Codex-facing API
OpenAI-compatible /v1 endpoints, generated catalog, and local capability boundary.
Runtime architecture
The hard part is preserving Codex agent semantics.
CodeSeeX classifies requests, adapts DeepSeek provider behavior, owns hosted tools, and keeps usage explainable without becoming a second transcript store.
- Request classes stay distinct: user turns, service requests, compaction, handoff loops, and final responses.
- DeepSeek tool protocol is consumed at the provider boundary and converted into standard tool events.
- Usage and logs explain what happened without storing full prompts by default.
Runtime layers
Runtime layers
Codex Desktop -> local adapter -> agent bridge -> provider adapter -> manager runtime.
Agent bridge
Request classification, client tool handoff, context compilation, and replay controls.
Provider adapter
DeepSeek streaming, reasoning, DSML tool protocol parsing, and usage normalization.
Manager runtime
Status, logs, usage, settings, balance, update checks, and tray operations.
Context and state boundaries
Context and state boundaries
Codex owns the conversation transcript. CodeSeeX keeps only current-process bridge state and bounded diagnostics needed to finish requests and explain runtime behavior.
- Full-context requests are used for the current upstream call, not duplicated as durable conversation storage.
- Tool facts are scoped evidence for CodeSeeX-owned execution, not a replacement for the Codex transcript.
- Tool output, visible thinking, binary payloads, and compact material are bounded before replay.
Provider protocol handling
DeepSeek-specific protocol text is treated as provider protocol, not user-visible prose. When tools are available it becomes standard tool calls; otherwise it is consumed at the safety boundary.
Usage that follows user work
Billable upstream calls are grouped into user sessions with model iterations, tool phases, client handoffs, service requests, cache hits, misses, output, latency, and estimated cost.
Logs for causality, not raw dumps
Logs for causality, not raw dumps
Safe diagnostic logs show categories, severity, request IDs, source probes, context warnings, and protocol events without exposing prompt payloads by default.
Ready to route Codex through DeepSeek with a real agent boundary?
Start with the generated TOML, keep the manager open for observability, and use the docs when you need to reason about tools, state, or privacy.