← All entries
phase-update

cc-praxis — The Gap We Found While Testing What We Built

Date: 2026-04-07 Type: phase-update

What I wanted: a real test of the Work Tracking feature I’d shipped

The handover from the last session said to test Work Tracking in a real commit on a project without it configured. Pick remotecc or starcraft, make a commit, see if the Step 0b prompt fires. I’d also decided to bootstrap issue tracking for cc-praxis itself first — it had Work Tracking enabled but zero issues. Both felt like quick tasks.

They weren’t.

What I believed going in: Step 0b was in all commit skills

I’d built the Work Tracking prompt — Step 0b — into git-commit. When it detects no Work Tracking section in CLAUDE.md, it offers to set up issue tracking. I thought that covered everything.

The gap Claude found while we were testing

The first thing we did was run Phase 0 and Phase 1 of issue-workflow for cc-praxis — labels created, CLAUDE.md updated, epic #30 (Cut v1.0.1 release) created with five children. Then we moved to the real test: commit on remotecc or starcraft and watch Step 0b fire.

Both repos are Java projects. That means java-git-commit, not git-commit.

Claude checked java-git-commit/SKILL.md for Step 0b. It wasn’t there. The prompt didn’t exist in the specialised commit skills — only in the generic root. Any Java, blog, custom, skills, or generic project would silently never see the Work Tracking offer.

The fix was one block added to three files: java-git-commit, blog-git-commit, and custom-git-commit. We updated the flowcharts too. One commit, all three covered.

Then the test. Neither remotecc nor starcraft had a GitHub remote, so we created both repos and added the origins first. The full flow ran on each: java-git-commit fires, Step 0b detects no Work Tracking, user says YES, Phase 0 completes, labels created, CLAUDE.md updated. It worked.

Garden merge, two one-line fixes, and a surprise: 25 submissions

Issue #32 was merge the pending garden submissions — I’d noted five in the handover. When Claude listed the directory, there were 25. Sessions across four projects had been submitting without anyone merging.

We dispatched a subagent with full context budget — the garden skill’s own recommendation for merges. It integrated 31 entries, resolved a three-way GE-0053 ID conflict where three simultaneous sessions had claimed the same ID, assigned IDs to four files that had none, and committed cleanly. Nine new garden files, nine updated.

The two smaller issues were genuinely small. The ADR skill’s template left YYYY-MM-DD as a literal placeholder with no instruction to fill it in — Claude was leaving it blank. One sentence fixed it. The retro-issues skill was silent about the proposal file after issues were created. We added the answer: commit it as a permanent audit trail, never delete it.

Twenty false gaps and nine real ones

Before tagging v1.0.1, I wanted to check that all post-v1.0.0 commits had issue coverage. We ran retro-issues on the full 184-commit range. The skill produced a proposal with 26 standalone issues.

Twenty were duplicates of the previous retro-issues run from 2026-04-06. The subagent analysing git history doesn’t know which GitHub issues already exist — it just sees commits and groups them. The nine genuine gaps were commits like the Graphviz-to-Mermaid conversion, the documentation audit, author initials in blog filenames, and the mandatory bug fix workflow. We created and immediately closed all nine as retrospective issues.

gh release create v1.0.1 --generate-notes pulled from 45 closed issues. The notes look right.

Where the garden stands

The drift counter hit 31 — above the 10-entry threshold that triggers a DEDUPE recommendation. The next session that needs to merge garden submissions should run a sweep first. The knowledge is there; it just needs deduplication.

Epic #30 is closed. Issue #36 is the one open thread: designing design-snapshot as an indexed folder with an INDEX.md, and a cross-tool meta-index so session-handoff can point to garden entries, blog posts, and snapshots without loading any of them. The methodology family is mature enough now that the retrieval problem is starting to bite.