ai-playbook-configure

/ai-playbook-configure

Configure edits Plan, Build, Verify and escalated repair, plus five stage-owned job bindings, including portable custom sources. QA inherits Verify; Coordinator is observed display-only. Personal defaults and named presets are local. Evidence-backed recommendations are read-only; unavailable-model replacements edit only the draft and still require explicit Apply. This skill never launches a model or edits runtime state.

Procedure

1. Read and discover. Identify the project and installed playbook locator. Use model-router's existing discovery procedure without its selection/launch steps. Exclude unavailable, unauthenticated, blocked, identity-unverifiable or role-ineligible routes. Supply a current-availability adapter to the helper's --discovery-command, not a static discovery file. Include the current chat's observed identity as optional coordinator route metadata in that same fresh discovery, not a user-supplied model or a guessed catalogue default. If unavailable, display unknown; do not switch the Coordinator. On each request, genuinely recheck role + model + runner + reasoning through the existing authority surface, echo the request ID and timestamp the observation inside the helper's clock window. Never redate cached results or treat guidance's 24-hour interval as access evidence. Keep bounded discovery local, not in shareable settings. Read the helper contract at {playbook-path}/v0.5/scripts/playbook-config.md using the installed playbook locator, not a path relative to the copied skill. The mockup and generated availability claims are not evidence. Supply an explicit user-local --preferences-dir outside the project on every conversation request; never put a private destination in shared configuration. Pass only known goal, task, risk and billing facts in read's context. Read existing project context instead of asking it again; unknown billing is a valid answer. Run the helper's read action. On blocked/recovery-required, report its corrective action and offer Reload or Not now. Completion criterion: current values, origins and verified routes. Official evidence uses a bounded external local cache, optionally selected with --evidence-dir on every helper request. The live helper defaults to a private user cache; never copy it into project settings or shared evidence. Reuse successful guidance, reasoning/capability and pricing checks for at most 24 hours only at Configure or existing lane-discovery checkpoints. First use, newly observed model/version and typed Refresh check immediately. Display claim-specific successful dates, fingerprints and material changes. Unchanged content produces no invented notice. A failed check retains its successful date and labels stale/incomplete; new-model advice needs adequate evidence. No background watcher, unrelated-turn check or package installation.
2. Present project defaults. Ask only the returned missing_context questions through typed Goal <context>, Billing api, Billing subscription, Billing mixed or Billing unknown, and Guided / Expert. Goal/billing stay out of project writes; billing saves only after explicit local approval. Known local presentation/billing skips those questions on reopening. Show all four model/runner/reasoning choices and effective origins, QA inheriting Verify, observed Coordinator authority/date or unknown, retained repair constraints, exact destination .playbook-config.json and whether adoption migrates legacy preferences. Legacy state stays byte-for-byte intact, becoming inert default history after adoption. Offer typed Edit, Edit Plan, Edit Build, Edit Verify, Edit Repair, Apply, Explain <role>, Not now. Expert presentation is compact, never a different proposal or weaker validation. Label every row as a retained edition/project starting choice or an explicit user edit, not a recommendation. Display the separate recommendations rows for all four roles: exact model, runner, supported reasoning, project/task fit, risk and limitations. No eligible recommendation means unknown, not an invented fallback. The helper retrieves official sources at initial live read; Explain reuses that proposal without retrieval or invocation. Provider guidance, published API rates and comparable local outcomes are separate claims, each with its own source URL, successful check date and uncertainty. An incomplete check's retrieval time is not a successful evidence date. General reasoning guidance never proves account access. Show API currency, token unit and billing route; show estimates only with labelled input/output/retry assumptions. Subscription/mixed routes show observed usage/limits or unknown, never API prices as a subscription bill. Do not promise savings or a cheapest verified outcome, run candidates, grant approval, or silently rewrite any default. Keep local evidence and billing out of shared config and tracked transcripts. Task coding / Task analysis and Risk ordinary / Risk high correct task advice only. For an unbootstrapped project, hand the complete role draft to the existing bootstrap approval preview. Do not run a second bootstrap or write project settings here before that gate; reopen after bootstrap and review again. Completion criterion: the user sees scope, origins, migration and destination.
3. Edit and preview. Pass replies through the helper's reply action with the previous proposal. When its step is replacement, show the unavailable saved value and exactly one verified suitable alternative, including role, runner/reasoning, task fit, costs and evidence limits. Offer typed Accept replacement, Choose another model or Not now. Acceptance edits only the affected role and returns to the complete before/after preview. Choosing another opens that role's ordinary editor. With no suitable alternative, show the limitation and restore-access/Refresh/explicit-editor recovery path; do not offer acceptance. Multiple unavailable roles are handled in role order on subsequent discovery/Apply checks, retaining all unrelated drafts. Pass ordinary editing replies with the previous proposal. Edit offers numbered roles. Edit <role> displays verified complete alternatives numbered from 1, showing model, runner and reasoning. Pick model offers numbered model, then supported runner, then supported reasoning, and complete identity if otherwise ambiguous. Never silently choose among providers or optional identity metadata. A number edits only the draft and returns the complete before/after preview; show retained roles and constraints too. Explain <role> identifies that selected route, availability observation and limitations without selecting or launching anything. Explain defaults to Build and reports discovery authority/date and explicitly unknown suitability/comparable cost; no paid benchmark. With no alternatives, offer Reload or Not now instead of inventing a route. Native controls may mirror, never replace, typed replies at every level, with immediate text fallback. Back returns to the full proposal without losing other edits; an invalid reply retains the previous draft under retained_proposal and offers Back. Edit and Edit <role> from that blocked result also reuse the retained draft, never implicitly reload and discard other edits. Reload intentionally discards unsaved draft changes. Guided / Expert show a separate local preference before/after preview with its exact destination and resolved_destination behind directory aliases. If directory identity changed, show the corrective message and require the new preview before another Apply; retain unrelated role drafts. Only Apply preference saves it using the same durability/completion/recovery protocol as project defaults; no project draft is saved by that action. Return to the complete role proposal afterward. Not now saves nothing pending; if presentation was explicitly saved earlier, explain that it remains saved while the project draft is cancelled. Completion criterion: reviewed draft or cancellation, with nothing saved. Skill editing uses Edit skills or Edit skills <job> for alignment, specification, implementation, code review or application QA. Show all skill_proposal and skill_changes before/after rows in the full proposal. Numbered skill_options retain the same verified source prefix/provenance. Choose <number> edits one job; Choose <number>,<number> retains supported alignment context → decisions composition order. Explain skills <job> shows inputs, outputs, effects, owner and eligibility without invocation. Unknown, changed, colliding or incompatible sources require an explicit fallback or block, never warning acknowledgement. Read ../../scripts/skill-bindings.md for contracts, current upstream exclusions and already eligible project QA proof. Configuration does not create or maintain a harness. Presets shows one Recommended entry plus named user presets. Recommended retains the starting draft; evidence-backed suggestions are displayed separately and do not select a model. No static cost tiers or availability-as-suitability claim. Load defaults or Load preset <name> changes only the draft, showing model/skill origins, before/after changes and destinations. Save defaults or Save preset <name> previews personal reusable data using the same validated model/skill schema. Apply preference saves only that personal data; other projects and active work remain unchanged. Billing <mode> previews a local billing preference; unknown consumption remains unknown and API prices are not subscription bills. Presentation edits preserve saved billing, defaults and presets. Missing dependencies or unknown fields require explicit review, not silent omission. Custom sources use a canonical project-relative identity or a logical identity resolved through an external machine-local binding/audit store. Show (Custom) unless existing stage evidence verifies a more specific source label. Missing local resolution/evidence requires explicit recovery, never a source rewrite or automatic fallback. Local JSON assertions are not real independent execution proof; the owning stage authenticates retained qualification at its existing gate. Never generate that proof by invocation.
4. Apply the reviewed proposal. Only after typed Apply, repeat genuine authoritative discovery through that adapter and pass the same proposal to the helper. Never silently create and Apply a refreshed proposal. Changed inputs require Reload and another preview/Apply. Unavailability returns the replacement conversation with the draft retained. Acceptance requires a refreshed preview and another explicit Apply; never substitute or save the old proposal. Report applied/unchanged/blocked/recovery-required exactly with destination and corrective message. Not now writes no pending defaults or presets; distinguish earlier explicitly saved presentation and any local read-side discovery refresh. Stop for retained recovery evidence; never remove it to make a save appear successful. When the preview includes a personal change, project Apply uses paired transaction journals and content completion for both destinations. Report every destination and the true result; independent saves are not paired success. Exact rollback still reports recovery-required, and concurrent edits/captured bytes must never be overwritten. Keep journals and private evidence local; never put billing, binding-store paths or unreviewed exception text in shareable settings/results. The reviewed personal destination remains visible in the local review/recovery flow. New-project seeding belongs to the bootstrap skill's existing preview and exact seed approval, never to an implicit Configure bootstrap or stage/model launch. Completion criterion: validated save, no-op, cancellation or named blocker, with no model dispatch or build.

Return playbook_result with outcome applied/unchanged/blocked, next_stage: null, and ordered required_actions from the corrective result. Existing lane gates remain selection, approval and launch authority.