mnemon-dev · Grivn · May 15, 2026 · May 15, 2026
diff --git a/Makefile b/Makefile
@@ -10,7 +10,7 @@ ifeq ($(GOBIN),)
   GOBIN     := $(shell go env GOPATH)/bin
 endif
 
-.PHONY: deps build install uninstall test unit vet harness-validate codex-app-eval codex-app-eval-suite codex-memory-deep-eval docker-build docker-run compose-up compose-down compose-dev release-snapshot clean help
+.PHONY: deps build install uninstall test unit vet harness-validate codex-app-eval codex-app-eval-suite codex-memory-deep-eval codex-skill-deep-eval docker-build docker-run compose-up compose-down compose-dev release-snapshot clean help
 
 .DEFAULT_GOAL := help
 
@@ -57,6 +57,9 @@ codex-app-eval-suite: ## Run real Codex app-server memory/skill scenario suite
 codex-memory-deep-eval: ## Run deep real Codex app-server memory regression suite
 	python3 scripts/codex_app_server_eval.py --suite --suite-name memory-deep
 
+codex-skill-deep-eval: ## Run deep real Codex app-server skill regression suite
+	python3 scripts/codex_app_server_eval.py --suite --suite-name skill-deep
+
 # ── Containers / Deployment ──────────────────────────────────────────
 
 docker-build: ## Build runtime Docker image

diff --git a/docs/harness/eval/CODEX_APP_SERVER.md b/docs/harness/eval/CODEX_APP_SERVER.md
@@ -37,6 +37,16 @@ The deep memory suite adds noisy recall filtering, stale-memory supersession,
 uncertain-preference rejection, secret-like value rejection, and multi-turn
 continuity through persisted `MEMORY.md`.
 
+For longer skill-loop regression, run:
+
+```bash
+make codex-skill-deep-eval
+```
+
+The deep skill suite adds transient evidence skip, missing-skill evidence,
+approved active skill creation, host-surface preservation, and proposal-first
+curation checks, plus reviewable skill authoring drafts.
+
 To trigger a real Codex turn, opt in explicitly:
 
 ```bash

diff --git a/docs/zh/harness/eval/CODEX_APP_SERVER.md b/docs/zh/harness/eval/CODEX_APP_SERVER.md
@@ -35,6 +35,16 @@ make codex-memory-deep-eval
 deep memory suite 会额外覆盖：带噪声的相关 recall、过期 memory 覆盖、
 不确定偏好拒绝、疑似 secret 值拒绝，以及通过持久化 `MEMORY.md` 完成多轮连续性。
 
+更长的 skill loop 回归可以运行：
+
+```bash
+make codex-skill-deep-eval
+```
+
+deep skill suite 会额外覆盖：跳过临时 evidence、记录 missing-skill evidence、
+执行已批准的 active skill 创建、保护 host skill surface，以及 proposal-first
+curation 不直接激活 skill，并验证 reviewable skill draft 的 authoring。
+
 如果需要触发真实 Codex turn，可以显式开启：
 
 ```bash

diff --git a/harness/eval/README.md b/harness/eval/README.md
@@ -32,6 +32,12 @@ Run the longer memory regression suite with:
 make codex-memory-deep-eval
 ```
 
+Run the longer skill-loop regression suite with:
+
+```bash
+make codex-skill-deep-eval
+```
+
 To run an actual Codex turn, use:
 
 ```bash
@@ -72,3 +78,12 @@ The `memory-deep` suite extends memory coverage with:
 - rejecting uncertain preference changes
 - rejecting secret-like values and generic restatements of existing safety policy
 - multi-turn continuity through persisted `MEMORY.md`
+
+The `skill-deep` suite extends skill-loop coverage with:
+
+- skipping transient one-off workflow evidence
+- recording missing-skill evidence as JSONL
+- applying an explicitly approved active skill creation
+- preserving the host skill surface during canonical skill changes
+- producing proposal-first curation output without activating skills
+- drafting reviewable skill content without activating it
diff --git a/harness/hosts/claude-code/projector.sh b/harness/hosts/claude-code/projector.sh
@@ -261,7 +261,7 @@ export MNEMON_SKILL_LOOP_USAGE_FILE="${CANONICAL_MODULE_DIR}/skills/.usage.jsonl
 export MNEMON_SKILL_LOOP_PROPOSALS_DIR="${CANONICAL_MODULE_DIR}/proposals"
 export MNEMON_SKILL_LOOP_HOST_SKILLS_DIR="${host_skills_dir}"
 export MNEMON_SKILL_LOOP_REVIEW_MIN_EVENTS="\${MNEMON_SKILL_LOOP_REVIEW_MIN_EVENTS:-20}"
-export MNEMON_SKILL_LOOP_PROTECTED_SKILLS="\${MNEMON_SKILL_LOOP_PROTECTED_SKILLS:-skill_observe,skill_curate,skill_manage,memory_get,memory_set}"
+export MNEMON_SKILL_LOOP_PROTECTED_SKILLS="\${MNEMON_SKILL_LOOP_PROTECTED_SKILLS:-skill_observe,skill_curate,skill_author,skill_manage,memory_get,memory_set}"
 EOF
   chmod 0755 "${CONFIG_DIR}/mnemon-skill-loop/env.sh"
 }
@@ -322,13 +322,15 @@ install_skill_loop() {
     "${CANONICAL_MODULE_DIR}/reports" \
     "${HOST_SKILLS_DIR}/skill_observe" \
     "${HOST_SKILLS_DIR}/skill_curate" \
+    "${HOST_SKILLS_DIR}/skill_author" \
     "${HOST_SKILLS_DIR}/skill_manage" \
     "${CONFIG_DIR}/agents" \
     "${CONFIG_DIR}/hooks/mnemon-skill-loop"
   write_skill_projection_env
 
   install_file "${MODULE_DIR}/skills/skill_observe.md" "${HOST_SKILLS_DIR}/skill_observe/SKILL.md" 0644
   install_file "${MODULE_DIR}/skills/skill_curate.md" "${HOST_SKILLS_DIR}/skill_curate/SKILL.md" 0644
+  install_file "${MODULE_DIR}/skills/skill_author.md" "${HOST_SKILLS_DIR}/skill_author/SKILL.md" 0644
   install_file "${MODULE_DIR}/skills/skill_manage.md" "${HOST_SKILLS_DIR}/skill_manage/SKILL.md" 0644
   install_file "${MODULE_DIR}/subagents/curator.md" "${CONFIG_DIR}/agents/mnemon-skill-curator.md" 0644
 
@@ -398,6 +400,7 @@ uninstall_skill_loop() {
   rm -rf "${CONFIG_DIR}/hooks/mnemon-skill-loop"
   rm -rf "${host_skills_dir}/skill_observe"
   rm -rf "${host_skills_dir}/skill_curate"
+  rm -rf "${host_skills_dir}/skill_author"
   rm -rf "${host_skills_dir}/skill_manage"
   rm -f "${CONFIG_DIR}/agents/mnemon-skill-curator.md"
   rm -rf "${CONFIG_DIR}/mnemon-skill-loop"

diff --git a/harness/hosts/codex/projector.sh b/harness/hosts/codex/projector.sh
@@ -289,6 +289,7 @@ install_skill_loop() {
     "${CANONICAL_MODULE_DIR}/reports" \
     "${HOST_SKILLS_DIR}/skill_observe" \
     "${HOST_SKILLS_DIR}/skill_curate" \
+    "${HOST_SKILLS_DIR}/skill_author" \
     "${HOST_SKILLS_DIR}/skill_manage" \
     "${CONFIG_DIR}/mnemon-skill-loop"
   write_runtime_env "${CONFIG_DIR}/mnemon-skill-loop" "MNEMON_SKILL_LOOP_ENV" "MNEMON_SKILL_LOOP_DIR"
@@ -301,13 +302,16 @@ export MNEMON_SKILL_LOOP_ARCHIVED_DIR="${CANONICAL_MODULE_DIR}/skills/archived"
 export MNEMON_SKILL_LOOP_USAGE_FILE="${CANONICAL_MODULE_DIR}/skills/.usage.jsonl"
 export MNEMON_SKILL_LOOP_PROPOSALS_DIR="${CANONICAL_MODULE_DIR}/proposals"
 export MNEMON_SKILL_LOOP_HOST_SKILLS_DIR="${HOST_SKILLS_DIR}"
+export MNEMON_SKILL_LOOP_PROTECTED_SKILLS="${MNEMON_SKILL_LOOP_PROTECTED_SKILLS:-skill_observe,skill_curate,skill_author,skill_manage,memory_get,memory_set}"
 EOF
 
   install_file "${MODULE_DIR}/skills/skill_observe.md" "${HOST_SKILLS_DIR}/skill_observe/SKILL.md" 0644
   install_file "${MODULE_DIR}/skills/skill_curate.md" "${HOST_SKILLS_DIR}/skill_curate/SKILL.md" 0644
+  install_file "${MODULE_DIR}/skills/skill_author.md" "${HOST_SKILLS_DIR}/skill_author/SKILL.md" 0644
   install_file "${MODULE_DIR}/skills/skill_manage.md" "${HOST_SKILLS_DIR}/skill_manage/SKILL.md" 0644
   append_codex_runtime_note "${HOST_SKILLS_DIR}/skill_observe/SKILL.md" "MNEMON_SKILL_LOOP_DIR" "${CONFIG_DIR}/mnemon-skill-loop/env.sh"
   append_codex_runtime_note "${HOST_SKILLS_DIR}/skill_curate/SKILL.md" "MNEMON_SKILL_LOOP_DIR" "${CONFIG_DIR}/mnemon-skill-loop/env.sh"
+  append_codex_runtime_note "${HOST_SKILLS_DIR}/skill_author/SKILL.md" "MNEMON_SKILL_LOOP_DIR" "${CONFIG_DIR}/mnemon-skill-loop/env.sh"
   append_codex_runtime_note "${HOST_SKILLS_DIR}/skill_manage/SKILL.md" "MNEMON_SKILL_LOOP_DIR" "${CONFIG_DIR}/mnemon-skill-loop/env.sh"
 
   write_host_manifest "${CONFIG_DIR}"
@@ -356,6 +360,7 @@ uninstall_skill_loop() {
   local host_skills_dir="${MNEMON_SKILL_LOOP_HOST_SKILLS_DIR:-${HOST_SKILLS_DIR:-${CONFIG_DIR}/skills}}"
   rm -rf "${host_skills_dir}/skill_observe"
   rm -rf "${host_skills_dir}/skill_curate"
+  rm -rf "${host_skills_dir}/skill_author"
   rm -rf "${host_skills_dir}/skill_manage"
   rm -rf "${CONFIG_DIR}/mnemon-skill-loop"
   if [[ "${PURGE_LIBRARY}" == "1" ]]; then

diff --git a/harness/modules/skill-loop/GUIDE.md b/harness/modules/skill-loop/GUIDE.md
@@ -22,7 +22,9 @@ Record evidence when a session shows one of these signals:
 - a skill should be protected, pinned, restored, staled, or archived
 
 Skip evidence for one-off commands, transient progress, raw chat logs, secrets,
-or facts better stored as memory.
+or facts better stored as memory. Do not record evidence merely because a
+single command succeeded or because the current prompt mentions the skill loop;
+there must be a reusable workflow or lifecycle signal.
 
 ## Lifecycle
 

diff --git a/harness/modules/skill-loop/README.md b/harness/modules/skill-loop/README.md
@@ -20,6 +20,7 @@ harness/modules/skill-loop/
 ├── skills/
 │   ├── skill_observe.md
 │   ├── skill_curate.md
+│   ├── skill_author.md
 │   └── skill_manage.md
 ├── subagents/
 │   └── curator.md
@@ -43,6 +44,7 @@ harness/modules/skill-loop/
 | `hooks/*.md` | Four lifecycle reminders. Prime syncs active skills; Nudge records evidence; Compact may trigger review; Remind is no-op by default. |
 | `skills/skill_observe.md` | Online evidence capture protocol. |
 | `skills/skill_curate.md` | Protocol for starting a curator review. |
+| `skills/skill_author.md` | Protocol for drafting reviewable `SKILL.md` content. |
 | `skills/skill_manage.md` | Approved lifecycle mutation protocol. |
 | `subagents/curator.md` | Background reviewer that proposes create, patch, consolidate, stale, archive, or restore actions. |
 | Host adapter | Host-specific projection lives outside the module under `harness/hosts/<host>/`. |
@@ -90,6 +92,7 @@ The key split is:
 GUIDE.md decides when skill evolution behavior is useful.
 skill_observe.md records evidence only.
 curator.md reviews evidence and proposes changes.
+skill_author.md drafts skill content for review.
 skill_manage.md applies approved changes to canonical state.
 prime.sh projects active canonical skills into the host skill surface.
 ```

diff --git a/harness/modules/skill-loop/env.sh b/harness/modules/skill-loop/env.sh
@@ -21,4 +21,4 @@ export MNEMON_SKILL_LOOP_USAGE_FILE="${MNEMON_SKILL_LOOP_USAGE_FILE:-${MNEMON_SK
 export MNEMON_SKILL_LOOP_PROPOSALS_DIR="${MNEMON_SKILL_LOOP_PROPOSALS_DIR:-${MNEMON_SKILL_LOOP_DIR}/proposals}"
 export MNEMON_SKILL_LOOP_HOST_SKILLS_DIR="${MNEMON_SKILL_LOOP_HOST_SKILLS_DIR:-${MNEMON_SKILL_LOOP_CONFIG_DIR}/skills}"
 export MNEMON_SKILL_LOOP_REVIEW_MIN_EVENTS="${MNEMON_SKILL_LOOP_REVIEW_MIN_EVENTS:-20}"
-export MNEMON_SKILL_LOOP_PROTECTED_SKILLS="${MNEMON_SKILL_LOOP_PROTECTED_SKILLS:-skill_observe,skill_curate,skill_manage,memory_get,memory_set}"
+export MNEMON_SKILL_LOOP_PROTECTED_SKILLS="${MNEMON_SKILL_LOOP_PROTECTED_SKILLS:-skill_observe,skill_curate,skill_author,skill_manage,memory_get,memory_set}"
diff --git a/harness/modules/skill-loop/module.json b/harness/modules/skill-loop/module.json
@@ -21,6 +21,7 @@
     "skills": [
       "skills/skill_observe.md",
       "skills/skill_curate.md",
+      "skills/skill_author.md",
       "skills/skill_manage.md"
     ],
     "subagents": [

diff --git a/harness/modules/skill-loop/skills/skill_author.md b/harness/modules/skill-loop/skills/skill_author.md
@@ -0,0 +1,56 @@
+---
+name: skill_author
+description: Draft or revise high-quality SKILL.md content for approved or proposed Mnemon skill-loop changes.
+---
+
+# skill_author
+
+Use this skill when a curator proposal, user request, or approved lifecycle
+change needs a concrete `SKILL.md` draft.
+
+## Boundary
+
+This skill authors skill content only. It does not decide lifecycle placement
+and does not activate, stale, archive, restore, or delete skills.
+
+Write drafts under:
+
+```text
+$MNEMON_SKILL_LOOP_PROPOSALS_DIR
+```
+
+Approved lifecycle placement is applied later with `skill_manage.md`.
+
+## Procedure
+
+1. Confirm the target skill id is hyphen-case: lowercase letters, numbers, and
+   `-`.
+2. Confirm the skill captures a reusable procedure, not project facts,
+   preferences, credentials, raw transcripts, or one-off task context.
+3. Draft a complete `SKILL.md` with:
+   - YAML frontmatter containing `name` and `description`
+   - a short trigger-oriented description
+   - a clear boundary section
+   - a concise procedure section
+   - safety or validation notes only when they change behavior
+4. Keep the skill focused. Prefer one workflow per skill.
+5. Use project-neutral language. Do not embed current branch names, temporary
+   tokens, credentials, private URLs, or task-specific facts.
+6. Save the draft as a proposal artifact such as:
+
+```text
+$MNEMON_SKILL_LOOP_PROPOSALS_DIR/<skill-id>.SKILL.md
+```
+
+7. Leave `skills/active`, `skills/stale`, `skills/archived`, and host skill
+   surfaces unchanged unless the user explicitly asks to use `skill_manage.md`
+   after approval.
+
+## Quality Checklist
+
+- The description tells the host when to use the skill.
+- The body teaches reusable judgment or procedure the model would not reliably
+  infer from the current task alone.
+- The content is short enough to load on demand.
+- The skill avoids duplicated policy already covered by `GUIDE.md`.
+- The draft is safe to review before activation.
diff --git a/harness/modules/skill-loop/skills/skill_curate.md b/harness/modules/skill-loop/skills/skill_curate.md
@@ -29,7 +29,9 @@ It does not directly apply lifecycle changes. Approved changes are applied with
    - `.usage.jsonl`
    - existing proposals
 3. Request proposals for create, patch, consolidate, stale, archive, or restore
-   actions only when evidence supports them.
+   actions only when evidence supports them. When a proposal needs concrete
+   skill content, use `skill_author.md` to draft reviewable `SKILL.md` content
+   under the proposals directory.
 4. Keep the output proposal-first. Do not enable a new active skill in the
    current session unless the user explicitly approves and the host supports it.
 

diff --git a/harness/modules/skill-loop/skills/skill_manage.md b/harness/modules/skill-loop/skills/skill_manage.md
@@ -25,6 +25,7 @@ $MNEMON_SKILL_LOOP_ARCHIVED_DIR
 ## Allowed MVP Operations
 
 - create an approved skill under `active/<skill-id>/SKILL.md`
+- apply approved `SKILL.md` content drafted by `skill_author.md`
 - patch an existing skill in its current lifecycle directory
 - consolidate duplicated skills with an approved replacement
 - move `active -> stale`
@@ -38,7 +39,8 @@ $MNEMON_SKILL_LOOP_ARCHIVED_DIR
 1. Read the approved proposal and confirm the intended operation.
 2. Check `MNEMON_SKILL_LOOP_PROTECTED_SKILLS`; do not modify protected skills
    unless the approval explicitly covers the exception.
-3. Keep skill ids filesystem-safe: lowercase letters, numbers, `_`, and `-`.
+3. Keep new user-facing skill ids hyphen-case: lowercase letters, numbers, and
+   `-`. Existing protocol skill ids may keep their established underscore names.
 4. Apply the smallest canonical change under the lifecycle directories.
 5. Prefer moving to `archived` over deletion.
 6. Do not edit the host skill surface directly. Let Prime regenerate it.

diff --git a/harness/modules/skill-loop/subagents/curator.md b/harness/modules/skill-loop/subagents/curator.md
@@ -4,6 +4,7 @@ description: Reviews Mnemon skill evidence and proposes skill lifecycle changes.
 tools: Read, Write, Edit, Bash, Grep, Glob
 skills:
   - skill_observe
+  - skill_author
   - skill_manage
 ---
 
@@ -44,7 +45,8 @@ Run curator review when:
 2. Inspect active, stale, and archived skills.
 3. Review usage evidence and existing proposals.
 4. Identify only evidence-backed opportunities:
-   - create a skill for a repeated workflow
+   - create a skill for a repeated workflow, using `skill_author` for draft
+     `SKILL.md` content when useful
    - patch a misleading, outdated, or incomplete skill
    - consolidate duplicated skills
    - move low-value active skills to stale