cc-praxis — Closing the Gaps
Date: 2026-04-14 Type: phase-update
The session started with a parked idea. Blog entries had needed a type taxonomy since April 13 — articles vs. diary notes — sitting in IDEAS.md. I brought Claude in and we designed it in under an hour, brainstormed the spec, then shipped via TDD subagents. Fourteen existing blog entries got backfilled with entry_type: note, subtype: diary, projects: [cc-praxis]. While we were in there, yaml_utils.py got replaced — the homemade YAML parser couldn’t handle list fields, so it was swapped out for PyYAML. Validator count went from 17 to 18.
Mid-session I noticed a gap in the workspace model. Artifact routing had essentially one destination with different syntax: the project repo. If a project wants zero methodology artifacts in its own repo, there’s nowhere for ADRs and DESIGN.md to accumulate between epics. The workspace is already a git repo with main — it just wasn’t wired as a routing target.
I spent time with Claude designing the fix: three routing layers (built-in default → ~/.claude/CLAUDE.md → workspace CLAUDE.md), three targets (project, workspace, alternative <path>). We specced it, implemented it in epic-close and workspace-init, then ran a systematic consistency review. Six issues found — including a SHA baseline ratchet bug where routing design → workspace would silently corrupt the next epic’s three-way merge if the baseline wasn’t updated.
The health sweeps ran twice. First pass: the garden skill documented in five README sections but not existing — we replaced it with forage months ago, never cleaned up the docs. Web app stats bar showed 44/17/295 instead of 48/19/446. The blog INDEX.md listed 1 of 14 posts. Second pass: CLAUDE.md’s quick reference had flowcharts in the commit tier when it runs in push. We fixed all of it, plus count drift across README, QUALITY.md, CLAUDE.md, and SKILL-MANAGER.md.
Test coverage was the day’s largest chunk. Before today: 446 tests, 3 of 22 validators covered. I brought Claude in to dispatch TDD subagent batches with one instruction: “read the source first, then write tests.” That instruction turned out to matter considerably.
Claude’s agents came back with findings from reading the source. validate_links.py uses requests.get, not requests.head — most link validators use HEAD to avoid downloading content, so the mock target assumption was wrong. Any test patching requests.head would pass silently while testing nothing. validate_cross_document.py writes to stderr, not stdout. validate_examples.py silently skips any JSON block matching {[^}]*} as a template marker — you need array-format JSON to reach the WARNING path. Three forage entries written on the spot.
One debugging session took longer than it should. We ran generate_web_app_data.py from a git worktree and it reported “CHAINING_TRUTH: updated (48 skills)”. The test still failed with empty builds_on lists. The script was finding the file, writing to it, reporting success — but every list was []. The problem: the session’s CWD was the worktree directory. Path.cwd() returned the worktree root. The worktree inherits CLAUDE.md from the main repo, so find_skills_root() stopped walking upward there, glob('*/SKILL.md') found nothing, and the script wrote empty data and logged a success. Nothing in the output signalled any of this.
Final: 1016 tests, 19 of 19 registered validators covered. Six forage entries submitted — three gotchas, one technique, two undocumented.