[Jacob953]

Change

Closes #1.

Move the DeepSeek Harness integration developed in kagent-dev/kagent#3195 into harnesses/deepseek/ as an independent community experiment maintained by @Jacob953.

• Connect DSH through ACP and the existing BYO/A2A contract, with persistent conversations and workspaces.
• Support HTTP MCP tools, protected-tool approvals, DATA-restored public tasks, settled conversation forks, and Linux process cleanup.
• Include an optional standalone PostgreSQL-backed approval service, deployment examples, tests and source attribution.
• Map persisted root-session model calls to kagent usage telemetry, including cache tokens, model identity, timing, partial usage and cancellation; exclude restored history from new reporting.
• Explicitly disable session-log and plugin-inventory uploads while retaining local persistence; cover the existing completion barrier with delayed/failed session/close regression tests.

Targets DSH 0.2.0-rc.2 and kagent 0a57368aca33, including the merged restoration contract from #3165. No shared kagent runtime/controller or DSH source changes are included.

Incremental output remains unsupported. Model usage is emitted after native-process cleanup and is best effort: subagent calls and crash-time backfill are unsupported, and invalid or oversized logs produce a warning without failing the conversation. Native ACP permissions still require a live process. Pending tasks cannot be checkpointed, uncertain remote outcomes are not automatically replayed, and shared recovery limits in #3010 remain. Supported configuration and conditions for adopting other upstream fixes are documented in the README.

Build

Passed from a clean source export, with the harness directory as the entire context and no credentials or build arguments:

docker build --platform linux/amd64 --tag byo-ext/deepseek:pr harnesses/deepseek

The October 19 follow-up rebuilt the AMD64 image from a clean source export and passed non-root configuration/version/supervisor and readiness checks. The ARM64 production build passed during October 16 migration acceptance.

Manual verification

• October 19: full Linux arm64 bridge race suite passed 144 unique test/subtest cases (52 top-level tests), with no failures, skips or race reports. Usage parser: 22 cases passed. Host short race tests and host/Linux lint also passed. Details are recorded in VERIFICATION.md.
• Real DSH usage checks cover cache accounting, process reconstruction, approval restore, stream cancellation and corrupt logs. The production A2A server and SDK exporters also passed a local OTLP HTTP/protobuf test: model spans retain their task/context and execution segment, resumed execution links to its origin, and repeated exports or late approvals add no tokens.
• October 16: all 15 independent Kind cluster scenarios passed, including persistent conversations/workspaces, MCP approvals, Worker replacement after a DATA snapshot, and settled forks.

Tests used real DSH with local model/MCP and controller fixtures, plus PostgreSQL where required. The OTLP receiver decodes network payloads; it is not a deployed collector/backend. Executor reconstruction retains one SDK provider and does not verify cross-process metric aggregation. No live provider calls or full AMD64 runtime regression were performed, and the October 19 follow-up did not repeat cluster acceptance. Historical results, reproducible commands and coverage boundaries are in VERIFICATION.md.

Checklist

• My commits include DCO sign-offs (git commit -s).
• I followed the Code of Conduct, including its guidance on AI assistance.
• I updated the relevant documentation and, for a new harness, the catalog.
• I included no credentials or private configuration.

AI assistance was used for implementation, testing, documentation and this description.

[Jacob953]

@EItanya, a note on the collaboration behind the recent updates: @foxj77's independent investigation and our discussion in kagent#3194 informed several improvements in this PR. Thank you, @foxj77, for sharing the findings and spike evidence.

• Usage reporting: his mapping notes and cache-accounting caveat informed the session-log integration. I checked the accounting against our pinned 0.2.0-rc.2 / llm-pi-ai path and added coverage for compressed logs, historical model context, cancellation and recovery without duplicate reporting. The OTLP export test also verifies token totals and span ownership after approval restoration.
• Upload defaults: his findings prompted explicit disabling of session-log and plugin-inventory uploads, with request-level regression checks while retaining local persistence.
• Persistence: our follow-up confirmed that the existing session/close flush addresses the completion window he identified. I added delayed/failed-close regression tests around that barrier.

We also clarified the distinction between adapter-owned MCP approvals and native ACP permissions: the durable MCP proxy is already implemented here; native permissions still require a live process.

His spike used 0.2.1-alpha.5 / llm-deepseek on macOS. This PR's Linux validation and remaining limits are recorded separately in VERIFICATION.md. The implementation, validation and ongoing maintenance remain my responsibility; this acknowledgment does not imply a formal review or maintenance commitment from @foxj77.

AI assistance: Codex helped with implementation, testing and this update.