Real Codex and Claude adapters assert no durable session identity, so #302 reattachment cannot hold

[taras]

What is wrong

The workflow Agent session mapping (#302, specs/workflow-workspace-spec.md §8.5) commits only from a canonical, tagged provider assertion. ACPX supplies that from the ACP session/new response's _meta, reading the first of agentSessionId or sessionId:

const AGENT_SESSION_ID_META_KEYS = ["agentSessionId", "sessionId"];

Neither published adapter puts anything there.

• @agentclientprotocol/codex-acp returns { sessionId, models, modes, …configOptions } from newSession — no _meta.
• @agentclientprotocol/claude-agent-acp likewise returns no session-identity _meta.

So retainedSession(store, placementKey)?.agentSessionId is always undefined, and the profile refuses before the first Prompt:

the provider established an Agent session without asserting a durable identity, so this run cannot record which conversation it is having.

How it was found

Running a real workflow through the compiled dist/xmd against a real (embedded) Codex adapter, while verifying #622. Reproduction:

deno compile --node-modules-dir=none --cached-only --frozen --exclude-unused-npm \
  --allow-all --include packages/code-review-agent --output dist/xmd packages/cli/src/compiled.ts

# a git repo whose flows/agent.md is: <Session name="review"><Prompt>hello</Prompt></Session>
CODEX_PATH=packages/acp/tests/fixtures/fake-codex-app-server.cjs \
XMD_WORKFLOW_RUNS=<tmp>/runs \
  ./dist/xmd workflow start --id=smoke-1 flows/agent.md

The adapter materializes, launches, and completes ACP initialize and session/new. The run then fails at the assertion check above.

Why every test passes today

packages/cli/tests/support/fake-acp.ts fabricates one:

agentSessionId: `agent-session:${input.sessionKey}`

Every WAL, WSR and WSL case substitutes that fake runtime, so the contract has never been exercised against a real adapter. The gap is in the contract's fit to real adapters, not in the retention code.

Scope

This blocks end-to-end portable Agent use: no workflow run can establish a session against a real Codex or Claude adapter, so no Prompt is ever sent and nothing is retained.

It does not belong to #622, whose acceptance is transport and retention of a checkpoint once a Prompt completes, and whose production evidence is compositional (Tier EA for the adapters, the compiled binary for materialization and launch, T2–T6 for retention through XMD). #622's compiled proof may terminate at exactly this refusal.

What a fix has to decide

• whether the identity comes from an upstream adapter change (a _meta.sessionId on session/new, alongside the Retain provider checkpoint tokens for workflow Agent Prompts #622 prompt-response work), or from a different canonical assertion this host accepts;
• what an existing run does when it retained a mapping under one scheme; and
• a regression that exercises the contract against a real adapter rather than a fabricated identity — the absence of which is why this went unnoticed.

Context

• packages/acp/vendor/acpx/generated/live-checkpoint-ClPCSdrW.js — extractAgentSessionId
• packages/cli/src/workflow-agent.ts — assertionsFor, established
• specs/workflow-workspace-spec.md §8.5
• Closed issue Run workflow Agents through XMD-mediated observations #302 owns the contract; this is the narrowly scoped repair under it.