[Bug Bash] Guest operations (read/write/shell/MCP) on a deleted or non-existent environment fail with Unimplemented instead of NotFound

Start in internal/apiservice/wakeretry.go, especially ensureAwake around lines 80–92, and trace how the wake probe handles gRPC status codes and deleting actor state. Reproduce with a deleted or non-existent environment using read, write, shell, or MCP tools/call; finish when these requests report NotFound and deleting actors fail without waiting for the full wake budget.

Summary & Root Cause

Running read, write, shell, or MCP tools/call against a deleted or non-existent environment fails with gRPC code Unimplemented (ate-env: env: rpc error: code = Unimplemented desc = actor ate-env/<id> not found) instead of NotFound (ate-env: env: not found: ...).

Root Cause

1. In internal/apiservice/wakeretry.go:80-89, ensureAwake probes the guest via atenet-router and assumes any error code other than codes.Unavailable came from a live guest:

   _, err := probe.GetProcess(ctx, &ateenvv1alpha.GetProcessRequest{ProcessId: "ate-env-wake-probe"})
   if status.Code(err) != codes.Unavailable {
       // Anything else, NotFound for the probe id included, came from
       // the guest itself.
       return nil
   }
   state, serr := s.actorState(ctx, atespace, id)
   

2. When an actor does not exist (or has finished deleting), atenet-router responds with a plain HTTP 404 Not Found (actor <atespace>/<id> not found) without a grpc-status trailer.
3. Per the gRPC over HTTP/2 specification, grpc-go maps an HTTP 404 response without grpc-status to codes.Unimplemented (12), not codes.Unavailable or codes.NotFound.
4. Because status.Code(err) == codes.Unimplemented != codes.Unavailable, ensureAwake returns nil immediately without calling s.actorState(ctx, atespace, id) (whose doc comment at :122-124 explicitly notes: "A deleted environment surfaces here as NotFound (through toGRPCError), which is the right answer to a caller still holding its handle").
5. The caller's subsequent RPC (ReadFile, WriteFile, StartProcess) then hits the same HTTP 404 from atenet-router and surfaces raw codes.Unimplemented to the client.

Versions

• agent-substrate/env: 40a6c63 (also present on main @ d579fd0)
• agent-substrate/substrate: ddb46651

Steps to Reproduce

Test Case 1: Read from a deleted environment (Bug Bash Flow — Step 9: Delete)

ate-env create $USER
echo "hello $(date)" | ate-env $USER write /data/note.txt

# Step 9 from the bug bash doc:
ate-env delete $USER
ate-env $USER read /data/note.txt

• Expected: Fails with NotFound:

  ate-env: env: not found: environment "chenyiwang" not found
  

• Actual: Fails with codes.Unimplemented (exit code 1):

  ate-env: env: rpc error: code = Unimplemented desc = actor ate-env/chenyiwang not found
  

Test Case 2: Run read, write, or shell on a non-existent environment (Bad Inputs)

ate-env does-not-exist read /data/note.txt
echo "hi" | ate-env does-not-exist write /data/note.txt
ate-env does-not-exist shell 'echo hi'

• Expected: All three commands fail with NotFound (ate-env: env: not found: ...).
• Actual: All three commands fail with codes.Unimplemented (exit code 1):

  ate-env: env: rpc error: code = Unimplemented desc = actor ate-env/does-not-exist not found
  

Test Case 3: Call MCP tools/call on a deleted or non-existent environment

curl -s http://localhost:7777/v1alpha/envs/does-not-exist/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"shell","arguments":{"command":"echo hi"}}}'

• Expected: Error message indicates the environment was not found (not found).
• Actual: Tool result reports code = Unimplemented:

  {"jsonrpc":"2.0","id":1,"result":{"content":[{"type":"text","text":"env: rpc error: code = Unimplemented desc = actor ate-env/does-not-exist not found"}],"isError":true}}
  

Suggested Fix

1. In ensureAwake (internal/apiservice/wakeretry.go:80-89), do not treat codes.Unimplemented from the wake probe as a live guest response (since atenet-router returns HTTP 404 -> codes.Unimplemented when the target actor is unknown, and GetProcess is implemented by ate-env-guest), or fall through to s.actorState(ctx, atespace, id) so a missing/deleted actor surfaces as codes.NotFound.
2. In ensureAwake (internal/apiservice/wakeretry.go:90-92), also fail fast when state == ateapipb.ActorState_ACTOR_STATE_DELETING instead of polling until DefaultWakeBudget (2m) expires.