[Brian Krabach (bkrabach)]

Summary

Migrate the default Amplifier Agent worker integration from private Engine/CLI/host-resolution internals to the public amplifier_agent binding (create_agent, create_session, start_turn, events, and close). The public session carries provider/model and effort selection, inherited history, status-file transport, cancellation, and parent-usage bridging. The dependency resolution pins engine 0.22.0, Core 2.0.1, and Foundation 21ad50fa40f7acff913cbf6615228ac359f7dedd as the narrow co-resolution exception; the github-copilot extra remains available. Graph dispatch and the existing nlspec/owned contract surfaces are unchanged.

Telemetry now reports actual event selections and cumulative usage snapshots; per-call counts and timing remain unknown where the public API does not expose them. Test isolation keeps the 17 real-routing cases and adds adjacent-cycle restoration checks plus an explicit dev-dependency discriminator.

Verification

• Integrated pipeline-runner suite: 297 passed, 1 existing opt-in remote-entry skip on Python 3.11 and again on Python 3.13.
• Adapter suite: 66 passed, 5 skips (2 live-credential tests and 3 documented unsupported SDK targets). Observability: 94 passed; status: 16 passed; timing: 11 passed.
• Root/module lock checks, Ruff check and format, diff checks, and source/scope/semantic-leak review: PASS.
• Retained graph/conformance report: 2,378 passed, 237 skipped; this is not a new manager qualification run.
• Earlier installed-consumer DTU qualification at 8ba837abae57a73ab811c24acdfe52715ef4e7d5 passed 13/13 ordinary-word two-node checks. This test-isolation increment was not re-run in DTU, so it is not claimed as fresh runtime qualification.

Checklist

• nlspec evidence: §1.4 permits backend implementations behind the CodergenBackend boundary; §4.5 retains the status-file and outcome contract. The surrounding §3.3 edge-selection algorithm and §4.6 human-wait behavior are unchanged; no graph-dispatch change is proposed.
• Tests pass: see the named suite results above; the existing opt-in and documented skips are called out explicitly.
• ruff check / ruff format --check clean.
• Diff scope matches the description above.
• N/A — specs/EXTENSIONS.md, ledger/, and contracts/ are unchanged; this is a backend replacement within the existing §1.4/§4.5 allowance, not a graph or contract extension.

Notes for reviewers

The backend migration does not claim parity for max_turns, caller-tools passthrough, or unavailable per-call metrics; an explicit positive turn cap fails rather than being silently ignored, and the coding-agent worker remains the alternative when caller tools are required. Two-hex recall remains unknown even when history is delivered; the final ordinary-word pass is not evidence of universal recall.

Related documentation correction: amplifier-bundle-attractor PR #361 removes stale fallback guidance and links back here for the runtime behavior.