E01 — Retrospective

Dates: 2026-10-12 to 2026-10-13. Story: RET-001. Participants: Product Owner, Claude Code.

E01 delivered shapes drawn from numbers in eight features, in six days. The retrospective changed how the project works more than any feature did: no more waiting for servers, measures of every command, a local library of sources, expert reviewing agents, a cleaner repository and a playground that is easier to use.

1. Facts

Measure	E01
Features delivered	8 (F01, F02, F07, F03, F09, F08, F04, F06); F10 moved to E14, F05 to E15
Stories	70, among them 9 spikes, 9 checkpoints, 9 audits, 9 validations, a review and this retrospective
Commits	544 before the history restart
ADRs	39 (two written at this retrospective)
Library	121 modules — model 50, geometry 41, math 18, io 7, render 5; 88 exported functions
Tests	744, in 13.6 s; 100 % coverage of src/; check:all in about 50 s locally
CI	about 2 min per run; at least 40 min spent waiting for it in explicit loops
Active time, by feature	F01 4 h 28, F08 3 h 29, F04 3 h 25, F06 2 h 24, F03 2 h 22, F02 2 h 20, F07 1 h 59, F09 1 h 17 (npm run metrics)
Most frequent commands	cat, python3, grep, sed (reading and editing files); git push 46 s median, the pre-push check

The durations before the metrics hooks are estimates from the transcripts: they include the time spent writing each command.

2. Logged items, decided

The improvement log held 88 rows and the feedback log 17; the retrospective opened with one table of them all, decided row by row by the Product Owner.

Theme	Item	Decision	Where it went
Branches and CI	a develop branch; no CI on work branches; checks run twice; releases for every doc push	done	ADR-0038
Branches and CI	faster CI	done through ADR-0038: one run per release; jobs merged	ci.yml
Dependencies	read release notes before upgrading; a cooldown on fresh versions	done: 7 days	docs/tooling/versions-and-security.md, .ncurc.cjs
Dependencies	Dependabot version updates	dropped; security updates kept	ADR-0038
Metrics	log every command with its time; time and tokens per story and task; reports	done	docs/tooling/metrics.md
Metrics	an exact token counter; a session timer; the call running	done: the status line	docs/tooling/metrics.md
Metrics	a report of time and tokens at the end of each turn	done	npm run metrics:session
Metrics	a watchdog relaunching a stalled run	later: the status line shows a silent session; no reliable relaunch exists	improvement log
Sources	a local library of the pages read, searched before going online	done: SQLite full-text search behind an MCP server	ADR-0039
Agents	expert agents with their own documents, in the background	done: math, verification, functional analysis, documentation	docs/tooling/agent.md
Agents	a domain-modelling expert	part of the functional-analyst	.claude/agents/functional-analyst.md
Agents	a setup portable to other tools	done: AGENTS.md; documentation bundled offline comes with the library	AGENTS.md
Agents	several passes on the configuration	done in this retrospective; again every four to five epics	docs/process/agile.md
Agents	a quiet terminal during runs	done: commits and status line only	/run
Agents	the bash guard denying quoted commands	fixed: here-document bodies are skipped	.claude/hooks/guard-bash.ts
Agents	a Stop hook to keep runs going	not now: runs did not stop early in E01	—
Repository	the Product Owner's words removed from every page	done	docs/conventions/writing.md
Repository	the old history replaced by one commit	done, with a summary kept	docs/process/history.md
Repository	stricter commit messages	not needed: no subject starts with a capital; ids inside are allowed	—
Repository	package.json and configuration files tidied	done; overrides checked, still needed	docs/tooling/dependencies.md
Documentation	modern writing rules; diagrams as code; every page reworked	done: scanning rules, Mermaid, every folder reread	docs/conventions/writing.md
Process	the agile method on one clear page	done	docs/process/agile.md
Process	a start-of-project interview, the agent as functional analyst	done	docs/playbook/analysis.md
Process	a complete experience report and a starter kit	done	this record, docs/playbook/setup.md
Process	ask early which things are kinds of one thing; read rules in their natural sense	done	docs/playbook/analysis.md
Process	a rules review every four to five epics and at the end, harden by default	adopted	docs/process/agile.md
Process	the epics' order on the backlog's front page	done	docs/backlog/README.md
Process	parallel threads; several projects at once	later, after several epics	improvement log
Process	thinking sessions on what stays public with a paid edition	later, before the pro edition	improvement log
Playground	a simpler interface: a floating toolbar, grouped panels, a pipeline column, a reset	done	playground/index.html
Playground	a short guided test per feature, the full recheck at the epic review	done: ?guide=epic	/review, /run
Playground	the playground asking its own questions, answered by voice	later: E12 or the editor applications	improvement log
Playground	step 72: the shape seemed to leave the canvas	fixed: a margin of 40	playground/show-shape.ts
Playground	"press Enter" in the steps while the drawing follows the typing	kept: a point's X and Y apply on Enter, so that a minus sign moves nothing yet	docs/guide/e01-draw-shapes.md
Playground	fields red after choosing a rectangle or a polygon from a free shape	fixed	playground/write-whole-shape-fields.ts
Tools	knip	done	knip.json
Tools	a property test for every math and geometry function	done as a fitness function; nine functions listed until EN-020	tests/fitness.test.ts
Tools	a coverage floor for the playground	done	vite.config.ts
Tools	StrykerJS	later: stryker-js issue 6210 still open on 2026-10-12	docs/tooling/dependencies.md
Tools	a dev container	not now: one machine, Node.js pinned	—
Tools	published previews of feature branches	not now: local previews suffice	—
Tools	generated pages: board, traceability, commits per story; tests in a real browser	later: E12	E12
Code	shared test helpers, one path writer, one pass per corner, folders of model/	planned: F11, first feature of E14	docs/backlog/features/E14-F11-code-base-ready-for-curves.md
Code	separate tests from code, folders that make sense	researched by SP-010 before any move	F11
Code	a state module for the playground, radii per shape kind, boundingBox on large lists, concave free shapes, old files without an anchor	later, with the features that need them	improvement log
Code	references verified without a date	dated when their features come	docs/references.md
Product	ideas for later epics: fill the parent box, a shape menu, imports of large SVG, a design system, memory of removed radii	kept in the domain and the epics	docs/domain/, docs/backlog/epics/
Product	an agile workspace application for working with agents	a separate private project, after this one	—

The feedback log's 17 critiques became the habits sheet of Working with the agent.

3. What helped, what hurt

Side	Helped	Hurt
Product Owner	to be filled with the Product Owner (RET-001.T2)	to be filled with the Product Owner (RET-001.T2)
Agent	spikes before code: no story blocked on a missing source; the auditor found real bugs (signed zeros, a spike consumed by its fillet); the guided test in the playground; questionnaires with numeric examples	waiting for the CI; questions asked twice; reading the same pages again; a long guided test that grew with every feature; no measure of time

4. Evolvability review

Question	Answer
1. Size	model/ holds 50 modules, playground/ 80: folders planned in F11 (SP-010, EN-022). Tests 13.6 s, check:all about 50 s.
2. Tools	knip's trigger was reached: adopted. StrykerJS still blocked (issue 6210). Mermaid adopted for diagrams.
3. Fitness functions	a property test for every math and geometry function, now tested; the layer imports were already a lint rule.
4. Patterns	repeated test helpers and two path writers: F11 (EN-021).
5. Agent configuration	AGENTS.md and CLAUDE.md rewritten; five expert agents; skills updated; the guard's false positive fixed.
6. Process	cadences kept; the guided test shortened per feature; a rules review every four to five epics; Markdown backlog still fit.

5. Changes adopted

• ADR-0038: develop, local checks, one CI into main, merge commits for features and releases, a 7-day cooldown.
• ADR-0039: the local source library behind an MCP server.
• Metrics hooks, the status line, the session report, npm run metrics read at each retrospective.
• AGENTS.md; expert agents math-expert, verifier, functional-analyst, doc-editor.
• Writing rules for scanning readers and for the Product Owner's decisions; Mermaid diagrams.
• The playground's toolbar, panels, pipeline column, reset and short guided tests.
• knip, the property fitness function, the playground's coverage floor; package.json tidied; repository-wide tests in tests/.
• F11 in E14 for the code debt; the history restarted at 0.1.0.

6. Playbook updates

• Principles: measure, read once, never wait for a server, write decisions not words.
• Starter kit: fourteen steps, the rhythm, what to read first, the cases met in E01.
• Start-of-project interview: new.
• Working with the agent: the configuration table, the Product Owner's lessons habit by habit.
• Lessons: seven new rows.

7. Best practices read for agent-built projects

Source	What it says that we use
REF-CC-MEMORY	short instruction files; AGENTS.md shared with other tools through an import
REF-CC-HOOKS	hooks for what must always happen; duration_ms on every tool call
REF-CC-SUBAGENTS	reviewers in a fresh context, with their own tools and documents
REF-CC-STATUSLINE	the session's facts on a status line refreshed every second
REF-MCP-TOOLS	a small tool server over stdio
REF-NNG-SCANNING	readers scan: concise, scannable, objective text