fix(agent): honor omit_skills_catalog in sub-agent prompts

scripts/ci/rust-coverage-changed.sh

@@ -242,8 +242,11 @@ compile_raw_coverage_target() {
 run_full() {
   log "running FULL instrumented suite (reason: $1)"
   llvm_cov clean --workspace
-  llvm_cov --no-report --no-fail-fast -p openhuman --lib
-  llvm_cov --no-report --no-fail-fast -p openhuman --bins
+  # The full lib/bin suites share process-global configuration, event-bus, and
+  # provider state. Keep them isolated just like the integration targets below
+  # so parallel libtest cases cannot leak state into one another.
+  llvm_cov --no-report --no-fail-fast -p openhuman --lib -- --test-threads=1
+  llvm_cov --no-report --no-fail-fast -p openhuman --bins -- --test-threads=1
   while IFS= read -r target; do
     [ -n "${target}" ] || continue
     log "running full-suite integration target: ${target}"
@@ -407,7 +410,7 @@ llvm_cov clean --workspace
 if [ "${#lib_filters[@]}" -gt 0 ]; then
   log "running scoped lib unit tests with filters: ${lib_filters[*]}"
   # libtest ORs multiple positional filters — one run covers all domains.
-  run_counted llvm_cov --no-report --no-fail-fast -p openhuman --lib -- "${lib_filters[@]}"
+  run_counted llvm_cov --no-report --no-fail-fast -p openhuman --lib -- "${lib_filters[@]}" --test-threads=1
 fi
 
 if [ "${#test_targets[@]}" -gt 0 ]; then

src/openhuman/agent/harness/builtin_definitions.rs

@@ -58,6 +58,7 @@ pub(crate) fn test_main_def() -> AgentDefinition {
         omit_identity: true,
         omit_memory_context: true,
         omit_safety_preamble: true,
+        omit_skills_catalog: true,
         omit_profile: true,
         omit_memory_md: true,
         model: ModelSpec::Inherit,
@@ -105,6 +106,7 @@ pub(crate) fn test_inherit_echo_def() -> AgentDefinition {
         omit_identity: true,
         omit_memory_context: true,
         omit_safety_preamble: true,
+        omit_skills_catalog: true,
         omit_profile: true,
         omit_memory_md: true,
         model: ModelSpec::Inherit,
@@ -145,6 +147,7 @@ pub(crate) fn test_inherit_parallel_worker_def() -> AgentDefinition {
         omit_identity: true,
         omit_memory_context: true,
         omit_safety_preamble: true,
+        omit_skills_catalog: true,
         omit_profile: true,
         omit_memory_md: true,
         model: ModelSpec::Inherit,

src/openhuman/agent/harness/definition_part_01.rs

@@ -85,6 +85,10 @@ pub struct AgentDefinition {
     #[serde(default = "defaults::true_")]
     pub omit_safety_preamble: bool,
 
+    /// If `true`, the installed skills catalog is stripped from the prompt.
+    #[serde(default = "defaults::true_")]
+    pub omit_skills_catalog: bool,
+
     /// If `true`, the user's `PROFILE.md` (generated by the onboarding
     /// enrichment pipeline — LinkedIn scrape, etc.) is NOT injected into
     /// the rendered prompt. Defaults to `true` so sub-agents stay lean:

src/openhuman/agent/harness/definition_tests.rs

@@ -9,6 +9,7 @@ fn make_def(id: &str) -> AgentDefinition {
         omit_identity: true,
         omit_memory_context: true,
         omit_safety_preamble: true,
+        omit_skills_catalog: true,
         omit_profile: true,
         omit_memory_md: true,
         model: ModelSpec::Inherit,

src/openhuman/agent/harness/session/builder/factory.rs

@@ -525,11 +525,17 @@ impl Agent {
         // body and get wrapped by [`SystemPromptBuilder::for_subagent`].
         let mut prompt_builder = match target_def {
             Some(def) => match &def.system_prompt {
-                PromptSource::Dynamic(build) => SystemPromptBuilder::from_dynamic(*build),
+                PromptSource::Dynamic(build) => {
+                    SystemPromptBuilder::from_dynamic_with_skills_catalog(
+                        *build,
+                        !def.omit_skills_catalog,
+                    )
+                }
                 PromptSource::Inline(text) => SystemPromptBuilder::for_subagent(
                     text.clone(),
                     def.omit_identity,
                     def.omit_safety_preamble,
+                    !def.omit_skills_catalog,
                 ),
                 PromptSource::File { path } => {
                     let prompt_root = config.workspace_dir.join("agent").join("prompts");
@@ -566,6 +572,7 @@ impl Agent {
                         body_text,
                         def.omit_identity,
                         def.omit_safety_preamble,
+                        !def.omit_skills_catalog,
                     )
                 }
             },

src/openhuman/agent/harness/subagent_runner/ops/runner.rs

@@ -11,7 +11,8 @@ use std::sync::Arc;
 use std::time::Instant;
 
 use crate::openhuman::agent::context::prompt::{
-    render_subagent_system_prompt_with_format, PromptContext, PromptTool, SubagentRenderOptions,
+    render_subagent_system_prompt_with_format_and_workflows, PromptContext, PromptTool,
+    SubagentRenderOptions,
 };
 use crate::openhuman::agent::file_state::with_file_state_agent_id;
 use crate::openhuman::agent::harness::agent_graph::{AgentTurnRequest, AgentTurnUsage};
@@ -59,24 +60,7 @@ use super::provider::{
     resolve_subagent_source, user_is_signed_in_to_composio, LazyToolkitResolver,
 };
 
-/// Runtime spawn-hierarchy gate decision for one delegation hop.
-///
-/// `parent_def` is the resolved parent agent definition (looked up from the
-/// global registry by its definition id) or `None` when the parent can't be
-/// resolved — e.g. a dynamically-named agent (model-council juror) or a custom
-/// agent absent from the registry, or any context where the registry isn't
-/// initialised. A `None` parent yields `Ok(())`: we skip rather than mask, the
-/// same defensive posture the loader takes for unknown child ids.
-///
-/// A **worker** parent is also exempted. At runtime a worker only reaches the
-/// spawn chokepoint via the documented collapsed `delegate_to_integrations_agent`
-/// path (→ `integrations_agent`, itself a worker) — a shape the loader
-/// intentionally leaves untouched. Re-denying it here would turn valid custom
-/// worker agents that use `{ skills = "*" }` into runtime failures. The
-/// worker-leaf authoring rule stays enforced statically at boot, and the
-/// per-parent allowlist gate blocks any other worker spawn.
-///
-/// For chat / reasoning parents the hop is checked against
+/// Validate the runtime spawn-hierarchy gate for one delegation hop.
 /// [`validate_tier_transition`] (the single source of truth shared with the
 /// boot loader walk); a forbidden hop is logged and becomes a
 /// [`SubagentRunError::TierViolation`]. Logging lives here (rather than at the
@@ -1295,6 +1279,7 @@ async fn run_typed_mode(
         definition.omit_safety_preamble,
         definition.omit_profile,
         definition.omit_memory_md,
+        definition.omit_skills_catalog,
     );
 
     let narrowed_integrations: Vec<crate::openhuman::agent::context::prompt::ConnectedIntegration> =
@@ -1400,14 +1385,21 @@ async fn run_typed_mode(
 
     let system_prompt = match &definition.system_prompt {
         PromptSource::Dynamic(build) => {
-            build(&prompt_ctx).map_err(|e| SubagentRunError::PromptLoad {
+            let mut body = build(&prompt_ctx).map_err(|e| SubagentRunError::PromptLoad {
                 path: format!("<dynamic:{}>", definition.id),
                 source: std::io::Error::other(e.to_string()),
-            })?
+            })?;
+            if !definition.omit_skills_catalog {
+                body.push_str("\n\n");
+                body.push_str(&crate::openhuman::agent::prompts::render_skills_catalog(
+                    &parent.workflows,
+                ));
+            }
+            body
         }
         PromptSource::Inline(_) | PromptSource::File { .. } => {
             let archetype_prompt_body = load_prompt_source(&definition.system_prompt, &prompt_ctx)?;
-            render_subagent_system_prompt_with_format(
+            render_subagent_system_prompt_with_format_and_workflows(
                 &parent.workspace_dir,
                 &model,
                 &allowed_indices,
@@ -1417,6 +1409,7 @@ async fn run_typed_mode(
                 render_options,
                 parent.tool_call_format,
                 &narrowed_integrations,
+                &parent.workflows,
                 agents_md.global.as_deref(),
                 agents_md.local.as_deref(),
             )

src/openhuman/agent/harness/subagent_runner/ops_tests.rs

@@ -10,6 +10,7 @@ fn make_def_named_tools(names: &[&str]) -> AgentDefinition {
         omit_identity: true,
         omit_memory_context: true,
         omit_safety_preamble: true,
+        omit_skills_catalog: true,
         omit_profile: true,
         omit_memory_md: true,
         model: ModelSpec::Inherit,

src/openhuman/agent/library/ops_tests.rs

@@ -10,6 +10,7 @@ fn definition() -> AgentDefinition {
         omit_identity: true,
         omit_memory_context: true,
         omit_safety_preamble: true,
+        omit_skills_catalog: true,
         omit_profile: false,
         omit_memory_md: true,
         model: ModelSpec::Hint("reasoning".to_string()),

src/openhuman/agent/orchestration/tools/spawn_parallel_agents_tests.rs

@@ -167,6 +167,7 @@ fn definition_with_tool_scope(
         omit_identity: true,
         omit_memory_context: true,
         omit_safety_preamble: true,
+        omit_skills_catalog: true,
         omit_profile: true,
         omit_memory_md: true,
         model: ModelSpec::Inherit,

src/openhuman/agent/prompts/builder.rs

@@ -120,6 +120,7 @@ impl SystemPromptBuilder {
         archetype_prompt_text: String,
         omit_identity: bool,
         omit_safety_preamble: bool,
+        include_skills_catalog: bool,
     ) -> Self {
         let mut sections: Vec<Box<dyn PromptSection>> =
             vec![Box::new(ArchetypePromptSection::new(archetype_prompt_text))];
@@ -143,12 +144,9 @@ impl SystemPromptBuilder {
         if !omit_safety_preamble {
             sections.push(Box::new(SafetySection));
         }
-        // Skills catalogue and connected integrations are rendered by
-        // the individual agent's `prompt.rs` when that agent needs
-        // them (integrations_agent for the skill-executor voice,
-        // orchestrator/welcome for the delegator voice). The shared
-        // builder intentionally does not emit them — keeping
-        // agent-specific prose scoped to the agent that owns it.
+        if include_skills_catalog {
+            sections.push(Box::new(SkillsCatalogSection));
+        }
         sections.push(Box::new(WorkspaceSection));
 
         Self { sections }
@@ -182,26 +180,25 @@ impl SystemPromptBuilder {
     pub fn from_dynamic(
         builder: crate::openhuman::agent::harness::definition::PromptBuilder,
     ) -> Self {
-        Self {
-            sections: vec![
-                Box::new(DynamicPromptSection::new(builder)),
-                // Project instructions (AGENTS.md). The ~26 dynamic
-                // `agents/<id>/prompt.rs` builders (orchestrator / main chat,
-                // welcome, integrations_agent, …) hand-assemble their own body
-                // via the `render_*` helpers and none of them individually call
-                // `render_agents_md`, so the pre-loaded AGENTS.md layers on
-                // `PromptContext` would otherwise be silently dropped for the
-                // primary agent. Inject the shared section centrally here —
-                // mirroring how `build()` appends the grounding contract for all
-                // dynamic builders — so every dynamic agent inherits the same
-                // AGENTS.md injection as the `with_defaults` / `for_subagent`
-                // chains. Rendered after the agent's own body (as trailing
-                // standing guidance) and before the central grounding suffix.
-                // Empty (skipped) when neither layer carries content or the
-                // `agents_md_enabled` gate is off.
-                Box::new(AgentsInstructionsSection),
-            ],
+        Self::from_dynamic_with_skills_catalog(builder, false)
+    }
+
+    /// Build a dynamic prompt and optionally append the shared skills catalog.
+    ///
+    /// Dynamic built-in agents are normally responsible for assembling their
+    /// own prompt body, so the catalog gate must be carried from the agent
+    /// definition into this wrapper rather than inferred from the body.
+    pub fn from_dynamic_with_skills_catalog(
+        builder: crate::openhuman::agent::harness::definition::PromptBuilder,
+        include_skills_catalog: bool,
+    ) -> Self {
+        let mut sections: Vec<Box<dyn PromptSection>> =
+            vec![Box::new(DynamicPromptSection::new(builder))];
+        if include_skills_catalog {
+            sections.push(Box::new(SkillsCatalogSection));
         }
+        sections.push(Box::new(AgentsInstructionsSection));
+        Self { sections }
     }
 
     pub fn add_section(mut self, section: Box<dyn PromptSection>) -> Self {

src/openhuman/agent/prompts/mod.rs

@@ -18,9 +18,10 @@ pub use render_helpers::{
     inject_snapshot_content, inject_workspace_file, inject_workspace_file_capped,
     memory_date_label, render_agents_md, render_ambient_environment, render_datetime,
     render_grounding, render_identity, render_runtime, render_safety,
-    render_subagent_system_prompt, render_subagent_system_prompt_with_format, render_tools,
-    render_user_files, render_user_identity, render_user_memory, render_user_reflections,
-    render_workspace, sync_workspace_file,
+    render_subagent_system_prompt, render_subagent_system_prompt_with_format,
+    render_subagent_system_prompt_with_format_and_workflows, render_tools, render_user_files,
+    render_user_identity, render_user_memory, render_user_reflections, render_workspace,
+    sync_workspace_file,
 };
 
 #[cfg(test)]

src/openhuman/agent/prompts/mod_tests_part_01_tests.rs

@@ -69,7 +69,7 @@ fn grounding_contract_appended_to_every_build_path() {
     assert!(defaults.contains(marker));
 
     // 2. Sub-agent static chain.
-    let sub = SystemPromptBuilder::for_subagent("role".into(), true, true)
+    let sub = SystemPromptBuilder::for_subagent("role".into(), true, true, false)
         .build(&ctx)
         .unwrap();
     assert!(sub.contains(marker));

src/openhuman/agent/prompts/mod_tests_part_02_tests.rs

@@ -70,7 +70,7 @@ fn render_subagent_system_prompt_renders_workspace_tail() {
 fn subagent_render_options_invert_definition_flags() {
     // (omit_identity, omit_safety_preamble,
     //  omit_profile, omit_memory_md)
-    let options = SubagentRenderOptions::from_definition_flags(true, false, false, false);
+    let options = SubagentRenderOptions::from_definition_flags(true, false, false, false, true);
     assert!(!options.include_identity);
     assert!(options.include_safety_preamble);
     assert!(options.include_profile);
@@ -110,6 +110,7 @@ fn render_subagent_system_prompt_honors_identity_safety_and_skills_flags() {
             include_safety_preamble: true,
             include_profile: false,
             include_memory_md: false,
+            include_skills_catalog: false,
         },
         ToolCallFormat::Json,
         &[],
@@ -193,6 +194,7 @@ fn render_subagent_system_prompt_injects_profile_md_even_when_identity_omitted()
             include_safety_preamble: false,
             include_profile: true,
             include_memory_md: false,
+            include_skills_catalog: false,
         },
         ToolCallFormat::PFormat,
         &[],
@@ -292,6 +294,7 @@ fn render_subagent_system_prompt_frames_memory_md_as_background() {
             include_safety_preamble: false,
             include_profile: false,
             include_memory_md: true,
+            include_skills_catalog: false,
         },
         ToolCallFormat::PFormat,
         &[],
@@ -340,6 +343,7 @@ fn render_subagent_system_prompt_omits_memory_framing_when_no_memory_content() {
             include_safety_preamble: false,
             include_profile: false,
             include_memory_md: true,
+            include_skills_catalog: false,
         },
         ToolCallFormat::PFormat,
         &[],
@@ -379,6 +383,7 @@ fn render_subagent_system_prompt_injects_profile_md_when_identity_included() {
             include_safety_preamble: false,
             include_profile: true,
             include_memory_md: false,
+            include_skills_catalog: false,
         },
         ToolCallFormat::PFormat,
         &[],
@@ -452,6 +457,7 @@ fn narrow_agent_with_omit_identity_still_loads_profile_md() {
         true,  // omit_safety_preamble
         false, // omit_profile   — opts IN to PROFILE.md
         false, // omit_memory_md — opts IN to MEMORY.md too
+        true,  // omit_skills_catalog
     );
 
     let tools: Vec<Box<dyn Tool>> = vec![Box::new(TestTool)];
@@ -498,7 +504,7 @@ fn narrow_subagent_definition_flags_skip_profile_md() {
     .unwrap();
 
     // Mirrors e.g. `critic/agent.toml` — all omit_* default-true.
-    let options = SubagentRenderOptions::from_definition_flags(true, true, true, true);
+    let options = SubagentRenderOptions::from_definition_flags(true, true, true, true, true);
 
     let tools: Vec<Box<dyn Tool>> = vec![Box::new(TestTool)];
     let rendered = render_subagent_system_prompt(
@@ -553,6 +559,7 @@ fn render_subagent_system_prompt_injects_memory_md_when_enabled() {
             include_safety_preamble: false,
             include_profile: false,
             include_memory_md: true,
+            include_skills_catalog: false,
         },
         ToolCallFormat::PFormat,
         &[],

src/openhuman/agent/prompts/mod_tests_part_03_tests.rs

@@ -28,6 +28,7 @@ fn profile_md_and_memory_md_are_capped_at_user_file_max_chars() {
             include_safety_preamble: false,
             include_profile: true,
             include_memory_md: true,
+            include_skills_catalog: false,
         },
         ToolCallFormat::PFormat,
         &[],
@@ -71,6 +72,7 @@ fn rendered_subagent_system_prompt_is_byte_stable_across_repeat_calls() {
         include_safety_preamble: false,
         include_profile: true,
         include_memory_md: true,
+        include_skills_catalog: false,
     };
 
     let first = render_subagent_system_prompt(
@@ -166,6 +168,7 @@ fn for_subagent_builder_injects_user_files_even_when_identity_omitted() {
         "You are a specialist agent.".into(),
         true, // omit_identity  — drops SOUL/IDENTITY preamble
         true, // omit_safety_preamble
+        false,
     );
     let rendered = builder.build(&ctx).unwrap();