Merge pull request #216 from ScriptedAlchemy/codex/hermes-parity-r6-consolidation

feat: stage managed skill consolidations

Author: ScriptedAlchemy

src/automation/managed_skill_model.rs

@@ -192,6 +192,16 @@ pub struct ManagedSkillPendingUpdate {
     pub body_markdown: String,
     #[serde(default)]
     pub support_files: Vec<ManagedSupportFile>,
+    /// Lifecycle state the skill transitions to when this staged change is
+    /// approved. `None` keeps the historical behavior (promote to `Active`).
+    /// Staged consolidations set `Some(Archived)`; skill content is always
+    /// preserved on disk (archive, never delete).
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub resulting_state: Option<ManagedSkillState>,
+    /// Reviewer-facing reason recorded when the change was staged (used by
+    /// consolidation proposals).
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub staged_reason: Option<String>,
 }
 
 impl ManagedSkillPendingUpdate {

src/automation/managed_skill_validation.rs

@@ -5,9 +5,9 @@ use crate::errors::{Result, TraceDecayError};
 
 use super::managed_skill_format::target_key;
 use super::managed_skill_model::{
-    ManagedSkill, ManagedSkillPendingUpdate, ManagedSkillUpdate, ManagedSupportFile,
-    SkillInstallTarget, MAX_MANAGED_SKILL_BODY_BYTES, MAX_MANAGED_SUPPORT_FILES,
-    MAX_MANAGED_SUPPORT_FILE_BYTES,
+    ManagedSkill, ManagedSkillPendingUpdate, ManagedSkillState, ManagedSkillUpdate,
+    ManagedSupportFile, SkillInstallTarget, MAX_MANAGED_SKILL_BODY_BYTES,
+    MAX_MANAGED_SUPPORT_FILES, MAX_MANAGED_SUPPORT_FILE_BYTES,
 };
 
 const ALLOWED_SUPPORT_ROOTS: &[&str] = &["references", "templates", "scripts", "assets"];
@@ -241,6 +241,13 @@ pub(crate) fn validate_managed_pending_update(
             "managed skill staged_at must be a positive timestamp".to_string(),
         ));
     }
+    if let Some(resulting_state) = pending.resulting_state {
+        if resulting_state != ManagedSkillState::Archived {
+            return Err(config_error(
+                "managed skill pending update resulting_state must be archived".to_string(),
+            ));
+        }
+    }
     let skill = ManagedSkill {
         metadata: pending.metadata.clone(),
         body_markdown: pending.body_markdown.clone(),

src/automation/managed_skills.rs

@@ -351,12 +351,65 @@ pub async fn stage_managed_skill_update(
         metadata: staged.metadata.clone(),
         body_markdown: staged.body_markdown.clone(),
         support_files: staged.support_files.clone(),
+        resulting_state: None,
+        staged_reason: None,
     };
     save_pending_update(profile_root, id, &pending).await?;
     record_skill_patch(profile_root, &staged, "staged_update".to_string()).await?;
     Ok(pending.into_skill())
 }
 
+/// Stages an archive transition for a managed skill as a pending update that
+/// must be approved (or discarded) through the normal review lifecycle.
+/// Skill content is untouched: approving only flips the state to `Archived`,
+/// keeping the body and support files recoverable on disk. Pinned skills are
+/// exempt, matching the Hermes curator.
+pub async fn stage_managed_skill_archive(
+    profile_root: &Path,
+    id: &str,
+    base_checksum: &str,
+    reason: Option<String>,
+) -> Result<ManagedSkill> {
+    let skill = load_managed_skill(profile_root, id).await?;
+    if base_checksum != skill.metadata.checksum {
+        return Err(config_error(format!(
+            "base_checksum for managed skill id '{id}' is stale"
+        )));
+    }
+    if skill.pending_update.is_some() {
+        return Err(config_error(format!(
+            "managed skill '{id}' already has a pending update"
+        )));
+    }
+    if skill.metadata.pinned {
+        return Err(config_error(format!(
+            "managed skill '{id}' is pinned and exempt from staged archive"
+        )));
+    }
+    if skill.metadata.state == ManagedSkillState::Archived {
+        return Err(config_error(format!(
+            "managed skill '{id}' is already archived"
+        )));
+    }
+
+    let mut staged = skill.clone();
+    staged.pending_update = None;
+    staged.set_state(ManagedSkillState::PendingApproval);
+    staged.touch();
+    let pending = ManagedSkillPendingUpdate {
+        base_checksum: base_checksum.to_string(),
+        staged_at: current_metadata_timestamp(),
+        metadata: staged.metadata.clone(),
+        body_markdown: staged.body_markdown.clone(),
+        support_files: staged.support_files.clone(),
+        resulting_state: Some(ManagedSkillState::Archived),
+        staged_reason: reason,
+    };
+    save_pending_update(profile_root, id, &pending).await?;
+    record_skill_patch(profile_root, &staged, "staged_archive".to_string()).await?;
+    Ok(pending.into_skill())
+}
+
 pub async fn discard_pending_managed_skill_update(
     profile_root: &Path,
     id: &str,
@@ -433,13 +486,17 @@ pub async fn approve_managed_skill(profile_root: &Path, id: &str) -> Result<Mana
     let approved = match skill.pending_update {
         None => set_managed_skill_state(profile_root, id, ManagedSkillState::Active).await?,
         Some(pending) => {
+            let resulting_state = pending.resulting_state.unwrap_or(ManagedSkillState::Active);
+            let patch_target = match resulting_state {
+                ManagedSkillState::Archived => "approve_staged_archive",
+                _ => "approve_staged_update",
+            };
             let mut promoted = pending.into_skill();
-            promoted.set_state(ManagedSkillState::Active);
+            promoted.set_state(resulting_state);
             promoted.refresh_checksum();
             remove_pending_update(profile_root, id).await?;
             save_managed_skill(profile_root, &promoted).await?;
-            record_skill_patch(profile_root, &promoted, "approve_staged_update".to_string())
-                .await?;
+            record_skill_patch(profile_root, &promoted, patch_target.to_string()).await?;
             promoted
         }
     };

src/automation/runner.rs

@@ -20,7 +20,8 @@ use super::managed_skills::list_managed_skills;
 use super::run_ledger::{AutomationRunLedgerRecord, AutomationTrigger};
 use super::session_reflector::validate_fact_proposals;
 use super::skill_usage::{
-    ingest_project_analytics_events, stale_skill_recommendations, summarize_skill_usage,
+    ingest_project_analytics_events, skill_overlap_candidates, stale_skill_recommendations,
+    summarize_skill_usage, DEFAULT_SKILL_OVERLAP_LIMIT,
 };
 use super::skill_writer::{
     activation_policy as skill_writer_activation_policy, skill_improvement_recommendations,
@@ -578,39 +579,41 @@ async fn finalize_skill_writer_success(
     proposed_ops: &Value,
     proposals: &[Value],
 ) -> Result<(Value, AutomationRunLedgerRecord)> {
-    let (created_skills, updated_skills, rejected_skills) =
-        match validate_and_apply_skill_proposals(
-            profile_root,
-            run_id,
-            proposals,
-            config.auto_enable_skills,
-        )
-        .await
-        {
-            Ok(result) => result,
-            Err(err) => {
-                finalizer
-                    .append_failed_record(
-                        response.model.clone(),
-                        evidence_hash,
-                        Some(proposed_ops.clone()),
-                        err.to_string(),
-                    )
-                    .await?;
-                return Err(err);
-            }
-        };
-    let accepted_count = created_skills.len() + updated_skills.len();
-    let rejected_count = rejected_skills.len();
+    let proposal_outcome = match validate_and_apply_skill_proposals(
+        profile_root,
+        run_id,
+        proposals,
+        config.auto_enable_skills,
+    )
+    .await
+    {
+        Ok(result) => result,
+        Err(err) => {
+            finalizer
+                .append_failed_record(
+                    response.model.clone(),
+                    evidence_hash,
+                    Some(proposed_ops.clone()),
+                    err.to_string(),
+                )
+                .await?;
+            return Err(err);
+        }
+    };
+    let accepted_count = proposal_outcome.created.len()
+        + proposal_outcome.updated.len()
+        + proposal_outcome.consolidations.len();
+    let rejected_count = proposal_outcome.rejected.len();
     let report = json!({
         "status": if config.auto_enable_skills { "auto_enabled" } else { "needs_approval" },
         "dry_run": true,
         "task": "skill_writer",
         "evidence_hash": evidence_hash,
         "activation_policy": activation_policy,
-        "created_skills": created_skills,
-        "updated_skills": updated_skills,
-        "rejected_skills": rejected_skills,
+        "created_skills": proposal_outcome.created,
+        "updated_skills": proposal_outcome.updated,
+        "staged_consolidations": proposal_outcome.consolidations,
+        "rejected_skills": proposal_outcome.rejected,
         "skill_improvement_recommendations": evidence
             .get("skill_improvement_recommendations")
             .cloned()
@@ -626,6 +629,7 @@ async fn finalize_skill_writer_success(
             "skills": proposed_ops.get("skills").cloned().unwrap_or_else(|| json!([])),
             "created_skills": report.get("created_skills").cloned().unwrap_or_else(|| json!([])),
             "updated_skills": report.get("updated_skills").cloned().unwrap_or_else(|| json!([])),
+            "staged_consolidations": report.get("staged_consolidations").cloned().unwrap_or_else(|| json!([])),
             "rejected_skills": report.get("rejected_skills").cloned().unwrap_or_else(|| json!([])),
         })),
         accepted_count,
@@ -634,6 +638,7 @@ async fn finalize_skill_writer_success(
     record.applied_ops = Some(json!({
         "created_skills": report.get("created_skills").cloned().unwrap_or_else(|| json!([])),
         "updated_skills": report.get("updated_skills").cloned().unwrap_or_else(|| json!([])),
+        "staged_consolidations": report.get("staged_consolidations").cloned().unwrap_or_else(|| json!([])),
     }));
     record.rejected_ops = report.get("rejected_skills").cloned();
     record.validation_report = Some(json!({
@@ -865,11 +870,14 @@ async fn build_skill_writer_evidence(
             }))
         })
         .unwrap_or_default();
+    let overlap_candidates =
+        skill_overlap_candidates(&existing_skills, DEFAULT_SKILL_OVERLAP_LIMIT);
     let skill_improvement_recommendations = skill_improvement_recommendations(
         &hits,
         &skill_usage_summaries,
         &stale_recommendations,
         &underused_tool_families,
+        &overlap_candidates,
     );
     let evidence = json!({
         "evidence_mode": evidence_mode_label(recent_session_slices.is_some()),
@@ -882,6 +890,7 @@ async fn build_skill_writer_evidence(
         "skill_usage_summaries": skill_usage_summaries,
         "stale_recommendations": stale_recommendations,
         "underused_tool_families": underused_tool_families,
+        "skill_overlap_candidates": overlap_candidates,
         "skill_improvement_recommendations": skill_improvement_recommendations,
         "existing_managed_skills": existing_skills
             .iter()
@@ -1360,7 +1369,7 @@ fn build_skill_writer_prompt(evidence: &Value) -> String {
         "\n",
         "An empty skills array is a real option when the session ran smoothly with no corrections and produced no new technique, but do not reach for it as a default.\n",
         "\n",
-        "Response contract: Return only JSON with a skills array of managed skill creates or updates. New skills may omit action or use action=create and must include id, title, summary, category, body_markdown, optional targets, optional support_files with text content, and reason. Targets, when present, must be an array using cursor, codex, claude, agents, opencode, kimi, or kiro; Hermes is host-owned and must not be targeted. Updates must use action=update or action=patch, include id and base_checksum, and include at least one changed field among title, summary, category, targets, body_markdown/body, support_files, or pinned. For updates, support_files is a complete replacement list, not a partial file patch. Activation is controlled only by the runner policy; do not assume activation from your response.\n",
+        "Response contract: Return only JSON with a skills array of managed skill creates or updates. New skills may omit action or use action=create and must include id, title, summary, category, body_markdown, optional targets, optional support_files with text content, and reason. Targets, when present, must be an array using cursor, codex, claude, agents, opencode, kimi, or kiro; Hermes is host-owned and must not be targeted. Updates must use action=update or action=patch, include id and base_checksum, and include at least one changed field among title, summary, category, targets, body_markdown/body, support_files, or pinned. For updates, support_files is a complete replacement list, not a partial file patch. Consolidations: when skill_overlap_candidates shows overlapping managed skills, you may propose action=merge (include id for the surviving skill, base_checksum, source_skill_id, source_base_checksum, reason, and optional merged title/summary/category/targets/body_markdown/support_files) or action=archive (include id, base_checksum, reason). Consolidations are always staged for human approval and archive-only; content is never deleted. Never propose merge or archive for pinned or user-authored skills. Activation is controlled only by the runner policy; do not assume activation from your response.\n",
     );
     format!(
         "{POLICY}{}",

src/automation/skill_usage.rs

@@ -8,10 +8,15 @@ use crate::errors::{Result, TraceDecayError};
 use crate::tracedecay::current_timestamp;
 
 mod analytics;
+mod overlap;
 mod recommendations;
 
 pub(crate) use analytics::analytics_import_key_for_request;
 pub use analytics::{ingest_analytics_events, ingest_project_analytics_events};
+pub use overlap::{
+    skill_overlap_candidates, SkillOverlapCandidate, DEFAULT_SKILL_OVERLAP_LIMIT,
+    SKILL_OVERLAP_CONTENT_THRESHOLD, SKILL_OVERLAP_TITLE_THRESHOLD,
+};
 pub use recommendations::{skill_improvement_recommendations, stale_skill_recommendations};
 
 const SKILL_USAGE_LEDGER_FILENAME: &str = "skill_usage.json";