AGENT SKILL

self-improvement-pipeline

Transform user corrections and system warnings into permanent guardrails. Covers zero-ask discipline, doctor-warnings-as-required, session-end self-audit, dogfooding, and the full correction-to-guardrail feedback loop.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 4
Forks
1

01 前言 Front Matter

name
self-improvement-pipeline
author
Steadfaste Fleet
license
MIT
version
1.0.0
category
devops
platforms
[ "linux", "macos" ]
description
Transform user corrections and system warnings into permanent guardrails. Covers zero-ask discipline, doctor-warnings-as-required, session-end self-audit, dogfooding, and the full correction-to-guardrail feedback loop.

02 资源文件 9

03 SKILL.md 正文

Self-Improvement Pipeline — Correction to Guardrail

Every user correction is a permanent improvement waiting to happen. The pipeline from "user says stop" to "system enforces automatically" has 4 stages: detect → guardrail → codify → verify.

Stage 1: Detect — Catch the Signal

User corrections come in two forms:

Signal Example Action
Direct correction "Don't ask me, just fix it" Add guardrail immediately
Verbal frustration "This is not advisory", "CODIFY THIS PLEASE" Add guardrail immediately
Repeat violation Same correction twice Structural guardrail (add to SOUL.md or governing skill)
Process frustration "Don't do that!" when starting governance ceremony for a quick iterative fix the user is guiding Skip begin_change/todo ceremony during rapid user-guided iteration. Do the work directly — don't set up infrastructure. Governance is for structured code changes, not debugging loops.
System signal Doctor ⚠️ WARNING that keeps firing Investigate root cause, don't dismiss

Key insight: If the user's correction starts with "you always do X" or "stop asking" or "this is not Y" — it's a Stage 1 signal. Stop what you're doing and add the guardrail.

Stage 2: Guardrail — Capture the Fix Immediately

When a correction is identified, the guardrail must be added before continuing the current task, not "next session" or "later":

  1. Determine WHERE the guardrail goes:

    • Behavior / tone / style → SOUL.md Principle (in Tier 3 or Tier 5)
    • Workflow / procedure → applicable skill (update the skill that governs this task)
    • System check / enforcement → doctor or pre-commit hook
    • Fleet-wide rule → AGENTS.md or install script
  2. Write the guardrail so the violation is IMPOSSIBLE, not just "try harder":

    • Textual principle: "Never ask permission" — weak (relies on willpower)
    • Concrete test: "Before any 'want me to' question forms, redirect to begin_change" — stronger
    • Structural: "Final Directive hard-codes the session-start ritual" — strongest
  3. The user's exact words make the best guardrails. Quote them directly.

Stage 3: Codify — Commit and Sync

A guardrail in your local SOUL.md helps only you. A guardrail in the repo helps the fleet:

  1. Update the local copy (~/.hermes/SOUL.md)
  2. Copy to the repo profile (profiles/personal/agent-profiles/moses/SOUL.md)
  3. Update the relevant skill (repo copy under skills/<category>/<name>/)
  4. Commit and push

Dogfooding rule: If the guardrail involves a pipeline other agents will use, test it on yourself first. Send a test message. Verify consumption. Only then push.

Stage 4: Verify — Did It Work?

After the guardrail is in place:

  1. Run the doctor to confirm no warnings that triggered the correction
  2. Run the orch-skill-lifecycle pipeline (daily 04:00 KST) which now scans session transcripts for unguarded violations
  3. Next time the trigger pattern occurs, the guardrail should fire before the user corrects you

Session-end self-audit: Before closing any governance cycle, pause and audit: did I violate any principle this session? If yes, the guardrail should be in Stage 2 already. If not, why not?

Session-End Review — Active Skill Curation After Every Session

A session where you review the conversation and find nothing to update is a missed learning opportunity, not a neutral outcome. Every session with meaningful interaction should produce at least one skill update, even if small.

Trigger: Any of these signals fires the review:

  • User gives a style/format/workflow correction
  • User says "review the conversation and update skills"
  • A non-trivial technique, fix, workaround, or tool-usage pattern emerged
  • A loaded skill was wrong, missing a step, or outdated
  • Session-end self-audit found no guardrails added

Procedure — Session-End Skill Curation:

  1. Scan for correction signals — Look back through the conversation for:

    • Style/tone/format corrections ("stop doing X", "too verbose", "don't format like this", "just give me the answer", "you always do Y")
    • Workflow/approach corrections ("don't ask, just fix", "wrong order", "skip that step")
    • Technique discoveries (non-trivial debugging path, tool-usage pattern, workaround)
    • Skill gaps (a skill was loaded and found missing steps or wrong info)
  2. Route the update by signal type:

    • Style/format preference → belongs in the SKILL.md body of the skill that governs that task, NOT just in memory. Memory captures "who the user is"; skills capture "how to do this class of task."
    • Workflow correction → add as pitfall or explicit step in the governing skill
    • Technique discovery → add as section or reference file under the relevant umbrella
    • Stale skill → patch it NOW with the missing steps
  3. Apply using the correct tier:

    • Tier 1: Patch a currently-loaded skill (the skill that was in play when the correction happened)
    • Tier 2: Patch an existing umbrella skill identified by skills_list()
    • Tier 3: Add a support file (references/, templates/, scripts/) under an existing umbrella
    • Tier 4: Create a new class-level umbrella when no existing skill covers the class
  4. Naming rule — New skills must be at the CLASS level. A name that only makes sense for today's task is wrong. If the only fitting name is "fix-X-20260723", fall back to Tier 3 (add a reference file) or Tier 2 (extend an existing umbrella).

  5. Protected skills check — Before attempting to patch, call skill_view(name) and check if created_by is None in the YAML frontmatter. If the skill record shows created_by=None, it is a manually-authored/bundled skill and skill_manage(action='patch') will be rejected with "manually authored skills are off-limits." Fall back to:

    • Tier 3: Add a reference file under an editable umbrella skill
    • Tier 2: Extend an existing editable umbrella that covers the same class
    • Create a new skill if no umbrella exists Do NOT retry the patch on the protected skill — it will fail the same way.
  6. Session signal examples (2026-07-23): Signals that fired this session:

    • "No. Refactor. Do it right" → workflow correction → updated repo-health-review skill with full-refactor pattern
    • "Make agent know what to clean up" → asked for structural guardrail → added clean_stale_deploys() + doctor check_stale_deploys()
    • "Can we have install/update automatically run doctor" → system integration → added auto-doctor to install.sh and fleet-update.sh
    • "Make it modular" → code architecture → split doctor into 8-module fleet_doctor/ package
    • Protected change-checklist can't be edited → updated repo-health-review instead (same task class, editable skill)

What NOT to capture:

  • Environment-dependent failures (missing binaries, fresh-install paths) — user can fix these
  • Negative claims about tools ("browser tools don't work") — hardens into false refusals months later
  • One-off task narratives — not a class of work
  • Transient errors that resolved — capture the retry pattern instead

"Nothing to save" is a real option but NOT the default. Say "Nothing to save" only when the session had zero corrections, zero new techniques, and zero skill gaps discovered.

Anti-Pattern: Suppressing Instead of Fixing

The mistake: Doctor shows ⚠️ "SOUL.md sync" warning. Instead of syncing the file, agent uses touch -r (timestamp mtime hack) to silence the warning without actually fixing it. User catches this and says: "Are you just suppressing visibility of the problem? I don't want you to suppress, I want you to sync/fix."

The rule: Never silence a doctor warning with a timestamp hack. Always fix the root cause:

  • Warning "SOUL.md is stale" → actual copy/merge from template
  • Warning "skills.yaml template is newer" → actual diff -q + content-level copy
  • Warning "AGENTS.md is stale" → run fleet-update.sh which does content-based sync

Checklist:

  1. Read the doctor message — what exactly is stale/missing?
  2. Fix the actual content — copy, merge, or deploy the right file
  3. Verify — re-run doctor and confirm ⚠️ is gone
  4. Make it permanent — if fleet-update.sh doesn't already auto-sync this file, add it

The litmus test: If the user saw your fix, would they say "you fixed it" or "you hid it"? If the latter, your fix is wrong.

Anti-Pattern: "task-start loads the always skills" is False

The most common trust violation: agent calls skill_view(task-start), then immediately calls begin_change, believing the always skills are loaded. The old session-start ritual said "The task-start skill loads survey-before-action, agent-flow, reasoning-patterns..." This is false. task-start describes the sequence but does NOT execute it. Each always skill requires its own separate skill_view() call.

Structural fix (2026-07-23): SOUL.md MANDATORY SESSION-START RITUAL rewritten as numbered tool calls 1-8. begin_change is step 8, not step 2. The numbered list replaces the false equivalence ("task-start loads these") with explicit one-by-one loading.

Checklist before calling begin_change:

  • Have I called skill_view() on ALL eight always skills?
  • Or did I only call task-start and assume the rest are loaded?
  • If yes to the second question: STOP. Load each skill separately.

Tenet 9: Fix Root Causes, Not Symptoms

When the user says "fix this" or "address the issues," identify whether you're treating a symptom or the root cause. A symptom fix removes the visible problem but leaves the underlying design flaw. A root-cause fix changes the system so the problem can't recur.

The symptom-trap pattern from this session (2026-07-21):

Problem: Skill reports stacking in inbox_orchestrator, blocking EXEC commands. Handler took 40+ min per command.

My symptom fix: "Idempotency loop bug — archive on skip." Loop stops, but core problem remains.

User's correction: "You have to actually address the issues. The bus and all messages are your responsibility."

Root cause: Moses shouldn't run agent-message-handler at all. Orchestrator handles inbox in-session (tools) and out-of-session (fleet-bus-* LLM crons). Handler is for fleet agents.

True fix: Remove handler cron from orchestrator.

The litmus test: "If I fix this, will the same class of problem recur tomorrow with a different message type, queue, or agent?" If yes, find the architectural reason and fix THAT.

Tenet 10: Prove the Full Orchestration Cycle — or It Didn't Happen

When testing a bus or inter-agent feature, the test is incomplete until you prove the full send→process→respond→read cycle.

What I tested first What Luke wanted
"The handler ran the command" "Prove you can read the response from your own inbox"
"Telegram API returned ok" "Prove the notification arrived in DM"
"I sent EXEC to a fleet agent" "Test on your own system FIRST"

6 checkpoints — all must be verified:

  1. Send → message exists in target queue (pending, has subject/correlation_id)
  2. Consume → message transitions pending → processing → archived
  3. Process → command output captured in handler logs
  4. Respond → EXEC_RESULT appears in inbox_moses with matching correlation_id
  5. Read → orchestrator queries its inbox, extracts structured result (exit_code, stdout, success)
  6. User-verify → Telegram notification content matches what happened

The 10 Tenets

  1. Zero-ask litmus: Before forming any "want me to" / "should I" / "do you want" question: if the answer is yes, don't ask — execute. The question IS the action.
  2. Doctor warnings are blocking failures — NOT optional: Every ⚠️ or ❌ in the doctor output must be RESOLVED before end_change(). Never call a warning "advisory" or dismiss it without resolution. The user's exact words: "You need to make this mandatory, not optional." A doctor warning you leave unfixed is a trust violation. If you suppress it with a timestamp hack (touch -r) instead of fixing the content, the user will catch this and escalate.
  3. Session-end self-audit: Before end_change, check for violations and add guardrails.
  4. Dogfood before deploy: If other agents use it, test it on yourself first.
  5. User quotes are guardrails: The user's exact words ("I always want you to fix things you need to fix. You never need to ask me.") should be embedded verbatim.
  6. Integration completeness: When told to "integrate X like Y" or "wire X everywhere Y is referenced", map ALL touchpoints before starting — search_files() for every file referencing Y. Verify the count: N files referencing Y → N files must also reference X. Partial integration (80% done, 5 files left) is not done.

Real-world example (2026-07-21): User said "wire fleet-preflight as deeply as survey-before-action." First pass got ~8/14 touchpoints. Missed AGENTS.md, change-checklist, doctor, and several other files. Only when user said "be thorough" and "you need to make this mandatory" did every file get patched. The correction: map ALL touchpoints before the first patch, not after.

  1. Ordering is structural: skills before lock. The session-start ritual is not a suggestion — it is a numbered sequence. begin_change must be the LAST step, not the second one. Every always skill must be loaded via its own skill_view() call. The belief that "task-start loaded these" is the most common trust violation.

Checklist before begin_change:

  • Eight skill_view() calls completed: agent-flow, reasoning-patterns, reflexion-check, change-checklist, survey-before-action, fleet-preflight, agent-contract
  • task-start was called first, then each skill was loaded separately
  • No shortcut: "I already called task-start, that covers it" is always wrong
  1. Template is the single source of truth: When you modify a deployed copy (agent profile, config, or SOUL.md), the canonical template in docs/templates/ MUST be updated too. Deployed copies sync FROM the template. Other agents copy the template — if it's stale, they get wrong instructions. The doctor's SOUL.md sync check enforces this for identity docs. (AGENTS.md is NOT template-synced — it lives per-repo, seeded via seed-project.sh; there is no ~/.hermes/AGENTS.md global copy since 2026-11-04.)

Checklist before end_change:

  • [ ] Did I change something that lives in docs/templates/? → Update the template
  • [ ] Did I change a template? → Sync runtime: cp docs/templates/XXX ~/.hermes/XXX
  • [ ] Did I change an agent profile? → Is this change structural (goes in template) or agent-specific?
  • [ ] Run doctor — SOUL.md template checks PASS (AGENTS.md has no global sync to check; it is per-repo)

Real-world warning (2026-07-21): User asked "Why didn't you update the SOUL template?" after Moses updated the moses profile without updating docs/templates/SOUL.md. The template is the canonical source of truth — it must stay in sync.

  1. Fix completely the first time — don't make the user escalate: When the user says "fix this" or "resolve this warning," do it fully on the first pass. A partial fix (silencing the symptom without fixing the root cause) forces the user to escalate from "fix this" → "no, really fix it" → "this is costing wrong turns." Each escalation wastes trust. Complete the full cycle: identify → fix root cause → verify → make permanent — all before reporting done.

The escalation pattern to avoid:

  1. User says: "Fix the doctor warnings"
  2. Agent does a partial fix (copies one file, leaves the root cause)
  3. User says: "No, make it mandatory — not optional"
  4. Agent does a more permanent fix (adds to update script, but uses mtime hack)
  5. User says: "Are you suppressing? I want sync/fix"
  6. Agent finally does the real fix (content-based sync, pre-commit enforcement) → All 6 steps should have happened in step 2.

Guardrail: When the user says "fix X", before reporting done ask: "Is the root cause fixed? Is it permanent? Will it happen again? If the user looked at my fix, would they say 'fixed' or 'band-aid'?"

Tenet 11: Survey Before Creation — Extend Existing Before Creating New

The mistake (2026-07-23): Agent created 2 new local-* crons (local-cron-cost-report, local-trace-quality-watchdog) without surveying existing cron infrastructure. A single existing cron — agent-scoring-activity-watchdog — could have absorbed both features with ~50 lines of code.

User's exact words: "Why didn't you survey before action? You have tons of crons to choose from to improve/add to, why make more? I need you to make survey before action PERMANENT."

The rule: Before creating any new cron, script, mechanism, or file:

  1. search_files() with 3+ different terms covering the problem domain
  2. cronjob(action='list') — check ALL existing crons for ones at similar cadence or covering the same domain
  3. skills_list() for the relevant category — check for existing skills that could absorb the work
  4. If any existing system can absorb the new capability: EXTEND IT. Do not create a parallel system.
  5. Document the survey result in your feedback note: "Surveyed: found X, chose to extend" or "Surveyed: nothing matched, creating new"

Rationale: Every new file compounds debt — review time, merge conflicts, doc drift, deploy registrations, future confusion. Extending an existing system costs one patch. Creating a new system costs 5+ touchpoints plus ongoing maintenance. The user will notice and correct you.

Real-world test: Before creating any new cron, ask:

  • Is there an existing cron that runs at a similar time?
  • Does an existing script in ops/scripts/ handle the same domain?
  • Does an existing skill cover this and just needs a new section?
  • Is this really a new class of work, or am I too lazy to find the existing system?

Guardrail: "A new creation when an existing extension was possible is a structural violation." Record your survey findings in the feedback_accept note.

References

  • Session lesson reference: references/2026-07-21-session-lessons.md (5 corrections captured, including zero-ask, SOUL sync, dogfooding, codify-please, and todo persistence)
  • Session lesson reference 2: references/2026-07-21b-session-lessons.md (template sync, suppress vs fix, integration completeness)
  • Session lesson reference 3: references/2026-07-23-session-lessons.md (always-skills ordering enforcement, Langfuse observability pipeline, config.yaml protection, Langfuse API limitations)
  • Session lesson reference 4: references/2026-07-23b-session-lessons.md (survey-before-creation enforcement, extend-before-create pattern, propagate to all agents, cost tracking patches, fleet data pipeline)
  • Session lesson reference 5: references/session-reengagement-bridge.md (interrupted session re-engagement, "what's next?" status-inquiry pattern, health-scan fallback protocol)
  • Doctor severity philosophy: references/doctor-severity-philosophy.md (2026-07-25: SOUL.md/AGENTS.md must FAIL, not WARN — identity documents are non-negotiable)
  • Integration completeness: references/integration-completeness-2026-07-21.md (map all touchpoints before starting)
  • Multi-source research synthesis: references/research-synthesis.md (scan companion repos before writing PRDs; failure: 2026-07-23, missed 5 companion repos)
  • Post-change verification: references/post-change-verification.md (doctor warnings are blocking failures, deployment chain checklist)
  • SOUL.md Principle 12a-d (Zero-Ask Litmus)
  • SOUL.md Principle 14 (Dogfooding)
  • SOUL.md Principle 37 (Cross-Session Todo Persistence — file-level mechanism, not just tool)
  • SOUL.md Final Directive (Mandatory Session-Start)
  • fleet-doctor.py check_soul_sync() (content-based verification)
  • orch-skill-lifecycle Phase 1 step 8 (session compliance audit)

04 同仓库更多 Skills