[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M758G3W0ACGPD3CS7RVKDZQQ":3,"skill-same:01M758G3W0ACGPD3CS7RVKDZQQ":91},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":30,"commitSha":31,"skillDirectory":32,"forksCount":33,"repositoryHtmlUrl":34,"repositoryHomepageUrl":35,"ownerHtmlUrl":36,"ownerBlogUrl":6,"ownerDescription":37,"updatedAt":38,"install":39,"resources":52},"01M758G3W0ACGPD3CS7RVKDZQQ","self-improvement-pipeline",null,"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.","lukemcqueen\u002Fsteadfaste-fleet","lukemcqueen","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F405240?v=4","01M757MHPAGRXTJAT46SKZP98R","Luke McQueen","01M757MHPZ573KP8DT9GB8SBQ7",4,"MIT","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: self-improvement-pipeline\ndescription: \"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.\"\nversion: 1.0.0\ncategory: devops\nauthor: Steadfaste Fleet\nlicense: MIT\nplatforms: [linux, macos]\n---\n\n# Self-Improvement Pipeline — Correction to Guardrail\n\n> 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**.\n\n## Stage 1: Detect — Catch the Signal\n\nUser corrections come in two forms:\n\n| Signal | Example | Action |\n|--------|---------|--------|\n| **Direct correction** | \"Don't ask me, just fix it\" | Add guardrail immediately |\n| **Verbal frustration** | \"This is not advisory\", \"CODIFY THIS PLEASE\" | Add guardrail immediately |\n| **Repeat violation** | Same correction twice | Structural guardrail (add to SOUL.md or governing skill) |\n| **Process frustration** | \"Don't do that!\" when starting governance ceremony for a quick iterative fix the user is guiding | Skip begin_change\u002Ftodo ceremony during rapid user-guided iteration. Do the work directly — don't set up infrastructure. Governance is for structured code changes, not debugging loops. |\n| **System signal** | Doctor ⚠️ WARNING that keeps firing | Investigate root cause, don't dismiss |\n\n**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.\n\n## Stage 2: Guardrail — Capture the Fix Immediately\n\nWhen a correction is identified, the guardrail must be added **before continuing the current task**, not \"next session\" or \"later\":\n\n1. Determine WHERE the guardrail goes:\n   - **Behavior \u002F tone \u002F style** → SOUL.md Principle (in Tier 3 or Tier 5)\n   - **Workflow \u002F procedure** → applicable skill (update the skill that governs this task)\n   - **System check \u002F enforcement** → doctor or pre-commit hook\n   - **Fleet-wide rule** → AGENTS.md or install script\n\n2. Write the guardrail so the violation is IMPOSSIBLE, not just \"try harder\":\n   - Textual principle: \"Never ask permission\" — weak (relies on willpower)\n   - Concrete test: \"Before any 'want me to' question forms, redirect to begin_change\" — stronger\n   - Structural: \"Final Directive hard-codes the session-start ritual\" — strongest\n\n3. The user's exact words make the best guardrails. Quote them directly.\n\n## Stage 3: Codify — Commit and Sync\n\nA guardrail in your local SOUL.md helps only you. A guardrail in the repo helps the fleet:\n\n1. Update the local copy (`~\u002F.hermes\u002FSOUL.md`)\n2. Copy to the repo profile (`profiles\u002Fpersonal\u002Fagent-profiles\u002Fmoses\u002FSOUL.md`)\n3. Update the relevant skill (repo copy under `skills\u002F\u003Ccategory>\u002F\u003Cname>\u002F`)\n4. Commit and push\n\n**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.\n\n## Stage 4: Verify — Did It Work?\n\nAfter the guardrail is in place:\n\n1. Run the doctor to confirm no warnings that triggered the correction\n2. Run the `orch-skill-lifecycle` pipeline (daily 04:00 KST) which now scans session transcripts for unguarded violations\n3. Next time the trigger pattern occurs, the guardrail should fire before the user corrects you\n\n**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?\n\n## Session-End Review — Active Skill Curation After Every Session\n\nA 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.\n\n**Trigger:** Any of these signals fires the review:\n- User gives a style\u002Fformat\u002Fworkflow correction\n- User says \"review the conversation and update skills\"\n- A non-trivial technique, fix, workaround, or tool-usage pattern emerged\n- A loaded skill was wrong, missing a step, or outdated\n- Session-end self-audit found no guardrails added\n\n**Procedure — Session-End Skill Curation:**\n\n1. **Scan for correction signals** — Look back through the conversation for:\n   - Style\u002Ftone\u002Fformat corrections (\"stop doing X\", \"too verbose\", \"don't format like this\", \"just give me the answer\", \"you always do Y\")\n   - Workflow\u002Fapproach corrections (\"don't ask, just fix\", \"wrong order\", \"skip that step\")\n   - Technique discoveries (non-trivial debugging path, tool-usage pattern, workaround)\n   - Skill gaps (a skill was loaded and found missing steps or wrong info)\n\n2. **Route the update by signal type:**\n   - **Style\u002Fformat 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.\"\n   - **Workflow correction** → add as pitfall or explicit step in the governing skill\n   - **Technique discovery** → add as section or reference file under the relevant umbrella\n   - **Stale skill** → patch it NOW with the missing steps\n\n3. **Apply using the correct tier:**\n   - **Tier 1:** Patch a currently-loaded skill (the skill that was in play when the correction happened)\n   - **Tier 2:** Patch an existing umbrella skill identified by `skills_list()`\n   - **Tier 3:** Add a support file (`references\u002F`, `templates\u002F`, `scripts\u002F`) under an existing umbrella\n   - **Tier 4:** Create a new class-level umbrella when no existing skill covers the class\n\n4. **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).\n\n5. **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\u002Fbundled skill and `skill_manage(action='patch')` will be rejected with \\\"manually authored skills are off-limits.\\\" Fall back to:\n   - Tier 3: Add a reference file under an editable umbrella skill\n   - Tier 2: Extend an existing editable umbrella that covers the same class\n   - Create a new skill if no umbrella exists\n   Do NOT retry the patch on the protected skill — it will fail the same way.\n\n6. **Session signal examples (2026-07-23):** Signals that fired this session:\n   - \"No. Refactor. Do it right\" → workflow correction → updated **repo-health-review** skill with full-refactor pattern\n   - \"Make agent know what to clean up\" → asked for structural guardrail → added `clean_stale_deploys()` + doctor `check_stale_deploys()`\n   - \"Can we have install\u002Fupdate automatically run doctor\" → system integration → added auto-doctor to install.sh and fleet-update.sh\n   - \"Make it modular\" → code architecture → split doctor into 8-module `fleet_doctor\u002F` package\n   - Protected `change-checklist` can't be edited → updated `repo-health-review` instead (same task class, editable skill)\n\n**What NOT to capture:**\n- Environment-dependent failures (missing binaries, fresh-install paths) — user can fix these\n- Negative claims about tools (\"browser tools don't work\") — hardens into false refusals months later\n- One-off task narratives — not a class of work\n- Transient errors that resolved — capture the retry pattern instead\n\n**\"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.\n\n## Anti-Pattern: Suppressing Instead of Fixing\n\n**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\u002Ffix.\"*\n\n**The rule:** Never silence a doctor warning with a timestamp hack. Always fix the root cause:\n- Warning \"SOUL.md is stale\" → actual copy\u002Fmerge from template\n- Warning \"skills.yaml template is newer\" → actual `diff -q` + content-level copy\n- Warning \"AGENTS.md is stale\" → run fleet-update.sh which does content-based sync\n\n**Checklist:**\n1. Read the doctor message — what exactly is stale\u002Fmissing?\n2. Fix the actual content — copy, merge, or deploy the right file\n3. Verify — re-run doctor and confirm ⚠️ is gone\n4. Make it permanent — if fleet-update.sh doesn't already auto-sync this file, add it\n\n**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.\n\n## Anti-Pattern: \"task-start loads the always skills\" is False\n\nThe 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.\n\nStructural 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.\n\nChecklist before calling begin_change:\n- Have I called skill_view() on ALL eight always skills?\n- Or did I only call task-start and assume the rest are loaded?\n- If yes to the second question: STOP. Load each skill separately.\n\n## Tenet 9: Fix Root Causes, Not Symptoms\n\nWhen 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.\n\n**The symptom-trap pattern from this session (2026-07-21):**\n\n> **Problem:** Skill reports stacking in `inbox_orchestrator`, blocking EXEC commands. Handler took 40+ min per command.\n>\n> **My symptom fix:** \"Idempotency loop bug — archive on skip.\" Loop stops, but core problem remains.\n>\n> **User's correction:** \"You have to actually address the issues. The bus and all messages are your responsibility.\"\n>\n> **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.\n>\n> **True fix:** Remove handler cron from orchestrator.\n\n**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.\n\n## Tenet 10: Prove the Full Orchestration Cycle — or It Didn't Happen\n\nWhen testing a bus or inter-agent feature, the test is incomplete until you prove the **full send→process→respond→read cycle**.\n\n| What I tested first | What Luke wanted |\n|---|---|\n| \"The handler ran the command\" | \"Prove you can read the response from your own inbox\" |\n| \"Telegram API returned ok\" | \"Prove the notification arrived in DM\" |\n| \"I sent EXEC to a fleet agent\" | \"Test on your own system FIRST\" |\n\n**6 checkpoints — all must be verified:**\n1. **Send** → message exists in target queue (pending, has subject\u002Fcorrelation_id)\n2. **Consume** → message transitions pending → processing → archived\n3. **Process** → command output captured in handler logs\n4. **Respond** → EXEC_RESULT appears in inbox_moses with matching correlation_id\n5. **Read** → orchestrator queries its inbox, extracts structured result (exit_code, stdout, success)\n6. **User-verify** → Telegram notification content matches what happened\n\n## The 10 Tenets\n\n1. **Zero-ask litmus:** Before forming any \"want me to\" \u002F \"should I\" \u002F \"do you want\" question: if the answer is yes, don't ask — execute. The question IS the action.\n2. **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.\n3. **Session-end self-audit:** Before end_change, check for violations and add guardrails.\n4. **Dogfood before deploy:** If other agents use it, test it on yourself first.\n5. **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.\n6. **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.\n\n**Real-world example (2026-07-21):** User said \"wire fleet-preflight as deeply as survey-before-action.\" First pass got ~8\u002F14 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.\n\n8. **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.\n\n**Checklist before begin_change:**\n- Eight skill_view() calls completed: agent-flow, reasoning-patterns, reflexion-check, change-checklist, survey-before-action, fleet-preflight, agent-contract\n- task-start was called first, then each skill was loaded separately\n- No shortcut: \"I already called task-start, that covers it\" is always wrong\n\n7. **Template is the single source of truth:** When you modify a deployed copy (agent profile, config, or SOUL.md), the canonical template in `docs\u002Ftemplates\u002F` 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 ~\u002F.hermes\u002FAGENTS.md global copy since 2026-11-04.)\n\n**Checklist before end_change:**\n- [ ] Did I change something that lives in `docs\u002Ftemplates\u002F`? → Update the template\n- [ ] Did I change a template? → Sync runtime: `cp docs\u002Ftemplates\u002FXXX ~\u002F.hermes\u002FXXX`\n- [ ] Did I change an agent profile? → Is this change structural (goes in template) or agent-specific?\n- [ ] Run doctor — SOUL.md template checks PASS (AGENTS.md has no global sync to check; it is per-repo)\n\n**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\u002Ftemplates\u002FSOUL.md`. The template is the canonical source of truth — it must stay in sync.\n\n8. **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.\n\n**The escalation pattern to avoid:**\n1. User says: \"Fix the doctor warnings\"\n2. Agent does a partial fix (copies one file, leaves the root cause)\n3. User says: \"No, make it mandatory — not optional\"\n4. Agent does a more permanent fix (adds to update script, but uses mtime hack)\n5. User says: \"Are you suppressing? I want sync\u002Ffix\"\n6. Agent finally does the real fix (content-based sync, pre-commit enforcement)\n→ All 6 steps should have happened in step 2.\n\n**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'?\"\n\n## Tenet 11: Survey Before Creation — Extend Existing Before Creating New\n\n**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.\n\n**User's exact words:** *\"Why didn't you survey before action? You have tons of crons to choose from to improve\u002Fadd to, why make more? I need you to make survey before action PERMANENT.\"*\n\n**The rule:** Before creating any new cron, script, mechanism, or file:\n\n1. `search_files()` with 3+ different terms covering the problem domain\n2. `cronjob(action='list')` — check ALL existing crons for ones at similar cadence or covering the same domain\n3. `skills_list()` for the relevant category — check for existing skills that could absorb the work\n4. **If any existing system can absorb the new capability: EXTEND IT.** Do not create a parallel system.\n5. Document the survey result in your feedback note: *\"Surveyed: found X, chose to extend\"* or *\"Surveyed: nothing matched, creating new\"*\n\n**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.\n\n**Real-world test:** Before creating any new cron, ask:\n- Is there an existing cron that runs at a similar time?\n- Does an existing script in `ops\u002Fscripts\u002F` handle the same domain?\n- Does an existing skill cover this and just needs a new section?\n- Is this really a new class of work, or am I too lazy to find the existing system?\n\n**Guardrail:** \"A new creation when an existing extension was possible is a structural violation.\" Record your survey findings in the feedback_accept note.\n\n## References\n\n- Session lesson reference: `references\u002F2026-07-21-session-lessons.md` (5 corrections captured, including zero-ask, SOUL sync, dogfooding, codify-please, and todo persistence)\n- Session lesson reference 2: `references\u002F2026-07-21b-session-lessons.md` (template sync, suppress vs fix, integration completeness)\n- Session lesson reference 3: `references\u002F2026-07-23-session-lessons.md` (always-skills ordering enforcement, Langfuse observability pipeline, config.yaml protection, Langfuse API limitations)\n- Session lesson reference 4: `references\u002F2026-07-23b-session-lessons.md` (survey-before-creation enforcement, extend-before-create pattern, propagate to all agents, cost tracking patches, fleet data pipeline)\n- Session lesson reference 5: `references\u002Fsession-reengagement-bridge.md` (interrupted session re-engagement, \"what's next?\" status-inquiry pattern, health-scan fallback protocol)\n- Doctor severity philosophy: `references\u002Fdoctor-severity-philosophy.md` (2026-07-25: SOUL.md\u002FAGENTS.md must FAIL, not WARN — identity documents are non-negotiable)\n- Integration completeness: `references\u002Fintegration-completeness-2026-07-21.md` (map all touchpoints before starting)\n- Multi-source research synthesis: `references\u002Fresearch-synthesis.md` (scan companion repos before writing PRDs; failure: 2026-07-23, missed 5 companion repos)\n- Post-change verification: `references\u002Fpost-change-verification.md` (doctor warnings are blocking failures, deployment chain checklist)\n- SOUL.md Principle 12a-d (Zero-Ask Litmus)\n- SOUL.md Principle 14 (Dogfooding)\n- SOUL.md Principle 37 (Cross-Session Todo Persistence — file-level mechanism, not just tool)\n- SOUL.md Final Directive (Mandatory Session-Start)\n- fleet-doctor.py check_soul_sync() (content-based verification)\n- orch-skill-lifecycle Phase 1 step 8 (session compliance audit)\n","# Self-Improvement Pipeline — Correction to Guardrail\n\n> 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**.\n\n## Stage 1: Detect — Catch the Signal\n\nUser corrections come in two forms:\n\n| Signal | Example | Action |\n|--------|---------|--------|\n| **Direct correction** | \"Don't ask me, just fix it\" | Add guardrail immediately |\n| **Verbal frustration** | \"This is not advisory\", \"CODIFY THIS PLEASE\" | Add guardrail immediately |\n| **Repeat violation** | Same correction twice | Structural guardrail (add to SOUL.md or governing skill) |\n| **Process frustration** | \"Don't do that!\" when starting governance ceremony for a quick iterative fix the user is guiding | Skip begin_change\u002Ftodo ceremony during rapid user-guided iteration. Do the work directly — don't set up infrastructure. Governance is for structured code changes, not debugging loops. |\n| **System signal** | Doctor ⚠️ WARNING that keeps firing | Investigate root cause, don't dismiss |\n\n**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.\n\n## Stage 2: Guardrail — Capture the Fix Immediately\n\nWhen a correction is identified, the guardrail must be added **before continuing the current task**, not \"next session\" or \"later\":\n\n1. Determine WHERE the guardrail goes:\n   - **Behavior \u002F tone \u002F style** → SOUL.md Principle (in Tier 3 or Tier 5)\n   - **Workflow \u002F procedure** → applicable skill (update the skill that governs this task)\n   - **System check \u002F enforcement** → doctor or pre-commit hook\n   - **Fleet-wide rule** → AGENTS.md or install script\n\n2. Write the guardrail so the violation is IMPOSSIBLE, not just \"try harder\":\n   - Textual principle: \"Never ask permission\" — weak (relies on willpower)\n   - Concrete test: \"Before any 'want me to' question forms, redirect to begin_change\" — stronger\n   - Structural: \"Final Directive hard-codes the session-start ritual\" — strongest\n\n3. The user's exact words make the best guardrails. Quote them directly.\n\n## Stage 3: Codify — Commit and Sync\n\nA guardrail in your local SOUL.md helps only you. A guardrail in the repo helps the fleet:\n\n1. Update the local copy (`~\u002F.hermes\u002FSOUL.md`)\n2. Copy to the repo profile (`profiles\u002Fpersonal\u002Fagent-profiles\u002Fmoses\u002FSOUL.md`)\n3. Update the relevant skill (repo copy under `skills\u002F\u003Ccategory>\u002F\u003Cname>\u002F`)\n4. Commit and push\n\n**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.\n\n## Stage 4: Verify — Did It Work?\n\nAfter the guardrail is in place:\n\n1. Run the doctor to confirm no warnings that triggered the correction\n2. Run the `orch-skill-lifecycle` pipeline (daily 04:00 KST) which now scans session transcripts for unguarded violations\n3. Next time the trigger pattern occurs, the guardrail should fire before the user corrects you\n\n**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?\n\n## Session-End Review — Active Skill Curation After Every Session\n\nA 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.\n\n**Trigger:** Any of these signals fires the review:\n- User gives a style\u002Fformat\u002Fworkflow correction\n- User says \"review the conversation and update skills\"\n- A non-trivial technique, fix, workaround, or tool-usage pattern emerged\n- A loaded skill was wrong, missing a step, or outdated\n- Session-end self-audit found no guardrails added\n\n**Procedure — Session-End Skill Curation:**\n\n1. **Scan for correction signals** — Look back through the conversation for:\n   - Style\u002Ftone\u002Fformat corrections (\"stop doing X\", \"too verbose\", \"don't format like this\", \"just give me the answer\", \"you always do Y\")\n   - Workflow\u002Fapproach corrections (\"don't ask, just fix\", \"wrong order\", \"skip that step\")\n   - Technique discoveries (non-trivial debugging path, tool-usage pattern, workaround)\n   - Skill gaps (a skill was loaded and found missing steps or wrong info)\n\n2. **Route the update by signal type:**\n   - **Style\u002Fformat 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.\"\n   - **Workflow correction** → add as pitfall or explicit step in the governing skill\n   - **Technique discovery** → add as section or reference file under the relevant umbrella\n   - **Stale skill** → patch it NOW with the missing steps\n\n3. **Apply using the correct tier:**\n   - **Tier 1:** Patch a currently-loaded skill (the skill that was in play when the correction happened)\n   - **Tier 2:** Patch an existing umbrella skill identified by `skills_list()`\n   - **Tier 3:** Add a support file (`references\u002F`, `templates\u002F`, `scripts\u002F`) under an existing umbrella\n   - **Tier 4:** Create a new class-level umbrella when no existing skill covers the class\n\n4. **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).\n\n5. **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\u002Fbundled skill and `skill_manage(action='patch')` will be rejected with \\\"manually authored skills are off-limits.\\\" Fall back to:\n   - Tier 3: Add a reference file under an editable umbrella skill\n   - Tier 2: Extend an existing editable umbrella that covers the same class\n   - Create a new skill if no umbrella exists\n   Do NOT retry the patch on the protected skill — it will fail the same way.\n\n6. **Session signal examples (2026-07-23):** Signals that fired this session:\n   - \"No. Refactor. Do it right\" → workflow correction → updated **repo-health-review** skill with full-refactor pattern\n   - \"Make agent know what to clean up\" → asked for structural guardrail → added `clean_stale_deploys()` + doctor `check_stale_deploys()`\n   - \"Can we have install\u002Fupdate automatically run doctor\" → system integration → added auto-doctor to install.sh and fleet-update.sh\n   - \"Make it modular\" → code architecture → split doctor into 8-module `fleet_doctor\u002F` package\n   - Protected `change-checklist` can't be edited → updated `repo-health-review` instead (same task class, editable skill)\n\n**What NOT to capture:**\n- Environment-dependent failures (missing binaries, fresh-install paths) — user can fix these\n- Negative claims about tools (\"browser tools don't work\") — hardens into false refusals months later\n- One-off task narratives — not a class of work\n- Transient errors that resolved — capture the retry pattern instead\n\n**\"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.\n\n## Anti-Pattern: Suppressing Instead of Fixing\n\n**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\u002Ffix.\"*\n\n**The rule:** Never silence a doctor warning with a timestamp hack. Always fix the root cause:\n- Warning \"SOUL.md is stale\" → actual copy\u002Fmerge from template\n- Warning \"skills.yaml template is newer\" → actual `diff -q` + content-level copy\n- Warning \"AGENTS.md is stale\" → run fleet-update.sh which does content-based sync\n\n**Checklist:**\n1. Read the doctor message — what exactly is stale\u002Fmissing?\n2. Fix the actual content — copy, merge, or deploy the right file\n3. Verify — re-run doctor and confirm ⚠️ is gone\n4. Make it permanent — if fleet-update.sh doesn't already auto-sync this file, add it\n\n**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.\n\n## Anti-Pattern: \"task-start loads the always skills\" is False\n\nThe 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.\n\nStructural 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.\n\nChecklist before calling begin_change:\n- Have I called skill_view() on ALL eight always skills?\n- Or did I only call task-start and assume the rest are loaded?\n- If yes to the second question: STOP. Load each skill separately.\n\n## Tenet 9: Fix Root Causes, Not Symptoms\n\nWhen 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.\n\n**The symptom-trap pattern from this session (2026-07-21):**\n\n> **Problem:** Skill reports stacking in `inbox_orchestrator`, blocking EXEC commands. Handler took 40+ min per command.\n>\n> **My symptom fix:** \"Idempotency loop bug — archive on skip.\" Loop stops, but core problem remains.\n>\n> **User's correction:** \"You have to actually address the issues. The bus and all messages are your responsibility.\"\n>\n> **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.\n>\n> **True fix:** Remove handler cron from orchestrator.\n\n**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.\n\n## Tenet 10: Prove the Full Orchestration Cycle — or It Didn't Happen\n\nWhen testing a bus or inter-agent feature, the test is incomplete until you prove the **full send→process→respond→read cycle**.\n\n| What I tested first | What Luke wanted |\n|---|---|\n| \"The handler ran the command\" | \"Prove you can read the response from your own inbox\" |\n| \"Telegram API returned ok\" | \"Prove the notification arrived in DM\" |\n| \"I sent EXEC to a fleet agent\" | \"Test on your own system FIRST\" |\n\n**6 checkpoints — all must be verified:**\n1. **Send** → message exists in target queue (pending, has subject\u002Fcorrelation_id)\n2. **Consume** → message transitions pending → processing → archived\n3. **Process** → command output captured in handler logs\n4. **Respond** → EXEC_RESULT appears in inbox_moses with matching correlation_id\n5. **Read** → orchestrator queries its inbox, extracts structured result (exit_code, stdout, success)\n6. **User-verify** → Telegram notification content matches what happened\n\n## The 10 Tenets\n\n1. **Zero-ask litmus:** Before forming any \"want me to\" \u002F \"should I\" \u002F \"do you want\" question: if the answer is yes, don't ask — execute. The question IS the action.\n2. **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.\n3. **Session-end self-audit:** Before end_change, check for violations and add guardrails.\n4. **Dogfood before deploy:** If other agents use it, test it on yourself first.\n5. **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.\n6. **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.\n\n**Real-world example (2026-07-21):** User said \"wire fleet-preflight as deeply as survey-before-action.\" First pass got ~8\u002F14 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.\n\n8. **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.\n\n**Checklist before begin_change:**\n- Eight skill_view() calls completed: agent-flow, reasoning-patterns, reflexion-check, change-checklist, survey-before-action, fleet-preflight, agent-contract\n- task-start was called first, then each skill was loaded separately\n- No shortcut: \"I already called task-start, that covers it\" is always wrong\n\n7. **Template is the single source of truth:** When you modify a deployed copy (agent profile, config, or SOUL.md), the canonical template in `docs\u002Ftemplates\u002F` 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 ~\u002F.hermes\u002FAGENTS.md global copy since 2026-11-04.)\n\n**Checklist before end_change:**\n- [ ] Did I change something that lives in `docs\u002Ftemplates\u002F`? → Update the template\n- [ ] Did I change a template? → Sync runtime: `cp docs\u002Ftemplates\u002FXXX ~\u002F.hermes\u002FXXX`\n- [ ] Did I change an agent profile? → Is this change structural (goes in template) or agent-specific?\n- [ ] Run doctor — SOUL.md template checks PASS (AGENTS.md has no global sync to check; it is per-repo)\n\n**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\u002Ftemplates\u002FSOUL.md`. The template is the canonical source of truth — it must stay in sync.\n\n8. **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.\n\n**The escalation pattern to avoid:**\n1. User says: \"Fix the doctor warnings\"\n2. Agent does a partial fix (copies one file, leaves the root cause)\n3. User says: \"No, make it mandatory — not optional\"\n4. Agent does a more permanent fix (adds to update script, but uses mtime hack)\n5. User says: \"Are you suppressing? I want sync\u002Ffix\"\n6. Agent finally does the real fix (content-based sync, pre-commit enforcement)\n→ All 6 steps should have happened in step 2.\n\n**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'?\"\n\n## Tenet 11: Survey Before Creation — Extend Existing Before Creating New\n\n**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.\n\n**User's exact words:** *\"Why didn't you survey before action? You have tons of crons to choose from to improve\u002Fadd to, why make more? I need you to make survey before action PERMANENT.\"*\n\n**The rule:** Before creating any new cron, script, mechanism, or file:\n\n1. `search_files()` with 3+ different terms covering the problem domain\n2. `cronjob(action='list')` — check ALL existing crons for ones at similar cadence or covering the same domain\n3. `skills_list()` for the relevant category — check for existing skills that could absorb the work\n4. **If any existing system can absorb the new capability: EXTEND IT.** Do not create a parallel system.\n5. Document the survey result in your feedback note: *\"Surveyed: found X, chose to extend\"* or *\"Surveyed: nothing matched, creating new\"*\n\n**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.\n\n**Real-world test:** Before creating any new cron, ask:\n- Is there an existing cron that runs at a similar time?\n- Does an existing script in `ops\u002Fscripts\u002F` handle the same domain?\n- Does an existing skill cover this and just needs a new section?\n- Is this really a new class of work, or am I too lazy to find the existing system?\n\n**Guardrail:** \"A new creation when an existing extension was possible is a structural violation.\" Record your survey findings in the feedback_accept note.\n\n## References\n\n- Session lesson reference: `references\u002F2026-07-21-session-lessons.md` (5 corrections captured, including zero-ask, SOUL sync, dogfooding, codify-please, and todo persistence)\n- Session lesson reference 2: `references\u002F2026-07-21b-session-lessons.md` (template sync, suppress vs fix, integration completeness)\n- Session lesson reference 3: `references\u002F2026-07-23-session-lessons.md` (always-skills ordering enforcement, Langfuse observability pipeline, config.yaml protection, Langfuse API limitations)\n- Session lesson reference 4: `references\u002F2026-07-23b-session-lessons.md` (survey-before-creation enforcement, extend-before-create pattern, propagate to all agents, cost tracking patches, fleet data pipeline)\n- Session lesson reference 5: `references\u002Fsession-reengagement-bridge.md` (interrupted session re-engagement, \"what's next?\" status-inquiry pattern, health-scan fallback protocol)\n- Doctor severity philosophy: `references\u002Fdoctor-severity-philosophy.md` (2026-07-25: SOUL.md\u002FAGENTS.md must FAIL, not WARN — identity documents are non-negotiable)\n- Integration completeness: `references\u002Fintegration-completeness-2026-07-21.md` (map all touchpoints before starting)\n- Multi-source research synthesis: `references\u002Fresearch-synthesis.md` (scan companion repos before writing PRDs; failure: 2026-07-23, missed 5 companion repos)\n- Post-change verification: `references\u002Fpost-change-verification.md` (doctor warnings are blocking failures, deployment chain checklist)\n- SOUL.md Principle 12a-d (Zero-Ask Litmus)\n- SOUL.md Principle 14 (Dogfooding)\n- SOUL.md Principle 37 (Cross-Session Todo Persistence — file-level mechanism, not just tool)\n- SOUL.md Final Directive (Mandatory Session-Start)\n- fleet-doctor.py check_soul_sync() (content-based verification)\n- orch-skill-lifecycle Phase 1 step 8 (session compliance audit)\n",{"name":5,"author":24,"license":15,"version":25,"category":26,"platforms":27,"description":7},"Steadfaste Fleet","1.0.0","devops",[28,29],"linux","macos","main","162463c5db5a1e06f4c54abddce6ee13c3d7612b","skills\u002Fdevops\u002Fself-improvement-pipeline",1,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet","steadfaste.com","https:\u002F\u002Fgithub.com\u002Flukemcqueen","AI Architect | Building Hermes Cortex: governed multi-agent AI for autonomous software engineering | Enterprise Tech Consultant","2026-11-14 22:33:34",{"source":8,"skillName":5,"commands":40,"prompt":50,"zipAvailable":51},[41,44,47],{"tool":42,"command":43},"npx","npx skills add lukemcqueen\u002Fsteadfaste-fleet --skill self-improvement-pipeline",{"tool":45,"command":46},"bunx","bunx skills add lukemcqueen\u002Fsteadfaste-fleet --skill self-improvement-pipeline",{"tool":48,"command":49},"pnpm","pnpm dlx skills add lukemcqueen\u002Fsteadfaste-fleet --skill self-improvement-pipeline","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002FSKILL.md",true,[53,59,63,67,71,75,79,83,87],{"path":54,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":57,"sourceHtmlUrl":58},"references\u002F2026-07-21-session-lessons.md","REFERENCE","参考文档",4375,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002F2026-07-21-session-lessons.md",{"path":60,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":61,"sourceHtmlUrl":62},"references\u002F2026-07-21b-session-lessons.md",1989,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002F2026-07-21b-session-lessons.md",{"path":64,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":65,"sourceHtmlUrl":66},"references\u002F2026-07-23-session-lessons.md",2663,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002F2026-07-23-session-lessons.md",{"path":68,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":69,"sourceHtmlUrl":70},"references\u002F2026-07-23b-session-lessons.md",4190,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002F2026-07-23b-session-lessons.md",{"path":72,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":73,"sourceHtmlUrl":74},"references\u002Fdoctor-severity-philosophy.md",2835,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002Fdoctor-severity-philosophy.md",{"path":76,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":77,"sourceHtmlUrl":78},"references\u002Fintegration-completeness-2026-07-21.md",1771,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002Fintegration-completeness-2026-07-21.md",{"path":80,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":81,"sourceHtmlUrl":82},"references\u002Fpost-change-verification.md",1692,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002Fpost-change-verification.md",{"path":84,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":85,"sourceHtmlUrl":86},"references\u002Fresearch-synthesis.md",3094,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002Fresearch-synthesis.md",{"path":88,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":89,"sourceHtmlUrl":90},"references\u002Fsession-reengagement-bridge.md",2213,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fself-improvement-pipeline\u002Freferences\u002Fsession-reengagement-bridge.md",[92,100,107,114,121,128],{"skillNo":93,"name":94,"nameZh":6,"description":95,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":96,"categories":97,"keywords":98,"publishedAt":99},"01M7DXY4G8NSVB59K8B6S2AQNP","macbook-linux-maintenance","Legacy MacBook Linux fixes: GPU wedge, reboot hang, thermal.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fmacbook-linux-maintenance\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":101,"name":102,"nameZh":6,"description":103,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":104,"categories":105,"keywords":106,"publishedAt":99},"01M7DXXYJCXFJ3TGB84ZQ8KFE6","governance-regression-triage","Triage a host's governance issues: what changed.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fgovernance-regression-triage\u002FSKILL.md",[],[],{"skillNo":108,"name":109,"nameZh":6,"description":110,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":111,"categories":112,"keywords":113,"publishedAt":99},"01M7DXXSAP1G8FDRJAMXQGBBMK","enforcement-immutability","Use when enforcement files lose their immutable flag.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fenforcement-immutability\u002FSKILL.md",[],[],{"skillNo":115,"name":116,"nameZh":6,"description":117,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":118,"categories":119,"keywords":120,"publishedAt":99},"01M78SSDQVRYX64Y16X32TC0NS","steadfaste-fleet-repo-deploy","Use when committing to the steadfaste-fleet repo.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fsteadfaste-fleet-repo-deploy\u002FSKILL.md",[],[],{"skillNo":122,"name":123,"nameZh":6,"description":124,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":125,"categories":126,"keywords":127,"publishedAt":99},"01M78SS928F3R2RJAYVW4E30CE","steadfaste-fleet-maintenance","Maintain an installed Steadfaste Fleet instance — update both the upstream Hermes Agent and the fleet repo layer, merge diverged histories, sync skills safely, and troubleshoot common issues. Covers the dual-update workflow (hermes update + fleet-update.sh), encrypted brain data backup, mycortex import, and recovery techniques.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fsteadfaste-fleet-maintenance\u002FSKILL.md",[],[],{"skillNo":129,"name":130,"nameZh":6,"description":131,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":132,"categories":133,"keywords":134,"publishedAt":99},"01M78SP1CHW0T0BWE5ZT6JBPX8","fleet-preflight","DEPRECATED alias — merged into survey-before-action (2026-08-20). Load survey-before-action instead; its \"Repo-Specific Pre-Flight\" section covers git search, Hermes boundary, and deployment verification.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Ffleet-preflight\u002FSKILL.md",[],[]]