Project Continuity

---
name: project-continuity
description: Maintain AI-KIT's own context, template boundaries, and installer verification when kit structure or procedures change.
---

# AI-KIT maintenance continuity

Read [contribution rules](../../../CONTRIBUTING.md). Read optional local `../../../PROJECT_CONTEXT.md`, `../../../AGENTS.md` and `../../../WIKI.md` when present; otherwise verify facts from public source/configuration. These three root files stay local and untracked, outside wheel/Docker payloads; their template counterparts stay public. Check distributions with and without local copies and preserve their on-disk bytes when untracking.
Keep the root context/history about this kit; keep template project facts uninitialized.
For a procedure change, update current maintainer commands, this pointer contract, and relevant implementation docs in the same task.
When verified facts or agreements change, synchronize this skill's scope/procedure/pointers and WIKI with the context; keep actual facts in context.
Validate templates/links/registry, the agent registry (integrations/agents.json), and behavioral installation scenarios; record root CHANGELOG entries. Keep the root SKILL.md (the repository as a skill) consistent with the CLI, presets, extras, and policies.
Review explicit semantic acceptance and future-upgrade conflicts whenever installer baseline behavior changes.
Check candidate draft preservation, state/settings validation, and adapter retirement conflicts when changing installation. Exercise installed Git visibility and portable linked-path refusal; distinguish synthetic fixtures from real application trials. Verify the conditional Docker trigger in installed AGENTS and preserve edited entry points during upgrades.
Exercise file-versus-directory native rule layouts on both probe and selected-write paths: simulate POSIX ENOTDIR on Windows, preserve existing file markers and rerun stability, and refuse blocked selected parents before writes. Do not hide permission failures or linked-path refusals during optional adapter probes; see [file-parent compatibility](../../../docs/ADAPTER_PATH_FIX.md).
Exercise manifest-based profile detection and project-fact drafts (version/commands into PROJECT_CONTEXT and project.json), scoped Node/Python ignore rules, the installed-project doctor report, opt-in extras (session-start, guards, CI) including key-level client settings merges, legacy whole-file migration, and hook execution, client detection markers, interactive/check onboarding, the adr/changelog helpers, and the offline router helper (route/configure with provider selection) when installer behavior changes.
For generated client configuration, exercise private/team settings files and mode switches, owned-entry moves, project.json precedence, secret lists, ignore blocks, every scoped rule format with root modules and skipped paths, presets that turn extras off, and workflow skill copies; recheck client formats against primary docs before changing them.
For routing, compare the implementation ladder with canonical POLICY, exercise English/Russian operation/risk regressions and override precedence, provider-review status/sources, and distinguish recommendations from executed model controls. For stack profiles, exercise manifest evidence, module scoping, and command drafting; for Filament specifically test Composer dependency/lock evidence and source/native skill lifecycle. For optional metrics, verify failure/recovery accounting, missing-value coverage, duplicate/lock/path guards, private/team journal visibility, schema 1/2 compatibility, experiment/arm/task identity, per-provider analyze output, and upgrade preservation. Exercise context snapshots and doctor warnings for new/removed modules, profile drift, and manifest changes. Exercise adapter reports plus explicit retirement of unchanged files, backups/state updates, and modified-file conflicts. Exercise opaque evaluator pair coverage without task-text persistence. Use the [evaluation protocol](../../../docs/EVALUATION.md) for separately authorized application trials; synthetic fixtures establish no savings.
Record actual OS/runtime and unavailable checks in the implementation record; keep source visibility checks separate from client activation.
For measurement-led changes, test schema 1/2/3 compatibility, fixed A/B controls, immutable Git blobs and isolated repeated workspaces, missing/estimated telemetry, semantic-review gates, interrupted runs and idempotent private exports. Exercise resume identity checks, per-unit failure recording and configurable kit-arm agents/presets. Check that worker Git discovery cannot reach the source through parents: a ceiling at cwd itself is ineffective. Keep worker requests separate from observed execution. Exercise routing modes/pins/profiles, distinct-case quality floors, matched history, unsupported controls, immutable proposal history and explicit backed-up restore. Exercise module context scope, duplicate content, critical rules/ADRs above byte budgets and linked paths. Follow [the staged plan](../../../docs/DEVELOPMENT_PLAN.md) and [worker contract](../../../docs/BENCHMARK.md); defer live adapters/orchestration until their mechanisms and total coordination costs are verified.
Resolve temporary fixture roots before deriving paths used by guards, assertions, or injected failures. Test equivalent path spellings; keep cleanup bound to original disposable directories. [Windows CI regression](../../../docs/WINDOWS_CI_FIX.md) records the relevant scenario.
For reference upgrades, check the actual published tree and old paths left by overlays. Compatibility pointers link to canonical owners; never hide broken legacy links from the checker.
When editing onboarding docs, verify examples against the actual CLI/defaults and preserve links to policy owners. Keep installation, native discovery, and measured execution claims distinct. For [OpenRouter documentation](../../../README.md#working-with-openrouter), recheck primary API/cache/usage docs; distinguish shipped context manifests from proposed request assembly/runtime, prompt caching from response reuse, and observed usage from assumed savings. Root README and Docker Hub Overview have no size quota; retain scoped instruction budgets and normal link/language checks.
For [PyPI releases](../../../docs/PYPI.md), verify metadata, VERSION, licenses, description, resources and private/path exclusions; build the wheel from sdist and smoke-install offline. Quote Twine globs; report OS gaps. A JSON 404 does not prove name eligibility. Test release/manual identity guards without uploads; preserve CI names and old tags. Published stable releases prepare PyPI artifacts; manual runs default to validation. Require exact version/tag/repository gates, owner approval and future tag coverage; agents must not approve deployments. OIDC belongs only in upload jobs consuming validated same-run artifacts. Keep setup, upload and live MCP activation separate.
For repository administration or changes to existing required CI names, consult [repository maintenance](../../../CONTRIBUTING.md#repository-maintenance), verify live protection and emitted names, and keep required checks aligned. New local jobs establish no hosted result or required-check update. Remote changes need explicit authorization; Core governs commits/pushes.
For every task assess [Docker image impact](../../../docs/DOCKER.md#change-impact-and-image-refresh), including packaged instructions and defaults. Follow canonical [image freshness](../../../template/ai-kit/CONTAINERS.md#image-freshness); rebuild/test affected final sources as ai-kit:dev or record why no rebuild is needed. For the image, verify actual client/daemon, pinned build inputs, VERSION/label identity, closed context, wheel resources, non-root bind permissions, read-only/offline preview and apply, preservation, conflicts, rerun and SIGTERM. Validate the Job schema; distinguish cluster/runtime coverage. Record test cache/volume ownership and size; follow canonical cleanup without shared-builder pruning. Keep image publication separate.
For [dialogue memory](../../../docs/DIALOGUE_MEMORY.md), test all supported imports, selected ChatGPT branches, omissions, untrusted roles, pending prose, explicit keyed conflicts/resolutions, source-bound review, immutable reruns, both exports and byte limits. Exercise linked/blocked paths, private/team upgrade preservation, Git/build/package exclusions and installed CLI execution. Do not infer semantic agreement from no keyed conflicts or claim live sync/upload from local export. Preserve raw sources and default project context.
For optional upstream skills, follow [security-audit](../../../docs/SECURITY_AUDIT.md) and [Strix integration](../../../docs/STRIX.md): preserve pinned bytes/license and newline attributes, verify bounded tree coverage/digests and nested-path aliases, explicit-only metadata, absent default/preset payloads, independent persistent choices, local adaptations/conflicts, links, visibility and wheel resources. Exempt only verified external examples from authored model/language/budget checks. Keep installation separate from audits/runtime setup; report unsupported validators. For Strix, check writable copy isolation, explicit full/diff scope, budget/data destination and no implicit cloud fallback; integration tests must start no scan, model or container.
For [Context7](../../../docs/CONTEXT7.md), check opt-in skill/registry selection, independent persistence, source/native copies, local conflicts and unchanged native MCP configuration. Keep normal skill discovery, version-aware sanitized queries and remote activation separate. Recheck client schemas/authentication against primary docs; a source revision does not pin the hosted service.
For [Web Design](../../../docs/WEB_DESIGN.md), check whole-module default absence, multi-entry skill selection, independent persistence, created-once project documents, native copies/retirement, local conflicts and linked/blocked paths. Exercise workflow context scope, required design memory above budgets, plain HTML without assumed stack facts, private raw output/team visibility and offline wheel resources. Keep TypeUI/SkillUI setup and real browser/Laravel acceptance separate from installer evidence; adopt external guidance only within the selected stack/visual system.
For a requested release, synchronize VERSION, root changelog, README, WIKI and maintainer context; keep kit history out of templates. Recheck CLI/version, kit, isolated archives, offline installation and prior-version upgrade preservation. Distinguish checks before/after labeling. Follow [the owner-approved release path](../../../docs/V1_4_3.md): preserve tags, wait for hosted build checks and report pending human approval; never approve PyPI for the owner. Git authorization follows Core. Release notes must distinguish installed guidance from external tools, live client/model setup and generated-site acceptance; verify security examples against pinned contracts.
For every Docker image release, follow the [Docker Hub checklist](../../../CONTRIBUTING.md#docker-hub-releases) and update the complete [Overview](../../../docs/DOCKER_HUB_OVERVIEW.md); synchronize version/features/examples before building and verified publication identity afterward.
For the current Docker/privacy release, follow [v1.4.1 preparation](../../../docs/V1_4_1.md); keep local image identity distinct from a registry digest and source-bound human review distinct from automatic conversation synchronization. After explicitly authorized publication, read back the registry digest, platform and version/configuration; preserve the prepared image and record later publication-only documentation separately. Do not assign additional tags implicitly.
Do not bootstrap the reference repository as an application or imply native/runtime checks passed from static validation.