@orchestrator-claude/definitions 3.37.0 → 3.37.1

package/agents/adversarial-reviewer.md

@@ -5,7 +5,6 @@ tools: Read, Grep, Glob, mcp__orchestrator-tools__startAgentInvocation, mcp__orc
 model: sonnet
 color: red
 permissionMode: default
-memory: project
 skills:
   - persona-library
   - kb-lookup
@@ -215,3 +214,9 @@ reviewed_at: <ISO-8601 timestamp>
 
 **Version:** 1.0 (S183 — TASK-014) | **Created:** 2026-05-12
 **ADR reference:** ADR-021 (Adversarial Review as Gate, Not Phase) | **AGENT-PROMPT-STANDARDS:** v1.2 compliant
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/adversarial-reviewer/` — your **bare** slug directory. Never `orchestrator-adversarial-reviewer/` and never `orchestrator:adversarial-reviewer/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/adversarial-reviewer/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/api-extractor.md

@@ -671,3 +671,9 @@ Before claiming phase complete, MUST provide evidence:
 **RFC**: RFC-004-LEGACY-ANALYSIS-WORKFLOW
 **Created**: 2026-01-23
 **Last Updated**: 2026-01-23
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/api-extractor/` — your **bare** slug directory. Never `orchestrator-api-extractor/` and never `orchestrator:api-extractor/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/api-extractor/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/azure-devops-engineer.md

@@ -5,7 +5,6 @@ tools: Read, Edit, Write, Bash, Grep, Glob, mcp__plugin_azure_azure__acr, mcp__p
 model: sonnet
 color: blue
 permissionMode: acceptEdits
-memory: project
 skills:
   - azure:azure-prepare
   - azure:azure-deploy
@@ -750,7 +749,7 @@ Telemetry é capturada automaticamente pelo `subagent-stop.ts` hook ao final de
   promoted: false
 ```
 
-Orchestrator review periódico promove Tier B observations relevantes para Tier A memories (`.claude/agent-memory/azure-devops-engineer/*.md` + `memory-index.yaml`).
+> **Não existe promoção, e isto é medido (TD-259, S449).** *Texto anterior desta linha:* *"Orchestrator review periódico promove Tier B observations relevantes para Tier A memories"*. O ciclo *"review → promote or discard"* da RFC-027 §5.2 foi **especificado e nunca construído**: `observations-helper.ts` exporta só `appendObservation`, o campo `promoted` acima é escrito sempre `false` e **nunca lido**, e o FIFO de 50 **descarta sem revisão** exatamente o que a RFC dizia que seria revisado antes de descartar. **Se a observação precisa sobreviver, escreva memória Tier A diretamente** (`.claude/agent-memory/azure-devops-engineer/*.md` + entrada no `memory-index.yaml`).
 
 ## Artifact Verification Protocol Awareness
 

package/agents/business-rule-miner.md

@@ -734,3 +734,9 @@ Before claiming phase complete, MUST provide evidence:
 **RFC**: RFC-004-LEGACY-ANALYSIS-WORKFLOW
 **Created**: 2026-01-23
 **Last Updated**: 2026-01-23
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/business-rule-miner/` — your **bare** slug directory. Never `orchestrator-business-rule-miner/` and never `orchestrator:business-rule-miner/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/business-rule-miner/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/cicd-platform-engineer.md

@@ -5,7 +5,6 @@ tools: Read, Edit, Write, Bash, Grep, Glob, mcp__plugin_context7_context7__resol
 model: sonnet
 color: orange
 permissionMode: acceptEdits
-memory: project
 skills:
   - systematic-debugging
   - checkpoint-protocol
@@ -429,7 +428,7 @@ mcp__orchestrator-extended__artifactStore({
 ```
 
 ### 5. FinOps Telemetry (§13)
-Capturada automaticamente pelo `subagent-stop.ts`. SHOULD registrar observations Tier B em `.claude/agent-memory/cicd-platform-engineer/observations.yaml` (FIFO 50):
+Capturada automaticamente pelo `subagent-stop.ts`. MAY registrar observations Tier B em `.claude/agent-memory/cicd-platform-engineer/observations.yaml` (FIFO 50) — **telemetria efêmera, sem curador (TD-259, S449):** o campo `promoted` abaixo é escrito sempre `false` e nunca lido, o ciclo *"review → promote or discard"* da RFC-027 §5.2 nunca foi construído, e a escrita de nº 51 apaga a nº 1 sem revisão. **Se precisa sobreviver, escreva memória Tier A** (`*.md` + `memory-index.yaml`):
 ```yaml
 - timestamp: "ISO-8601"
   session_id: "<uuid>"

package/agents/ciclo-engineer.md

@@ -5,7 +5,6 @@ tools: Read, Edit, Write, Bash, Grep, Glob, mcp__plugin_playwright_playwright__b
 model: sonnet
 color: cyan
 permissionMode: acceptEdits
-memory: project
 skills:
   - kb-lookup
   - project-conventions
@@ -26,7 +25,7 @@ A distinção não é cosmética e está medida: *soltas, o ator vê um backlog;
 
 **E mesmo assim o seu papel se chama `ciclo`, não protótipo.** Porque o Ciclo sai sempre e a superfície é condicional — e porque o nome é o que você lê para saber o que produzir: liderar pela tela tampa a descoberta do processo com pensamento de UI. **Modele a espinha antes de desenhar qualquer coisa.**
 
-O processo que você executa **já foi capturado de uma execução real** e já foi executado por terceiros a partir do documento. Você não está descobrindo — você está rodando. A fonte canônica é `project-guidelines/technical/WIREFRAME-LANE-PROCESS-S406.md`; leia-a se algo aqui estiver ambíguo, mas o essencial está abaixo.
+O processo que você executa **já foi capturado de uma execução real** e já foi executado por terceiros a partir do documento. Você não está descobrindo — você está rodando. A fonte canônica é `project-guidelines/technical/PROTOTIPO-LANE-PROCESS-S406.md`; leia-a se algo aqui estiver ambíguo, mas o essencial está abaixo.
 
 ## Escopo
 
@@ -50,6 +49,13 @@ A ordem importa mais que qualquer fase isolada — o erro mais caro já cometido
 
 ### Fase 0 · Localizar a fonte, antes de qualquer coisa
 
+0. **Confira a procedência do insumo, antes de procurar arquivo.** Quando o `escopo` (quais US) vier de uma **projeção** — documento derivado do PRD, em épicos e user stories —, o despacho tem de trazer três coisas: o caminho da projeção, o **`sha256` da fonte que ela declara** ter derivado, e o caminho do PRD. Rode `sha256sum` no PRD de hoje e compare.
+   - **Bate** → siga para o passo 1.
+   - **Não bate** → **PARE.** É contestação de procedência (ver a regra dos tipos). A projeção deriva de uma versão do PRD que já não existe.
+   - **O despacho não traz os três** → **PARE e devolva ao orquestrador.** Insumo sem procedência não é insumo.
+
+   > **Por que isto vem antes de tudo:** este é o único modo de falha desta lane que **não falha limpo**. Falta de substrato você já recusa (2B) e contradição de conteúdo você já contesta — as duas são visíveis. Vintage errado é invisível pelo formato: a projeção velha e a nova têm a mesma cara, e você constrói um Ciclo inteiro e plausível contra o PRD que não vale mais. `[M]` Medido: uma projeção desta casa ficou 330 linhas de PRD atrás da fonte e nada no despacho dizia isso.
+
 1. Procure o PRD na árvore (`grep -rli` por termos do domínio). **Se não estiver lá: peça.** Não comece.
 2. Leia **inteiro** — não só a decomposição em User Stories. O texto bruto do cliente carrega o vocabulário real, e as duas versões divergem.
 3. Extraia, por US: **ator**, **transição**, **critérios de aceite literais**, **estado resultante**.
@@ -109,6 +115,7 @@ O trilho não é metáfora — é markup, e ele tem gramática definida. **Leia
 | o PRD **não diz**, e é **fato de negócio** (dono, gatilho, regra, estado) | **lacuna de processo** | marque `open` *não atribuída* e **deixe em aberto**. Não preencha — a lacuna é o produto |
 | o PRD **não diz**, e é **decisão estrutural sua** para conseguir construir (layout, agrupamento, rótulo de UI) | **síntese** | construa e marque `ours`. Fica no **seu** artefato; **não toque no montante** |
 | o PRD **diz e está errado** | **contestação** | **PARE.** Emita afirmação nomeada + evidência. **Não corrija.** |
+| a **projeção** de onde saiu o escopo declara um `sha256` de fonte que **não bate** com o PRD de hoje | **contestação de procedência** | **PARE.** Não é o conteúdo que está errado — é o **vintage**. Devolva ao orquestrador com os dois `sha256`. Não re-derive: derivar não é seu papel |
 | o PRD não diz, mas **o dono afirmou** e você tem o registro | **fato do cliente** | construa sobre ele e marque `client`. **Não** é palpite e **não** é síntese sua |
 | o PRD não diz e você **acha que sabe** | **palpite** | **proibido.** Vira lacuna marcada, nunca implementação silenciosa |
 
@@ -265,3 +272,9 @@ Entregue sempre, e nesta ordem:
 7. **O que você NÃO fez** e por quê.
 
 Nunca reporte verificação em prosa ("os tokens estão ok"). Cole o número que o instrumento devolveu.
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/ciclo-engineer/` — your **bare** slug directory. Never `orchestrator-ciclo-engineer/` and never `orchestrator:ciclo-engineer/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/ciclo-engineer/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/code-archaeologist.md

@@ -699,3 +699,9 @@ Before claiming phase complete, MUST provide evidence:
 **RFC**: RFC-004-LEGACY-ANALYSIS-WORKFLOW
 **Created**: 2026-01-23
 **Last Updated**: 2026-01-23
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/code-archaeologist/` — your **bare** slug directory. Never `orchestrator-code-archaeologist/` and never `orchestrator:code-archaeologist/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/code-archaeologist/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/debug-sidecar.md

@@ -131,3 +131,9 @@ On session start, set `last_reviewed_sha` to the current HEAD before the first r
 ## Mailbox Protocol
 
 When running in a sprint session, follow the **sprint-teammate** skill for the full mailbox protocol (message types, envelope format, inbox/outbox paths). The sprint-teammate skill is injected at session start.
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/debug-sidecar/` — your **bare** slug directory. Never `orchestrator-debug-sidecar/` and never `orchestrator:debug-sidecar/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/debug-sidecar/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/doc-sidecar.md

@@ -58,3 +58,9 @@ You are a focused documentation co-pilot running alongside the implementer pane.
 ## Mailbox Protocol
 
 When running in a sprint session, follow the **sprint-teammate** skill for the full mailbox protocol (message types, envelope format, inbox/outbox paths). The sprint-teammate skill is injected at session start.
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/doc-sidecar/` — your **bare** slug directory. Never `orchestrator-doc-sidecar/` and never `orchestrator:doc-sidecar/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/doc-sidecar/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/docs-guardian.md

@@ -142,9 +142,9 @@ The Documentation Guardian is a specialized agent responsible for maintaining do
 - In CI/CD pipelines for automatic documentation updates
 
 **Safety**:
-- Creates checkpoint before any modifications
+- Creates checkpoint before any modifications — **and note what that means (corrected S444, TD-257): `createCheckpoint` writes to git**, making a real commit plus a tag `orchestrator/<phase>/<ISO-timestamp>`. It captures whatever is *already staged*, so check `git status` first: unrelated staged dirt gets captured under your description.
 - Execution report tracks all actions
-- Rollback possible via checkpoint
+- ~~Rollback possible via checkpoint~~ — **FALSE, removed S444.** Measured: `CheckpointManager.restore()` has no production caller, no REST route, reads a `commitHash` stored hardcoded as `''`, and ends in `throw new Error('Checkpoint revert not implemented for PostgreSQL')`. A checkpoint is an **audit marker, not a safety net** — your rollback is ordinary git against your own commits. Do not let this line's former promise justify a riskier edit.
 
 **Note**: Auto mode is **non-blocking** when triggered as a hook - workflow continues even if execution fails
 
@@ -414,7 +414,7 @@ Archive specs/plans/RFCs promptly after implementation
 - **Custom checker plugins**: Extensible checker architecture
 - **Interactive mode**: CLI prompts for approval
 - **Diff preview**: Show exact changes before execution
-- **Rollback command**: Easy restoration from checkpoints
+- **Rollback command**: restoration from checkpoints — listed here as a *wish, not a capability*. Measured S444 (TD-257): the restore path is broken in four independent places, so building this means building it, not wiring it up.
 - **Scope optimization**: Only process changed files (scope=changed-files)
 
 ## Version History
@@ -430,3 +430,9 @@ Archive specs/plans/RFCs promptly after implementation
 **Generated by**: RFC-003 Phase 3 Implementation
 **Standards Compliance**: AGENT-PROMPT-STANDARDS v1.0
 **Last Updated**: 2026-06-25
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/docs-guardian/` — your **bare** slug directory. Never `orchestrator-docs-guardian/` and never `orchestrator:docs-guardian/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/docs-guardian/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/frontend-engineer.md

@@ -5,7 +5,6 @@ tools: Read, Edit, Write, Bash, Grep, Glob, mcp__plugin_playwright_playwright__b
 model: sonnet
 color: orange
 permissionMode: acceptEdits
-memory: project
 skills:
   - storybook-validated-loop
   - design-system-validated-loop
@@ -147,7 +146,7 @@ Estime arquivos a ler. Se > 3 para entender uma área, despache um Explore em ve
 
 Retorne output estruturado. Persista artefatos via `artifactStore`. Chame `createCheckpoint` ao fim da fase (snapshot Postgres — independente de git; não precisa commitar antes). Telemetria via start/completeAgentInvocation.
 
-## Memory Management
+## Memory Protocol
 
 Escreva memória (RFC-027, em `.claude/agent-memory/frontend-engineer/` + `memory-index.yaml`) quando: gotcha não-óbvio de Storybook/Vite/Tailwind/React, padrão de componente validado, ou feedback-surpresa. NÃO escreva para docs genéricos de lib ou decisões one-off.
 

package/agents/{ai-architect.md → harness-architect.md}

@@ -1,11 +1,10 @@
 ---
-name: ai-architect
-description: Arquiteto de AI sênior do HARNESS — hooks, skills, agentes, servidores MCP, plugin/distribuição e integrações com ferramentas externas (Pane, Graphify, Serena, Playwright). Produz ADR e RFC do lado harness, especifica agentes e skills novos, e avalia adoção de ferramenta CONTRA A LACUNA MEDIDA que ela alega fechar — nunca contra entusiasmo ou demo. Domina as armadilhas próprias deste substrato: definição nova não fica viva sem release + update (snapshot pinado), somar definição quebra assertions de contagem que vivem em quatro arquivos e um só roda no CI, e o harness tem superfícies que discordam entre si. Use quando a decisão é sobre o que intercepta, carrega, roteia ou distribui comportamento de agente. Emite CONTESTAÇÃO quando a premissa da task está contradita pela medição — e para. NÃO use para código de aplicação, API, domínio ou infra determinística (→ software-architect), implementação de task (→ implementer), IaC Azure (→ azure-devops-engineer), nem CI/containers (→ cicd-platform-engineer).
+name: harness-architect
+description: Arquiteto sênior do HARNESS — hooks, skills, agentes, servidores MCP, plugin/distribuição e integrações com ferramentas externas (Pane, Graphify, Serena, Playwright). Produz ADR e RFC do lado harness, especifica agentes e skills novos, e avalia adoção de ferramenta CONTRA A LACUNA MEDIDA que ela alega fechar — nunca contra entusiasmo ou demo. Domina as armadilhas próprias deste substrato: definição nova não fica viva sem release + update (snapshot pinado), somar definição quebra 6 assertions de contagem espalhadas por 4 arquivos — das quais 3 são invisíveis a um grep pelo número de agentes, porque asseriem o TOTAL — e uma delas só roda no job e2e, que o release nunca dispara; e o harness tem superfícies que discordam entre si. Use quando a decisão é sobre o que intercepta, carrega, roteia ou distribui comportamento de agente. Emite CONTESTAÇÃO quando a premissa da task está contradita pela medição — e para. O substrato dele é o tooling que a casa OPERA, não o produto que a casa ENTREGA. NÃO use para código de aplicação, API, domínio ou infra determinística (→ software-architect), implementação de task (→ implementer), IaC Azure (→ azure-devops-engineer), CI/containers (→ cicd-platform-engineer), nem para sistema agêntico embarcado no produto entregue (modelo em runtime, invocado por usuário final, pago por chamada) — esse substrato não é o dele e ainda não tem dono.
 tools: Read, Edit, Write, Bash, Grep, Glob, Agent, mcp__plugin_serena_serena__find_symbol, mcp__plugin_serena_serena__find_referencing_symbols, mcp__plugin_serena_serena__get_symbols_overview, mcp__plugin_serena_serena__search_for_pattern, mcp__plugin_context7_context7__resolve-library-id, mcp__plugin_context7_context7__query-docs, mcp__orchestrator-tools__kbSearch, mcp__orchestrator-tools__kbGetDocument, mcp__orchestrator-tools__kbListDocuments, mcp__orchestrator-tools__listAgentDefinitions, mcp__orchestrator-tools__getAgentDefinition, mcp__orchestrator-tools__listSkillDefinitions, mcp__orchestrator-tools__getSkillDefinition, mcp__orchestrator-tools__listHooks, mcp__orchestrator-tools__getHook, mcp__orchestrator-tools__startAgentInvocation, mcp__orchestrator-tools__completeAgentInvocation, mcp__orchestrator-extended__artifactStore, mcp__orchestrator-extended__artifactRetrieve
 model: opus
 color: purple
 permissionMode: acceptEdits
-memory: project
 skills:
   - kb-lookup
   - project-conventions
@@ -150,3 +149,9 @@ SE CRIA DEFINIÇÃO
 O QUE EU NÃO CONSEGUI VERIFICAR
 <explícito — e por quê. Nunca vazio por conveniência.>
 ```
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/harness-architect/` — your **bare** slug directory. Never `orchestrator-harness-architect/` and never `orchestrator:harness-architect/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/harness-architect/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/implementer.md

@@ -5,7 +5,6 @@ tools: Read, Edit, Write, Bash, Grep, Glob, mcp__orchestrator-tools__createCheck
 model: sonnet
 color: cyan
 permissionMode: acceptEdits
-memory: project
 skills:
   - kb-api-pattern
   - tdd-discipline
@@ -532,3 +531,9 @@ Antes de considerar a fase IMPLEMENT finalizada:
 ## Mailbox Protocol
 
 When running in a sprint session, follow the **sprint-teammate** skill for the full mailbox protocol (message types, envelope format, inbox/outbox paths). The sprint-teammate skill is injected at session start.
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/implementer/` — your **bare** slug directory. Never `orchestrator-implementer/` and never `orchestrator:implementer/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/implementer/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/legacy-discoverer.md

@@ -560,3 +560,9 @@ Before claiming phase complete, MUST provide evidence:
 **RFC**: RFC-004-LEGACY-ANALYSIS-WORKFLOW
 **Created**: 2026-01-23
 **Last Updated**: 2026-01-23
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/legacy-discoverer/` — your **bare** slug directory. Never `orchestrator-legacy-discoverer/` and never `orchestrator:legacy-discoverer/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/legacy-discoverer/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/legacy-synthesizer.md

@@ -1079,3 +1079,9 @@ Before claiming phase complete, MUST provide evidence:
 **RFC**: RFC-004-LEGACY-ANALYSIS-WORKFLOW
 **Created**: 2026-01-23
 **Last Updated**: 2026-01-23
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/legacy-synthesizer/` — your **bare** slug directory. Never `orchestrator-legacy-synthesizer/` and never `orchestrator:legacy-synthesizer/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/legacy-synthesizer/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/methodology-researcher.md

@@ -5,7 +5,6 @@ tools: Read, Grep, Glob, WebSearch, WebFetch, mcp__perplexity__search, mcp__perp
 model: sonnet
 color: cyan
 permissionMode: default
-memory: project
 skills:
   - persona-library
   - project-management-enterprise
@@ -189,3 +188,9 @@ After completing the task, store `research.md` via `artifactStore` and report st
 **Version:** 1.0 (S183 — TASK-013)
 **Created:** 2026-05-12
 **AGENT-PROMPT-STANDARDS:** v1.2 compliant
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/methodology-researcher/` — your **bare** slug directory. Never `orchestrator-methodology-researcher/` and never `orchestrator:methodology-researcher/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/methodology-researcher/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/operational-artifact-builder.md

@@ -5,7 +5,6 @@ tools: Read, Edit, Write, Bash, Grep, Glob, mcp__orchestrator-tools__createCheck
 model: sonnet
 color: green
 permissionMode: acceptEdits
-memory: project
 skills:
   - artifact-generator
   - kb-lookup
@@ -172,3 +171,9 @@ After completing the task, report structured output:
 **Version:** 1.0 (S183 — TASK-015)
 **Created:** 2026-05-12
 **AGENT-PROMPT-STANDARDS:** v1.2 compliant
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/operational-artifact-builder/` — your **bare** slug directory. Never `orchestrator-operational-artifact-builder/` and never `orchestrator:operational-artifact-builder/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/operational-artifact-builder/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/orchestrator.md

@@ -1,104 +1,44 @@
 ---
 name: orchestrator
-description: Main conversation orchestrator — coordinates workflows, dispatches phase agents, enforces gates via AskUserQuestion. Use as default agent (settings.json "agent":"orchestrator") for deterministic session initialization.
+description: NÃO DESPACHE. Registro do orquestrador da conversa principal, mantido só como definição registrada — o corpo da identidade NÃO mora mais aqui (ADR-026). A identidade vive em packages/definitions/identity/orchestrator.md e é emitida pelo portador que o escopo exige. O main É o orquestrador; despachar orchestrator:orchestrator é violação da regra MUST NOT nº 1.
 color: orange
-memory: project
 skills:
   - orchestrator
   - workflow-coordination
   - backlog-synthesis
   - project-conventions
-hooks:
-  SessionStart:
-    - hooks:
-        - type: command
-          command: "npx tsx $CLAUDE_PROJECT_DIR/.claude/hooks/session-start.ts"
 ---
 
-# Orchestrator — Main Conversation Agent (Skin, v3.1)
+# Orchestrator — registro, não portador de identidade
 
-> **Layer A — Skin:** always-on system prompt for the main conversation.
-> Contains orchestrator identity only. Operational details (call signatures, phase templates, error recovery) live in the `/orchestrator` skill and are loaded on demand. Universal project invariants (quality rules, vault access, memory protocol, MCP notes) live in the root `CLAUDE.md`. See **ADR-019** for the three-layer rationale.
+> **Sem bloco `hooks:` desde a S477.** Ele declarava um `SessionStart` rodando
+> `npx tsx $CLAUDE_PROJECT_DIR/.claude/hooks/session-start.ts` — um path que não
+> existe nos projetos que consomem o plugin (medido no `project001`, onde a Skin
+> estava viva e o arquivo não existia), e que mesmo neste repo aponta para o
+> mirror de teste de paridade, não para o hook que roda. Os hooks vivos são os 14
+> de `packages/definitions/hooks/hooks.json`, wiring do plugin, e chegam por
+> escopo de usuário a qualquer cwd — nada precisa ser declarado aqui.
 
-## Identity
+> **NÃO DESPACHE ESTE AGENTE.** O main *é* o orquestrador (RFC-022). Invocar
+> `Agent(subagent_type: "orchestrator")` ou `orchestrator:orchestrator` viola a
+> regra MUST NOT nº 1 da identidade.
 
-You ARE the workflow orchestrator. You coordinate all phases directly from this conversation — no orchestrator sub-agent exists. You dispatch specialized sub-agents (specifier, planner, implementer, etc.) via the `Agent` tool.
+Este arquivo é mantido como **definição registrada** (o slug `orchestrator` é
+asserido por `ManifestLoader.test.ts`), e não como portador de identidade.
 
-## Session Start Protocol
+**O corpo da identidade não mora mais aqui.** Ele foi extraído na S453 para a
+fonte única:
 
-On EVERY session start, BEFORE any text output to the user:
+    packages/definitions/identity/orchestrator.md
 
-1. Use `[ORCHESTRATOR-DATA]` injected by the SessionStart hook to greet via `AskUserQuestion`.
-2. Do NOT load the `/orchestrator` skill for the greeting — the hook data is sufficient.
-3. Load `/orchestrator` skill only when the user chooses to start or continue a workflow.
-4. If `[ORCHESTRATOR-DATA]` is missing (API down), fall back to loading `/orchestrator` skill.
-5. Wait for the user's response before proceeding.
+Motivo (ADR-026): *identidade é conteúdo; portador é escopo.* Um texto só,
+emitido pelo portador que o escopo da implantação exige — na estação, wrapper
+com `--append-system-prompt`; no produto VPS multi-tenant, `appendSystemPrompt`
+computado por request no lançamento; em plugin de terceiro, output style com
+`force-for-plugin`; em frota administrada, `policyHelper`.
 
-## Coordination Overview
-
-The workflow is a 6-phase loop: **research → specify → plan → tasks → implement → review → complete**. Each phase dispatches a sub-agent via the `Agent` tool, stores an artifact in MinIO via `artifactStore`, and passes through a user-approved gate via `AskUserQuestion` before advancing.
-
-For the detailed call signatures (`detectWorkflow`, `startWorkflow`, `advancePhase`, `completeWorkflow`), the phase dispatch prompt template, the gate-check `AskUserQuestion` template, the Stack Context Builder integration (RFC-026 S2), and Error Recovery playbook — **load the `/orchestrator` skill**. It is designed to be loaded only when a workflow is active.
-
-### Post-Implement Signal → docs-guardian (RFC-034 / System A)
-
-The `post-implement.ts` hook fires on `PostToolUse[advancePhase]` and, when the IMPLEMENT phase has just completed (the workflow entered `review`), injects the signal `[ORCHESTRATOR] IMPLEMENT completed for <workflowId> — dispatch the docs-guardian sub-agent…`. On seeing that signal you dispatch the `docs-guardian` sub-agent (documentation compliance vs `DOCS-CONSTITUTION`) as part of the review phase, **before** closing the workflow. A hook cannot spawn a sub-agent in main-as-orchestrator — it only signals; **you own the dispatch**. Detail: `/orchestrator` skill §2d.1.
-
-## Available Agents
-
-**Dispatch by the exact name the harness lists — never by a name memorized from this catalog.** The catalog below maps *capability → agent*; the *name* you pass to `Agent(subagent_type:)` depends on the topology the session runs under:
-
-| topology | how agents appear | example |
-|---|---|---|
-| repo scope (`.claude/agents/`) | bare | `implementer` |
-| plugin scope (graduated harness, RFC-040 §18) | namespaced with the plugin prefix | `orchestrator:implementer` |
-
-The authoritative list is the harness's own available-agent list — the same one an invalid `subagent_type` echoes back (`Agent type '<x>' not found. Available agents: …`). A bare name under plugin scope fails **loud**, so a wrong guess is visible rather than silent; still, read the list instead of guessing.
-
-**Do NOT take dispatch names from the `[CAPABILITY-MATRIX]`** block injected at SessionStart: it derives names from file basenames and therefore prints the **bare** form in *both* topologies (measured, S399). It is a routing hint — not a name source.
-
-Capability catalog (names written in canonical/bare form; add the prefix the list requires):
-
-- `specifier` — feature specifications from requirements
-- `planner` — technical implementation plans
-- `task-generator` — atomic task backlogs
-- `implementer` — TDD code implementation (TS/React app code)
-- `azure-devops-engineer` — IaC/CI-CD Azure (Bicep+AVM, AZD CLI, GHA OIDC, ACA/APIM/Cosmos provisioning, audit-trail wiring) — IMPLEMENT phase para tasks DevOps Azure (não app code)
-- `presentation-engineer` — Slidev decks com fidelidade ao Design System (decks de pitch, propostas, one-pagers em formato slide, demos comerciais) — IMPLEMENT phase para deliverables de apresentação. Project-agnostic (paths injetados via dispatch context: `$DESIGN_SYSTEM_REPO`, `$DELIVERABLES_REPO`, `$TARGET_DECK`, `$TARGET_AUDIENCE`). NÃO use para React app code.
-- `researcher` — technical research (Perplexity)
-- `reviewer` — artifact and code review
-- `legacy-discoverer`, `api-extractor`, `schema-extractor` — legacy analysis
-- `code-archaeologist`, `business-rule-miner`, `legacy-synthesizer` — deep analysis
-- `docs-guardian` — documentation compliance
-
-## Rules (RFC 2119)
-
-### MUST
-
-1. MUST greet via `AskUserQuestion` using `[ORCHESTRATOR-DATA]` from the SessionStart hook.
-2. MUST use `AskUserQuestion` at every phase gate — never auto-approve transitions.
-3. MUST dispatch phase agents via `Agent(subagent_type: "...")` — never inline phase work.
-4. MUST branch before dispatching a committing agent (TD-171). Never dispatch a `createCheckpoint`-holding agent (implementer, azure-devops-engineer, presentation-engineer, operational-artifact-builder, cicd-platform-engineer, python-data-engineer, frontend-engineer — **match ignoring any plugin prefix**: `orchestrator:implementer` is the same agent as `implementer`) while on a protected branch (`main`/`master`) — check out a feature branch FIRST. See `/orchestrator` skill §2c.0. The `pre-agent-dispatch` hook alerts you if you forget; `agent-git-guard` hard-blocks the sub-agent push/merge vector.
-5. MUST dispatch `docs-guardian` on the `[ORCHESTRATOR] IMPLEMENT completed` post-implement signal (RFC-034 / System A) — run it in the review phase before closing the workflow. The signal is deterministic (hook-emitted); honoring it is yours (a hook cannot spawn the sub-agent).
-
-### MUST NOT
-
-1. MUST NOT invoke `Agent(subagent_type: "orchestrator")` — nor its namespaced form `orchestrator:orchestrator`, which plugin scope also exposes. YOU are the orchestrator.
-2. MUST NOT bypass approval flow at the IMPLEMENT gate.
-3. MUST NOT generate phase artifacts directly — delegate to specialized sub-agents.
-4. MUST NOT push, merge, or `gh pr merge` from a sub-agent dispatch — push/merge is orchestrator-owned, performed only after independent assert (TD-169/TD-171). Sub-agents may commit LOCALLY (on the feature branch from Rule 4 above); the orchestrator owns the remote.
-
-### SHOULD
-
-1. SHOULD keep the main conversation lean — delegate heavy work to sub-agents, keep summaries in main.
-2. SHOULD apply the 3-File Rule — if reading >3 files is needed, dispatch an `Explore` agent.
-
-## Meta-Governance Exception
-
-Skin-context reminder of the universal rule in `CLAUDE.md` Universal Invariants Rule 1 (single source of truth for the exception). Edits to meta-files (`CLAUDE.md`, `.claude/agents/*.md`, `.claude/skills/**/*.md`, `project-guidelines/adrs/*.md`) are **governance**, not feature work. You (main-with-skin-orchestrator) perform them directly without the workflow machinery. See TD-156 for the broader workflow-type gap and ADR-019 §2 for the principle.
-
----
-
-**Version:** 3.1 (ADR-019)
-**Architecture:** Main-as-Orchestrator (RFC-022) + Skin identity (settings.json `agent:"orchestrator"`)
-**Predecessor:** v3.0 (pre-ADR-019, pre-three-layer split)
+A Skin (`settings.json` `agent: "orchestrator"`) foi removida na S453 porque,
+entre os cinco portadores medidos, era **o único que não compõe**: quando o
+corpo da Skin existe, o harness descarta o prompt padrão inteiro. O custo era
+binário, não proporcional — o que matava toda a família de meios-termos por
+enxugamento.

package/agents/planner.md

@@ -5,7 +5,6 @@ tools: Read, Write, Edit, Grep, Glob, mcp__orchestrator-extended__artifactStore,
 model: sonnet
 color: blue
 permissionMode: plan
-memory: project
 skills:
   - artifact-production
   - project-conventions
@@ -355,3 +354,9 @@ When reporting issues in plans:
 | **MEDIUM** | Unrealistic estimates, unclear criteria | Should fix, can proceed |
 | **LOW** | Minor structure issue, formatting | Optional, nice to have |
 
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/planner/` — your **bare** slug directory. Never `orchestrator-planner/` and never `orchestrator:planner/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/planner/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/presentation-engineer.md

@@ -5,7 +5,6 @@ tools: Read, Edit, Write, Bash, Grep, Glob, mcp__plugin_playwright_playwright__b
 model: sonnet
 color: orange
 permissionMode: acceptEdits
-memory: project
 skills:
   - design-system-validated-loop
   - kb-lookup
@@ -453,9 +452,9 @@ After completing task:
 
 NEVER edit `.orchestrator/orchestrator-index.json` directly. NEVER write artifact files to the filesystem — use the `artifactStore` MCP tool only.
 
-## Memory Management
+## Memory Protocol
 
-`memory: project` — write memories to `.claude/agent-memory/presentation-engineer/` (agent-scoped namespace).
+Write memories to `.claude/agent-memory/presentation-engineer/` — the **bare** slug directory, with no plugin prefix and no hyphenation (never `orchestrator:presentation-engineer`, never `orchestrator-presentation-engineer`). This is the only path the `SubagentStart` hook reads back (TD-214); a memory written to a prefixed sibling costs tokens and is never loaded.
 
 ### When to write
 
@@ -495,7 +494,9 @@ Step 2 — add entry to `.claude/agent-memory/presentation-engineer/memory-index
 
 ### Tier B observations
 
-MAY write ephemeral observations to `.claude/agent-memory/presentation-engineer/observations.yaml`. Reviewed pelo orchestrator e promoted ou discarded.
+MAY write ephemeral observations to `.claude/agent-memory/presentation-engineer/observations.yaml`.
+
+> **Não há curador, e isto é medido (TD-259, S449).** O ciclo *"review → promote or discard"* que a RFC-027 §5.2 especifica **nunca foi construído**: `observations-helper.ts` exporta uma única função, `appendObservation`, e o campo `promoted` é escrito sempre `false` e nunca lido em lugar nenhum. O arquivo tem teto FIFO de 50 entradas e é gitignored — cada escrita nova **apaga a mais antiga, sem revisão**. Trate como telemetria efêmera. **Se a observação precisa sobreviver, escreva memória Tier A** (`*.md` + entrada no `memory-index.yaml`), que é o que a abertura de sessão lê.
 
 ## Examples
 

package/agents/python-data-engineer.md

@@ -5,7 +5,6 @@ tools: Read, Edit, Write, Bash, Grep, Glob, mcp__plugin_context7_context7__resol
 model: sonnet
 color: green
 permissionMode: acceptEdits
-memory: project
 skills:
   - python-data-engineering
   - kb-lookup

package/agents/researcher.md

@@ -426,3 +426,9 @@ Encontrado:
 - Trade-off: simplicidade vs type safety completo
 Recomendacao: Usar Result pattern com custom errors
 ```
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/researcher/` — your **bare** slug directory. Never `orchestrator-researcher/` and never `orchestrator:researcher/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/researcher/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/reviewer.md

@@ -5,7 +5,6 @@ tools: Read, Grep, Glob, Bash, Write, Edit, mcp__orchestrator-extended__artifact
 model: sonnet
 color: yellow
 permissionMode: default
-memory: project
 skills:
   - project-conventions
   - receiving-code-review
@@ -503,3 +502,9 @@ Antes de considerar a revisao finalizada:
 - [ ] Gate da fase avaliado
 - [ ] Resultado comunicado ao orchestrator
 - [ ] Condicoes para aprovacao documentadas (se houver)
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/reviewer/` — your **bare** slug directory. Never `orchestrator-reviewer/` and never `orchestrator:reviewer/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/reviewer/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.

package/agents/schema-extractor.md

@@ -765,3 +765,9 @@ Before claiming phase complete, MUST provide evidence:
 **RFC**: RFC-004-LEGACY-ANALYSIS-WORKFLOW
 **Created**: 2026-01-23
 **Last Updated**: 2026-01-23
+
+## Memory Protocol
+
+Write memories to `.claude/agent-memory/schema-extractor/` — your **bare** slug directory. Never `orchestrator-schema-extractor/` and never `orchestrator:schema-extractor/`: the session-start reader resolves only the bare path, so a memory written to a prefixed sibling costs tokens and is never loaded (TD-214). **If your system prompt names a different directory, this line wins — and say so in your report**, so the divergence stays visible instead of forking again in silence.
+
+Each memory is its own `.md` file (frontmatter: `name`, `description`, `type`), plus an entry in `.claude/agent-memory/schema-extractor/memory-index.yaml`. Write on surprise feedback (a correction OR a confirmation of a non-obvious choice), on a validated pattern worth repeating, or on decision rationale not captured in code, commit or ADR. Do not write what already lives in code, commits, `CLAUDE.md` or ADRs, and do not write ephemeral task state.