ginee v0.26.0

Highlights

• Specialist cardinals can resume each other directly. Frontend ↔ QA iteration cycles (and other single-defect peer fixes) no longer relay through team-lead — 7 main-thread turns per cycle collapse to 2. Over 5–10 iterations on a single task, main-thread context savings approach the warm-reuse savings of #90 / D43. Round-trip cap default 3 per phase; escalation to team-lead on contradictory return, multi-cardinal surface, or scope change (#189).
• The wrapper-anchor pattern is now an official primitive. New core/protocols/adapter-wrapper-pattern.md names .claude/agents/<role>.md (and adapter siblings) as the canonical Tier-1 / Class-A entry of the compliance playbook (#135). Design invariant: hardenable properties (tools: · model: · description: · phase-participation:) live in frontmatter; charter prose MUST NOT restate the host-enforced rule. Per-adapter availability matrix (Claude / Cursor / Copilot / Codex / generic) ships with the protocol.
• Every dispatch now carries a scope-size estimate. Team-lead emits ≤15m / 15-60m / >60m + one-line signal before every cardinal dispatch — surfaces on dispatch-prompts, sub-issue bodies, and the iteration-protocol load decision. Previously trigger-circular ("loaded on >15m, with no step deciding") (#168).
• Phase-6 defects ship with a committed failing test. QA writes the reproducer; engineer's fix-oracle is "test passes locally". Permanent regression gate per defect class. Untestable defects (pure visual / human-in-loop / timing) flag testable: false + rationale (#184).
• Engineers self-verify before QA hand-off. Phase 4 acceptance raised from unit-only to a strict change-scoped fix-loop gate — every available suite (component / API / integration / E2E / pixel-check / script-quality / post-step health / deploy smoke per role) MUST run green before hand-off. QA Phase 5 reframed as backstop (independent re-execution) — MUST NOT first-pass discover (#177).
• SA realigned to classical-architect boundaries. phase-participation: [1, 2, 7] (excluded from 4/5/6); SA artefacts forbid implementation rendering (function/member identifiers · <file>:<line> citations · commit SHAs · handler-body snippets · wiring prescriptions). Two new Claude PreToolUse hooks (T14 / T15) enforce hard. Phase 7 governance review now conditional (#182).
• /ginee-iterate skill — review-cycle replies forward to the warm cardinal. Skill-runner no longer short-circuits the warm cardinal on small fixes — SendMessage relay preserves warm-reuse savings across 10–20 iteration cycles that previously bloated main-thread context by 50–100k tokens. Skill count 12 → 13 (#154).

Adopter migration

Existing installs run /ginee-update to apply. Update path is mechanically driven — install.{ps1,sh} --update-only nukes .agents/ginee/core/+adapters/+extras/, fetches the v0.26.0 ref, re-copies _shared/agents/*.md → .claude/agents/ (or adapter equivalent). The new SendMessage in specialist tools: arrays propagates automatically.

Notes

• No local/framework.config.yaml changes needed. Peer round-trip cap default is 3; override via iteration.peer-roundtrip-cap if needed. Estimation gate, defect-reproducer, self-verify all opt-out via existing compliance.disabled keys.
• Warm-reuse prerequisite remains — Claude adopters need CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in .claude/settings.json § env for SendMessage to be available; restart Claude Code after change. Adopters without warm-reuse fall back to team-lead-relay (no behaviour change).
• SA boundary enforcement is HARD — PreToolUse hooks (T14 / T15) block Task(subagent_type=solution-architect) in Phase 4/5/6 contexts and SA-owned-path edits carrying implementation rendering. Per-tactic opt-out: compliance.disabled: [pretooluse-task-hook] / [pretooluse-sa-artefact-hook].
• Full specs: migrations/ — estimation-gate-fires.md · defect-reproducer-contract.md · engineer-self-verify.md · sa-boundary-tightening.md · ginee-iterate-skill.md. The wrapper-anchor pattern doesn't ship a migration doc — the spec itself (core/protocols/adapter-wrapper-pattern.md) is the canonical reference.