Debugger

One TSV beside the Pi session file, <session>.problems.tsv, with columns ts, kind, problem, fix, status, target, trace. Never in a repo. Never committed.

Two writers.

• The extension (pi/extensions/debugger) appends a row for every failed bash, read, or subagent call, skipping lookups that fail by design (rg, grep, fd, find, ls, git diff) and repeating an identical failure once per session. No model tokens.
• The model appends the rows only it can see with one log_problem call: kind is wrong-assumption, correction, wrong-path, leftover, or drift; problem is one line; fix is one line or omitted while open; target is the ppstack skill, playbook, script, or Pi config file that should change so it does not recur. One call, keep working. Do not discuss the row unless it blocks the work.

Skip routine progress and guesses that were right. Do not grep the tree for TODOs to fill the log. Only the parent logs; a child names its snag in its report.

/debugger prints the log with no model turn. scripts/log.sh <file> <kind> <problem> <fix> <status> <target> [trace] writes a row by hand with the same columns.

Nothing here is promoted anywhere. The human reads the log and decides what changes. reflect is separate and mines the transcript on its own.