[AI:Improver] 2026W28

[dzykovic]

Weekly Optimization - 2026W28

What this PR contains

• Optimization report in .operator/data/retrospectives/2026W28.md
• Prompt tuning: modifications to .operator/context/*.md and .operator/analyst/*.md based on PR feedback
• Task optimization: priority changes or cancellations in .operator/data/tasks/

Review guide

1. Check the optimization report for patterns found
2. Review agent prompt changes - are the new rules correct and helpful?
3. Review task changes - are reprioritizations/cancellations reasonable?

Merge to apply prompt improvements and task updates.

---

Generated by Operator - Improver Agent

[Copilot]

Pull request overview

Weekly “Improver” optimization update that refreshes the repository’s Operator prompt/rule documentation and records the week’s retrospective, aiming to reduce drift between prompts, rules, and actual engine behavior.

Changes:

• Updated planner and improver quality guidance (frontmatter quoting + AOP fenced EMIT syntax).
• Refined engine layer-graph/context statements to reflect current dependencies.
• Added the 2026W28 optimization retrospective including a body-update lifecycle record.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
.operator/planner/quality-gates.md	Adds a rule about frontmatter quoting when embedding sample work-item frontmatter blocks.
.operator/improver/quality-log.md	Documents required fenced AOP === EMIT … === / === END EMIT === syntax for lifecycle records.
.operator/data/retrospectives/2026W28.md	New weekly retrospective capturing prompt tweaks, task clarifications, and observed patterns.
.operator/context/engine.md	Updates the engine layer graph dependency notes for infra/ and logging/.
.operator/analyst/consistency.md	Updates analyst consistency checks to reflect corrected package/layer boundary wording.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[dzykovic]

Applied review feedback.

[dzykovic]

[AI-REVIEWER] PASS. This weekly improver PR is net-positive and every assertion is fact-checked against source: the core/zod runtime correction (packages/core/package.json + schemas/**), the infra/ and logging/ layer-graph corrections (verified against actual imports), the AOP fenced-block-only requirement (agent-output-protocol.ts:41 START_MARKER), and the frontmatter-quoting rule (buildFrontmatter at work-items.ts:1108) all match the code exactly. The report's watching-notes are also accurate — bundled improver.md still carries the buggy EMIT: colon form and authoritative typescript.md:105-106 is still stale, both correctly tracked as separate findings rather than smuggled into this PR.

Changes touch only .operator/** data files (no TS/code, no gate, no primitives boundary), are single-revert reversible, English-only, and carry no customer names.

Non-blocking observation (not held): the report's "14 pending findings" tally reads slightly high against 13 finding files on disk, but it drives no reconciliation action and findings may partly live in KV, so it is within report-bookkeeping noise. Cleared for the owner's merge.

[dzykovic]

[AI-REVIEWER] PASS on the merits — but routed to ai:manual for an owner decision, not auto-merged.

This PR edits engine-consumed prompts, not just records, so I treated every changed line as a rule and fact-checked it against the tree. All of it is true:

• @operator/core genuinely carries runtime (packages/core/package.json declares zod; schemas/** construct schema values, plus error classes).
• infra/ imports @operator/core, a type-only Logger from logging/ (workspace-init.ts:7), Node built-ins, and better-sqlite3 only under infra/local/ (sqlite-guard.ts).
• logging/logger.ts imports a type-only OperationContext from core, plus pino and node:stream.
• The AOP parser's START_MARKER (agent-output-protocol.ts:42) requires the fenced === EMIT <type> === form, so the new rule about colon EMIT: lines being ignored is correct.
• buildFrontmatter (work-items.ts:1108) double-quotes exactly the fields the new frontmatter-quoting gate names.

It does not revert PR #7's correction. I checked specifically: #7 is still open, and this edits the same consistency.md line in the same corrective direction. No silent regression.

English-only (the only non-ASCII characters are layer-graph arrows ← ⇏ →, which the rule explicitly permits as typographic symbols), no customer or sandbox name, single-revert safe, no protected surface.

Why this needs you rather than the gate

The change is correct, and I would defend it — but three parts of it are judgements a person should make, not a machine:

1. It tightens a quality gate — a new frontmatter-quoting rule in .operator/planner/quality-gates.md. Every future planner run is measured against it.
2. It rewrites the analyst discovery prompt (.operator/analyst/consistency.md) layer-boundary wording, which steers what the analyst flags on every future cycle.
3. It records a body-update against task T20260704-7B62ACD9, whose PR (#7) is still open. Landing this while that PR is in flight is a sequencing call.

None of those is a defect. They are exactly the class of change where the honest answer to "should a machine land this unreviewed?" is no.

One note, recorded not held

.operator/data/retrospectives/2026W28.md (~line 85) says "14 pending findings"; the tree has 13 — 11 pending, plus F20260704-0001 and -0002 correctly in-progress with child tasks in flight. Line 96's "12 engine-code findings" is likewise imprecise. The record's conclusion (no parent/child lifecycle drift, no reconciliation needed) is nonetheless correct, so the wrong tally drives no wrong action. Worth fixing so the record reads true; not worth blocking on.

[dzykovic]

Superseded by #41 (2026W31), which rewrites the same two prompt files with a strictly more accurate layer graph (adds logging/ types-only and optional pino-pretty). The only unique payload left here is the 2026W28 report plus an EMIT body-update for T20260704-7B62ACD9 — a task that completed on 2026-07-05, so the update has no target. Branch is conflicting and five weeks stale.