Merge develop into main: GT-603 agent turn ledger - #83
Merged
Conversation
…ico (GT-603) El ledger de turnos agénticos estaba escrito, probado y SIN REGISTRAR: `IAgentExecutionPort` no aparecía en ningún contenedor ni en ningún endpoint, y `/assistant/converse` —la única vía por la que un agente tocaba el corpus de un cliente— proxyaba al runtime sin persistir nada. La afirmación diferencial del producto (distinguir un acto humano de uno agéntico en el expediente) era falsa en el código. Y era la ficha que CADUCA: `audit_entries` es append-only por trigger desde COH-014, así que toda fila escrita antes de que existiera el discriminador queda inatribuible para siempre — el UPDATE que la tiparía está prohibido por la base. - `AuditActor`/`AuditActorType`: human | agent | system, más `unknown` reservado a lo anterior a la migración y EXCLUIDO de toda ruta de escritura. - `AuditEntry.Create` exige la atribución como parámetro obligatorio: los 8 sitios que escriben en el expediente declaran ahora quién escribe. Se añade `Rehydrate` para poder LEER `unknown` sin poder escribirlo. - Migración `AddAuditActorAttribution`: `actor_type` NOT NULL + `agent_id`, `model_id`, `session_id`. El DEFAULT `'unknown'` vive sólo dentro del `ADD COLUMN` (fast default: sin reescritura, sin disparar el trigger) y se RETIRA acto seguido, para que ninguna fila nueva entre sin declarar su actor. Los tres triggers append-only quedan intactos. - `AssistantEndpoints` pasa por el puerto: scope comprobado, turno asentado y sólo entonces se deja actuar al runtime. La respuesta se devuelve intacta. - `AgentTurnAuditor` CONFIRMA la escritura (`SaveEntitiesAsync`): sin ella, «el turno se pudo registrar» era una afirmación sobre memoria que nadie persistía. - Un turno sin identidad de agente se rechaza: no es atribuible, y en un registro que no admite corrección es preferible no escribirlo a escribirlo mal. - `audit-trail.robot.mjs` verifica extremo a extremo, contra la API viva, que un turno agéntico aterriza tipado como `agent` con su agente identificado y que un acto humano en el MISMO expediente sigue diciendo `human`. Verificado: 1034 pruebas + 14 de arquitectura en verde; las pruebas vivas de esquema contra PostgreSQL real (EVOLITH_CORE_LIVE=1) y RoboSoft `audit-trail` 25/25 contra tracker-api en marcha. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ment never installs `Deploy check (kind + Helm + smoke)` was the only red job on this branch, and the only red assertion in it was `POST /assistant/converse responds 2xx` — status 502. Every ledger assertion around it passed: the turn is in the trail, typed `agent`, with the acting agent identified, the human principal preserved, the scope declared and the prompt absent. That is not a flake and not a regression. `AssistantEndpoints` seats the turn through `IAgentExecutionPort` BEFORE the runtime is allowed to act, then proxies; the chart wires `AgentRuntime__BaseUrl` to `evolith-agent-runtime`, a service the kind smoke does not install, so the proxy leg returns 502 `AgentRuntime.Failed` while the seating — the thing this scenario exists to prove — succeeds. So the check asserted the availability of an absent dependency. It now asserts what the deployment can actually answer for: the turn was seated. 2xx when a runtime replies, 502 when none is deployed, and still red on 4xx (scope or attribution refused, no turn seated) or 503 (`AgentExecution.NotAuditable` / `AgentTurnAuditor.*` — the ledger refused the entry). Not fixed by enabling `MockFallback`: that would turn the ledger assertions green against a stub reply. `smoke-app.sh` pins `CoreApi__MockFallback=false` for the same reason. Installing the runtime in the smoke environment is the real closure and belongs with the deployment, not with this gap. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…za (GT-616) y ledger (GT-603) El PR estaba DIRTY, y por eso GitHub no ejecutaba ni un check: no podía calcular el merge. El conflicto es real, no textual. Mientras esta rama estaba abierta, `develop` recibió GT-616 y le puso una traza de gobernanza al mismo endpoint — pero sobre la vía ANTIGUA, `gateway.ConverseAsync`, que es exactamente la que GT-603 sustituye por `IAgentExecutionPort`. Quedarse con un lado perdía algo real: el de develop, el asiento en el expediente; el de la rama, la traza. Se conservan los dos. La traza envuelve al PUERTO, no al gateway. Desde GT-603 el turno se asienta antes de que el runtime actúe, así que una traza que empezara después no vería el único tramo capaz de rechazar el turno — que es el que hay que poder buscar. Las tres salidas quedan etiquetadas: `ex.Code` en el rechazo de perímetro, `resultado.Error` SIN traducir cuando falla el puerto, y `ok` en el éxito. El error sin traducir es deliberado: `AgentExecution.NotAuditable` y `AgentRuntime.Failed` salen ambos como 5xx, pero en la traza son dos historias distintas —el expediente rechazó el asiento, o el runtime no respondió— y colapsarlas al código HTTP borraría justo la que hay que poder distinguir. Verificado: `Tracker.Presentation` compila con 0 avisos y 245 tests de Assistant/AgentExecution/Tracing/Governance en verde. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
GT-603: the audit ledger can finally say whether a human or an agent did it
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Promotes
developtomainafter the branch-hygiene pass on this satellite.What lands
#81 — GT-603: the audit ledger can say whether a human or an agent acted.
/assistant/conversewas the only path a non-human executor touched a customer's corpus through, and it persisted nothing — so the product's differentiating claim was false in code. The turn now goes throughIAgentExecutionPort, which checks the scope, seats the entry and only then lets the runtime act.Two commits were added on top to get the PR mergeable:
Deploy check (kind + Helm + smoke)was the branch's only red job and its only red assertion wasPOST /assistant/converse responds 2xx— status 502. Every ledger assertion around it passed. The chart wiresAgentRuntime__BaseUrltoevolith-agent-runtime, a service the kind smoke does not install, so the proxy leg fails while the seating succeeds. The check now asserts what the deployment can answer for — that the turn was seated — and stays red on 4xx (scope or attribution refused) or 503 (the ledger refused the entry). Not fixed withMockFallback: that would green the ledger assertions against a stub, andsmoke-app.shpinsCoreApi__MockFallback=falsefor the same reason.DIRTY, which is why GitHub ran no checks at all on it. While the branch was open,developlanded GT-616 and put a governance trace on this same endpoint — over the oldgateway.ConverseAsyncpath that GT-603 replaces. Keeping either side alone lost something real. The trace now wraps the port, not the gateway: the turn is seated before the runtime acts, so a span starting later would miss the only leg that can refuse it. All three exits are tagged, with the port's failure tagged untranslated soAgentExecution.NotAuditableandAgentRuntime.Failedstay distinguishable in traces even though both leave as 5xx.Known red check, pre-existing on this trunk
conformfails, and it fails ondevelopandmaintoo — since97916eb("close contract and capability parity gaps"), which made the Core checkout mandatory and hard-fails withexit 1whenCORE_REPO_TOKENis absent. That secret is not configured on this repository, so every run since 2026-08-01T08:29 is red. It is unrelated to #81 and this merge neither causes nor worsens it.Branch hygiene in the same pass
feat/eag-01-consume-ums-sdkdeleted: its premise is blocked upstream (T-040 — theUms.Sdk.*packages are unpublished, 404 on nuget.org), its merge intodevelopwas reverted twice (7d5eed2,bc20da1), and the substance of EAG-01 is now carried by ADR T-053, whose interim HS256 path is already implemented ondevelop.🤖 Generated with Claude Code