_agents-overview

Agents Overview

List of 34 Agents defined in bkit and their roles (v2.1.13) v2.1.38: QA measurement release — agent count unchanged (34, model matrix invariant). Three QA agent definitions changed, none of them a count or a model: qa-test-planner had Write on its disallowed list while its stated role is producing test plan documents, so the plan it exists to write could not be written anywhere; qa-test-generator now reads that plan and stops when it is absent; and qa-lead sequences planner before generator (the generator's input is the planner's output), dispatches qa-monitor for runtime evidence, and records a live Chrome MCP probe. CC recommended: v2.1.220.

v2.1.37: Permission-mode awareness release — agent count unchanged (34, model matrix invariant). No agent definition, model, or count change; this release is scoped to the hook decision layer (bkit now reads permission_mode) and to guardrail rule precision. CC recommended: v2.1.220.

v2.1.37: Permission-mode awareness + CC v2.1.228–v2.1.232 response — agent count unchanged (34, model matrix invariant). Six agent definitions changed, none of them a count or a model: 23 frontmatter declarations across six keys Claude Code does not read were removed or moved to a field it does. Three agents whose skills preloading had never happened — the key was skills_preload, which appears nowhere in Claude Code's documentation — now preload the five skills they declared, including code-analyzer's three, which the PDCA Check phase depends on. cc-version-researcher drops the withdrawn sixth differentiation (ENH-432), and sprint-orchestrator records that "await completion" may now span turns under fork mode (ENH-478).

v2.1.36: Guardrail precision release — agent count unchanged (34, model matrix invariant). No agent definition, model, or count change; this release is scoped to Destructive Detector rule precision (issue #148), config wiring, and the runtime-immutability decision recorded in ADR 0016. CC recommended: v2.1.220.

v2.1.35: Correction release — agent count unchanged (34, model matrix invariant). No agent definition, model, or count change; this release is scoped to git-worktree detection correctness, the advisory text that described it, and the repository-wide child_process argv migration. CC recommended: v2.1.220.

v2.1.34: Reachability release — hook events 22 -> 21, blocks 25 -> 24. Hook timeout is measured in SECONDS and bkit had been writing milliseconds, so a declared 10000 on Stop meant 2h46m rather than 10s and a hung hook had no effective cancellation — the real cause behind issue #139, whose symptom alone was patched in v2.1.30. The FileChanged handler had never run once since v2.1.1 (an if rule cannot take | alternation, if is evaluated only on tool events, and FileChanged's matcher names literal files rather than globs), so PDCA document-change detection moved to PostToolUse Write/Edit where it demonstrably fires. once: true on SessionStart was ignored outside skill frontmatter and is gone. SKILL.md edits were never linted and unified-write-post never ran on Edit, because one if rule and a Write-only matcher covered half the cases. Every finding was reproduced against a real Claude Code runtime, and a new host-integration layer asserts from the outside that registered hooks actually dispatch. CC recommended: v2.1.220.

v2.1.33: Enforcement release — agent count unchanged (34, model matrix invariant). bkit-impact-analyst corrects a stale directory reference (mcp-servers/ → servers/). No agent definition, model, or count change; this release is scoped to hook enforcement, scope rules, session UX and CI gating. CC recommended: v2.1.220. v2.1.32: CC v2.1.219/220 compatibility — agent count unchanged (34, model matrix invariant). cto-lead and pm-lead correct their nested-spawn notes: Claude Code v2.1.217 disabled nested subagent spawning by default and v2.1.219 re-enabled it at depth 3, so the previous claim that Task() is "blocked by CC's nested spawn restriction" no longer holds. One-level dispatch is now documented as bkit's own convention rather than a platform guarantee, with CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH offered for a deterministic bound. sprint-master-planner gains a rule against delegating back to the lead that invoked it, closing two declared spawn cycles that were unreachable at depth 1. CC recommended: v2.1.220. v2.1.30: Stop-hook stdin-block hardening (#139) — agents unchanged; a runtime-reliability fix in lib/core/io.js / scripts/unified-stop.js (bounded stdin read so the Stop hook cannot stall on a held-open pipe). No agent definition, model, or count change. v2.1.26: Fable cost retune — the 3 high-frequency PDCA verifiers (gap-detector, design-validator, pdca-iterator) move fable→opus (they run in the repeated Check/iterate loop; Opus 4.8 is strong at verification at half Fable's $10/$50 cost). Matrix now: 6 fable (leads + sprint verifier) / 10 opus / 16 sonnet / 2 haiku. Fable stays on the long-horizon orchestrators where its planning/delegation edge compounds. v2.1.25: Claude 5 Model Alignment + Issue Response — 4-tier role-based model matrix: 9 fable (verification & orchestration core) / 7 opus (deep reasoning & security) / 16 sonnet (implementers) / 2 haiku (monitors). 16 reassignments (9 opus→fable, 1 opus→sonnet sprint-report-writer, 6 sonnet→haiku pdca-eval-* — subsequently REMOVED from agents/ per #128/ADR 0014, deprecation registry at test/contract/deprecation-registry.json). Descriptions compacted −44% per #129 (compact 8-language triggers; "Do NOT use for" moved to body). Model floor: fable requires CC ≥ v2.1.170 (SessionStart advisory ENH-368 below it). CC recommended: v2.1.198. v2.1.24: Skill namespace hardening (#125/#126) — agents unchanged (40 agent files: 34 active + 6 deprecated pdca-eval-* tombstones). v2.1.13: Sprint Management agents — added 4 sprint agents (sprint-orchestrator, sprint-master-planner, sprint-qa-flow, sprint-report-writer). Removed 6 pdca-eval-* agents (Korean-only frontmatter + v1.6.1 stale baseline + 0 spawn sites). Total 36 → 34. cto-lead/pm-lead/qa-lead extended with sprint Task spawn patterns + body sections. pdca-iterator/product-manager/gap-detector/self-healing/pipeline-guide/qa-monitor descriptions extended for sprint awareness (관점 1-1). v2.1.11: 4 Sprints × 20 FRs Integrated Enhancement — Agents unchanged (36); Sprint γ adds lib/application/pdca-lifecycle/ pilot referenced by pdca-iterator and gap-detector workflows. CC recommended: v2.1.118+ (79 consecutive compatible releases). v2.1.10: Sprint 5a~7 complete — cto-lead body expanded (5 Task spawn blocks + Task(pm-lead) / Task(qa-lead) / Task(pdca-iterator) added to frontmatter, G-T-01/02). Enterprise teammates 5→6 (G-T-03). Bulk @version 2.0.0 → 2.1.10 refresh across 79 files (66 lib + 13 scripts). CC recommended: v2.1.117+ (75 consecutive compatible releases). v2.1.9: CC v2.1.116 response — ENH-253/254/259/263 (zero-script-qa fork verification, defense-in-depth security docs, custom skill warning, Docs=Code 15-file correction). CC recommended: v2.1.116+ (74 consecutive compatible, v2.1.115 skipped). v2.1.8: Issue #81 hotfix - agents unchanged. Hook/lib layer focus (lib/core/context-budget.js, session-ctx-fp.js). CC recommended: v2.1.111+ (72 consecutive compatible). v2.1.7: Issue #79 hotfix, PDCA workflow stabilization.

v1.4.1: Added Context Engineering perspective - Role-based Behavioral Rules Layer v1.5.0: Claude Code Exclusive v1.5.4: Model distribution verified: 7 opus / 7 sonnet / 2 haiku, 9 acceptEdits / 7 plan v1.5.6: Auto-memory integration, multi-agent memory optimization (CC v2.1.59) v1.5.7: /simplify + /batch PDCA flow integration, English conversion for stop script outputs v1.5.8: Studio Support awareness - Path Registry, state directory migration, 186 exports v1.5.9: Executive Summary module, AskUserQuestion Preview UX, ENH-74~81, 199 exports v1.6.0: PM Agent Team (5 agents), 21 total agents v1.6.1: 3-Tier Agent Security Model (9 acceptEdits agents with tiered disallowedTools), CTO Orchestration Redesign, 208 exports, CC v2.1.71 v1.6.2: Agent frontmatter effort/maxTurns native support (29 agents), CC v2.1.78 v1.6.2+: cc-version-researcher + bkit-impact-analyst agents (31 agents), CC version analysis workflow v2.0.6: self-healing agent (opus) added (32 agents, 11 opus / 19 sonnet / 2 haiku)