feat(modal): per-session custom image + honeycomb-compat Railway deploy hand-off

[listar2000]

What & why

A hand-off branch for standing up a standalone agent-sdk server on Railway dedicated to honeycomb
seed/task generation (isolated from the shared staging/production servers, which serve other clients).
Branched off current main (774ade1).

1. feat(modal): per-session custom image (ec3b65d)

Boot a non-default image on the modal supervisor path via a per-session image on the session-create
recipe: Recipe.image → _sessions_create_eager → ModalSandboxSession._create_kwargs →
_get_image(image, dockerfile), resolving an im-... filesystem-snapshot id (Image.from_id), a
docker registry ref (Image.from_registry), or a materialized dockerfile (Image.from_dockerfile),
never memoizing the per-session image into the shared module global. Adds the matching image= kwarg
to the Agent client and to _CLONABLE_FIELDS.

Only the modal provider honors it; the daytona/default path is unaffected (image defaults to
None). This was previously carried as an uncommitted local patch on the box running honeycomb's
generation server — it powers honeycomb's --profile aws LocalStack generation (a combined
LocalStack + agent-sdk-runtime image). Committing it upstream so the deploy branch is self-contained.
The image field 3-way auto-merged cleanly alongside main's new no_reap recipe field.

2. Deploy hand-off docs (c3d1b66)

• HANDOFF.md — Railway deploy runbook: required/forbidden env, the per-Daytona-org snapshot decision,
  post-deploy verification, the known long-turn stream-drop caveat (Railway edge cap / RAM) + its
  client-side durable fix, a no-app-auth security note, and the client→server contract to preserve.
• .env.example — env template (DATABASE_URL + DAYTONA_API_KEY required; model creds must NOT be
  set — they're stripped server-side and forwarded per-session by the client).

3. fix(usage): codex cached/thought tokens on the done frame (bdaf04c)

Codex reports token usage on its terminal PromptResponse (camelCase cachedReadTokens/thoughtTokens),
not via a usage_update event, so UsageStats silently dropped a codex run's cache-read and reasoning
tokens. Accumulate both additively (claude snake_case variants tolerated), and add
_meta.quota.token_count to the SSE done-usage probe paths so the client can reconstruct token counts
even on subscription auth (where total_cost_usd stays 0). The client-side field additions are
back-compat (default 0); the SSE probe change is server-side and takes effect on the next server build.

Notes

• Daytona generation needs none of the modal code (it's dormant there); the branch is deployable
  as-is for the daytona path.
• .runtime-snapshot-tag is unchanged (agent-sdk-46606ca) — no snapshot rebuild implied.
• No railway.toml / Dockerfile changes (single-worker session-affinity pin kept intentionally).

🤖 Generated with Claude Code

https://claude.ai/code/session_01EzkMPi5tAr3e4T5HBQ14Dt

[Chanbinski2]

Thanks for the PR @listar2000, two things:

1. Since we want to create a separate server of agent-sdk for honeycomb, can you create this PR to point to another branch, like 'honeycomb' for example?
2. If your changes regarding modal is just for honeycomb, I think the commits here can just move to the 'honeycomb' branch?

Let me know.

[Chanbinski2]

Also if this is all you need for honeycomb, I can just create a new branch and then set up the credentials on railway. + I'm not sure if railway has a path for codex oauth, so we might need to add that. Let me know what works best for you.

[listar2000]

@Chanbinski2 Thanks for the help. I will work locally to get two more features in before getting this to a standalone branch (the current merge to main is just a marker).

• Daytona support for running oracle AWS task image
• Codex OAuth support for running evaluations