cogikyo/dotfiles/papercuts

Papercuts

Apply mechanize before proposing a fix for a recurring Hole in Report.

Authority

Collab owns this procedure. Do not dispatch a child unless the evidence window is too large to classify here. Do not edit permissions, skills, or agents unless the user asks after the report.

Scope

Default to today and the current workspace. Treat $ARGUMENTS as the window: a session id, directory, agent, or time bound.

Start with the sessions tool and any log evidence supplied by the user. Select sessions by directory and time, then inspect tools with tool: "shell" for failed calls. Use part with the returned <message ID>#<index> address to read the command, error, and output. Include subagent failures when the question concerns agent use or dispatch. Compare later calls from the same command family to see whether a legal form succeeded. Tool history does not prove every permission ask; report missing permission evidence instead of inferring it from a failed command.

Ignore read-only git that already has an allowed twin unless the agent never switched to it.

Classify

Cluster by command family, not by every line.

For each cluster, pick one verdict:

ask is only a candidate when Collab, attended, sometimes needs the command and discard or history rewrite is not the default.

git add without -- is Recovered when git add -- follows. Worktree git checkout -- and git restore -- are Proper deny unless the user later needed that discard. git -C on an already-allowed read is an Agent miss when the agent never switches to workdir=.

Report

Return a short diagnosis, not a log dump.

• Name the sessions and agents that produced the clusters.
• Give one example command per cluster, a count, and the verdict.
• For Agent miss, name the legal form.
• For Hole, name the missing allow pattern.
• Propose a fix only when a Hole is real. Do not apply it here.

Omit empty verdict sections. If nothing failed in the window, say that and stop.

Epistemology handoff

For each reported command family, include the relevant session IDs, time range, diagnosis, and whether the evidence suggests a skill or instruction-content issue. Make that judgment from the command classification and do not widen into general transcript mining. When content improvement is indicated, append a paste-ready scope in this form:

/epistemology sessions <ids>; time <range>; focus on <command family>; papercuts diagnosis: <diagnosis>

Epistemology owns any later transcript analysis and content proposal.