lightweight retrospective audit for agentic coding sessions on problem/quality/research/security/task-organization
tegwick
776579c660
The fleet already has a session-closing ritual — the operator asks for a hall-of-helix entry at wind-down — but it is unspecified, and PQRST is not part of it. The one estimate in the hall was added by hand as `30/23/18/19/10`, a format the spec does not define. Two tasks added: T04 — Specify the closing routine in a new CLOSING.md, quoting the operator's wind-down sentence so an agent recognises the situation, and link it from README.md, ENTRY.md, and the AGENTS.md Close protocol. Takes a position on two things the routine is silent on: the estimate covers the substantive session and excludes the closing ritual itself (otherwise the ritual inflates T on every session that has a seat), and the prompt is reached by path with only the output block inlined. Constrains the change against both repos' INTENT.md: a record on a seat, never a score on a worker. T05 — Carry the record in entries as a quoted canonical signature in frontmatter plus a `## PQRST estimate` section holding the full stored record, with templates/entry.md, ENTRY.md, and check-entries.py updated to match. Required for agent-session seats after adoption; human seats exempt; the 102 existing seats grandfathered. Explicitly forbids backfilling estimates for sessions nobody observed — that is the fabricated evidence spec R4 forbids. Only the one manual estimate is normalised, and it already sums to 100. T01 gains a sixth pilot question — whether a session estimating itself inflates P — since the hall entry is written by the same session being audited. T02 now evaluates hall-of-helix as the default answer to the storage question rather than weighing it from scratch, with its sampling bias stated. T03 picks up the ritual-scope and self-estimation findings. Execution order T01 → T04 → T05 → T02 → T03 is stated in the workplan; task ids are identity, not sequence. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SQ6oF1DtVDKcD1FCpvRVLx Assistant: claude-code Assistant-Model: opus Assistant-Process: 272883@bnt-lap001 Assistant-Session: f40c8f53-fb65-4980-9d29-bcdb3dd946f7 |
||
|---|---|---|
| history | ||
| spec | ||
| workplans | ||
| .custodian-brief.md | ||
| .gitignore | ||
| .repo-classification.yaml | ||
| AGENTS.md | ||
| INTENT.md | ||
| PqrstPrompt.md | ||
| README.md | ||
| SCOPE.md | ||
| WORK-RECORDS.md | ||
pqrst-practice
A lightweight retrospective audit for closing agentic coding sessions: a forced 100% allocation of a session's effective engineering effort across five stable dimensions — Problem, Quality, Research, Security, Task organization.
Where did the effective engineering effort of this session actually go?
Start here
| I want to… | Read |
|---|---|
| Close a session right now | PqrstPrompt.md — paste it, done |
| Understand the model | spec/PqrstEstimationPractice.md (v0.1) |
| Know why it exists, and what it must not become | INTENT.md |
| Know what belongs in this repo | SCOPE.md |
| Work in this repo as an agent | AGENTS.md |
What a record looks like
PQRST-Estimate
P: 25%
Q: 15%
R: 45%
S: 0%
T: 15%
Sum: 100%
Confidence: medium
Signature: P25 Q15 R45 S0 T15
Dominant factors: Existing auth middleware was undocumented and contradicted the
README, so most attention went to reconstructing the request path before
changing behavior.
The numbers are the index. The Dominant factors sentence is the content — a
signature stored without it is not auditable.
What this is not
Not a productivity score, not telemetry, not a planning target. There is no good profile: an exploration session is R-heavy, a hardening session Q-heavy, a credential integration S-heavy. The practice fails the moment anyone tries to score well on it.
This repository defines the practice and supplies the prompt. Collectors,
validators, dashboards, and record storage belong elsewhere — see
SCOPE.md.