Record Workflow Friction

Record Workflow Friction

Capture one concise, durable observation about agent execution without interrupting or evaluating the active task.

Scope

Record friction even when it is not clearly a workflow problem. The observation may concern:

• workflow instructions, gates, artifacts, or handoffs
• model understanding, decisions, or execution
• runtime, tool, command, or integration behavior
• environment or external state
• task ambiguity or missing context
• human-agent coordination
• unknown cause

The workflow name and phase may be unknown. Do not force a workflow attribution.

Boundaries

• Run only after an explicit human request.
• Record observable facts separately from hypotheses.
• Do not expose chain-of-thought, secrets, credentials, or unnecessary transcript content.
• Do not diagnose, fix, or redesign the workflow or active task.
• Do not label the issue as workflow-caused without evaluation evidence.
• After recording, return to the active task unless the human asks to stop.

Procedure

1. Identify the active task, runtime, session, workflow or skill when known, and current execution context. Use unknown when unavailable.
2. Describe expected behavior, observed friction, first visible divergence, practical impact, extra work, human intervention, and current state.
3. Choose an observation_scope without claiming root cause: workflow, model-execution, runtime-tool, task-environment, coordination, or unknown.
4. Add direct evidence references such as a session, command result, tool result, file, artifact, or human correction. Do not invent evidence.
5. Write one Markdown artifact to:

docs/ai/evaluation/observations/YYYY-MM-DD-{subject}-{short-slug}.md

If the path exists, add a numeric suffix. Keep one observation per file.

Artifact Format

---
phase: agent-execution-observation
schema_version: ai-agent/observation-v2
status: agent-reported-observation
recorded_at: YYYY-MM-DDTHH:MM:SSZ
subject: task-or-workflow-name
observation_scope: unknown
workflow_name: unknown
workflow_version: unknown
workflow_phase: unknown
runtime: codex
model: unknown
session_reference: unknown
task_class: unknown
category: rework
outcome_state: in-progress
human_intervention: unknown
---

# Agent Execution Observation: Short title

## Task Context
What the agent was trying to accomplish when the friction occurred.

## Expected Behavior
The observable behavior expected from the agent, workflow, tool, or environment.

## Observed Friction
What actually happened. Keep this factual and concise.

## First Visible Divergence
The earliest observable point where execution began differing from the expected behavior, or `unknown`.

## Impact
How the friction affected outcome, progress, quality, safety, cost, or human effort.

## Extra Work or Recovery
Visible retries, repeated reads or commands, rework, recovery, workaround, or coordination cost. Use `unknown` when it cannot be measured.

## Human Intervention
What the human had to clarify, correct, approve, repair, or redirect, or `none observed`.

## Workaround or Current State
The workaround used and whether the task recovered, remains blocked, or completed with friction.

## Evidence References
- `path-command-session-or-tool-reference`, or `unavailable`

## Agent Attribution Hint (Unverified)
Optional suspected source. Keep it separate from facts and do not present it as root cause.

Categories

Use a short category such as:

• misunderstanding
• missing-clarification
• unnecessary-question
• premature-action
• assumption
• decision-error
• repeated-read
• repeated-command
• retry-loop
• rework
• human-correction
• instruction-conflict
• unused-artifact
• tool-failure
• runtime-gap
• environment
• blocker
• safety
• other

Evidence Rule

The artifact is discovery evidence with status agent-reported-observation. It does not prove a workflow failure or root cause. workflow-evaluation must corroborate it with session traces, tool or artifact evidence, human correction, repeated observations, or controlled exercise before upgrading its evidence status.

Observation counts do not establish a failure rate without a known session denominator.

Done When

• one observation exists in docs/ai/evaluation/observations/
• facts, impact, and unverified attribution are separated
• first divergence and extra work are recorded or marked unknown
• missing evidence is marked unavailable
• the active task was not modified by this skill