mistakes

name	mistakes
description	Ledger of known mistakes on gowkhtmltopdf and the rules that prevent them — stdlib-only violations, semantic-breaking lint renames, parallel-agent build breakage, release/version errors, stale-binary verification, and fix-regression cycles. Use before writing engine code, planning a release, or launching parallel work.

Mistakes

Evidence from the session-history scan (170 sessions, Aug 3–15). Each mistake below carries the rule that prevents it.

1. Stdlib-only promise broken by one-off "exception"

The project premise (pure Go, no third-party libraries, per every planning session and README) was abandoned for the first hard problem: an SVG logo fix added tdewolff/canvas + go-text/typesetting (~30 indirect deps) as a go.mod "exception".

Rule: Any dependency addition is a project-policy change. It requires amending the canonical plan/README first and explicit user sign-off — never a silent "exception". If a feature fights the constraint, solve it in-engine first (the builtin rasterizer existed before being deleted).

2. Lint cleanup introduced 16+ semantic bugs

A golangci-lint "no behavior changes" pass swapped return orders, silently dropped display:flex/grid, swapped jpegDims W/H, and broke LengthToPt float precision — the repair wave then spawned 3 more subagents. Scripted renames added more damage: "DOUBLE RENAME BUG", globalCfglobalCfg, broken rune literals, struct-field keys renamed in composite literals, idx == idx shadowing.

Rule: Never auto-rename or bulk---fix code. gopls rename only for single-identifier, mechanically-safe renames, each followed by go build + the package tests. Prefer excluding a noisy linter in .golangci.yml over bulk edits. "No behavior changes" is verified by diff review, not assumed.

3. Parallel agents on coupled packages broke the build repeatedly

Concurrent sessions editing settings/pdf/outline APIs produced ColorMode undefined (×2), ScanFontDir undefined, forced git stash dances, sleep-poll loops, and a plan-file reference (vertical.go) that did not exist on the branch.

Rule: One agent owns one package; agents never edit packages a sibling imports. Wave-planning assigns ownership explicitly. No two agents run golangci-lint --fix on the same tree. If a sibling's change breaks the build, stop and coordinate — do not work around it.

4. Release errors from missing checks

• Wrong go.mod module path (module gowkhtmltopdf instead of the GitHub path) broke go install at release time and forced deleting/recreating the published v0.2.2 tag.
• Version bump committed without running tests (cli.Version still 0.2.1 while VERSION said 0.2.2).
• CI version-stamp check failed 3× (awk $$ escaping) and version was maintained in two places.

Rule: Use skills/release-note/SKILL.md and the release checklist: module path correct, go build ./... && make test green, all version files changed together (VERSION, cli.Version, CHANGELOG, docs, frontend), go install smoke-tested before tagging. Tags and releases are immutable once published.

5. Verification against stale artifacts

PDF/UA compliance validators ran against an old binary ("The CLI binary wasn't rebuilt — still serving the old code"), and stale golden PDFs/benchmark corpora were used as ground truth.

Rule: Before any verification: rebuild (make build/make samples), confirm binary timestamp, and confirm golden outputs were regenerated from the current generator. Never trust committed PDFs as proof of current engine behavior.

6. Fixes regressed each other with no visual gate

Spacing fixes caused full-document text overlap ("throughout the whole freaking application there is too much overlaps text now"); underline fixes broke page 14; border fixes broke page 3. Regressions were found only by the user re-inspecting.

Rule: Every visual fix is verified with a screenshot-diff/regression check (full-document, not just the fixed fixture) before commit. If a fix touches layout metrics, re-render the whole fixture corpus and eyeball diffs. No fix ships "locally green" without the corpus pass.

7. Engine approximations presented as real CSS

The "subgrid lite" copy-inherit behavior was engine-invented, and the agent admitted "it is not supposed to match a browser" — only after the user asked whether it was real CSS.

Rule: Never claim a feature matches a CSS spec the engine does not implement. Docs and answers distinguish "engine behavior" from "CSS-standard behavior". When the engine diverges from browsers, say so explicitly and mark it in the fidelity docs.

8. Engine-vs-template fix policy was absent

Fixture-31 went CSS fix → engine fix → engine revert → template CSS, with the skill updated only after the churn.

Rule: For look-only layout issues, prefer template/author CSS first. Engine fixes are for structural bugs (pagination, borders, overflow), and each engine fix states its blast radius. Decide before editing, and record the decision in the checklist.

9. Planning/scope mistakes

• Committed another application's plan files into this repo, then deleted and rewrote them.
• Ticket #35 was created, fully rewritten (subprocess→cgo), then updated a 3rd time — intent needed 2–3 explanations.
• Scope creep: a Makefile/generate.go session edited unrelated convert.go.

Rule: Verify plan files belong to this repo before committing. Ask one clarifying question when a ticket's approach contradicts the user's stated intent. Touching files outside the task's ownership requires explicit justification.

10. Session-hygiene mistakes

• 5 failed skill loads from wrong paths in one session.
• Subagent violated an explicit no-git instruction.
• Empty sessions created (0 messages) before subagent storms.

Rule: Skill paths are relative to the repo skills/ root — verify with a read before loading. The git ban is in every subagent prompt (see skills/avoidable-work/SKILL.md). Do not create a session without intent.