Friction Triage

/friction-triage — Turn friction signals into routed, boarded GitHub issues

Codifies the manual sweep: read signals/friction → group recurrences → classify actionable-vs-upstream → write issue bodies → file in the owning repo → add to the Grimnir Roadmap board → log the sweep to Munin.

This is the workflow layer. The data layer lives in hugin (src/friction-mcp.ts collector, scripts/friction-report.mjs aggregator). Read friction via the munin-memory MCP in-session (keyless); only shell out to the hugin script when you want raw aggregate counts.

Model

Triage is judgment-heavy (grouping recurrences, actionable-vs-upstream calls, writing issue bodies). Prefer Opus. Do not downshift to a small model for the grouping/classification step.

Step 1 — Read the corpus

1. memory_list namespace signals/friction to enumerate events (previews + tags).
2. Determine the sweep window: read meta/friction-triage (status key) for the last sweep date. Only triage events after that date — older ones were already filed. If no prior sweep, take everything.
3. memory_read_batch the full content of the in-window, open events (need the summary/detail/tool_name fields, not just tags).

Step 2 — Filter out the non-actionable

Skip an event if any of:

• Tag status:completed, fixed-by:*, or the content has a **Resolution:** note → already handled.
• Key starts with smoke-test / tailscale-smoke → test artifacts, not real friction.
• Already carries an issue:* tag from a prior sweep (see Step 8) → already filed.

Step 3 — Group recurrences

Cluster the surviving events by tool / subsystem (tool_name + user_tags). Multiple events with the same root cause become one consolidated issue (e.g. 6 tool:codex failures → one "Codex review reliability" issue listing each failure mode). Note distinct failure modes within a cluster as sub-points.

Step 4 — Classify: actionable vs upstream

For each cluster decide if it's Magnus-fixable:

• File it if the fix lives in a repo Magnus owns.
• Do NOT file — flag only if it's a Claude Code harness bug (e.g. StructuredOutput param-drop, core tool transport), an external MCP/provider bug (e.g. playwright MCP internals), or already covered by a CLAUDE.md rule (e.g. concurrent git-tree → worktree rule). List these in the summary as "not filed — upstream/covered" with a one-line reason.

Step 5 — Route to the owning repo

Map each cluster to where the fix lives (not where the friction surfaced):

| Cluster signal | Owning repo | |---|---| | codex / review-pr-codex, submit-task signing, friction-mcp, hugin orchestration, delegate/broker | hugin | | memoryupdatestatus, munin core, consolidation, openrouter key, memory_* tools | munin-memory | | ratatoskr / Telegram notify | ratatoskr | | m5 / gateway / llama-swap / serving / inference eval | gille-inference | | drone-lab specifics | drone-lab |

All repos are under Magnus-Gille/. If a cluster doesn't match, ask the user which repo owns it rather than guessing. Verify the slug exists: gh repo view Magnus-Gille/ --json name -q .name.

Step 6 — Write issue bodies to files

Always write the body to a file and use --body-file — never inline. Issue bodies contain backticks, `` markup, and shell metacharacters that break inline command substitution (a real past failure). Write each to the scratchpad dir.

Each body includes: Problem · Evidence (dated event summaries from the detail fields) · Proposed fix(es) · Source (the signals/friction keys). Keep it self-contained so the issue stands without Munin access.

Step 7 — File issues + add to the board

For each cluster, in a single bounded loop (capture URLs, don't dump full output):

url=$(gh issue create --repo "Magnus-Gille/" --title "" --body-file "")
gh project item-add 1 --owner Magnus-Gille --url "$url"

The Grimnir Roadmap is project 1 (--owner Magnus-Gille); new items auto-land in Todo. Omit --label unless you've confirmed the label exists in that repo (missing labels fail the create).

Then verify all landed on the board (item-add prints nothing useful on success):

gh project item-list 1 --owner Magnus-Gille --format json -L 300 | \
  python3 -c 'import sys,json; d=json.load(sys.stdin); [print((c:=it["content"]).get("repository","").split("/")[-1], c.get("number"), it.get("status")) for it in d["items"]]'

Step 8 — Log the sweep to Munin

1. memory_log to meta/friction-triage (tags ["milestone","decision","friction"]): the issues filed (repo#num + one-line each), the not-filed/upstream items, and any cross-cutting finding.
2. memory_write meta/friction-triage/status (tag active): compact index table of open issues with priorities, the sweep date (so the next run's Step 1 window is correct), and next steps.
3. Optional, closes the loop: tag each filed source event so the next sweep skips it — memory_write the event back with an added issue:# tag (use the patch.tags_add form). Skip if it adds too many round-trips; the sweep-date window already prevents double-filing.

Step 9 — Report

Give the user: a table of issues filed (repo#num, title, severity, event count), what was deliberately not filed and why, and the highest-leverage issue to start on. Offer to begin on the top one.

Quick mode

/friction-triage report — Steps 1–4 only: read, filter, group, classify, and present the proposed issue table without filing anything. Use to preview a sweep before committing to tickets.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Magnus-Gille
• Source: Magnus-Gille/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.