[medium] Release 1.1.0: Codex activation and evidence-based agent handoffs

[ytvee]

Summary

Release candidate 1.1.0, retaining 21 skills and the lightweight WebDev workflow. Related to #71. This updates the existing release PR; it does not merge, tag, publish or change user-project installations.

Superpowers-inspired task/report packets, task-wide Git review evidence, scoped repair re-review, shared retry budgets and structural-versus-behavioral evaluation remain. No mandatory TDD, universal worktrees, new runtime, provider, dependency or delegation for trivial tasks.

Workload-aware model routing

• Separate task risk, model capability and reasoning effort. Inline/tools for trivial work; Luna low/medium for bounded extraction and mechanical edits; Terra medium for established implementation; Sol high for ambiguous cross-module work and material review; Astra high for consequential architecture and critical review; Astra xhigh only for exceptional coupled architecture. These are dated, evidence-gated recommendations, not fixed universal bindings or comparative performance claims.
• Keep four base roles backward compatible. New full onboarding normally adds wdk_worker_light and wdk_architect; three light/deep specialties are optional. Distinct roles prevent recommending conflicting fixed model/effort pairs under one name.
• light is not an effort value. max/ultra require actual support, justified need, budget and compatible execution semantics; no automatic escalation for environment, path or permission errors.
• Preserve existing customized bindings during upgrades. Additive configuration refuses unowned paths, collisions and implicit role removal. Both formats, rollback and idempotence are tested. Per-role fingerprints support evidence carry-forward only with verified unchanged effective client/auth/permissions/tools/dispatch context; changed bindings invalidate old evidence.
• Full onboarding includes scoped configuration and tiny read-only canaries without redundant confirmation. Facts-only/no-model-change, Plan Mode, upgrades and non-Codex paths do not configure models or run canaries. Primary/global settings, trust, security, MCP, concurrency and user-owned files remain protected.
• Update installation/bootstrap docs, local templates, release notes and acceptance checklist. Architecture, greenfield, design, debugging and review examples are covered by one progressively loaded matrix rather than repeated model tables in every skill.

Verification

• Local full validator passed: six generated targets, 48 planning smoke cases, schemas, versions, context budgets, structural/synthetic behavior checks and release archives.
• Installer: 45 tests, one Windows symlink-privilege skip. Review package: 16 tests, one equivalent skip. Real junction and journal ACL read-back checks passed without elevation. Markdown links passed.
• Independent code/instruction review found a carry-forward fingerprint ambiguity; fixed and re-reviewed clean in the inspected scope.
• Two actual candidate selection scenarios passed manual trace review, with zero file changes. The comparable baseline recommended multiple incompatible pairs for one fixed role; the candidate separated roles and avoided unnecessary xhigh. The inventory contains 33 scenarios, not 33 executed model evaluations.
• Final candidate b5e281bc842cafd1a3f9859f6ccb74aae3538176: all PR and push workflows passed. PR evidence: source/six targets and Linux/Windows regressions, Ruff format/lint, YAML and Markdown, links. Exact-head push workflows also passed: source/helpers, quality, links. Working tree is clean. The last commit only records the completed canary correction; its schema/link checks and all CI passed after the full local validation of the implementation.

Authenticated native dispatch

On Windows, authenticated Codex CLI/App Server 0.154.0-alpha.6.2 loaded nine registered project-local roles in the already user-trusted disposable fixture. No new global/trust/security changes were made. Parent 01a0a29d-fc8b-7370-8339-29aaa625b61a selected native agent_type, omitted explicit model/effort overrides and used fork_turns: none, at most two concurrent children.

Role	Actual child model / effort	Child ID
wdk_lookup	gpt-5.6-luna / low	01a0a29e-5852-7ed1-abbd-30b9e649b864
wdk_worker_light	gpt-5.6-luna / medium	01a0a29e-6da0-7d53-a3ba-523f88385be5
wdk_worker	gpt-5.6-terra / medium	01a0a29e-ab80-77a2-819c-6b7f1b510bba
wdk_reviewer_light	gpt-5.6-terra / high	01a0a29e-c4a2-73e3-94b5-3861959c6ef6
wdk_complex	gpt-5.6-sol / high	01a0a29f-433f-7b53-b347-3e3114613254
wdk_reviewer	gpt-5.6-sol / high	01a0a29f-598d-7cd1-95ec-d92953512163
wdk_architect	gpt-6-astra / high	01a0a29f-e22c-70b2-a050-45f55384c2ec
wdk_architect_deep	gpt-6-astra / xhigh	01a0a29f-f7d1-7260-9a78-a5ae9ccab3c3
wdk_reviewer_deep	gpt-6-astra / high	01a0a2a0-acc4-7a23-9428-77df3063c204

Actual child turn-context metadata matched all nine bindings. Every child read the relevant fixture successfully (exit 0). Parent/children were read-only, approval never; parent network access false. Primary remained Sol/medium; managed native configuration passed hash inspection afterward. Architecture tasks identified scope-aware caching, stale responses, sensitive persistence, safe rollback and the impossibility of guaranteeing isolation while unsafe old clients continue unchanged. Reviewers correctly rejected the deliberately insufficient repair; that rejection is expected, not a failed reviewer test.

The first lookup returned correct data but paraphrased instead of quoting exactly: native dispatch passed, exact-output acceptance failed. Bootstrap now requires literal acceptance checks and bounded same-role correction. One follow-up stopped before delegation after guessing a nonexistent profile path; this is retained as a harness failure, not an accepted run or reason to raise model cost. Exact-path correction then passed: parent 01a0a2a4-724f-7482-a18b-aa7fd5736595, child 01a0a2a4-cda5-7702-b6d5-859d8509583c, actual Luna/low, read exit 0, separate case-sensitive literal comparison true. Only that child was re-run; the original failure remains disclosed.

These results supersede the previous dated trust blocker on this tested surface only. They do not establish direct-binding mode on another app surface, complete architecture-skill quality, universal client activation, savings or production readiness of the synthetic design. Account catalog, profiles, paths, raw traces and recovery journals remain outside source and release archives. Merge, issue closure and publication remain separate actions.