Status skills on three axes + planners that close open decisions per batch (0.2.0)

[elmayii]

What

Two things the maintainer had been doing by hand for three months, moved into the package (Sprint 10, issue #8):

1. Planners close each batch's open decisions with the human before writing it. cdev-planner shows the batch grouping, then goes batch by batch — questions with suggested answers and one recommended → answers recorded → that batch written → next. No open decision, no question; the prompt is never handed back for confirmation. With no human present, a batch with open decisions is written BLOCKED, never READY. cdev-monorepo-planner applies the rule by path reference, per SYSTEM_BATCH.
2. cdev-status and cdev-monorepo-status — read-only readings of where a sprint stands on three axes (functional requirements, non-functional requirements, user stories), engineered at read time from the raw chain sprint → batch → handoff. The plan's format does not change: there is no requirement list to keep current.

Plus: field evidence and the RFC, live docs, version 0.2.0, and the sprint report (docs/develop/reports/s10.md — including what was deliberately not built and why).

Draft on purpose. The merge waits on two things:

• the maintainer's "accepted" summary on the RFC — [RFC] Planners close each batch's open decisions with the human before writing it #9
• S10-B06 closed. The field exercise was run on 2026-09-20 — by the agent, at the maintainer's request, on local clones of a real workspace (never the real trees) — and is recorded in AGENT_PROGRESS.md. What remains are the two judgments only the maintainer can make: is every requirement and story in the produced status one they recognise (B7), and is it the status they used to ask for by hand (B9)

Merge by merge commit, not squash: the branch carries one commit per batch by design.

Layer

Core (the planner rule — RFC: Discussion #9, field evidence: docs/develop/reports/s10-field-evidence.md) + skill (the two status skills: read-only, the loop is untouched, so no RFC; recorded in DECISIONS 2026-09-19) + docs.

Evidence

Everything below is recorded per check in docs/develop/AGENT_PROGRESS.md, newest-first.

• Planner rule — sandbox, fresh subagents, hand-conditioned fixture. Attended: fixture state captured at six stops — nothing written at the first, then sprint header, then one batch after each round of answers, across five batches; lettered options with one recommended; a permission question; a free-text answer taken; an overruled recommendation honoured; a fully specified batch written with no question. Unattended: the undecided batch BLOCKED with two named decisions and a suggested answer each. Baseline (main's planner, same scenario): one up-front round, then four batches at once and five assumptions its own report says "the human has NOT seen".

• cdev-status — sandbox, cold, fresh subagents: cold start (twice), since, not-derivable, unconditioned repo; every fixture left with a clean tree. The first round found three defects in the skill's own text (the repo's working rules counted as non-functional requirements; two figures per axis; length) — fixed and re-exercised. The not-derivable clause took four attempts; three are on record as the fixtures' fault, one of them discounted because the fixture's commit message leaked the answer.

• cdev-monorepo-planner pointer and cdev-monorepo-status: sandbox not-run, declared (scripts/sandbox.ps1 builds single repositories only) — field-exercised instead, headless with --plugin-dir, on clones of a real workspace. Planner, launched bare: pointer followed, grouping first, questions per SYSTEM_BATCH before each write (clone untouched after the first turn), system-level and permission decisions raised, an answer pushed back on where it collided with a repository's local protocol. Status: the first run fanned out to three subagents and died on the usage limit with nothing delivered; "read narrowly" brought it to ~4 USD and a delivered status that led with local truth — and over-corrected (it skipped contracts, decisions and local handoffs, and its own method note said so); now "slices, never fewer sources", regression-checked on a fixture, not yet re-run on a real workspace.

• claude plugin validate . passes

• Changed skills/profiles exercised in a sandbox fixture — single-repo skills yes; multi-repo skills not-run, covered by the pending field exercise (protocol §4, local line)

• CHANGELOG.md updated if behavior changed — ## [0.2.0]

Closes #8

🤖 Generated with Claude Code