Keep a series of owner-reviewed changes in iteration

[mzored]

Result

A series of small changes the owner reviews stays in iteration until they ask for delivery, whether or not they name a workflow. Before this, only a session that invoked skiphow-fast-fixes got that behavior.

Scope

• plugins/skiphow/skills/skiphow/SKILL.md: the iteration paragraph. The deferred-check regime no longer depends on "an explicitly bounded iteration session", a state nothing outside an invoked workflow could enter. It is keyed on what the owner says, holds back push, shared integration and delivery tests exactly as the fast-fixes rule does, and keeps one clause of deference so an invoked workflow's own agreement still outlasts a bare acceptance.
• evals/cases.json: one microcase, iteration-series-announced-once, for a series announced once and then continued with bare edits.
• docs/decisions.md (amends the 4.4.0 iteration-shape record), docs/evidence.md, docs/guide.md, docs/outcome-contract.md, site/evidence/index.html.
• VERSION, both plugin manifests, the three eval instrument files, CHANGELOG.md 4.7.3.

No new file, no new step, no mandatory gate or persistent state. skiphow-fast-fixes is unchanged.

Observed defect

Two installed 4.7.1 sessions in the maintainer's own projects on 2026-09-09. In the session where the owner announced a screen review in ordinary language inside a running longrun programme, a 105-minute stretch of screen edits produced twelve guarded-integration invocations, five merges into the shared development branch, and five whole-backend-suite runs of two and a half to four minutes, plus repairs to tests the next edit invalidated. In the session where the owner typed the workflow command, the deferral held and the one wide run followed an explicit request to integrate. Two projects and two tasks, so this is an observation, not a controlled pair. Session content is not retained in this repository.

Tests and package evidence

python scripts/check.py passes, including the corpus validator on the new case. git diff --check is clean. Versions are aligned across VERSION, both manifests, evals/cases.json, evals/cto-cases.json, evals/host-smoke.json and the site.

Blockers and unverified limits

• Cross-host review is blocked, not passed. codex exec returned a usage-limit error with a reset date of 2026-09-15, so the Codex round for this change did not run. An independent review on the same host stood in. It raised four qualifying findings, all confirmed against the files and fixed here: the kernel had stopped deferring to an invoked workflow and so contradicted skiphow-fast-fixes on an accepting turn; the canonical contract and the kernel disagreed on when a series ends; the decision record claimed an isolated variable two field sessions cannot establish; and the raised rule held back less than its source rule, leaving push unheld outside the workflow.
• The corrected behavior is UNVERIFIED on both hosts. No paid behavioral session was run; iteration-series-announced-once is the unrun scenario that would show it.
• The failing session had also been given the same rule by its own repository's instruction file and broke that too, after four compactions had removed both from context. No package text survives a compaction, so this change removes a readable defect in the text and removes nothing about context loss.

🤖 Generated with Claude Code