pl-analyze

PL · Analyze — incident analysis & process evaluation

Scripts named below run as bash "${PL_SKILL_DIR}/scripts/<name>.sh" ..., where PL_SKILL_DIR is the absolute path of the pl skill directory; set it in the same Bash call. The workspace layout (ai-doc/..., ledgers) is documented in the pl core's State files section; a source that does not exist in this project is skipped and noted in the analysis.

Sub-skill of the PL role, also usable standalone. Words such as journey, callsign, night shift and take-over are defined once, in the pl core's Terms glossary. The rules in this file are the PL role's defaults; a project may choose otherwise.

Invoke when something went wrong — an escaped bug, a stalled run, a run killed at its time limit, an estimate blowup, a protocol breach, a close-audit FAIL — or when the owner ("owner" means whoever owns the work, often the user) asks for a retrospective, a root-cause account of what went wrong, or for periodic process evaluation (patterns in estimate errors, recurring frictions). The deliverable is never just an explanation: it is a verified cause AND a fix landed in the right place. Core iron rules bind unchanged; the owner is addressed in the user's own language.

Method — evidence first, always from disk

Sessions lie and die; disk is the memory. Rebuild the timeline exclusively from:

• sno reach log --as <your-strict-address> --work <id> — the journey's full message provenance (sno reach export --work <id> --output <path> keeps the full thread bytes); command grammar is defined in ~/.local/lib/sno-reach/current/guide/agent-reach.md
• ai-doc/JOURNAL/routing-ledger.jsonl — routed/closed/correction events (board_closed lines are written by bash "${PL_SKILL_DIR}/scripts/todo.sh" close; the PL writes the journey's own routed, closed and correction lines by hand)
• ai-doc/ACTIVE/PL/exec-logs/ + dispatch files under ai-doc/ACTIVE/PL/dispatch/
• git log/diff of the journey's commits when the project uses git (the fence file lists the paths the journey may touch)
• the convergence series (the recorded remaining-work counts per fix cycle; bash "${PL_SKILL_DIR}/scripts/convergence-watch.sh" verdict --journey <id>)
• ~/.local/state/agent-spawns.jsonl + agent-callsigns.jsonl + tmux autopsy (an executor killed at its time limit leaves its session up — read the screen)

Reports, memories, and your own recollection are HYPOTHESES to check against this record, never evidence. Timestamp every step; a contradiction between two sources is a finding, not noise. N different-looking failures arriving together get ONE joined timeline BEFORE any is diagnosed separately — join the timestamps in the target machine's system log first; cascades collapse into one root event and red herrings die at the join — several apparently separate failures are often one shared cause. A status table built from disk beats a report built from a stale card — trust the freshest disk join, not the most confident narrator.

Root-cause classification → fix location

The class DECIDES where the fix lands — never write a note where a mechanism belongs:

Class	Signature	Fix lands in
Machinery bug	a script/skill deterministically did the wrong thing	the skill CODE (script fix + test), then reinstalled
Judgment lapse	the agent saw the signal and chose wrong / waited on a human	MECHANIZE it — an alarm/verdict/gate in scripts (e.g. a stalled-close alarm); a memory note is NOT a fix
Estimation error	estimate vs actual > 2×	the est/actual pair → calibration material for agentic-time-estimate (when installed; otherwise a plain note of the pair)
Knowledge trap	a known trap, hit again	the lessons file ai-doc/LESSONS.md (written only by the PL; a trap enters the index on its SECOND hit — see the pl core, Lessons file)
Process gap	no rule existed for the situation	retro proposal → owner approval → skill text patch (never self-edited mid-journey)

Design principle: the machine must lose patience before the owner does — every place the owner had to scold marks a missing mechanical control; an analysis is not done until that control is named (or shown to already exist).

Output shape

Owner-facing report, written in the user's own language, four sections: (1) conclusion — what broke, one line; (2) evidence chain — the 2–4 load-bearing facts, each with its path/command; (3) where the fix lands — the fix and its location per the table above; (4) recurrence prevention — the check/alarm/rule that now exists. Agent-facing: the same content as an English postmortem section in the journey's journal entry, plus ledger correction events wherever the durable record was wrong.

Routing the lessons

A finished analysis routes each lesson to the file that OWNS the behavior: dispatch-time rules → pl-dispatch; in-flight controls and alarms → pl-watch; close/probe rules → pl-audit; environment checks → pl-env; board/memory/Reach-substrate law → the pl core. A lesson without a named owner file is not landed.

Boundaries

• Script/machinery fixes land directly ONLY within the core iron-rule-1 exception: PL-owned control scripts under the pl skill's scripts/, with a test, then reinstalled and reported to the owner. Anything else that needs code — repo source, other skills' scripts — is dispatched to an executor, never hand-edited. SKILL text changes still ride the retro → owner-approval path (core boundary).
• Estimation analysis only FILES the calibration pair — rate/multiplier changes belong to the estimator skill's own validation loop (when installed).
• Analysis of a journey never reopens settled owner rulings; if the cause traces to a ruling, report the trace, don't relitigate.