bew/dotfiles/opencode-reflect-friction

Goal: surface all friction from this session — both the main conversation and any subagents that were launched — so every source of misalignment or repeated correction is captured in one report.

Step 1 — Discover subagents

Scan the conversation for task tool calls that produced a task_id. Collect: task_id, agent type, and brief description of what each subagent was launched to do.

If any task_ids are found:

• List them to the user: task_id, agent type, one-line purpose.
• Ask via question tool: "Which of these subagents should I collect friction from?" (allow "all", "none", or a subset)
• STOP. Wait for user's answer before doing anything else.

If none found, skip to Main conversation scan.

Step 2 — Collect subagent friction

For each selected subagent, resume it using the task tool with its task_id.

Prompt to send to the revived subagent:

You are being revived to report on your own session's friction.

Please load the opencode-reflect-friction skill and run it on your inner transcript.

If you cannot load a skill, here is the core friction prompt to run instead:

Review your conversation above. Identify moments where: the user pushed back or corrected you,

you had to redo something, you misunderstood an instruction, you missed a preference that should

have been known, or you hit a tool/environment failure that caused retries.

For each friction moment, note a brief description of what happened and which artefact or action

was responsible.

Output only a flat numbered list: N. [artefact-name or "general"] — what happened

If no friction found, output only: No friction detected.

Wait for each subagent's response before proceeding.

Step 3 — Main conversation scan

Review the main conversation above (excluding subagent tool result blocks already collected in Collect subagent friction). Identify moments where user pushed back, corrected agent, asked to redo something, or had to repeat a preference that should have been known.

If a focus hint was provided by the caller, weight analysis toward that area, artefact, or theme. Do not restrict scan to it.

Collect internally — do not output anything yet. If no friction found anywhere (main + all subagents), output only: No friction detected. and stop.

Proceed to Produce report.

Step 4 — Produce report

Friction moments

Short numbered list. One line each.

• Main conversation items: N. [artefact-name or "general"] — what happened
• Subagent items: N. [artefact-name or "general" | subagent-type via invoking-artefact] — what happened

Improvement ideas

For each friction moment, one brief idea of what to change in the relevant artefact (or AGENTS.md) to prevent it. Not fully concrete — enough to guide a focused rework session.

• Main conversation items: N. [artefact-name or "general"] — what to add/change/remove
• Subagent items: N. [artefact-name or "general" | subagent-type via invoking-artefact] — what to add/change/remove

Post-output notes

If the last 1–2 messages mention that the session was forked (e.g. "session forked"):

• Omit the fork note below.
• Instead, offer to directly proceed with applying the improvement ideas.

Otherwise, add at end:

To act on these, fork this session before proceeding with these updates.

If context has many old messages/tool outputs, also add:

Consider compressing context (saves tokens/money!)