@nikolasp98/minion 2026.6.14-dev.20260616061802 → 2026.6.16-dev

package/docs/superpowers/specs/2026-06-10-recon-quick-wins-and-silent-agent-design.md

@@ -0,0 +1,151 @@
+# Design: Recon Quick Wins + Hermes Silent Agent
+
+**Date:** 2026-06-10
+**Branch/Worktree:** `feat/recon-quick-wins` @ `/home/nikolas/Documents/CODE/minion-quick-wins`
+**Source:** Absorption picks from `recon-openclaw-hermes-agentzero-2026-06-10.md`
+**Scope decisions (user, 2026-06-10):** one worktree, two phases · silent agent heuristic-gated · **on by default** (with kill-switch + safety rails).
+
+> Reminder: minion is its own independent project. OpenClaw/Hermes/Agent Zero are peer idea-sources; we port designs, not couple repos.
+
+---
+
+## Goal
+
+Land the low-risk "quick wins" from the recon, then the strategic Hermes self-improvement pair (two-file memory + silent review agent), in one phased worktree. Each component is an independent atomic commit. The silent agent is the headline: it turns minion's existing single-shot post-turn memory extractor into a forked review sub-agent that also writes/patches skills.
+
+## Existing scaffolding we build on (already in minion)
+
+- `src/agents/pi-embedded-runner/lifecycle-hooks.ts` — "from Hermes's 4-hook model": `preLlmCall` / `postLlmCall` / `onSessionStart` / `onSessionEnd`. Trajectory + silent agent hook here.
+- `src/memory/extract-memories-runner.ts` — fire-and-forget post-turn extractor ("NEVER delays the user's next message — runs in a detached microtask"), gated by `shouldExtract(...)`. The silent agent **evolves** this; we reuse its gate + detached-microtask pattern.
+- `src/agents/system-prompt/system-prompt.ts` `buildAgentSystemPrompt(...)` — injection point for `behaviour.md` + two-file memory.
+- `src/plugin-sdk/` — `registerHttpRoute(...)` and `registerWebhookTarget(...)` already exist → webhooks/admin-http-rpc are pure extension work.
+- `src/agents/skills/` — workspace skills (`bundled-context.ts`, `workspace.ts`, `serialize.ts`) → progressive disclosure + `skill_write` plug in here.
+- pgvector `agent_memories` store stays as the **searchable long-term** memory; two-file model is the **always-injected working set**. Distinct roles, no conflict.
+
+---
+
+## Phase 1 — Independent quick wins (ship first, atomic commits)
+
+### 1.1 tool-call-repair
+
+- **What:** Normalize/repair malformed streaming tool-call JSON (partial JSON, trailing commas, concatenated objects, unescaped newlines) before parse. We already hit this with OpenRouter.
+- **Where:** `src/llm/tool-call-repair/` (internal module — minion's `packages/` are app-level, so not a published package). Consumed by the streaming tool-call accumulator in the LLM layer.
+- **Interface:** `repairToolCallArguments(raw: string): { value: string; repaired: boolean }` + a streaming accumulator helper.
+- **Tests:** fixture-driven (known-bad provider payloads → expected repaired JSON). Reference: OpenClaw `packages/tool-call-repair`.
+- **Config:** none (always-on; pure resilience, no behavior change on valid input).
+
+### 1.2 Trajectory recording
+
+- **What:** Per-session JSONL of turn events (model call meta, tool calls, content, usage) with size caps + pointer file + export. Foundation for replay/eval.
+- **Where:** new `src/trajectory/` — `types.ts`, `runtime.ts`, `writer.ts` (append + rotation), `export.ts`. Storage under the session state dir.
+- **Wiring:** `lifecycle-hooks` `postLlmCall` (+ tool dispatch) emit events; never blocks the loop.
+- **Config:** `trajectory.enabled` (default **off** — observability opt-in).
+- **Reference:** OpenClaw `src/trajectory/`.
+
+### 1.3 tokenjuice (tool-output reduction)
+
+- **What:** Reduce large/repetitive bash/exec tool output before it reaches the model (collapse repeated lines, head/tail windows, dedupe). Token savings on code-heavy agents.
+- **Where:** runner middleware `src/agents/pi-embedded-runner/tool-result-reduce.ts`, applied alongside the existing `tool-result-truncation.ts` before results are appended to history.
+- **Config:** `tools.outputReduce.enabled` (default **on** for shell/exec-class tools; conservative thresholds).
+- **Reference:** OpenClaw `extensions/tokenjuice`.
+
+### 1.4 webhooks extension
+
+- **What:** Authenticated inbound HTTP POST → trigger an agent/flow. Missing piece for CRM/booking/payment integration.
+- **Where:** `extensions/webhooks/` using `registerHttpRoute` + secret (HMAC/bearer) verification; maps a configured path → dispatch.
+- **Config:** per-route secret + target agent/flow in extension config.
+- **Reference:** OpenClaw `extensions/webhooks`.
+
+### 1.5 admin-http-rpc extension
+
+- **What:** Curated subset of gateway RPC methods over plain HTTP for ops/CI (health/status/config/channels/models/agents/cron/tasks/diagnostics) — no WS client needed.
+- **Where:** `extensions/admin-http-rpc/` using `registerHttpRoute`; **method allowlist** + bearer auth; read-leaning by default.
+- **Config:** bearer token + method allowlist.
+- **Reference:** OpenClaw `extensions/admin-http-rpc`.
+
+### 1.6 Progressive skill disclosure
+
+- **What:** Inject only skill **name + description** by default (~3k tok cap), load full SKILL.md on demand. Cuts skill token overhead.
+- **Where:** `src/agents/skills/` prompt-build path (`bundled-context.ts`/`serialize.ts`); add a `skill_view(name)` tool that returns full body (and `skill_view(name, path)` for support files).
+- **Config:** `skills.progressiveDisclosure` (default **on**; legacy full-load behind the flag for rollback).
+- **Reference:** Hermes `agent/skill_utils.py` three-level load.
+
+### 1.7 behaviour.md self-modification tool (Agent Zero)
+
+- **What:** Per-agent `behaviour.md` ruleset, injected into the system prompt; an `update_behaviour` tool lets the agent merge adjustments into it (via a utility model) — persists across restarts, human-readable/editable.
+- **Where:** `behaviour.md` in the agent workspace; injected in `buildAgentSystemPrompt`; tool in agent tools; merge via utility model.
+- **Config:** `agents.*.behaviourFile` (default **on**; empty file = no-op injection).
+- **Safety:** injection-scanned before write (reuse `src/memory/injection-scanner.ts`); bounded size.
+- **Reference:** Agent Zero `plugins/_memory/tools/behaviour_adjustment.py`.
+
+---
+
+## Phase 2 — Two-file memory + silent agent (the IP)
+
+### 2.1 Two-file memory model
+
+- **What:** Two frozen-injected flat files per agent: `MEMORY.md` (agent/env working notes, ~2,200 char budget) and `USER.md` (who the user is, ~1,375 char budget). Injected as one system-prompt block at session start (stabilizes the prefix cache). A `memory` tool with three actions: **add / replace / remove** via substring matching. Char budget enforced at write → forces consolidation/eviction (no unbounded growth).
+- **Where:** files in the agent workspace; injection in `buildAgentSystemPrompt`; `memory` tool in agent tools. The current single-MEMORY.md auto-extract target is replaced by this two-file model.
+- **Relationship to pgvector:** pgvector `agent_memories` remains the **searchable long-term store** (RAG/recall). Two-file is the **always-in-context working set**. The silent agent may write to both (flat for working set, sink for long-term).
+- **Config:** budgets configurable; on by default.
+- **Reference:** Hermes `agent/memory_manager.py`, `docs/.../memory.md`.
+
+### 2.2 skill_write capability + self-authored provenance
+
+- **What:** A guarded tool to create/patch workspace skills (write SKILL.md + support files), so the silent agent can persist learned skills. 4-tier write preference (patch loaded skill → patch umbrella → add support file → new umbrella).
+- **Where:** `src/agents/skills/` write path; `skill_write` tool. Reuse workspace skill storage.
+- **Provenance marker:** every skill written by `skill_write`/the silent agent is stamped in SKILL.md frontmatter with `metadata.minion.origin: self-authored` (+ author agent id + timestamp). This flag rides through the existing skill-entry serialization (same path that carries the `bundled` flag) so the hub can segregate it. Hand-authored/bundled skills are unaffected.
+- **Safety:** injection-scan skill content before write; size cap; writes only into the agent's own workspace skills dir.
+
+### 2.4 Hub changes (cross-repo: `minion_hub`)
+
+Two hub touches, both fed by gateway-side flags/events emitted from this worktree:
+
+**(a) Self-authored skill label**
+
+- **What:** Surface self-authored skills in each agent's **Skill Registry** with a distinct pill (e.g. `self-authored`, accent-colored) next to the existing `bundled` pill, for visual segregation. (See user screenshot: pills render inline after the skill name.)
+- **Gateway side (this worktree):** ensure the `origin: self-authored` flag is included in the skill entries the gateway sends to the hub (`listSkills`/skills RPC payload) — mirror how `bundled` is already propagated.
+- **Hub side (`minion_hub`, separate repo/branch):** in the Skill Registry component that renders the `bundled` `<Pill>`, add a `self-authored` variant gated on the new flag; optional filter facet alongside All/Enabled/Disabled. Svelte 5 runes, reuse existing pill styling.
+- **Reference:** existing `bundled` pill in the hub Skill Registry (screenshot 2026-06-10).
+
+**(b) Silent-agent "action" block in chat**
+
+- Render the new `action` chat message part (from §2.3) as an inline block in the hub chat thread, mirroring the existing "Reasoning" block component (italic, labeled header, screenshot 2026-06-10). Distinct label/icon ("Action" / self-improvement).
+- Gateway side (this worktree): emit the `action` part over the chat stream (mirror `reasoning`). Hub side: add the renderer + the shared protocol part type (`@minion-stack/shared`, if reasoning lives there).
+
+**Note:** §2.4 is the only `minion_hub` (SvelteKit/Bun) work — coordinated in a separate hub branch. Per cross-project impact: skill-format + chat-part change → gateway runtime + `@minion-stack/shared` + hub registry/chat.
+
+### 2.3 Silent agent (background self-review) — headline
+
+- **What:** After a heuristic-gated turn end, **fire-and-forget** spawn a review sub-agent fed the conversation snapshot, **tool-whitelisted to `[memory, skill_write]`** (enforced at the dispatch layer — full schema kept for cache parity, guarded at runtime, per Hermes). It runs the adapted Hermes `_SKILL_REVIEW_PROMPT`: distinguish class-level skills from user-profile facts, choose the right write tier, and explicitly _"do NOT capture transient errors / negative claims"_ — "nothing to save is valid and must not be the default."
+- **Where:** `src/agents/self-improvement/background-review.ts`. Reuses `extract-memories-runner`'s detached-microtask + `shouldExtract`-style gate.
+- **Surfacing (decided):** emit an **"action" block** into the hub chat stream — a new chat message part type `action` that mirrors the existing `reasoning` part (`src/gateway/open-responses.schema.ts` `type: "reasoning"`; rendered by the hub as the italic "Reasoning" block in the screenshot). The action block summarizes what the review did (e.g. _"Updated skill `clinic-intake`; added 1 user-profile fact"_) or notes a no-op. Carried as a chat event so it lands inline in the conversation like reasoning. Not silent/log-only.
+- **Trigger (decided):** heuristic-gated — substantive turn / message-count threshold / tool-call activity (reuse existing gate).
+- **Default (decided):** **on by default**, with safety rails:
+  - kill-switch config `agents.*.selfImprovement.enabled` (default true) + global `selfImprovement.enabled`.
+  - never blocks the user turn (detached); hard timeout; failures swallowed + logged.
+  - conservative prompt (bias to no-op); tool whitelist; char budgets; injection-scan every write.
+  - per-turn cost cap / debounce so it can't run on every trivial message.
+- **Reference:** Hermes `agent/background_review.py:1-609` (the `_SKILL_REVIEW_PROMPT` at L45-148 is the adopted IP).
+
+---
+
+## Cross-cutting
+
+- **Error handling:** every new subsystem is fail-safe — trajectory, tokenjuice, silent agent, behaviour merge all swallow+log errors and never break the user-facing turn.
+- **Config surface:** all gated under existing agent/gateway config; defaults chosen per decisions above.
+- **Testing:** unit tests per module (tool-call-repair fixtures, two-file budget enforcement, substring add/replace/remove, gate logic, reducer thresholds); e2e for the silent-agent loop with a mock LLM (assert it writes memory/skill and never blocks).
+- **`tsgo` + `pnpm check` + `pnpm test`** green before each phase merge.
+
+## Out of scope (deferred)
+
+- Skill lifecycle states + curator (Hermes) — later.
+- Custom CodeQL queries — meta-repo CI concern, separate.
+- talk/agent-consult, workboard, memory-wiki, tasks registry, video-gen, context-engine — Part-4 strategic items, separate worktrees.
+- Full re-baseline against OpenClaw — explicitly not a goal.
+
+## Build order
+
+1. tool-call-repair → 2. tokenjuice → 3. trajectory → 4. progressive skill disclosure → 5. behaviour.md tool → 6. webhooks → 7. admin-http-rpc → **(Phase 2)** 8. two-file memory → 9. skill_write (+ self-authored provenance) → 10. silent agent (+ emit `action` chat part) → 11. hub changes (`minion_hub`): self-authored pill + `action` block renderer.
+
+(1–7 independent; 8→9→10 ordered by dependency; 11 depends on 9's provenance flag + 10's `action` part, and is the only `minion_hub` work — coordinated in a separate hub branch.)

package/extensions/alert-watcher/index.js

@@ -2491,6 +2491,23 @@ function resolveBundledSkillsDir(opts = {}) {
 	} catch {}
 }
 //#endregion
+//#region src/agents/skills/compact-prompt.ts
+/**
+* Render a compact, name+description-only skills listing for progressive
+* disclosure. The full SKILL.md body is deliberately omitted; the agent is
+* instructed to load it on demand via the `skill_view` tool. This keeps the
+* system prompt small when many skills are available.
+*/
+function formatSkillsCompact(skills) {
+	if (skills.length === 0) return "";
+	return [
+		"<available_skills>",
+		...skills.map((s) => `- **${s.name}** — ${s.description.trim()}`),
+		"</available_skills>",
+		"Call the `skill_view` tool with a skill name to load its full instructions before using it."
+	].join("\n");
+}
+//#endregion
 //#region src/shared/config-eval.ts
 function isTruthyValue(value) {
 	if (value === void 0 || value === null) return false;
@@ -4239,6 +4256,35 @@ function applySkillsPromptLimits(params) {
 		truncatedReason
 	};
 }
+/**
+* Resolve whether progressive disclosure is enabled. Default is ON; only an
+* explicit `false` disables it (inlining full skill bodies in the prompt).
+*/
+function resolveProgressiveDisclosure(config, override) {
+	if (override !== void 0) return override;
+	return config?.skills?.progressiveDisclosure !== false;
+}
+/**
+* Map prompt-eligible skill entries to compact `{ name, description, body }`
+* records. The body is the full SKILL.md text, read from disk on demand; if the
+* file can't be read we fall back to an empty body (compact mode omits bodies
+* from the prompt anyway, so this only affects callers that inspect `body`).
+*/
+function toCompactSkills(skills) {
+	return skills.map((skill) => {
+		let body = "";
+		try {
+			body = fs.readFileSync(skill.filePath, "utf-8");
+		} catch {
+			body = "";
+		}
+		return {
+			name: skill.name,
+			description: skill.description ?? "",
+			body
+		};
+	});
+}
 function buildWorkspaceSkillSnapshot(workspaceDir, opts) {
 	const eligible = filterSkillEntries(opts?.entries ?? loadSkillEntries(workspaceDir, opts), opts?.config, opts?.skillFilter, opts?.eligibility, opts?.denyFilter);
 	const resolvedSkills = eligible.filter((entry) => entry.invocation?.disableModelInvocation !== true).map((entry) => entry.skill);
@@ -4250,7 +4296,7 @@ function buildWorkspaceSkillSnapshot(workspaceDir, opts) {
 	const prompt = [
 		remoteNote,
 		truncated ? `⚠️ Skills truncated: included ${skillsForPrompt.length} of ${resolvedSkills.length}. Run \`openclaw skills check\` to audit.` : "",
-		formatSkillsForPrompt(compactSkillPaths(skillsForPrompt))
+		resolveProgressiveDisclosure(opts?.config, opts?.progressive) ? formatSkillsCompact(toCompactSkills(skillsForPrompt)) : formatSkillsForPrompt(compactSkillPaths(skillsForPrompt))
 	].filter(Boolean).join("\n");
 	const skillFilter = normalizeSkillFilter(opts?.skillFilter);
 	const denyFilter = normalizeSkillFilter(opts?.denyFilter);

package/extensions/bluebubbles/index.js

@@ -1769,6 +1769,7 @@ const AgentSandboxSchema = z.object({
 	]).optional(),
 	perSession: z.boolean().optional(),
 	workspaceRoot: z.string().optional(),
+	tier0: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 	docker: SandboxDockerSchema,
 	browser: SandboxBrowserSchema,
 	prune: SandboxPruneSchema
@@ -1970,7 +1971,8 @@ const ToolsSchema = z.object({
 	exec: ToolExecSchema,
 	fs: ToolFsSchema,
 	subagents: z.object({ tools: ToolPolicySchema$1 }).strict().optional(),
-	sandbox: z.object({ tools: ToolPolicySchema$1 }).strict().optional()
+	sandbox: z.object({ tools: ToolPolicySchema$1 }).strict().optional(),
+	outputReduce: z.object({ enabled: z.boolean().optional() }).strict().optional()
 }).strict().superRefine((value, ctx) => {
 	addAllowAlsoAllowConflictIssue(value, ctx, "tools cannot set both allow and alsoAllow in the same scope (merge alsoAllow into allow, or remove allow and use profile + alsoAllow)");
 }).optional();
@@ -4736,6 +4738,7 @@ z.object({
 			importContacts: z.boolean().optional()
 		}).strict().optional(),
 		memorySync: z.object({ enabled: z.boolean().optional() }).strict().optional(),
+		trajectory: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 		personalAgentObservations: z.object({
 			enabled: z.boolean().optional(),
 			dbPath: z.string().optional()
@@ -4848,6 +4851,7 @@ z.object({
 	memory: MemorySchema,
 	skills: z.object({
 		allowBundled: z.array(z.string()).optional(),
+		progressiveDisclosure: z.boolean().optional(),
 		load: z.object({
 			extraDirs: z.array(z.string()).optional(),
 			watch: z.boolean().optional(),

package/extensions/feishu/index.js

@@ -2313,6 +2313,7 @@ const AgentSandboxSchema = z.object({
 	]).optional(),
 	perSession: z.boolean().optional(),
 	workspaceRoot: z.string().optional(),
+	tier0: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 	docker: SandboxDockerSchema,
 	browser: SandboxBrowserSchema,
 	prune: SandboxPruneSchema
@@ -2514,7 +2515,8 @@ const ToolsSchema = z.object({
 	exec: ToolExecSchema,
 	fs: ToolFsSchema,
 	subagents: z.object({ tools: ToolPolicySchema }).strict().optional(),
-	sandbox: z.object({ tools: ToolPolicySchema }).strict().optional()
+	sandbox: z.object({ tools: ToolPolicySchema }).strict().optional(),
+	outputReduce: z.object({ enabled: z.boolean().optional() }).strict().optional()
 }).strict().superRefine((value, ctx) => {
 	addAllowAlsoAllowConflictIssue(value, ctx, "tools cannot set both allow and alsoAllow in the same scope (merge alsoAllow into allow, or remove allow and use profile + alsoAllow)");
 }).optional();
@@ -5280,6 +5282,7 @@ z.object({
 			importContacts: z.boolean().optional()
 		}).strict().optional(),
 		memorySync: z.object({ enabled: z.boolean().optional() }).strict().optional(),
+		trajectory: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 		personalAgentObservations: z.object({
 			enabled: z.boolean().optional(),
 			dbPath: z.string().optional()
@@ -5392,6 +5395,7 @@ z.object({
 	memory: MemorySchema,
 	skills: z.object({
 		allowBundled: z.array(z.string()).optional(),
+		progressiveDisclosure: z.boolean().optional(),
 		load: z.object({
 			extraDirs: z.array(z.string()).optional(),
 			watch: z.boolean().optional(),

package/extensions/flows/index.js

@@ -2743,6 +2743,23 @@ function resolveBundledSkillsDir(opts = {}) {
 	} catch {}
 }
 //#endregion
+//#region src/agents/skills/compact-prompt.ts
+/**
+* Render a compact, name+description-only skills listing for progressive
+* disclosure. The full SKILL.md body is deliberately omitted; the agent is
+* instructed to load it on demand via the `skill_view` tool. This keeps the
+* system prompt small when many skills are available.
+*/
+function formatSkillsCompact(skills) {
+	if (skills.length === 0) return "";
+	return [
+		"<available_skills>",
+		...skills.map((s) => `- **${s.name}** — ${s.description.trim()}`),
+		"</available_skills>",
+		"Call the `skill_view` tool with a skill name to load its full instructions before using it."
+	].join("\n");
+}
+//#endregion
 //#region src/shared/config-eval.ts
 function isTruthyValue(value) {
 	if (value === void 0 || value === null) return false;
@@ -4491,6 +4508,35 @@ function applySkillsPromptLimits(params) {
 		truncatedReason
 	};
 }
+/**
+* Resolve whether progressive disclosure is enabled. Default is ON; only an
+* explicit `false` disables it (inlining full skill bodies in the prompt).
+*/
+function resolveProgressiveDisclosure(config, override) {
+	if (override !== void 0) return override;
+	return config?.skills?.progressiveDisclosure !== false;
+}
+/**
+* Map prompt-eligible skill entries to compact `{ name, description, body }`
+* records. The body is the full SKILL.md text, read from disk on demand; if the
+* file can't be read we fall back to an empty body (compact mode omits bodies
+* from the prompt anyway, so this only affects callers that inspect `body`).
+*/
+function toCompactSkills(skills) {
+	return skills.map((skill) => {
+		let body = "";
+		try {
+			body = fs.readFileSync(skill.filePath, "utf-8");
+		} catch {
+			body = "";
+		}
+		return {
+			name: skill.name,
+			description: skill.description ?? "",
+			body
+		};
+	});
+}
 function buildWorkspaceSkillSnapshot(workspaceDir, opts) {
 	const eligible = filterSkillEntries(opts?.entries ?? loadSkillEntries(workspaceDir, opts), opts?.config, opts?.skillFilter, opts?.eligibility, opts?.denyFilter);
 	const resolvedSkills = eligible.filter((entry) => entry.invocation?.disableModelInvocation !== true).map((entry) => entry.skill);
@@ -4502,7 +4548,7 @@ function buildWorkspaceSkillSnapshot(workspaceDir, opts) {
 	const prompt = [
 		remoteNote,
 		truncated ? `⚠️ Skills truncated: included ${skillsForPrompt.length} of ${resolvedSkills.length}. Run \`openclaw skills check\` to audit.` : "",
-		formatSkillsForPrompt(compactSkillPaths(skillsForPrompt))
+		resolveProgressiveDisclosure(opts?.config, opts?.progressive) ? formatSkillsCompact(toCompactSkills(skillsForPrompt)) : formatSkillsForPrompt(compactSkillPaths(skillsForPrompt))
 	].filter(Boolean).join("\n");
 	const skillFilter = normalizeSkillFilter(opts?.skillFilter);
 	const denyFilter = normalizeSkillFilter(opts?.denyFilter);
@@ -5379,6 +5425,7 @@ const AgentSandboxSchema = z.object({
 	]).optional(),
 	perSession: z.boolean().optional(),
 	workspaceRoot: z.string().optional(),
+	tier0: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 	docker: SandboxDockerSchema,
 	browser: SandboxBrowserSchema,
 	prune: SandboxPruneSchema
@@ -5580,7 +5627,8 @@ const ToolsSchema = z.object({
 	exec: ToolExecSchema,
 	fs: ToolFsSchema,
 	subagents: z.object({ tools: ToolPolicySchema }).strict().optional(),
-	sandbox: z.object({ tools: ToolPolicySchema }).strict().optional()
+	sandbox: z.object({ tools: ToolPolicySchema }).strict().optional(),
+	outputReduce: z.object({ enabled: z.boolean().optional() }).strict().optional()
 }).strict().superRefine((value, ctx) => {
 	addAllowAlsoAllowConflictIssue(value, ctx, "tools cannot set both allow and alsoAllow in the same scope (merge alsoAllow into allow, or remove allow and use profile + alsoAllow)");
 }).optional();
@@ -8208,6 +8256,7 @@ z.object({
 			importContacts: z.boolean().optional()
 		}).strict().optional(),
 		memorySync: z.object({ enabled: z.boolean().optional() }).strict().optional(),
+		trajectory: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 		personalAgentObservations: z.object({
 			enabled: z.boolean().optional(),
 			dbPath: z.string().optional()
@@ -8320,6 +8369,7 @@ z.object({
 	memory: MemorySchema,
 	skills: z.object({
 		allowBundled: z.array(z.string()).optional(),
+		progressiveDisclosure: z.boolean().optional(),
 		load: z.object({
 			extraDirs: z.array(z.string()).optional(),
 			watch: z.boolean().optional(),

package/extensions/gmail-calendar/index.js

@@ -2530,6 +2530,23 @@ function resolveBundledSkillsDir(opts = {}) {
 	} catch {}
 }
 //#endregion
+//#region src/agents/skills/compact-prompt.ts
+/**
+* Render a compact, name+description-only skills listing for progressive
+* disclosure. The full SKILL.md body is deliberately omitted; the agent is
+* instructed to load it on demand via the `skill_view` tool. This keeps the
+* system prompt small when many skills are available.
+*/
+function formatSkillsCompact(skills) {
+	if (skills.length === 0) return "";
+	return [
+		"<available_skills>",
+		...skills.map((s) => `- **${s.name}** — ${s.description.trim()}`),
+		"</available_skills>",
+		"Call the `skill_view` tool with a skill name to load its full instructions before using it."
+	].join("\n");
+}
+//#endregion
 //#region src/shared/config-eval.ts
 function isTruthyValue(value) {
 	if (value === void 0 || value === null) return false;
@@ -4278,6 +4295,35 @@ function applySkillsPromptLimits(params) {
 		truncatedReason
 	};
 }
+/**
+* Resolve whether progressive disclosure is enabled. Default is ON; only an
+* explicit `false` disables it (inlining full skill bodies in the prompt).
+*/
+function resolveProgressiveDisclosure(config, override) {
+	if (override !== void 0) return override;
+	return config?.skills?.progressiveDisclosure !== false;
+}
+/**
+* Map prompt-eligible skill entries to compact `{ name, description, body }`
+* records. The body is the full SKILL.md text, read from disk on demand; if the
+* file can't be read we fall back to an empty body (compact mode omits bodies
+* from the prompt anyway, so this only affects callers that inspect `body`).
+*/
+function toCompactSkills(skills) {
+	return skills.map((skill) => {
+		let body = "";
+		try {
+			body = fs$1.readFileSync(skill.filePath, "utf-8");
+		} catch {
+			body = "";
+		}
+		return {
+			name: skill.name,
+			description: skill.description ?? "",
+			body
+		};
+	});
+}
 function buildWorkspaceSkillSnapshot(workspaceDir, opts) {
 	const eligible = filterSkillEntries(opts?.entries ?? loadSkillEntries(workspaceDir, opts), opts?.config, opts?.skillFilter, opts?.eligibility, opts?.denyFilter);
 	const resolvedSkills = eligible.filter((entry) => entry.invocation?.disableModelInvocation !== true).map((entry) => entry.skill);
@@ -4289,7 +4335,7 @@ function buildWorkspaceSkillSnapshot(workspaceDir, opts) {
 	const prompt = [
 		remoteNote,
 		truncated ? `⚠️ Skills truncated: included ${skillsForPrompt.length} of ${resolvedSkills.length}. Run \`openclaw skills check\` to audit.` : "",
-		formatSkillsForPrompt(compactSkillPaths(skillsForPrompt))
+		resolveProgressiveDisclosure(opts?.config, opts?.progressive) ? formatSkillsCompact(toCompactSkills(skillsForPrompt)) : formatSkillsForPrompt(compactSkillPaths(skillsForPrompt))
 	].filter(Boolean).join("\n");
 	const skillFilter = normalizeSkillFilter(opts?.skillFilter);
 	const denyFilter = normalizeSkillFilter(opts?.denyFilter);

package/extensions/googlechat/index.js

@@ -1770,6 +1770,7 @@ const AgentSandboxSchema = z.object({
 	]).optional(),
 	perSession: z.boolean().optional(),
 	workspaceRoot: z.string().optional(),
+	tier0: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 	docker: SandboxDockerSchema,
 	browser: SandboxBrowserSchema,
 	prune: SandboxPruneSchema
@@ -1971,7 +1972,8 @@ const ToolsSchema = z.object({
 	exec: ToolExecSchema,
 	fs: ToolFsSchema,
 	subagents: z.object({ tools: ToolPolicySchema }).strict().optional(),
-	sandbox: z.object({ tools: ToolPolicySchema }).strict().optional()
+	sandbox: z.object({ tools: ToolPolicySchema }).strict().optional(),
+	outputReduce: z.object({ enabled: z.boolean().optional() }).strict().optional()
 }).strict().superRefine((value, ctx) => {
 	addAllowAlsoAllowConflictIssue(value, ctx, "tools cannot set both allow and alsoAllow in the same scope (merge alsoAllow into allow, or remove allow and use profile + alsoAllow)");
 }).optional();
@@ -4737,6 +4739,7 @@ z.object({
 			importContacts: z.boolean().optional()
 		}).strict().optional(),
 		memorySync: z.object({ enabled: z.boolean().optional() }).strict().optional(),
+		trajectory: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 		personalAgentObservations: z.object({
 			enabled: z.boolean().optional(),
 			dbPath: z.string().optional()
@@ -4849,6 +4852,7 @@ z.object({
 	memory: MemorySchema,
 	skills: z.object({
 		allowBundled: z.array(z.string()).optional(),
+		progressiveDisclosure: z.boolean().optional(),
 		load: z.object({
 			extraDirs: z.array(z.string()).optional(),
 			watch: z.boolean().optional(),

package/extensions/irc/index.js

@@ -1771,6 +1771,7 @@ const AgentSandboxSchema = z.object({
 	]).optional(),
 	perSession: z.boolean().optional(),
 	workspaceRoot: z.string().optional(),
+	tier0: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 	docker: SandboxDockerSchema,
 	browser: SandboxBrowserSchema,
 	prune: SandboxPruneSchema
@@ -1972,7 +1973,8 @@ const ToolsSchema = z.object({
 	exec: ToolExecSchema,
 	fs: ToolFsSchema,
 	subagents: z.object({ tools: ToolPolicySchema$1 }).strict().optional(),
-	sandbox: z.object({ tools: ToolPolicySchema$1 }).strict().optional()
+	sandbox: z.object({ tools: ToolPolicySchema$1 }).strict().optional(),
+	outputReduce: z.object({ enabled: z.boolean().optional() }).strict().optional()
 }).strict().superRefine((value, ctx) => {
 	addAllowAlsoAllowConflictIssue(value, ctx, "tools cannot set both allow and alsoAllow in the same scope (merge alsoAllow into allow, or remove allow and use profile + alsoAllow)");
 }).optional();
@@ -4738,6 +4740,7 @@ z.object({
 			importContacts: z.boolean().optional()
 		}).strict().optional(),
 		memorySync: z.object({ enabled: z.boolean().optional() }).strict().optional(),
+		trajectory: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 		personalAgentObservations: z.object({
 			enabled: z.boolean().optional(),
 			dbPath: z.string().optional()
@@ -4850,6 +4853,7 @@ z.object({
 	memory: MemorySchema,
 	skills: z.object({
 		allowBundled: z.array(z.string()).optional(),
+		progressiveDisclosure: z.boolean().optional(),
 		load: z.object({
 			extraDirs: z.array(z.string()).optional(),
 			watch: z.boolean().optional(),

package/extensions/llm-task/index.js

@@ -2491,6 +2491,23 @@ function resolveBundledSkillsDir(opts = {}) {
 	} catch {}
 }
 //#endregion
+//#region src/agents/skills/compact-prompt.ts
+/**
+* Render a compact, name+description-only skills listing for progressive
+* disclosure. The full SKILL.md body is deliberately omitted; the agent is
+* instructed to load it on demand via the `skill_view` tool. This keeps the
+* system prompt small when many skills are available.
+*/
+function formatSkillsCompact(skills) {
+	if (skills.length === 0) return "";
+	return [
+		"<available_skills>",
+		...skills.map((s) => `- **${s.name}** — ${s.description.trim()}`),
+		"</available_skills>",
+		"Call the `skill_view` tool with a skill name to load its full instructions before using it."
+	].join("\n");
+}
+//#endregion
 //#region src/shared/config-eval.ts
 function isTruthyValue(value) {
 	if (value === void 0 || value === null) return false;
@@ -4239,6 +4256,35 @@ function applySkillsPromptLimits(params) {
 		truncatedReason
 	};
 }
+/**
+* Resolve whether progressive disclosure is enabled. Default is ON; only an
+* explicit `false` disables it (inlining full skill bodies in the prompt).
+*/
+function resolveProgressiveDisclosure(config, override) {
+	if (override !== void 0) return override;
+	return config?.skills?.progressiveDisclosure !== false;
+}
+/**
+* Map prompt-eligible skill entries to compact `{ name, description, body }`
+* records. The body is the full SKILL.md text, read from disk on demand; if the
+* file can't be read we fall back to an empty body (compact mode omits bodies
+* from the prompt anyway, so this only affects callers that inspect `body`).
+*/
+function toCompactSkills(skills) {
+	return skills.map((skill) => {
+		let body = "";
+		try {
+			body = fs.readFileSync(skill.filePath, "utf-8");
+		} catch {
+			body = "";
+		}
+		return {
+			name: skill.name,
+			description: skill.description ?? "",
+			body
+		};
+	});
+}
 function buildWorkspaceSkillSnapshot(workspaceDir, opts) {
 	const eligible = filterSkillEntries(opts?.entries ?? loadSkillEntries(workspaceDir, opts), opts?.config, opts?.skillFilter, opts?.eligibility, opts?.denyFilter);
 	const resolvedSkills = eligible.filter((entry) => entry.invocation?.disableModelInvocation !== true).map((entry) => entry.skill);
@@ -4250,7 +4296,7 @@ function buildWorkspaceSkillSnapshot(workspaceDir, opts) {
 	const prompt = [
 		remoteNote,
 		truncated ? `⚠️ Skills truncated: included ${skillsForPrompt.length} of ${resolvedSkills.length}. Run \`openclaw skills check\` to audit.` : "",
-		formatSkillsForPrompt(compactSkillPaths(skillsForPrompt))
+		resolveProgressiveDisclosure(opts?.config, opts?.progressive) ? formatSkillsCompact(toCompactSkills(skillsForPrompt)) : formatSkillsForPrompt(compactSkillPaths(skillsForPrompt))
 	].filter(Boolean).join("\n");
 	const skillFilter = normalizeSkillFilter(opts?.skillFilter);
 	const denyFilter = normalizeSkillFilter(opts?.denyFilter);

package/extensions/matrix/{credentials-B1q3x-N0.mjs → credentials-BtiLgiNR.mjs}

@@ -1768,6 +1768,7 @@ const AgentSandboxSchema = z.object({
 	]).optional(),
 	perSession: z.boolean().optional(),
 	workspaceRoot: z.string().optional(),
+	tier0: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 	docker: SandboxDockerSchema,
 	browser: SandboxBrowserSchema,
 	prune: SandboxPruneSchema
@@ -1969,7 +1970,8 @@ const ToolsSchema = z.object({
 	exec: ToolExecSchema,
 	fs: ToolFsSchema,
 	subagents: z.object({ tools: ToolPolicySchema }).strict().optional(),
-	sandbox: z.object({ tools: ToolPolicySchema }).strict().optional()
+	sandbox: z.object({ tools: ToolPolicySchema }).strict().optional(),
+	outputReduce: z.object({ enabled: z.boolean().optional() }).strict().optional()
 }).strict().superRefine((value, ctx) => {
 	addAllowAlsoAllowConflictIssue(value, ctx, "tools cannot set both allow and alsoAllow in the same scope (merge alsoAllow into allow, or remove allow and use profile + alsoAllow)");
 }).optional();
@@ -4735,6 +4737,7 @@ z.object({
 			importContacts: z.boolean().optional()
 		}).strict().optional(),
 		memorySync: z.object({ enabled: z.boolean().optional() }).strict().optional(),
+		trajectory: z.object({ enabled: z.boolean().optional() }).strict().optional(),
 		personalAgentObservations: z.object({
 			enabled: z.boolean().optional(),
 			dbPath: z.string().optional()
@@ -4847,6 +4850,7 @@ z.object({
 	memory: MemorySchema,
 	skills: z.object({
 		allowBundled: z.array(z.string()).optional(),
+		progressiveDisclosure: z.boolean().optional(),
 		load: z.object({
 			extraDirs: z.array(z.string()).optional(),
 			watch: z.boolean().optional(),

package/extensions/matrix/index.js

@@ -1,5 +1,5 @@
-import { a as setMatrixRuntime, i as getMatrixRuntime } from "./credentials-B1q3x-N0.mjs";
-import { A as listMatrixAccountIds, I as resolveMatrixAuth, M as resolveMatrixAccount, N as resolveMatrixAccountConfig, S as sendPollMatrix, T as resolveMatrixRoomId, _ as RelationType, b as reactMatrixMessage, c as listMatrixDirectoryGroupsLive, d as removeMatrixReactions, f as fetchEventSummary, g as MsgType, h as EventType, i as normalizeMatrixUserId, j as resolveDefaultMatrixAccountId, l as listMatrixDirectoryPeersLive, m as summarizeMatrixRawEvent, n as probeMatrix, p as readPinnedEvents, r as normalizeMatrixAllowList, s as resolveMatrixRoomConfig, t as resolveMatrixTargets, u as listMatrixReactions, v as resolveMatrixActionLimit, x as sendMessageMatrix, y as resolveActionClient } from "./resolve-targets-Dw7tv-7K.mjs";
+import { a as setMatrixRuntime, i as getMatrixRuntime } from "./credentials-BtiLgiNR.mjs";
+import { A as listMatrixAccountIds, I as resolveMatrixAuth, M as resolveMatrixAccount, N as resolveMatrixAccountConfig, S as sendPollMatrix, T as resolveMatrixRoomId, _ as RelationType, b as reactMatrixMessage, c as listMatrixDirectoryGroupsLive, d as removeMatrixReactions, f as fetchEventSummary, g as MsgType, h as EventType, i as normalizeMatrixUserId, j as resolveDefaultMatrixAccountId, l as listMatrixDirectoryPeersLive, m as summarizeMatrixRawEvent, n as probeMatrix, p as readPinnedEvents, r as normalizeMatrixAllowList, s as resolveMatrixRoomConfig, t as resolveMatrixTargets, u as listMatrixReactions, v as resolveMatrixActionLimit, x as sendMessageMatrix, y as resolveActionClient } from "./resolve-targets-ml4iU4Bd.mjs";
 import { createRequire } from "node:module";
 import { MarkdownConfigSchema, ToolPolicySchema, emptyPluginConfigSchema, maybeInjectDisclosure, resolveDisclosureConfig } from "minion/plugin-sdk";
 import { DEFAULT_ACCOUNT_ID, PAIRING_APPROVED_MESSAGE, addWildcardAllowFrom, applyAccountNameToChannelSection, buildChannelConfigSchema, createActionGate, deleteAccountFromConfigSection, formatDocsLink, formatPairingApproveHint, jsonResult, mergeAllowFromEntries, normalizeAccountId, promptChannelAccessConfig, readNumberParam, readReactionParams, readStringParam, setAccountEnabledInConfigSection } from "openclaw/plugin-sdk";
@@ -1298,7 +1298,7 @@ const matrixPlugin = {
 		await previousLock;
 		let monitorMatrixProvider;
 		try {
-			monitorMatrixProvider = (await import("./matrix-B9amoffh.mjs")).monitorMatrixProvider;
+			monitorMatrixProvider = (await import("./matrix-CWlUSNSh.mjs")).monitorMatrixProvider;
 		} finally {
 			releaseLock();
 		}