Slow loop: periodic process review (drains devwork learnings, defer-rot, process-quality trend, conformance census)

jbr870 commented

2026-08-11 14:55:44 +00:00

Owner

What this is

A periodic scan over the accumulated artifact trail across many feature runs. Not a pipeline phase —
it runs on a different clock, reads what the runs left behind, and proposes changes. It approves
nothing by itself.

Sibling of the backlog-grooming trigger; both are periodic cross-run readers of the same trail, and they
may well be one skill.

What it drains

1. devwork-scoped learnings. Decision 8 routed the four learning scopes: project → CLAUDE.md,
stack:* → project-local .claude/skills/stack-<name>/SKILL.md, personal → deleted as an axis, and
devwork → this review. The routing is implemented and working; devwork learnings are correctly
recorded unhomed awaiting a reader — and the reader is what does not exist, so that scope is
accumulating with nothing draining it.

Filing an issue per learning was explicitly rejected: a project run should not have (and may not even
have) write access to the suite repo, and it would bypass the recurrence threshold — one observation
is a hypothesis, N across runs earns a skill edit.

2. Defer-rot watch. Inflow rises with automation while the stock grows. See the grooming issue for
the current numbers.

3. Process-quality trend — three instruments, all falling out of the UAT return path's
classification, all countable per feature rather than judged from a diff:

Instrument	Fires when	What it indicts
UAT escape rate	UAT finds a defect the validation cases passed	The validation cases are weak
Requirements-change rate	The feature matched the PREQ and the human wanted something else anyway	The PREQ process is weak
UAT iteration count	A feature loops through UAT repeatedly	Something upstream was wrong in a way nobody caught. Trend it, do not block on it

The middle one is the most valuable signal in the system: nothing else in the pipeline can detect a
wrong requirement. Every other check measures against the PREQ, so a wrong PREQ is invisible to all of
them. That is why the re-entry classification is recorded per finding rather than just acted on.

4. Conformance census — sdlc-process-observability-analysis.md §7: collect → conformance →
analyst → aggregate → curator proposes skill changes, human approves.

Why it runs here

No cross-repo write is needed: the scan runs in devwork-skills, reading the project forges.

Source: sdlc-suite-simplification-decisions.md decisions 7/8 + "The slow loop — parked".

## What this is A periodic scan over the accumulated artifact trail across many feature runs. **Not a pipeline phase** — it runs on a different clock, reads what the runs left behind, and *proposes* changes. It approves nothing by itself. Sibling of the backlog-grooming trigger; both are periodic cross-run readers of the same trail, and they may well be one skill. ## What it drains **1. `devwork`-scoped learnings.** Decision 8 routed the four learning scopes: `project` → `CLAUDE.md`, `stack:*` → project-local `.claude/skills/stack-<name>/SKILL.md`, `personal` → deleted as an axis, and **`devwork` → this review**. The routing is implemented and working; `devwork` learnings are correctly recorded `unhomed` awaiting a reader — **and the reader is what does not exist**, so that scope is accumulating with nothing draining it. Filing an issue per learning was explicitly rejected: a project run should not have (and may not even have) write access to the suite repo, and it would bypass the **recurrence threshold** — one observation is a hypothesis, N across runs earns a skill edit. **2. Defer-rot watch.** Inflow rises with automation while the stock grows. See the grooming issue for the current numbers. **3. Process-quality trend** — three instruments, all falling out of the UAT return path's classification, all countable per feature rather than judged from a diff: | Instrument | Fires when | What it indicts | |---|---|---| | **UAT escape rate** | UAT finds a defect the validation cases passed | The validation cases are weak | | **Requirements-change rate** | The feature matched the PREQ and the human wanted something else anyway | The PREQ process is weak | | **UAT iteration count** | A feature loops through UAT repeatedly | Something upstream was wrong in a way nobody caught. Trend it, do not block on it | **The middle one is the most valuable signal in the system: nothing else in the pipeline can detect a wrong requirement.** Every other check measures against the PREQ, so a wrong PREQ is invisible to all of them. That is why the re-entry classification is recorded per finding rather than just acted on. **4. Conformance census** — `sdlc-process-observability-analysis.md` §7: collect → conformance → analyst → aggregate → curator proposes skill changes, human approves. ## Why it runs here No cross-repo write is needed: the scan runs **in devwork-skills**, reading the project forges. *Source: `sdlc-suite-simplification-decisions.md` decisions 7/8 + "The slow loop — parked".*

jbr870 referenced this issue

2026-08-11 14:57:37 +00:00

Slow loop: backlog grooming trigger (the fan-in scan nothing fires) #9

jbr870 commented

2026-08-11 14:57:37 +00:00

Author

Owner

Sibling of #9 — see the note there on whether these are one skill.

jbr870 commented

2026-08-13 11:20:03 +00:00

Author

Owner

Process review 2026-08-13 — first /meta-retrospective run

The periodic process review this issue specifies now exists as a repo-local meta skill (/meta-retrospective, with /meta-amend as the light change lane and scripts/lint-conventions.sh as the mechanical gate). First run completed today over all 12 dogfood run reports + the 2026-08-07 aggregate + tracker learnings:

Filed: #28 (phase-outcome-post slowness, 4 runs), #29 (zsh nomatch/quoting baseline gap, 3+1 runs), #30 (integrate 4.5b flagless helper call — mechanical, verified at HEAD), #31 (glab-cli quick-ref traps, 2 runs); evidence comment on #14 (cross-adapter /resolve bypass).
Watchlisted: 7 single-source candidates (sdlc-retrospectives/watchlist.md).
Verified resolved at HEAD: ~15 finding clusters, incl. the merge-path gate hole, wait discipline, lane bootability, seeded-DB capture, readiness staleness/prereqs — the redesigns since 2026-08-07 absorbed them.
Record: sdlc-retrospectives/2026-08-13.md. Not yet drained: verity's forge learnings (no local checkout on this box).

Filed issues carry no priority:* — awaiting operator triage.

## Process review 2026-08-13 — first /meta-retrospective run The periodic process review this issue specifies now exists as a repo-local meta skill (`/meta-retrospective`, with `/meta-amend` as the light change lane and `scripts/lint-conventions.sh` as the mechanical gate). First run completed today over all 12 dogfood run reports + the 2026-08-07 aggregate + tracker learnings: - **Filed:** #28 (phase-outcome-post slowness, 4 runs), #29 (zsh nomatch/quoting baseline gap, 3+1 runs), #30 (integrate 4.5b flagless helper call — mechanical, verified at HEAD), #31 (glab-cli quick-ref traps, 2 runs); evidence comment on #14 (cross-adapter /resolve bypass). - **Watchlisted:** 7 single-source candidates (`sdlc-retrospectives/watchlist.md`). - **Verified resolved at HEAD:** ~15 finding clusters, incl. the merge-path gate hole, wait discipline, lane bootability, seeded-DB capture, readiness staleness/prereqs — the redesigns since 2026-08-07 absorbed them. - **Record:** `sdlc-retrospectives/2026-08-13.md`. Not yet drained: verity's forge learnings (no local checkout on this box). Filed issues carry no `priority:*` — awaiting operator triage.

jbr870 referenced this issue from a commit

2026-08-13 12:06:37 +00:00

meta: add the meta lane — /meta-retrospective, /meta-amend, and the conventions lint gate

jbr870 referenced this issue

2026-08-15 13:01:46 +00:00

Ship the suite as a Claude Code marketplace plugin (repo move, plugin root, CLAUDE_PLUGIN_ROOT paths, namespacing) #50

Rows
Columns

Slow loop: periodic process review (drains devwork learnings, defer-rot, process-quality trend, conformance census) #10

What this is

What it drains

Why it runs here

Process review 2026-08-13 — first /meta-retrospective run