feat(agents): serve a Fabric agent manifest at GET /manifest

[marcusds]

Summary

A Fabric agent server exposes only /health and /v1/chat/completions, so anything driving it — Studio, eval, intake — has to guess the contract it must respect. This adds GET /manifest, an allowlist projection of the AgentConfig the server already holds in app.state, describing the agent's operational contract.

Groundwork for external agents on Fabric (ASTD-314): once an agent can describe itself, registering one by URL can discover it instead of asking the user to type everything. Nothing consumes the endpoint yet — that is ASTD-383.

Related Issue

ASTD-382 (Linear). Parent: ASTD-314.

Changes

• fabric/manifest.py (new) — manifest models and build_agent_manifest(). Pure read model over AgentConfig, no I/O.
• fabric/server.py — builds the manifest once during lifespan startup, serves it at GET /manifest.
• fabric/serving_models.py — SESSION_ID_HEADER moved here so the manifest and the server can both read it without an import cycle. server.SESSION_ID_HEADER still resolves via re-export, so existing imports are unchanged.
• tests/unit/test_fabric_manifest.py (new) — 13 tests.

What the manifest carries, and why

Descriptive fields are the cheap part. These are the ones the platform gets wrong when it has to guess:

• environment.workspace_scope — open_session passes the same base_dir to ensure_local_workspace_dir and fabric.start_runtime on every session, so one workspace is shared by all of them. Reported as agent scope, which tells a caller invocations are not isolated. An eval that doesn't know this lets row 1 leak into row 200 and reads the result as variance.
• serving.max_concurrent_invocations — surfaces the semaphore FabricSessionManager already enforces, so a caller can size its batch instead of queueing or tripping the 503 path.
• telemetry.emits — keeps intake from double-instrumenting an agent that already ships ATIF.
• agent.revision — content hash, so a consumer can tell whether the agent changed underneath a set of results.
• tunable — knob names for nemo-optimization (prompt keys, harness settings keys). Names only.

Two shape notes: workspace_scope is an enum rather than a stateful bool because scope is a fact about the server that survives Fabric gaining per-session environments, where a bool would need redefining. And models includes harness-level models as harness:<name> — the common agent (including this repo's own fixture) declares no top-level models at all, so reading only config.models would report zero.

Redaction

The route is unauthenticated, so this is an allowlist, never a model_dump(). Excluded: api_key_env, base_url, McpServerConfig.url/.env, prompts values, instructions.system.content, and host filesystem paths (skills are emitted as basenames).

revision hashes the redacted projection rather than the raw config, so the fingerprint cannot become a side channel on the values it omits.

Two test layers back this: one sentinel value is planted in every dangerous config slot and asserted absent from the serialized manifest; and test_every_agent_config_field_is_classified walks AgentConfig.model_fields against projected/redacted sets, so a newly added config field fails the suite until someone classifies it.

Type of Change

• Code change (feature, bug fix, or refactor)
• Code change with documentation updates
• Documentation only
• Contributor tooling or automation
• CI, build, or test infrastructure

Quality Gates

• Tests added or updated for changed behavior
• Existing tests cover changed behavior — justification:
• Tests not applicable — justification:
• Documentation updated for user-visible behavior
• Documentation not applicable — justification: the endpoint has no consumer yet; user-facing docs land with ASTD-383, which is what actually surfaces it.

Verification

• Pull request title follows the repository's Conventional Commit format
• Every commit includes an appropriate Signed-off-by: trailer
• uv run pre-commit run -a passes, or any blocked checks are identified below
• Targeted tests pass, or tests are marked not applicable above
• No secrets, API keys, or credentials are included

Targeted validation:

• pytest plugins/nemo-agents/tests/unit/test_fabric_manifest.py plugins/nemo-agents/tests/unit/test_fabric_server.py → 39 passed
• pytest plugins/nemo-agents/tests/unit → 988 passed, 3 failed. All three failures (test_optimize_submit_targets_agents_route, test_deployments_list_defaults_to_table, test_jobs_discovered_via_entry_points) are pre-existing and caused by nemo_optimization not being installed in this venv — confirmed by re-running them with this branch's changes stashed and getting the identical three. test_service.py fails collection for the same reason, also pre-existing.
• ruff check and ruff format --check on the changed paths → clean
• ty check on the changed paths → clean
• uv run pre-commit run -a was not run repo-wide. Pre-commit ran at commit time over the changed files: ruff, ruff format, ty, copyright headers, plugin-import check, merge-conflict check — all passed.

Summary by CodeRabbit

• New Features

  • Added a public agent manifest describing runtime, serving, capabilities, models, telemetry, and configurable settings.
  • Added a GET /manifest endpoint for retrieving the manifest.
  • Included stable revision identifiers and redaction of sensitive configuration details.
  • Standardized the session ID header as X-Nemo-Session-Id.

• Tests

  • Added coverage for manifest generation, security redaction, revision stability, and endpoint responses.

[github-actions]

Suite	Lines Covered	Line Rate	Branch Rate
Unit Tests	31466/40086	78.5%	63.1%
Integration Tests	18326/38038	48.2%	20.8%

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 3a6498b5-ef29-43d2-b4b6-4e13a6b194f1

📥 Commits

Reviewing files that changed from the base of the PR and between 8944625 and df8998e.

📒 Files selected for processing (4)

• plugins/nemo-agents/src/nemo_agents_plugin/fabric/manifest.py
• plugins/nemo-agents/src/nemo_agents_plugin/fabric/server.py
• plugins/nemo-agents/src/nemo_agents_plugin/fabric/serving_models.py
• plugins/nemo-agents/tests/unit/test_fabric_manifest.py

---

📝 Walkthrough

Walkthrough

Changes

Fabric agent manifest

Layer / File(s)	Summary
Manifest contract and projection plugins/nemo-agents/src/nemo_agents_plugin/fabric/manifest.py	Defines public Pydantic manifest models. Projects redacted configuration, capabilities, models, telemetry, tunables, runtime data, and deterministic revisions.
Startup construction and HTTP endpoint plugins/nemo-agents/src/nemo_agents_plugin/fabric/server.py, plugins/nemo-agents/src/nemo_agents_plugin/fabric/serving_models.py	Builds and stores the manifest during startup. Adds GET /manifest and centralizes SESSION_ID_HEADER.
Manifest behavior validation plugins/nemo-agents/tests/unit/test_fabric_manifest.py	Tests projections, redaction, revision stability, configuration classification, and endpoint responses.

Sequence Diagram(s)

sequenceDiagram
  participant FabricServer
  participant AgentConfig
  participant FabricClient
  FabricServer->>AgentConfig: load configuration during startup
  FabricServer->>FabricServer: build and store AgentManifest
  FabricClient->>FabricServer: GET /manifest
  FabricServer-->>FabricClient: AgentManifest response

Loading

Possibly related PRs

• NVIDIA-NeMo/nemo-platform#838: Introduced the Fabric-backed agent and AgentConfig flow used by the manifest.
• NVIDIA-NeMo/nemo-platform#890: Added configuration sections projected by the manifest builder.
• NVIDIA-NeMo/nemo-platform#909: Introduced the Fabric server extended by the manifest endpoint.

Suggested reviewers: mikeknep, tylersbray, ajaythorve

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the primary change: serving a Fabric agent manifest through GET /manifest.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch astd-382-fabric-agent-manifest/mschwab

---

_{Comment @coderabbitai help to get the list of available commands.}