Mandatory pre-flight checklist before creating or modifying any file. Prevents redundant work by systematically checking for existing resources first. Includes repo-specific pre-flight (git search, Hermes boundary, deploy verification — formerly cortex-preflight) and a full-repo stale-artifact audit pattern.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Mandatory pre-flight checklist before creating or modifying any file. Prevents redundant work by systematically checking for existing resources first. Includes repo-specific pre-flight (git search, Hermes boundary, deploy verification — formerly cortex-preflight) and a full-repo stale-artifact audit pattern.
Mandatory pre-flight for every file create/modify operation.
Run this BEFORE writing any file, adding a cron, creating a skill, or modifying any existing resource. Do not skip, batch, or defer.
Survey
Phase 0a: Load Domain Skills for the Operation Type
Before searching existing resources, load domain knowledge for what you're about to build. The 7 always-skills (task-start, agent-flow, survey-before-action, etc.) teach you HOW to work. Domain skills teach you WHAT the craft requires — pitfalls, portability, conventions, gotchas.
Without domain skills, agents make preventable mistakes: writing .sh files with bash portability bugs, creating crons with wrong output formats, patching nginx with macOS paths on Linux.
The rule: for the first file you create or modify, check both axes — operation and extension — and load EVERY matching skill.
nginx-web-app-deployment (or nginx-security-pipeline if security-focused)
Configuring Docker / compose
docker-management, env-aware-compose-wrapper
Building a web app service
nginx-web-app-deployment, prevent-crash-looping
Writing tests (any language)
test-driven-development
Debugging a failure
root-cause-debugging (or systematic-debugging)
Performance diagnosis
linux-performance-diagnostics
Cross-agent feature / protocol
cross-agent-design
Installing packages
package-security
Creating a skill
hermes-agent-skill-authoring, pii-scrubbing
Writing documentation
documentation-auditing, doc-freshness
System security change
linux-server-hardening, security-audit
Deploying Langfuse
langfuse-self-hosted
Configuring Ollama
ollama-setup
Setting up CI/CD
ci-cd-pipeline
Building MCP server
mcp-server-building
SSH / auth / credential change
secure-credential-handling
Extension-Axis Triggers (secondary, when operation isn't specialized)
If the operation isn't in the table above, fall back to file extension:
File extension
Load these skills
.sh / .bash
shell-scripting
.py (module-level, not one-shot)
codebase-design, python-debugpy
.py (one-shot script)
error-handling
.conf (nginx)
nginx-web-app-deployment
.yaml / .yml (Docker)
docker-management
.yaml / .yml (general)
skills_list fallback
.md
documentation-auditing
Makefile
project-run-scripts
.json
skills_list fallback — no dedicated skill
.sql
skills_list fallback — known gap, no dedicated skill
.toml
skills_list fallback — known gap
.env
pii-scrubbing, secure-credential-handling
Discovery Fallback — Always Run
After loading the matched skills, always run a discovery sweep for skills you didn't know existed:
# Category derived from the primary file type or operation domain# e.g. 'devops' for .sh / cron / docker, 'software-development' for .py
skills_list(category="devops")
skills_list(category="software-development")
# If you load a skill that references others, load those too
related = skill_view(name="X").related_skills or []
for name in related:
skill_view(name=name)
This catches the skills the mapping table author didn't think of. A skill you never load can never save you from a mistake.
When to Load
This Phase 0a runs AFTER you've loaded the 7 always-skills (task-start, agent-flow, etc.) and AFTER you've classified the task with agent-flow (Step 5), but BEFORE you open a governance lock (begin_change). It supplements the "on-task skills" from the manifest.
Sequence:
✅ Always skills loaded (task-start)
✅ Task classified (agent-flow)
✅ Domain skills loaded ← you are here
→ begin_change (governance lock)
→ Search existing resources (Phase 1)
Why before begin_change: You cannot know what to lock if you don't know what domain knowledge you need. Loading domain skills first means your first change is informed.
1. Search for existing resources
Search these in order. Batch independent searches together.
Keywords: function name, tool name, purpose (e.g. "audit", "health", "failure")
2. Examine what exists
If a candidate match is found, read it before deciding. Assess:
Does it do some or all of what I need?
Can I add a flag/parameter to extend it?
Is it in the right location or needs moving?
3. Decision
If ...
Then ...
Covers 100%
Use as-is
Covers 80%+ with small patch
Patch the existing tool
Covers 50-80%
Refactor or extend
Covers <50%
Create new. Document why in commit.
4. Creating something new anyway
Document in commit why existing resources don't fit.
4b. User approval for deletions
If your plan includes deleting files or directories, present the deletion plan for approval BEFORE executing. Show what, why safe to remove (zero consumers, stale duplicate, etc.), and expected improvement. Wait for explicit confirmation before git rm or rm -rf.
4c. Make old names discoverable before deletion
Before deleting a renamed/merged artifact, ensure the surviving artifact makes the old name findable:
Skills: add aliases: [old-name] or tags: [old-name] in frontmatter
Scripts/tools: add # Formerly known as comment or --compat flag
Docs: add deprecation note at the top (e.g. > Moved to [new-path](new-path.md))
Git history: mention both names in commit message
Why: A deletion without discoverability orphans every inbox message, cron config, and agent reference that used the old name.
4d. Guardrail: verify target exists before editing
Before calling patch(), write_file(), or any tool that modifies a file, confirm the exact target path exists with read_file() or search_files(target="files") first. Patching a nonexistent path silently fails — the edit never lands. Always verify with a second method. search_files(target="files") matches basenames only, not directory names.
Run these after the generic survey above, before writing code — they are
specific to the Hermes Cortex repo and its deployment model.
1. Check git for missing files
search_files() only scans disk. If it finds nothing, the file may still
exist in git (committed but not deployed):
cd ~/hermes-cortex
git log --oneline --all -- "**/<pattern>*"# search git history
git show HEAD:<path-to-file> # view file in git but not disk
Common scenario: script exists in repo (ops/scripts/manage/foo.py)
but was never deployed to ~/.hermes-cortex/scripts/foo.py. Running
cortex-update.sh fixes this.
2. GOVERNANCE FILE WORKFLOW — ORCHESTRATORS ONLY
Non-orchestrators: stop here for these files — send an inbox message to
inbox_orchestrator requesting the change. If you ARE an orchestrator
(Moses, Esther): hooks, enforcer plugin, and skills follow this rule:
Fix the REPO SOURCE first in ~/hermes-cortex/ — never the deployed copy
Commit, push, then run cortex-update.sh to deploy
A fix applied to the deployed copy WILL be overwritten on next update
Specific files (always repo source, never deployed copy): ~/.hermes/plugins/governance-enforcer/__init__.py, ~/.hermes-cortex/scripts/pre-commit-score, ~/.hermes-cortex/hooks/* (symlinks), ~/.hermes/skills/<cat>/<name>/SKILL.md, all files with a register() entry in cortex-update.sh.
3. Hermes boundary check
File location
Action
In ~/hermes-cortex/
✅ Ours — modify freely
In ~/.hermes/ AND in repo skills/
✅ Ours — modify the repo copy, deploy
In ~/.hermes/ but NOT in repo
❌ Hermes default — do NOT touch
In ~/.hermes-cortex/state/*
✅ Live config — modify directly
In ~/.hermes/config.yaml
✅ Live config — modify directly
To extend a Hermes default skill, create a supporting skill in the repo
(see file-ownership-boundaries skill) instead of editing the default.
4. Verify deployed copies match repo
A source change is not deployed until cortex-update.sh runs:
grep -n "register.*<script-name>" ~/hermes-cortex/ops/scripts/cortex-update.sh
ls -la ~/.hermes-cortex/scripts/<script-name>
bash ~/hermes-cortex/ops/scripts/cortex-update.sh # if missing
5. Check what agent type you are
Agent type
Can do
orchestrator (Moses, Esther)
Fleet dispatch, bus operations, skill lifecycle
server-agent (Joseph, Kustos, Gisu)
Local maintenance, health reports
dev-agent (Titus)
Local reports, push-only bus
6. Check for stale deploy references — EVERY deploy location
Before renaming or removing a file, search every location that could
reference the old name (not just the obvious ones):
git status for foreign unstaged/untracked changes — a concurrent session's in-flight edits block the pre-push dogfood gate (it diffs the working tree). Coordinate first; never stash/clean a peer's work. Cross-ref: change-checklist Phase 0.
Deployment Pitfalls — cortex-update.sh Side Effects
Account for each before and after deploy:
SOURCE header breaks checksums.cortex-update.sh adds a # SOURCE:
header below the shebang to every deployed .sh/.py. Raw MD5 between
repo and deployed ALWAYS differs. Use the doctor's _content_md5()
(strips header) on deployed paths — never _md5().
Lock cleanup. cortex-update removes .governance-*.json locks whose
heartbeat exceeded TTL (>1h) plus legacy v1 locks. A fresh session-scoped
v2 lock survives. If your lock is gone after deploy, re-acquire with
begin_change() — and score any PENDING cycles from the purged lock first.
Skills-loaded marker survives deploys. Per-session marker files at
state/skills-loaded/<session-id> survive the enforcer plugin reload —
write tools keep working after cortex-update. (The old shared-marker race
is gone since 2026-08-01.)
Hook symlinks prevent drift. Deployed hooks (~/.hermes-cortex/hooks/)
are absolute symlinks to ~/.hermes-cortex/scripts/ sources. If a hook is
a file copy instead of a symlink, it drifts and the doctor flags it. The
repo .hermes-cortex/hooks/ tracks docs only, not deployables.
PENDING cycles accumulate. Every begin_change() creates a cycle;
when cortex-update purges the lock and you re-acquire, old cycles stay
PENDING. Score ALL PENDING cycles (feedback_accept/feedback_override)
before end_change() or the doctor reports a governance leak.
Critical rules
Search ops/scripts/ first — canonical location for installed scripts
Check skills_list() — if a skill already covers the workflow, use it
Check cronjob list — existing cron patterns may already solve the problem
Never trust a single zero-result search_files. When you get 0 results, verify with ls, find, git ls-tree, or git log before concluding the resource doesn't exist.
Never register user-owned files in the deploy map. A register() entry in cortex-update.sh makes a file a deploy target — it is overwritten whenever content differs. That is right for repo-managed files, destructive for files an agent personalizes (memory, user config). Memory (MEMORY.md/USER.md) is Hermes-owned and must NEVER be registered (2026-08-05: a memory seed register clobbered live memory 7× in a day; the doctor then FAILed on the healthy personalized state and suggested the destructive resync). cortex-update.sh now fails CLOSED on any target under /memories/ or ~/.hermes/ — the guard is physical, not a convention.
If no match found: note it mentally. Do not skip this step next time.
Post-Action Audit
Mandatory post-flight before calling end_change().
5. Search for stale communications
After modifying, deleting, or renaming files other agents reference, check pending outbound messages with mcp_agent_inbox_inbox_read(unread_only=True). If any reference now-stale paths or instructions, send a correction before releasing the governance lock.
6. Decide whether to share
I created/modified ...
I should put it in ...
A script
ops/scripts/ + register in cortex-update.sh
A skill
skills///
A cron pattern
AGENTS.md or cron-schedules.md
A workflow/technique
Save as a skill or reference
Machine/user specific
Keep private
Default posture: share.
7. Verify the guardrail
If this work responded to a mistake: did you implement a guardrail that prevents recurrence? Is it enforceable? Test it catches the same failure. If any answer is no, the loop isn't closed.
8. Test before shipping
Every change must be verified end-to-end before pushing to main. Verification means exercising the actual changed code path — not just staring at the diff.
What changed
Verification required
install-crons.sh cron
Run test cron via CLI, confirm registration, clean up
Shell script
bash -n <file> + live test of changed branch
Python script
Run with expected args, check output and exit code
Cron definition
Create test cron, verify in list, remove
Pre-commit/git hook
Commit through the RUNNING hook, not the repo copy
Any file with runtime copy
Verify the RUNTIME copy changed — not just the repo source
The check before git push: "Did I actually run the changed code path?" If only looked — stop. Run it.
Missing script at destination → cron creates but fails every tick
Absolute path that works on one machine but not others
Repo source changed but running copy stale — editing .hermes-cortex/hooks/pre-commit while core.hooksPath points to ~/.hermes-cortex/hooks/ where the old version still runs
9. Audit post-push communications
After pushing to main, verify outbound messages match the deployed state. If you described a feature as "available" but it's not on main at the receiver's pulled commit, send a correction.
Companion tools
scripts/agents-doc-audit.py — prune accumulated doc refs after surveys
Test paths — search tests for stale source paths (src/, deploy/, etc.)
Mid-survey user requests — tag as follow-up, finish current step, return immediately
Why this exists
Created after building check-agents-dot-md.py when agents-doc-audit.py already existed. A 15-second search would have prevented the redundant work.
Phase 0a was added per Luke's directive after Esther identified the gap: agents load always-skills (governance/process) but skip domain skills (craft knowledge). Without domain skills, agents make preventable mistakes — bash portability bugs, wrong cron output formats, path conventions for the wrong OS.
