release-planning - SKILL.md Agent Skill

name: release-planning description: This skill should be used to plan releases in rivet — assign which requirements/artifacts belong to which release and drive development from that plan — and to run the issue-driven delivery loop: an error, regression, or optimization comes in, gets evaluated, flows through the full verification chain, and ships in a planned release. Use it when scoping a release ("what goes in v0.X"), triaging an incoming issue/bug/optimization toward a release, asking "is v0.X ready to cut", or driving a roadmap. Builds on rivet's existing `release:` field + status lifecycle (draft→proposed→approved→implemented→verified→accepted); composes with the feature loop (build each item), traceability-audit (a release is ready when its items' V is closed), and release-execution (cut it). metadata: author: pulseengine.eu version: "0.1.0"

Release planning & the delivery loop

Two halves of one working model: plan which requirements land in which release (in rivet, so the V-model artifacts drive the roadmap — not a separate spreadsheet or only GitHub milestones), and deliver against that plan via an issue-driven loop (error/optimization in → evaluate → full chain → ship). This matches how the toolchain repos already work — synth plans by phase milestones and ships eval-first fixes ("mla fusion regresses flat_flight 255→257 on Cortex-M4 — gate it?"); loom plans by version milestones (v0.4/v0.5/v1.0) — but makes the plan live in rivet and the loop explicit.

Part A — Release planning in rivet

rivet artifacts already carry a release: field and a status lifecycle (draft → proposed → approved → implemented → verified → accepted). Use them as the plan, not GitHub milestones alone:

Scope a release = the set of requirements/decisions tagged release: vX.Y. Assign each requirement to the release it's targeted for; an unassigned requirement is backlog, not a commitment.
Readiness is a query, not an opinion — for release: vX.Y, how many of its artifacts are verified/accepted vs still approved/implemented? rivet (filter by release + status) + rivet coverage give the live burn-down. The release is cuttable when every artifact in its scope is verified and the V is closed (see [traceability-audit]).
Drive from it — "what must we implement for v0.X" is exactly the release: v0.X + status≠verified set. Pull work from there.
Mirror to GitHub milestones for visibility (synth phases, loom versions), but rivet's release: field is the source of truth — the milestone is a view.
Move scope deliberately — deferring a requirement to a later release is a scope decision (bump its release:), surfaced and logged, not silent.

Part B — The issue-driven delivery loop

An error, regression, or optimization arrives (a GitHub issue, a failing run, a benchmark delta). Run it as a loop, not an ad-hoc fix:

Evaluate — measure, don't guess. Reproduce and quantify. The synth pattern is the bar: a regression is a number (255→257 cycles), an optimization is a measured delta, a bug is a failing run. Root-cause it.
Classify against the plan. Is it a new requirement (capability gap), a regression against an existing verified requirement (the trace says what broke), or an optimization (a decision with a measured benefit + a constraint it must not violate)? Each maps to a rivet artifact.
Land it in rivet, assigned to a release. Create/extend the requirement/decision/test, set release: (this release if urgent, a planned one if not) and status. Link it into the trace (what it satisfies/affects) so a regression re-opens the right requirement's verification.
Run the full chain. [pulseengine-feature-loop] + [oracle-gate-a-change] — fix/implement, re-establish the oracle (tests, witness MC/DC, proofs via [proof-synthesis]); a perf optimization must keep the correctness oracle green while improving the measured number.
Close the V. [traceability-audit] — the artifact is verified only with passing tests at the right levels; the regression's original requirement is re-verified.
Ship in its release. When the release's scope is all verified, [release-execution] cuts it (V-model gate + falsification statement). Given the frequent cadence here, a single high-value fix can be its own release.

Worked shapes (from the repos)

Optimization (synth/loom): measured regression/gain → a decision artifact with the benchmark + the invariant it must preserve → implement → witness/proof oracle stays green AND the number improves → verified → release.
Bug (kiln): failing run → the requirement it violates is found via the trace → fix + a regression test that verifies it → verified → patch release.
Capability gap (loom "consume scry invariants"): new requirement → release: vX.Y → feature loop → verified → planned release.

Where this composes

[bootstrap-verification] — a new piece gets its first release plan here.
[pulseengine-feature-loop] / [oracle-gate-a-change] — execute each item.
[proof-synthesis] — discharge proof obligations an item carries.
[traceability-audit] — release readiness = the scope's V closed at all levels.
[release-execution] — cut the release once readiness holds.
[report-tool-friction] — if rivet can't express a planning query you need.

Anti-patterns

Planning only in GitHub milestones, divorced from the rivet trace — the roadmap then can't be driven by, or gated on, verification state.
Fixing without a measurement — an "optimization" with no before/after number, or a "bug fix" with no failing-then-passing test, isn't a closed loop.
Optimizing the number while letting the correctness oracle slip — perf work must keep proofs/tests/MC-DC green; a faster wrong answer is a regression.
Cutting a release before its scope is verified — readiness is the query, not the calendar; move scope out explicitly rather than shipping it unverified.
Untriaged issues — an issue with no rivet artifact and no release is invisible to the plan and resurfaces as a release-time surprise.