refactor(templates): reduce strands-http-python to a barebones agent

[Hweinstock]

Summary

Targets refactor. The base work this depends on has landed on refactor: aws#2170 (template rename + export-harness-python) and #20 (feat/wire-model-providers, non-Bedrock model providers on the scaffold path). This PR rebases directly onto refactor and shows only its own diff.

The base leaves agent-python-strands — the template the scaffold path (agentcore project create / add runtime) renders — rich (~715-line main.py, plus mcp_client/, skills/, model/mantle_compat.py), carrying dozens of flags no scaffold ever sets. The harness-export template was already split out into templates/export-harness-python (by aws#2170).

This PR makes agent-python-strands a barebones Strands agent:

1. main.py → ~130 lines: a single tool, a per-session_id in-process Agent cache, prompt extraction, and the stream entrypoint. Dropped every export-only flag (tools, execution limits, truncation, gateways, remote MCP, mounts, config bundle, payments, Mantle).
2. Deleted mcp_client/, skills/, model/mantle_compat.py from the scaffold template (they live only in export-harness-python now).
3. Multi-provider model rendering is retained and now live — model/load.py branches on Bedrock/Anthropic/OpenAI/Gemini/LiteLLM, and pyproject.toml pulls provider SDKs via strands-agents[<provider>] extras. With feat(runtime): support Anthropic, OpenAI, and Gemini model providers #20's provider wiring on refactor, the barebones template renders correct clients + AgentCore Identity wiring for non-Bedrock providers.
4. Trimmed the scaffold render context in runtime.ts to the keys the barebones template uses.

Net: +80 / −1170 lines across 11 files. export harness is untouched (it renders export-harness-python, provided by aws#2170).

No-memory session retention

An agent without memory reuses one Agent per session_id (in-process LRU cache), so successive invokes on the same session keep their conversation history within a live process — the same behavior the previous template had. This is best-effort (the cache resets on cold start / isn't shared across instances); for durable, cross-instance history, add memory (the session manager), preserved here.

Spec

Problem: The strands python template has too many parameters on the refactor branch, causing the template to be difficult to read, extend, and understand.

Definition of Done:

• the template now describes a barebones strands agent.
• the only rendering left is name, memory, and model provider. The scaffolded is dead-simple. All other flags and templates are dropped. Memory should still use our session manager.
• we still support codezip and containers.
• an agent without memory should still remember previous invokes, when invoked with the same sessionId. That is, we don't create a new agent if one already exists for that session.

Verification: Create a project with a CodeZip runtime, add a Container runtime, run the dev server for each and verify local invocation, deploy to AWS and invoke via the CLI — across all 3 memory options (combined into one project).

Note on the "no-memory context" item: the spec suspected a bug where a no-memory agent lost context on the same sessionId. Investigation showed the previous template already reused one Agent per session in-process, so a warm deployed runtime does remember across same-session invokes (confirmed on AWS). It's only best-effort (resets on cold start / not shared across instances) — not a defect worth a code change. This PR keeps that same session-reuse approach and drops the "bug" framing.

Follow-ups (from review): (a) export harness kept working — the base's export-harness-python template serves it, so this PR only barebones the scaffold template; (b) alternative model providers kept in the template — now wired end-to-end since #20 landed on refactor; (c) export template stays out of scaffolding — it's structurally unreachable (no resolver key or --template enum value maps to it; only export harness renders it); (d) rebased onto refactor; (e) session-reuse matches the previous template (no bug-fix claim).

Verification

Re-run after the aws#2170 rebase, against dev account 440744214761 / us-east-1, using the compiled binary (bun run compile).

Unit tests / build

• bun run typecheck → clean; oxlint / prettier --check → clean
• bun test → 2798 pass, 0 fail (incl. feat(runtime): support Anthropic, OpenAI, and Gemini model providers #20's non-Bedrock provider tests; snapshot updated: scaffold no longer emits mcp_client/, skills/, model/mantle_compat.py)
• Non-Bedrock scaffolding (live, stacked on feat(runtime): support Anthropic, OpenAI, and Gemini model providers #20): project create --model-provider open_ai (and gemini, anthropic) renders valid Python — e.g. OpenAI → OpenAIModel(model_id="gpt-4.1") with @requires_api_key Identity wiring; deps resolve to strands-agents[openai] / [gemini] / [anthropic] / plain for Bedrock. All py_compile-clean.

Scaffold (both build types × all 3 memory options)

One project, 6 runtimes: CodeZip + Container, each with memory = none | shortTerm | longAndShortTerm. All generated Python compiles.

Local dev (agentcore project dev --mode headless)

• CodeZip, no memory — same session: told "Bob / teal" → asked → "Your name is Bob and your favorite color is teal."
• Container, no memory — "21 + 21? Use your tool." → "The result of 21 + 21 is 42."; same-session codeword → "The codeword you gave me is ORANGE."

Deployed on AWS (agentcore project deploy + project invoke runtime)

All 6 runtimes invoked via the AWS InvokeAgentRuntime API; all retained context within a warm session ("Your name is Bob and your favorite fruit is mango.").

Durability contrast (new session id, same --user-id):

• long-term memory → "your favorite fruit is mango" (recalled across sessions via AgentCore Memory)
• no memory → "I truly do not know your favorite fruit… Each conversation I have starts fresh without memory of previous interactions."

How to reproduce

BIN=./dist/bin/agentcore-linux-x64            # from: bun run compile
D=/tmp/demo && rm -rf $D && mkdir -p $D && cd $D

$BIN project create --name demo --template agent-python-strands \
  --build CodeZip --memory none --runtime-name agentNone
cd demo
$BIN project add runtime --framework strands --language Python --build Container \
  --model-provider Bedrock --memory longAndShortTerm --name agentCtrLong

# local dev + invoke (headless binds :8080)
$BIN project dev --mode headless --agent agentNone &
curl -sN -X POST http://127.0.0.1:8080/invocations \
  -H 'Content-Type: application/json' \
  -H 'x-amzn-bedrock-agentcore-runtime-session-id: sess-000000000000000000000000000001' \
  -d '{"prompt":"My name is Bob. Remember it."}'
# invoke again on the SAME session → the warm process remembers (in-process per-session reuse)

# deploy + invoke via AWS
$BIN project deploy
$BIN project invoke runtime --name agentNone --session-id sess-... --payload '{"prompt":"Hello!"}'

Teardown

aws cloudformation delete-stack --stack-name AgentCore-demo-default --region us-east-1

[Hweinstock]

you can remove this block.

[Hweinstock]

delete this test, it should exist in the handlers not here.

[Hweinstock]

this test looks uselsss, remove it.

[Hweinstock]

imo templating in params here for litellm is unnecessary complexity. If a customer wants to user liteLLM I think its easier to take this default as the scaffolding and adjust from there rather than passing on command line.