feat(sensor): add DeepSeek Harness parser

[DvJiang]

What type of PR is this? (check all applicable)

• Refactor
• Feature
• Bug Fix
• Optimization
• Documentation Update

Related issue: N/A

What changed?

• Add a DeepSeek Harness parser for current v3 JSONL and Zstandard-compressed session logs.
• Normalize messages, tool calls and results, PTC sub-dispatches, approvals, provider/model context, and token usage into the Sensor schema.
• Preserve structured tool-result content and metadata, including compaction replacement provenance, while keeping the original execution result in normalized tool telemetry.
• Register the dsh source with AgentObserver and the CLI, including content-aware incremental exports for resumed sessions.
• Add parser documentation, a three-platform CI workflow, runtime zstandard dependency, and focused tests.

Why?

DeepSeek Harness stores local agent sessions in a structured event stream that ADR Sensor did not previously ingest. This adds read-only support using the same source-specific parser pattern as the existing agent integrations.

How did you test it?

• pytest tests/ -q on Python 3.11: 245 passed.
• pytest tests/test_dsh_parser.py -q on Python 3.9: 14 passed.
• Ruff checks and formatting checks pass for the changed parser and tests.
• uv lock --check passes.
• Source distribution and wheel builds complete successfully.

Potential risks

DeepSeek Harness may evolve its on-disk event schema. To avoid silently misinterpreting incompatible data, the parser only accepts v3 as the highest session generation and skips legacy or future generations. It reads session files without modifying them. For a torn trailing Zstandard frame, it follows the upstream recovery contract by retaining only complete newline-terminated JSONL records and discarding the incomplete tail.

[CLAassistant]

All committers have signed the CLA.

[barisozbas]

This adds useful DeepSeek Harness support, but I recommend addressing these issues before merging:

1. Preserve tool-result metadata and replacement history

[Parser line 257](https://github.com/uber/ADR/blob/8e0b61bdd01ae8c70b39f8906927587d9656b6da/Sensor/adr_sensor/parsers/
dsh_parser.py#L257)

Tool results are reduced to text/status/error, dropping attachment references and meta, which can contain contextual
filesystem diffs.

DSH compaction can also append a replacement result with the same call ID. The parser overwrites the original
execution result without retaining the replacement relationship.

Please retain structured result metadata and distinguish execution results from later model-context replacements. This
is separate from the text truncation already used by other Sensor parsers.

2. Match upstream recovery for incomplete Zstandard tails

[Parser line 434](https://github.com/uber/ADR/blob/8e0b61bdd01ae8c70b39f8906927587d9656b6da/Sensor/adr_sensor/parsers/
dsh_parser.py#L434)

The parser discards all decoded content from an incomplete final frame. However, DSH’s persistence contract recovers
complete newline-terminated JSONL records from that frame.

A frame missing only its final checksum byte can contain a fully decoded event that ADR currently drops.

Please recover complete records from the incomplete final frame and update the test that currently expects them to be
discarded.

3. Preserve distinct source session IDs

[Parser line 187](https://github.com/uber/ADR/blob/8e0b61bdd01ae8c70b39f8906927587d9656b6da/Sensor/adr_sensor/parsers/
dsh_parser.py#L187)

Stripping the optional session- prefix maps both session-review and review to dsh_review. These are valid
distinct custom/imported IDs, but the parser deduplicates them into one session.

Please namespace the complete source ID without stripping its prefix and add a regression test covering this
collision.

[DvJiang]

Thanks for the detailed review. I addressed all three points in 847c0b4:

1. Tool-result content, attachments, meta, errors, surfaceOp, and sourceEventSeqs are retained in session_context.tool_result_metadata. Compaction replacements no longer overwrite the normalized original execution result.
2. Torn trailing Zstandard frames now recover complete newline-terminated JSONL records, including the missing-checksum-byte case, while discarding only the incomplete record tail.
3. DSH session IDs now namespace the complete source ID, with a regression test covering session-review and review as distinct sessions.

The full Sensor suite passes on Python 3.11 (245 tests), the focused DSH suite passes on Python 3.9 (14 tests), and Ruff, lockfile, and package-build checks pass locally.

[barisozbas]

open findings resolved. LGTM