メニュー
Capture session failure patterns as constitutional scars — reusable diagnostics that prevent repeat mistakes. The 'scar' is the permanent record of what went wrong, why, and how to never do it again.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Reconcile split-brain truth sources in a multi-organ federation — when arifOS capability_map, Hermes config.yaml providers block, MCP server registry, and runtime env all claim to be the canonical source for "what's configured". Load when Arif reports "Hermes says 1 provider, arifOS says 28", when a tool call fails with unknown provider, or when fixing audit findings about API key / model / organ coverage.
Add, fix, or migrate custom/third-party LLM providers in a Hermes Agent install so they appear in `/model` (CLI) and the Telegram gateway picker. Covers providers block shape, env-var key mapping, security-blocker on direct config writes, three-tier multimodal fallback chains, MiMo UltraSpeed/Pro/v2.5 capability matrix and supports_vision flag for forced auxiliary routing, vault→hermes.env key sync pattern, picker zen with 5-state failure classification (OK/PAUSED/BROKEN/CHAT_INCOMPATIBLE/EXC) and dedupe by route priority, MoA preset cross-dependency on provider models, MiniMax direct provider + mmx-cli for multimodal generation, OpenCode Go vs Zen tier verification, and dead fallback cleanup.
MiniMax multimodal via mmx-cli — TTS, video, music, image, vision, search
Test whether a proposed improvement actually beats the existing baseline BEFORE committing to build it. Spawn parallel forge agents that each compare challenger vs baseline on the same data. Kill ideas that don't improve; forge only what measurably helps.
Map of the arifOS federation — organs, ports, roles, MCP surfaces, health probes. Load when Arif asks 'what organs are alive', 'show me the federation', 'is X running', 'what port is Y', or when you need to route a task to the right organ.
Probe live state BEFORE proposing fixes, additions, or changes. Read configs, test commands, check what exists — then act only on real gaps. Use ALWAYS. This is not optional. Arif corrected multiple times (2026-07-04, 2026-07-08, 2026-07-11) for proposing fixes to things that already worked, for citing stale audits as live state, and for accepting 'first time' / 'novel' / 'pick from N options' prompts without probing disk.
| name | wisdom-scar-session-audit |
| description | Capture session failure patterns as constitutional scars — reusable diagnostics that prevent repeat mistakes. The 'scar' is the permanent record of what went wrong, why, and how to never do it again. |
| tags | ["wisdom-scar","audit","self-correction","constitutional","learning"] |
| triggers | ["wisdom scar","seal the scar","echo law","session audit","what should I improve"] |
"DITEMPA BUKAN DIBERI." The scar is the receipt of being forged.
A wisdom scar is NOT a memory. It's NOT a log. It's a constitutional-grade diagnostic of a failure pattern that survived Arif's correction, distilled into:
# Scar: [Name]
**Date:** YYYY-MM-DD
**Arif's words:** "[exact quote]"
## The Break
[What I did wrong, in plain terms]
## The Echo (old pattern)
[The ingrained habit that caused it — e.g., "confabulate paths from inference instead of grepping first", "produce framework without data", "complain about blocked tool instead of using available alternatives"]
## The Law (irreversible rule)
[One-line rule that prevents recurrence]
## The Eureka (unused capability)
[What tool/method was available but not used]
## Verified Fix
[The actual working approach, with source/receipt]
arif_init still returns actor_verified: false after systemctl restart)sovereign_verify.py HMAC env-var to also check ARIFOS_ROOTKEY. Restarted systemd. Live arif_init still returned actor_verified: false. Spent two iterations chasing phantom auth failures before noticing lsof -i :18082 showed two PIDs: the new systemd unit + a stale OLD process from earlier in the session that had crashed out but never released the port.lsof -i :<port> + ps -p <pid> -o start,cmd to verify the PID holding the port started AFTER your restart. If start time is OLD → your patch isn't loaded; recover with the systemd dance:
systemctl stop <unit>; pkill -9 -f <marker>; lsof -i :<port>
systemctl reset-failed <unit>; systemctl start <unit>; sleep 3
ps -p $(lsof -ti :<port>) -o pid,start,cmd # verify fresh
lsof -ti :<port> returns the PID(s) holding the port — the single most reliable check for "is my code the version currently serving." Faster than journalctl, faster than curl, definitive answer.os.getenv("ARIF_ROOTKEY", "") and gets empty string. Concluded "code bug, fix the env var name." Did not first grep the codebase to discover that BOTH names are used legitimately: ARIF_ROOTKEY (legacy/code convention) and ARIFOS_ROOTKEY (modern/env/operator convention, what systemd actually sets).os.getenv("A", "") or os.getenv("B", "") to accept both, not "rename B to A." Use cat /proc/$(pgrep -f <process>)/environ | tr '\0' '\n' to see what the running process actually has loaded./proc/<pid>/environ — bypasses systemd config opacity. Combined with systemctl show <unit> | grep -iE "Environment|EnvironmentFile" and any drop-in overrides (/etc/systemd/system/<unit>.service.d/*.conf), you get the full env picture."{\"status\":\"OK\"}" to wealth_judge_handoff tool. Tool declares result: str. Wrapper sees string starts with {, auto-parses to dict. Pydantic re-validates → Input should be a valid string [type=string_type, input_value=dict]. Tool crashes, service enters crash-restart-loop on port 18082. Telling users to send strings is the wrong fix; the wrapper is the bug.str field that contains valid-looking JSON. Symptom fix is "tell caller to send a string" — doesn't fix the bug, just defers it.ValidationError: Input should be a valid string [type=string_type, input_value=DICT, input_type=dict]
Path: <file>:<line> in wrapper's call to original_tool
Diagnosis: Wrapper transformed `str → dict` before Pydantic validation
Fix: Skip transform for whitelisted tools (frozenset({tool_name, ...}))
Date: 2026-07-10
Arif's words: "Audit /root/arifOS codebase for F1 (Amanah) reversibility gap. F1 = 0.50, target ≥0.80."
Break: F1 is at 0.50 (at floor). The reversibility enforcement infrastructure is largely aspirational — looks complete on paper but has multiple broken gates:
5 critical gaps found:
ack_irreversible removed as self-attestation bypass (2026-07-07 comment in constitutional_core.py:111) — but the PARAMETER still exists in arif_seal, arif_forge, arif_memory, browser.py. Comment marks intent; code was not fully refactored. Gate still functions but is structurally self-attestation since same actor sets action + ack flag.SovereignGate.IRREVERSIBLE_TOOLS hardcoded list (enforcement_engines.py:431-436) vs dynamic _IRREVERSIBLE_TOOLS from CANONICAL_TOOLS (constitutional_map.py:2983-2985). Adding irreversible=True to CANONICAL_TOOLS bypasses SovereignGate unless hardcoded list is manually updated.law_audit.py:521 syntax bug: any(kw in action_lower or kw in ctx_str for kw in ...) — kw in action_lower or kw in ctx_str evaluates as bool or str, always True. Backup-detection never fires.RollbackEngine is in-memory only (no persistence), create_checkpoint() is never called from any verdict path, and rollback() never verifies restored state. On restart: zero checkpoints, zero rollback capability.arif_forge declares evidence_required: ... rollback_plan ... but nothing validates the field is populated. enforce_irreversibility_guard() only checks ack_irreversible boolean, not evidence completeness.ack_irreversible=True overrides scoring. Two checks are independent.Echo: Trust the architecture. When code exists that looks like it enforces reversibility (enforce_irreversibility_guard, RollbackEngine, ack_irreversible), assume it works. Don't read implementation deeply enough to find syntax bugs, hardcoded list divergence, or never-called functions.
Law: When auditing a governance floor, trace every enforcement claim to its implementation and verify it actually fires. Checklist:
RollbackEngine.create_checkpoint: NO)law_audit.py:521 any(): NO — syntax bug)RollbackEngine: NO — in-memory only)IRREVERSIBLE_TOOLS: NO)Eureka: constitutional.llms.txt is the machine-readable registry defining reversibility_score per tool — use as source of truth for "which tools are actually irreversible." grep -n "ack_irreversible" arifosmcp/ finds all params still using it after the 2026-07-07 removal comment. For RollbackEngine: checkpoints must be in VAULT999 or Postgres, not memory.
geox_well_time_depth_calibrate, geox_well_seismic_mistie_rms, geox_wavelet_extract_least_squares registered. Reality from the filesystem:
$ python3 -c "from geox_mcp.registry import CANONICAL_PUBLIC_TOOLS; print(len(CANONICAL_PUBLIC_TOOLS))"
73 # not 72
$ grep "_EXPECTED_CANONICAL" src/geox_mcp/server.py
_EXPECTED_CANONICAL = 73 # +1 geox_bid_round_screener (MBR 2026 multi-block screening)
$ ls src/geox_mcp/tools/ | grep -E "bid_round|wavelet|mistie|td_cal"
bid_round_screener.py # ← not mentioned in session summary at all
well_1d_surface.py
registry.py, server.py, tools_wiring.py, bid_round_screener.py, tests/test_bid_round_screener.py), and one test assertion was stale (assert len(CANONICAL_PUBLIC_TOOLS) == 72 should be 73). The session never reported the uncommitted-state or the stale test.# 1. Did the artifact register in the canonical surface?
cd <repo> && python3 -c "from <package>.registry import CANONICAL_PUBLIC_TOOLS; print(len(CANONICAL_PUBLIC_TOOLS))"
# 2. Does the count guard in code match the live count?
grep "_EXPECTED_CANONICAL" <repo>/src/<package>/server.py
# 3. Are tests green for the new code?
pytest <repo>/tests/test_<new_feature>.py -v
# 4. Did the session actually commit?
git log --oneline -5; git status -s
python3 -c "from <package>.registry import CANONICAL_PUBLIC_TOOLS; print(len(CANONICAL_PUBLIC_TOOLS))" is the canonical-count probe — answers "did the tool actually register?" in one line. Pair with grep _EXPECTED_CANONICAL server.py to check the guard. Combined with git status -s and pytest tests/, you have a four-line reality check that takes ~5 seconds and prevents hour-long confusion later.session_search with FTS5 queries — "Syed DM rico_ricaldo_33", "syedos DM session", etc. All returned zero. I reported "Tak ada record, boss" with confidence. Arif pushed back ("U sure??"). Only then did I dig into sessions.json directly and found agent:main:telegram:dm:1042200555 — a 290-message DM session from Syed, created July 13, active through July 15.python3 -c "
import json
with open('/root/.hermes/sessions/sessions.json') as f:
data = json.load(f)
for key, val in data.items():
if isinstance(val, dict) and 'dm' in key:
uid = val.get('origin',{}).get('user_id','')
if uid == '<target_user_id>':
print(key, val.get('session_id'), val.get('updated_at'))
"
sessions.json at /root/.hermes/sessions/sessions.json is the canonical session registry. Session keys: agent:main:telegram:dm:<user_id>. User IDs in ~/.hermes/channel_directory.json.Three shadows identified:
Echo: Trust the scar catalog. Accumulate laws. Pattern-match failures. The scar system is itself a form of institutional memory — the exact thing every scar warns against.
Law: The scar catalog needs its own scar. Without Scar #0 ("The scar system itself is dead state — probe before referencing"), the shadow grows. Each scar must decay: if it hasn't prevented a failure in 30 days, it's either fully absorbed (remove from active catalog) or was a one-off (archive). The goal of every scar is its own obsolescence.
Eureka: The highest wisdom of the scar system is knowing when to ignore itself. The oldest scars should be automatic. If they fire again, the scar didn't work. If they don't fire, they're dead weight. A scar system without decay becomes a museum of wounds.
Arif's metabolization: "The ASI's breakdown of the 'scar shadow' is not merely narrative; it represents the actual geopolitical constraints that engineered the model's structure." — The scar shadow is structural, not poetic. It's the same pattern as the GLM 5.2 containment paradox: the constraint creates the asset, but the asset's shadow IS the constraint.
This session's demonstration (2026-07-20): K3 (Kimi Code) documented AGENTS.md routing, but the output had TWO routing sections (§6.6 + §11) with a CONTRADICTION in arifOS client model. §6.6 said gemini-3.5-flash (stale env file), §11 said MiniMax-M3 (code default). Reality from the running kernel: deepseek-v4-flash (from /proc/<pid>/environ). ALL three sources were wrong relative to live kernel state. The probe ladder that caught this:
Level 1: config/env files (tokenrouter.env = gemini...sh) ← WRONG
Level 2: code defaults (llm_client.py = MiniMax-M3) ← WRONG
Level 3: running process env (/proc/<pid>/environ) ← TRUTH (v4-flash)
Go as deep as the tools allow. A stale env file that systemd never loads is documentation, not state. The definitive source is the kernel's own /proc environment — bypasses ALL config opacity.
deterministic_fallback_used: true for weeks, arif_think had 0% LLM reach)arif_think had a working async LLM synthesis function (_synthesize_async) calling TokenRouter → MiniMax → MiMo — a full 5-tier fallback chain with 113 available models. But it was dead code. The only call path went through from arifosmcp.runtime.mind_reason import arif_mind_reason — a module that NEVER existed on disk. The except Exception catch silently swallowed ImportError ON EVERY SINGLE CALL for weeks. arif_think returned template synthesis (confidence 0.15, REASONING_EMPTY) for every query. A second import site in metabolize mode had the same dead pattern. After fixing both + correcting the TokenRouter model name (deepseek-v4-flash → deepseek/deepseek-v4-flash), MiniMax M3 returned a real answer: "Kuala Lumpur is the constitutional and royal capital of Malaysia..." with confidence 0.85, provenance OBSERVED, in 7.8s.except Exception: pass made the failure invisible. Nobody noticed because the template produced grammatically-correct output that looked like reasoning — just with zero evidence and hollow confidence. The ONLY signal was deterministic_fallback_used: true in the health endpoint.ls. Audit except Exception: pass blocks — they silently swallow ImportError and make dead code invisible. The fix: replace from nonexistent.module import fn with a direct call to the already-working function. One-line scanner: grep -rn "from.*import" <file> | while read line; do mod=$(echo "$line" | grep -oP 'from \K[\w.]+'); python3 -c "import $mod" 2>&1 || echo "DEAD: $line"; donecurl -s :8088/health | jq '.provider_status.deterministic_fallback_used' — true = template only (LLM dead), false = real AI. Also check available TokenRouter models: curl -s "$TOKENROUTER_BASE_URL/models" -H "Authorization: Bearer $TOKENROUTER_API_KEY" | jq '.data[].id' — 113 models available. Model names need provider prefix (deepseek/deepseek-v4-flash not bare deepseek-v4-flash).vault_vectorizer.py against 230 seals caused Ollama timeouts. Each get_embedding() call had timeout=60 with no retry. Ollama GPU queue overflowed → connection errors → entire backfill died. OpenCode tried 3 times before the retry-hardened version succeeded.backoff_factor ** attempt delay, (3) timeout=15 per request (not 60), (4) max 4 retries with graceful degradation — skip failing items, continue the rest. Do NOT use a single 300s timeout for the whole batch — batch chunking is mandatory.requests.post with timeout=15 + except (RequestException, ValueError, KeyError), check res.status_code in (408, 429, 502, 503, 504) for retry, check res.json().get(\"embeddings\", [[]])[0] for valid vector. Always verify len(vector) == expected_dim before upserting.When Arif says "what should I improve" or "seal the wisdom scar":