- name
SKILL.md
技能檔案保留在發布者的儲存庫中。
在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add athola/claude-night-market --skill friction-detector -a cursor -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add athola/claude-night-market --skill friction-detector -a claude-code -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add athola/claude-night-market --skill friction-detector -a codex -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add athola/claude-night-market --skill friction-detector -a grok -y在 Grok Bot 聊天中傳送複製的訊息。Bot 會從 GitHub 取得這個 SKILL.md,並加入你所有 Bot 共用的技能中。
Install this skill: https://github.com/athola/claude-night-market/tree/HEAD/plugins/abstract/skills/friction-detector來自發布者
Detect friction signals during agent execution, track them across sessions, and graduate recurring patterns into permanent guidance. Bridges the gap between ephemeral session friction and durable CLAUDE.md rules.
Research backing: Claude Coach (hook-based friction detection with SQLite storage), alirezarezvani's self-improving-agent (three-tier MEMORY to CLAUDE.md graduation), and the ACE framework (arXiv: evolving playbooks from execution feedback, +10.6% on agent tasks).
Current gap: LEARNINGS.md exists but requires manual aggregation via
/abstract:aggregate-logs. This skill adds automatic friction detection and
a structured promotion path.
Weight scoring: High = 3, Medium = 2, Low = 1 points per occurrence. Weighted score determines graduation velocity.
Tier 1: Friction Log (ephemeral, per-session)
Location: ~/.claude/friction/sessions/{date}-{id}.json
Retention: 30 days, then pruned
Threshold: 1 occurrence, logged, no action
Tier 2: Pattern Candidate (persistent, LEARNINGS.md)
Location: ~/.claude/skills/LEARNINGS.md (friction section)
Threshold: 3+ occurrences across 2+ sessions
Action: flagged for review in next friction report
Tier 3: Graduated Rule (CLAUDE.md or skill update)
Threshold: reviewed + user-approved
Action: permanent guidance added to project/user config
Constraint: NEVER auto-modify CLAUDE.md
graduation_score = (weighted_count * recency_factor) / sessions_seen
recency_factor:
last 7 days = 1.0
8-14 days = 0.7
15-30 days = 0.4
31+ days = 0.1
Tier 2 threshold: graduation_score >= 6.0
Tier 3 proposal: graduation_score >= 12.0
Run at session end, at 80% context usage (via
conserve:clear-context), or after failed improvement
cycles (when abstract:metacognitive-self-mod detects
regression).
For each friction indicator found, wrap it in the shared session-capture envelope (ADR-0011) so downstream readers can ingest friction signals and trace-capture entries through one parser:
{
"schema_version": "session-capture/1",
"session_id": "2026-04-14-abc12345",
"timestamp": "2026-04-14T10:23:00Z",
"source": "friction-detector",
"payload": {
"signal_type": "retry_loop",
"description": "rg command failed 3x, fell back to grep",
"context": "searching for pattern in node_modules",
"weight": "medium"
}
}
Legacy files written before envelope adoption are read
as session-capture/0 (entire file treated as the
payload). See docs/adr/0011-session-capture-envelope.md
for the contract and migration path.
FRICTION_DIR=~/.claude/friction/sessions
mkdir -p "$FRICTION_DIR"
# Count prior occurrences of similar signals
if command -v rg &>/dev/null; then
rg -c "$SIGNAL_TYPE" "$FRICTION_DIR"/*.json 2>/dev/null || echo "0"
else
grep -rc "$SIGNAL_TYPE" "$FRICTION_DIR"/*.json 2>/dev/null || echo "0"
fi
Aggregate across session logs: sum weighted occurrences, apply recency decay, divide by session count, compare against tier thresholds.
Tier 2 crossing: append to LEARNINGS.md friction section. Tier 3 crossing: present proposal with evidence to user, wait for explicit approval before any modification.
Write session log to
~/.claude/friction/sessions/{date}-{session_id}.json
and update ~/.claude/friction/index.json.
Ignore these signals:
Decay factor: signals older than 30 days contribute only 10% of their original weight (see graduation formula recency_factor).
## Friction Report: Session {date}
### New Signals (Tier 1)
- [RETRY] `rg` command failed 3x, fell back to `grep`
- [RE-READ] Read SKILL.md 4 times (lost file structure context)
### Recurring Patterns (Tier 2 candidates)
- [CORRECTION] User corrected file path format 4x across 3 sessions
Score: 8.4 (threshold: 6.0)
Candidate: Add path format guidance to CLAUDE.md
### Graduation Proposals (Tier 3)
- [RULE] "Always use absolute paths in Read tool"
Evidence: 7 corrections across 5 sessions
Score: 14.2 (threshold: 12.0)
Action: Approve / Reject / Defer
### Noise Filtered
- 2 transient network timeouts (ignored)
- 1 user-initiated deep exploration (ignored)
Feeds into: LEARNINGS.md (Tier 2 patterns, same
format as /abstract:aggregate-logs),
abstract:skill-improver (priority scoring), and
abstract:metacognitive-self-mod (pipeline
effectiveness).
Consumes from: session transcripts,
aggregate_learnings_daily hook data, and the
performance tracker for trend correlation.
abstract:skill-authoring)/abstract:aggregate-logs)abstract:metacognitive-self-mod: improvement analysisabstract:skills-eval: evaluation criteria/abstract:aggregate-logs: manual LEARNINGS.md generationconserve:clear-context: triggers friction scan at 80%~/.claude/friction/sessions/{date}-{session_id}.json
via the session-capture/1 schemagraduation_score >= 12.0 generate a Tier 3
proposal. The skill does not auto-modify CLAUDE.mdDetect friction signals; graduate patterns into rules. Use for session retrospectives.
---name: friction-detectordescription: Detect friction signals; graduate patterns into rules. Use for session retrospectives.---# plugins/abstract/skills/friction-detector/SKILL.md常見問題
從發布者下載 plugins/abstract/skills/friction-detector/SKILL.md,或複製路徑後開啟儲存庫。PlugMCP 只收錄這項技能,不代管檔案。
發布者將它存放在 https://github.com/athola/claude-night-market。
Claude、Cursor、Codex 和 Grok Build 可以讀取你從發布者儲存庫加入的 SKILL.md。在聊天中傳送技能的 GitHub 資料夾連結,Grok Bot 就會安裝它。