fix(executors): robust deepseek-web tool-call parsing and agentic context retention

[BugsBag]

Summary

• Fixes tool (function-calling) support for the deepseek-web provider, used by agentic
  clients.
• Tool-call parsing: chat.deepseek.com has no native function calling, so OmniRoute
  serializes tools into a prompt and parses the text reply back into tool_calls. The
  shared parser only handled the canonical <tool>{json}</tool> form, so DeepSeek's many
  ad-hoc variants (<tool:name>, <tool_call> with type/params, name=/id=
  attributes, nested <tool><tool>…, XML <name>/<arguments> children, <parameter>
  style) were returned as plain text and stalled the agent loop. A new DeepSeek-specific
  parser recognizes all observed variants.
• Context retention: the upstream prompt was built from the last user message only and
  dropped assistant tool_calls and role:"tool" results, so the model had no memory of
  prior steps and restarted every turn (re-creating todos, re-listing files, re-checking
  uv, …). Tool requests now replay the full trajectory into the prompt.
• Surrounding text: explanatory text emitted alongside a tool call is now forwarded to
  the client (stream + non-stream) instead of being dropped.
• The shared webTools.ts parser is untouched, only exports added.

Validation

• npm run lint
• npm run test:unit
• npm run test:coverage
• Coverage is still >= 60% for statements, lines, functions, and branches

Tests Added Or Updated

• tests/unit/deepseek-web-tools-variants.test.ts (new) — type-checked unit tests for the
  parser (parseDeepSeekToolCalls) across every variant, the strict prompt
  (serializeDeepSeekToolPrompt), and the conversation builder
  (buildToolConversationPrompt).
• tests/unit/deepseek-web-tools-execute.test.ts (new) — executor-pipeline tests
  (fetch mocked): text + tool_calls emission (stream and non-stream) and full-trajectory
  context replay into the upstream prompt.

Coverage Notes

• Production files touched (open-sse/):
  • open-sse/translator/deepseekWebTools.ts (new) — 92.92% stmts / 79.03% branches /
    100% funcs / 92.92% lines, covered by deepseek-web-tools-variants.test.ts.
  • open-sse/executors/deepseek-web.ts — 89.59% / 71.81% / 96.87% / 89.59%, the new
    tool/prompt paths covered by deepseek-web-tools-execute.test.ts (existing suites
    cover the rest).
  • open-sse/translator/webTools.ts — 98.68% / 82.88% / 100% / 98.68%; change is
    export-keyword only, existing web-tools-translation* suites unaffected.
• No touched file moved below the gate; all are above 60% on every metric.

Notes

• The shared webTools.ts parser is intentionally left untouched (it works for other providers); the DeepSeek logic lives in a separate module that reuses
  webTools helpers via newly exported functions (no behavior change there).
• Tool requests now resend the full transcript each turn (stateless, correct for an agent
  loop). Very long sessions grow the prompt; a follow-up could cap it to the system prompt
  • first task + last N tool exchanges if upstream size limits are hit.
• No DB migrations, no feature flags, no new network surfaces.
• Manually validated end-to-end against a live DeepSeek-web agent session by the author.

[gemini-code-assist]

Code Review

This pull request introduces DeepSeek-web-specific tool-call translation and prompt building to handle various ad-hoc tool-use shapes emitted by DeepSeek, preventing amnesia in agentic loops by replaying the entire conversation trajectory. It also adds comprehensive unit tests for these new features. The review feedback correctly identifies a potential issue in the PARAM_TAG_RE regular expression where lazy matching could greedily swallow subsequent parameter tags, and provides a robust fix using a negative lookahead.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.