Dev-loop nova backend (AWS Bedrock)

FEAT-405 adds a single nova dev-loop backend that reaches AWS-Bedrock-hosted models through one AWS credential, spanning three seats in the dev-loop flow. This document is the operator reference: what it is, how to configure it, and what to expect until its soft dependency (FEAT-404) lands.

This feature is purely additive ([R3]): an operator who configures nothing sees byte-identical behaviour to before this feature — claude-code still develops, codex still reviews adversarially, and nodes/research.py is unmodified.

The three seats

Seat	Role	Default model	Shape	Transport
Development	dev-agent-pool worker (agent: "nova")	minimax.minimax-m2.5	tool loop	bedrock-mantle (OpenAI-compatible) → NovaCodeDispatcher
Adversarial review	read-only second opinion	us.amazon.nova-2-lite-v1:0	no tools, one call	Converse — NovaAdversarialReviewDispatcher / NovaClient.ask()
Mechanical (PR enrichment)	"Summary of changes" section	us.amazon.nova-2-lite-v1:0	no tools, one call	Converse — NovaClient.ask()

Both Converse seats default to Amazon's own Nova models, not to us.anthropic.* ids. Bedrock gates every Anthropic model behind a per-account "Anthropic use case details" form, so an account holding a valid Bedrock API key still gets ResourceNotFoundException: Model use case details have not been submitted for this account on the first call. Native Nova ids need no such form, which makes them the correct default for a Nova backend. Deployments that have completed the Anthropic form can point either key back at a us.anthropic.* id — the ids stay curated and selectable.

A Claude model cannot hold the nova development seat: Bedrock offers no Chat Completions API for the Anthropic family, so the dev seat is limited to models Bedrock serves over bedrock-mantle (MiniMax, Kimi, GLM). A Claude development seat is served by the existing claude-code backend instead.

Curated model ids (verified against the Bedrock model cards)

Model	Bedrock id	Notes
MiniMax M2.5	minimax.minimax-m2.5	dev seat default; output capped at 8K tokens
Kimi K2.5	moonshotai.kimi-k2.5	dev seat alternative; output capped at 16K tokens
Z.ai GLM-5	zai.glm-5	dev seat alternative; output capped at 128K tokens
Nova 2 Lite	us.amazon.nova-2-lite-v1:0	adversarial-review and mechanical-seat default; geo/global only (us./eu./jp./global.) — no in-region access, the us. prefix is required
Nova Pro	us.amazon.nova-pro-v1:0	Converse alternative for either no-tools seat (previous Nova generation)
Claude Opus 5	us.anthropic.claude-opus-5	selectable for the adversarial seat; requires the per-account Anthropic use-case form; no in-region access in us-west-2/us-east-2 — needs a geo/global prefix
Claude Haiku 4.5	us.anthropic.claude-haiku-4-5-20251001-v1:0	selectable for the mechanical seat; requires the per-account Anthropic use-case form
Claude Fable 5	global.anthropic.claude-fable-5	available via translate(); not a default for any seat

None of minimax.minimax-m2.5 / moonshotai.kimi-k2.5 / zai.glm-5 have a geo or global inference profile — they are never region-prefixed, regardless of region_prefix (see models/bedrock_models.py's REQUIRES_REGION_PREFIX allowlist).

Enabling the dev seat

Add nova to the pool's agent specs (DEV_LOOP_DEV_AGENTS, or a programmatic DevAgentPoolConfig), e.g.:

{"agents": [{"agent": "nova", "model": "minimax.minimax-m2.5", "count": 1}]}

build_dispatcher resolves it to (NovaCodeDispatcher, NovaCodeDispatchProfile), exactly like every other backend branch.

Enabling the adversarial seat

The adversarial seat defaults to codex (ADVERSARIAL_BACKEND, catalog.py) — unchanged. Set DEV_LOOP_ADVERSARIAL_BACKEND=nova to select the Nova reviewer instead. Any other value raises a clear error naming both valid options (codex, nova).

The Nova adversarial reviewer is read-only by construction: it is never handed a tool at all (not a sandboxed profile — the model literally has no tool configuration to invoke), and its verdict always reports files_modified == [] regardless of what the model claims.

Credentials — two paths

Bedrock Converse (adversarial + mechanical seats, via NovaClient) and bedrock-mantle (dev seat) use different credential mechanisms:

1. SigV4 (access/secret keypair or an AWS profile) — the standard BedrockConverseBase resolution: explicit aws_access_key/ aws_secret_key kwargs → the aws_id profile in parrot.conf::AWS_CREDENTIALS → the plain boto3 credential chain (env vars, shared credentials file, SSO, instance role, ...). Used by the adversarial and mechanical seats' Converse calls.
2. Bedrock API key (bearer token) — AWS_NOVA_API_KEY (conf), the same key BedrockConverseBase already uses. The nova dev seat's bedrock-mantle client uses only this path (a bearer token, not a SigV4 keypair) — see NovaCodeDispatcher._resolve_bedrock_api_key(). Missing it raises a clear DispatchExecutionError naming the key.

Config keys (conf.py)

Key	Default	Used by
DEV_LOOP_NOVA_CODE_MODEL	minimax.minimax-m2.5	dev seat (build_dispatcher's nova branch)
DEV_LOOP_NOVA_MANTLE_BASE_URL	derived from region (below)	dev seat's bedrock-mantle client
DEV_LOOP_NOVA_MANTLE_REGION	BEDROCK_AWS_REGION → AWS_REGION_NAME → us-east-1	derives the mantle base URL when DEV_LOOP_NOVA_MANTLE_BASE_URL is unset: https://bedrock-mantle.{region}.api.aws/v1
DEV_LOOP_NOVA_REVIEW_MODEL	us.amazon.nova-2-lite-v1:0	adversarial seat
DEV_LOOP_NOVA_MECHANICAL_MODEL	us.amazon.nova-2-lite-v1:0	mechanical seat (PR enrichment)
DEV_LOOP_ADVERSARIAL_BACKEND	codex	adversarial-seat selector — codex or nova
AWS_NOVA_API_KEY	unset	Bedrock API key (bearer) — required for the dev seat, also usable for Converse

PR-body enrichment (mechanical seat)

FeatureHandoffNode/DeploymentHandoffNode splice an optional ## Summary of changes section into the PR body, generated by one no-tools Converse call on DEV_LOOP_NOVA_MECHANICAL_MODEL. Enrich, never replace: the deterministic template (_build_title/_build_body) is untouched and is also the exact fallback — any LLM failure, timeout, or empty response degrades silently to the pre-FEAT-405 template output. Titles are never touched by the mechanical seat.

A permanent Bedrock condition (model not enabled for the account/region, Anthropic use-case form not submitted, bad model id) logs one actionable WARNING naming the model and the remedy — no stack trace, because the fallback is the designed behaviour. Genuinely unexpected failures still log with a full traceback.

Per-agent usage report

Every dev-loop run now also writes {run_id}.usage.json and {run_id}.usage.html alongside the existing {run_id}.bundle.json / {run_id}.report.md (under conf.OUTPUT_DIR/dev_loop_runs/), and folds a ## Usage markdown section into report.md. All three views (usage.json, the markdown section, usage.html) render from the same UsageReport model, so they cannot disagree. Unreported values render — (never a fabricated 0), and no pricing/cost figure appears anywhere.

Known limitation: UsageReport.agents is keyed by dev-loop node id ("development", "qa", ...), not by individual DevAgentPool worker — session_state.NodeId is a closed Literal of the flow's fixed node names, so a pool worker's own id ("development.w1", "development.w2", ...) cannot be represented in session state today. A future feature would need to widen that type (or add a separate per-worker channel) to report per-worker usage for multi-worker pools.

FEAT-404 soft dependency

FEAT-405 emits round-events for every dev-loop backend (ClientRoundEvent, from LLMCodeDispatcher's tool loop), but per-round accumulation inside BedrockConverseBase (i.e. the Converse-based adversarial and mechanical seats) is FEAT-404's job, shipped separately.

Until FEAT-404 lands: a Bedrock-backed seat's rounds/tokens render — in the usage report — this is expected, not a bug. nova dev-seat rounds (via bedrock-mantle, driven through LLMCodeDispatcher's loop) are unaffected and populate normally today.