationStore()
+  const unsubscribeLive = registerLiveDelegationEvents(api, liveStore)
   api.lifecycle.onDispose(() => {
-    for (const unsub of unsubLive) {
-      try {
-        unsub()
-      } catch {
-        /* ignore */
-      }
-    }
+    unsubscribeLive()
     liveStore.map.clear()
   })
 

package/src/plugins/tui/tsdown.config.ts

@@ -1,4 +1,5 @@
 import { readFileSync } from 'node:fs'
+import { appendFile, readdir, readFile } from 'node:fs/promises'
 import { join } from 'node:path'
 import { fileURLToPath } from 'node:url'
 import solid from 'rolldown-plugin-solid'
@@ -9,6 +10,21 @@ const pkg = JSON.parse(readFileSync(join(__dirname, 'package.json'), 'utf8')) as
   version: string
 }
 
+async function ensureDistArtifactsEndWithNewline(): Promise<void> {
+  const dist = join(__dirname, 'dist')
+  const entries = await readdir(dist, { withFileTypes: true })
+
+  await Promise.all(
+    entries
+      .filter((entry) => entry.isFile())
+      .map(async ({ name }) => {
+        const path = join(dist, name)
+        const contents = await readFile(path, 'utf8')
+        if (!contents.endsWith('\n')) await appendFile(path, '\n')
+      }),
+  )
+}
+
 export default defineConfig({
   entry: { tui: 'src/index.tsx', server: 'src/server.ts' },
   format: ['esm'],
@@ -33,4 +49,5 @@ export default defineConfig({
   deps: {
     neverBundle: [/^@opencode-ai\//, /^@opentui\//, /^solid-js/],
   },
+  onSuccess: ensureDistArtifactsEndWithNewline,
 })

package/src/routing.yml

@@ -133,7 +133,7 @@ agents:
     - artifact-management
     - context-compression
     - auto-continue
-    - incremental-implementation
+    - council-synthesis
   athena:
     role: Strategic planner
     reasoning_effort: high
@@ -188,8 +188,6 @@ agents:
     - pantheon_code_mode
     skills:
     - tdd-with-agents
-    - file-prompts
-    - streaming-patterns
     - git-workflow-and-versioning
     - incremental-implementation
   aphrodite:
@@ -211,7 +209,7 @@ agents:
     skills:
     - tdd-with-agents
     - visual-review-pipeline
-    - file-prompts
+    - incremental-implementation
   demeter:
     role: Database specialist
     reasoning_effort: medium
@@ -459,6 +457,8 @@ handoffs:
         (priority-aware, budget-allocated), archive IMPL/REVIEW to 02-progress-log.md,
         update _xref/index.md cross-references.'
 
+# Canonical explicit fallback map for Zeus timeout/retry policy. Evaluate
+# left-to-right; an absent agent entry means escalate (never guess a fallback).
 fallback_chains:
   apollo:
     - athena
@@ -518,7 +518,7 @@ background_delegation:
   prune_ttl_ms: 86400000    # 24h
   read_only_agents: [apollo, gaia]
   session_max: 3
-  retry_count: 1
+  retry_count: 1              # one retry after an initial transient failure; no retry loops
 # ── O5: permission.task — glob rules for which subagents an agent may
 # invoke (last matching rule wins; deny removes the agent from the delegate
 # tool description entirely — prevents circular delegation at the
@@ -551,7 +551,10 @@ hashline:
 full_auto:
   enabled: false                 # opt-in: goal continuations are OFF by default
   max_continuations: 25          # hard cap on goal continuations per session
-  # Wave 3 (PR #46): full-auto goal loop. One active goal per session
-  # (pending → in_progress → done); while active, idle sessions are
-  # continued automatically toward the objective. Documentation mirror —
-  # the plugin resolves behavior in code (GOAL_LOOP_DEFAULTS).
+  # Full-auto only continues a task/scope explicitly authorized by the user.
+  # It never authorizes auth/security gates, risky data/schema or destructive
+  # operations, global config, broad permissions, production deploy, push, or merge.
+  # Low-risk execution is proportional: no forced plan/discovery/council/artifact
+  # for a bounded fix; use focused verification. Documentation mirror — runtime
+  # behavior is implemented in the plugin (GOAL_LOOP_DEFAULTS).
+  # Wave 3 (PR #46): one active goal per session (pending → in_progress → done).

package/src/skills/README.md

@@ -4,14 +4,14 @@
 
 Skills are reference modules loaded on demand by agents. Each skill provides domain-specific knowledge — architecture patterns, security checklists, optimization strategies, and implementation guides. Agents load skills when their task matches the skill's description, keeping context focused and token-efficient.
 
-There are **20 skills** divided into **5 domains**.
+There are **21 skills** divided into **5 domains**.
 
 ## Skills by Domain
 
 ### Domain 1: Orchestration & Workflow
 
 1. **agent-coordination** — `agent-coordination/SKILL.md`
-   - Multi-agent orchestration with model routing, category delegation, and sprint management.
+   - Lightweight coordination for substantial tasks with multiple independent specialists.
    - Used by: Zeus
 
 2. **artifact-management** — `artifact-management/SKILL.md`
@@ -34,6 +34,10 @@ There are **20 skills** divided into **5 domains**.
    - Pin session objectives to prevent scope creep across long multi-agent sessions.
    - Used by: Zeus, Mnemosyne
 
+- **council-synthesis** — `council-synthesis/SKILL.md`
+  - Full structured council workflow, loaded only for `/pantheon` or a material multi-perspective decision.
+  - Used by: Zeus
+
 ### Domain 2: Development Tools
 
 7. **clonedeps** — `clonedeps/SKILL.md`
@@ -46,7 +50,7 @@ There are **20 skills** divided into **5 domains**.
 
 9. **incremental-implementation** — `incremental-implementation/SKILL.md`
     - Implement in thin vertical slices — one commit per task, testable, rollback-safe.
-    - Used by: Zeus, Hermes, Aphrodite, Demeter, Prometheus, Talos
+    - Used by: Hermes, Aphrodite, Demeter, Prometheus, Talos
 
 10. **loop-engineering** — `loop-engineering/SKILL.md`
     - Iterative refinement pattern for complex problems — solve, review, improve, repeat in controlled cycles.
@@ -129,7 +133,7 @@ The `SKILL.md` file contains YAML frontmatter with `name`, `description`, and pl
 | 6 | session-goal | Orchestration & Workflow | Zeus, Mnemosyne |
 | 7 | clonedeps | Development Tools | Hermes, Hephaestus |
 | 8 | git-workflow-and-versioning | Development Tools | Hermes, Prometheus, Iris |
-| 9 | incremental-implementation | Development Tools | Zeus, Hermes, Aphrodite, Demeter, Prometheus, Talos |
+| 9 | incremental-implementation | Development Tools | Hermes, Aphrodite, Demeter, Prometheus, Talos |
 | 10 | loop-engineering | Development Tools | All agents |
 | 11 | reflect | Development Tools | All agents |
 | 12 | simplify | Development Tools | All agents |
@@ -141,6 +145,7 @@ The `SKILL.md` file contains YAML frontmatter with `name`, `description`, and pl
 | 18 | spec-driven-development | Planning | Athena |
 | 19 | verification-planning | Planning | Athena, Zeus |
 | 20 | visual-review-pipeline | Frontend Development | Aphrodite |
+| 21 | council-synthesis | Orchestration & Workflow | Zeus |
 
 ---
 

package/src/skills/agent-coordination/SKILL.md

@@ -1,95 +1,25 @@
 ---
 name: agent-coordination
-description: "Multi-agent orchestration with model routing, category delegation, and sprint management. Use for coordinating Pantheon agents."
+description: "Coordinate two or more specialists for substantial multi-phase work. Skip for questions, bounded fixes, and one-agent tasks."
 context: fork
 globs: []
 alwaysApply: false
 ---
 
-# Agent Coordination
+# Agent coordination
 
-Master guide to orchestrating the multi-agent system. Includes category routing, model selection, and sprint management for rapid, TDD-driven feature development.
+Load only when the work genuinely needs multiple specialists or dependent phases. For the default routing, retry, and safety rules, follow Zeus and `pantheon://routing`.
 
----
-
-## Agent Hierarchy
-
-See `AGENTS.md` for the full agent table.
-
----
-
-## Category Routing
-
-Route tasks to optimized agents based on category instead of specifying agents manually:
-
-| Category | Agent | Use For |
-|----------|-------|---------|
-| **Deep** | Athena, Hephaestus | Complex planning, AI pipelines |
-| **Quick** | Talos, Apollo | Fast fixes, codebase search |
-| **Ultrabrain** | Zeus | Multi-agent orchestration, model routing |
-
----
-
-## Sprint Workflow
-
-```
-1. User describes feature
-2. @athena creates PLAN with phases and agents
-3. User approves PLAN (GATE 1)
-4. Zeus dispatches agents to phases (parallel when possible)
-5. Each agent: TDD cycle → writes IMPL artifact
-6. @themis reviews all phases (GATE 2)
-7. User reviews Themis findings
-8. If approved → user commits (GATE 3)
-9. @mnemosyne updates memory bank
-```
-
----
-
-## Parallel Execution Declaration
+## Lean coordination
 
-```
-🔀 PARALLEL EXECUTION — Phase 2
-Running simultaneously:
-- @hermes   → backend tests   → .tmp/IMPL-phase2-hermes.md
-- @aphrodite → frontend       → .tmp/IMPL-phase2-aphrodite.md
-- @demeter  → migrations      → .tmp/IMPL-phase2-demeter.md
-Themis reviews all three after completion.
-```
-
----
-
-## Artifact Protocol
-
-All phase outputs go to `.pantheon/memory-bank/.tmp/`. See `skill: artifact-management` for the complete protocol (who generates what, templates, lifecycle).
-
----
-
-## Safety Gates
-
-| Gate | When | Why |
-|------|------|-----|
-| **GATE 1** | After PLAN | User confirms scope |
-| **GATE 2** | After Themis review | User sees changes |
-| **GATE 3** | Before git commit | User controls history |
-
----
-
-## Model Selection
-
-Model routing is handled by Prometheus and the platform configuration. Zeus delegates model selection to the tier system (`fast` / `default` / `coding` / `premium`).
-
-- **Deep tasks** → High-quality models (Claude, GPT-4)
-- **Quick tasks** → Fast/cheap models (Haiku, 4o-mini)
-- **Visual tasks** → Multimodal models (GPT-4V, Claude Vision)
-- **Cost optimization** → Route to cheapest model that meets quality bar
-
----
+1. State the desired outcome and split only work with distinct ownership or dependencies. Ask @athena to plan only when scope, architecture, or acceptance criteria are materially unclear.
+2. Assign each bounded task to the most specific configured specialist. Include relevant context, expected result, and focused verification; do not make @apollo rediscover context the implementer can inspect directly.
+3. Dispatch independent tasks together only when native background delegation is enabled. Otherwise use normal `task()` calls. Do not create waves for a single task or poll without a returned task ID.
+4. Collect summaries once, resolve dependencies, and report evidence, remaining risk, and blockers. Add @themis, artifacts, or user approval only where the task's risk or requested workflow requires them.
 
-## Anti-Patterns
+## Boundaries
 
-- ❌ Zeus doing implementation work (orchestrates, doesn't code)
-- ❌ Skipping TDD cycle (RED → GREEN → REFACTOR)
-- ❌ Auto-committing without user approval
-- ❌ Parallel phases with dependencies (order matters)
-- ❌ Ignoring Themis CRITICAL findings
+- Zeus coordinates; implementation agents inspect, edit, and verify their assigned scope.
+- Preserve auth/security, data/schema, destructive-action, production, and permission gates.
+- Do not commit, push, merge, deploy, or broaden scope without explicit authorization.
+- Model selection comes from the active configured preset; do not invent a model tier or price.

package/src/skills/artifact-management/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: artifact-management
-description: "Structured artifact trail for feature implementations — plans, implementations, reviews, and decisions."
+description: "Structured artifacts for planned, multi-phase, or materially risky implementation work."
 context: fork
 globs: []
 alwaysApply: false
@@ -8,7 +8,7 @@ alwaysApply: false
 
 # Artifact Management
 
-Structured artifact trail system for documenting feature implementations. Defines what, where, and how agents produce phase outputs.
+Structured artifact trail for planned, multi-phase, or materially risky implementation work. Artifacts are not a prerequisite for simple fixes or read-only investigation.
 
 ---
 
@@ -55,6 +55,13 @@ All ephemeral artifacts live in `.pantheon/memory-bank/.tmp/` — gitignored and
 
 > **Zeus does NOT generate artifacts.** He orchestrates agents that generate them.
 
+## Proportional Use
+
+- Create PLAN/IMPL/REVIEW artifacts for planned multi-phase work or when risk, auditability, or an explicit request needs a durable phase trail.
+- A bounded, reversible fix should be investigated only as far as needed, implemented, verified with focused checks, and returned as a concise summary. Do not create PLAN, IMPL, or REVIEW artifacts for it.
+- Read-only lookup or trivial discovery needs no DISC artifact; use one only when findings are broad enough to be reused or drive a material decision.
+- The absence of an artifact never waives Themis review or human approval for auth/security, data/schema/migrations, production-impacting work, or sensitive actions.
+
 ---
 
 ## Templates
@@ -103,9 +110,9 @@ All ephemeral artifacts live in `.pantheon/memory-bank/.tmp/` — gitignored and
 
 ## Human Pause Points
 
-1. **After PLAN** → user reads `.tmp/PLAN-<feature>.md` and approves
-2. **After REVIEW** → user reads `.tmp/REVIEW-<feature>.md` and validates focus items
-3. **Before git commit** → user executes manually
+1. **After a required PLAN** → pause only when a material scope/risk decision needs approval.
+2. **After required REVIEW** → preserve auth/security/data/schema and other sensitive approval gates.
+3. **Before commit/push/merge/deploy or destructive/global action** → require explicit human approval.
 
 ---