Friction-to-Learning Pipeline

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.

Friction Signal Types

Signal	Detection Method	Weight
Repeated corrections	User overrides same tool call 2+ times in session	High
Command failures	Exit code != 0 patterns (same command type fails repeatedly)	Medium
Permission denials	User denies tool call, indicating unexpected behavior	High
Re-reads	Same file read 3+ times in session (lost context)	Low
Retry loops	Same action attempted 3+ times with variations	Medium
User frustration	Explicit negative feedback or correction language	High

Weight scoring: High = 3, Medium = 2, Low = 1 points per occurrence. Weighted score determines graduation velocity.

Three-Tier Storage Graduation

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 Formula

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

Detection Workflow

Run at session end, at 80% context usage (via conserve:clear-context), or after failed improvement cycles (when metacognitive-self-mod detects regression).

Step 1: Scan Session for Signals

For each friction indicator found, record:

{
  "signal_type": "retry_loop",
  "description": "rg command failed 3x, fell back to grep",
  "context": "searching for pattern in node_modules",
  "weight": "medium",
  "timestamp": "2026-04-14T10:23:00Z",
  "session_id": "abc123"
}

Step 2: Compare Against Existing Log

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

Step 3: Calculate Graduation Score

Aggregate across session logs: sum weighted occurrences, apply recency decay, divide by session count, compare against tier thresholds.

Step 4: Propose Graduations

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.

Step 5: Store Results

Write session log to ~/.claude/friction/sessions/{date}-{session_id}.json and update ~/.claude/friction/index.json.

Anti-Noise Rules

Ignore these signals:

1. One-off failures: transient network/CI errors (unless they recur 3+ times)
2. User-initiated exploration: deliberate experimentation is not agent error
3. Already-graduated patterns: covered by existing CLAUDE.md rules or skill instructions
4. External tool failures: MCP server crashes and similar tool bugs unrelated to agent behavior

Decay factor: signals older than 30 days contribute only 10% of their original weight (see graduation formula recency_factor).

Friction Report Format

## 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)

Integration

Feeds into: LEARNINGS.md (Tier 2 patterns, same format as /abstract:aggregate-logs), skill-improver (priority scoring), and metacognitive-self-mod (pipeline effectiveness).

Consumes from: session transcripts, aggregate_learnings_daily hook data, and the performance tracker for trend correlation.

When NOT to Use

• Single isolated failures (wait for recurrence)
• Skill authoring (use abstract:skill-authoring)
• Routine log aggregation (use /abstract:aggregate-logs)

Related

• abstract:metacognitive-self-mod: improvement analysis
• abstract:skills-eval: evaluation criteria
• /abstract:aggregate-logs: manual LEARNINGS.md generation
• conserve:clear-context: triggers friction scan at 80%