[GH-ISSUE #1636] [Question] Track DeepSeek Harness headless completion without process exit on Codex #553

New issue

Open

opened 2026-10-15 18:26:07 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-15 18:26:07 +00:00

Owner

Originally created by @N0zoM1z0 on GitHub (Oct 10, 2026).
Original GitHub issue: https://github.com/morluto/rea/issues/1636

Before submitting

Searched related agent integration reports, including #1587.
Removed credentials, private paths and target data.

Area

DeepSeek Harness headless client / model-provider cleanup.

Observed behavior

Native DeepSeek Harness 0.2.0-rc.2 can discover the installed REA skill, call its actual MCP server and produce correct Codex-backed fixture answers, but the headless process remains alive after emitting its final JSON event.

Two independent bundle profiles returned the routes, backend name, retry count and static-only limitation correctly. Two independent missing-target profiles asked for an app name/path with zero REA operations. All four native runs reached a final response but required stopping the owned scope at the 120-second ceiling. An earlier imported-CLI experiment also produced a correct final response and did not exit; the native repeats rule out attributing this solely to that wrapper.

No-REA controls

In two more fresh Git roots and HOME directories, configure only the documented Codex-backed model route. Do not install REA or any skill and do not register an MCP server. Run:

dsh --profile headless --json 'Reply exactly READY. Do not use tools.'

Both emit final: READY, make zero tool calls, and still require stopping the owned scope at a 45-second ceiling. Thus REA/MCP setup is not necessary to reproduce this condition. The session and process-completion results are recorded separately; a correct final answer is not an exit-success oracle.

Expected behavior

The native headless client exits after a completed one-shot response and releases its model-provider resources.

REA version / environment

REA source 1bba5545, skill 37 in the integration trials. No REA in the controls. Linux x64, Node 24.18.0, DeepSeek Harness 0.2.0-rc.2, real gpt-6-sol through its documented openai-codex adapter. Private HOME/XDG/cache directories, one CPU and 3 GiB process-tree memory ceiling. Authentication used access-only transport values; the original OAuth store was unchanged.

Ownership and follow-up

The reduced controls point at client/model-adapter lifecycle ownership, not a REA cleanup defect. Keep this as a compatibility record; investigate upstream disposal/connection ownership before considering a REA change. No REA production workaround, forced exit or timeout-policy change is proposed.

Verification limits

The real DeepSeek API model provider was not tested. These tests do not establish behavior for other models/providers, native engines or Windows/macOS. Native deterministic-loopback REA integration verifiers exit successfully and independently validate all 170 schemas and complete named-schema Evidence. Headless JSON tool-result previews are truncated; do not parse them as complete producer JSON. A follow-up inspection reused the installed client's frame scanner to decode every complete frame of the canonical session logs (no torn tails). All 18 REA call inputs and full outputs passed their named contracts, and final cited Evidence IDs matched returned Evidence. No transport credential was found in the complete decoded logs. The process-exit failures remain unchanged.

Refs #1587, #1627.

Originally created by @N0zoM1z0 on GitHub (Oct 10, 2026). Original GitHub issue: https://github.com/morluto/rea/issues/1636 ### Before submitting - [x] Searched related agent integration reports, including #1587. - [x] Removed credentials, private paths and target data. ### Area DeepSeek Harness headless client / model-provider cleanup. ### Observed behavior Native DeepSeek Harness 0.2.0-rc.2 can discover the installed REA skill, call its actual MCP server and produce correct Codex-backed fixture answers, but the headless process remains alive after emitting its `final` JSON event. Two independent bundle profiles returned the routes, backend name, retry count and static-only limitation correctly. Two independent missing-target profiles asked for an app name/path with zero REA operations. All four native runs reached a final response but required stopping the owned scope at the 120-second ceiling. An earlier imported-CLI experiment also produced a correct final response and did not exit; the native repeats rule out attributing this solely to that wrapper. ### No-REA controls In two more fresh Git roots and HOME directories, configure only the documented Codex-backed model route. Do not install REA or any skill and do not register an MCP server. Run: ```sh dsh --profile headless --json 'Reply exactly READY. Do not use tools.' ``` Both emit `final: READY`, make zero tool calls, and still require stopping the owned scope at a 45-second ceiling. Thus REA/MCP setup is not necessary to reproduce this condition. The session and process-completion results are recorded separately; a correct final answer is not an exit-success oracle. ### Expected behavior The native headless client exits after a completed one-shot response and releases its model-provider resources. ### REA version / environment REA source 1bba5545, skill 37 in the integration trials. No REA in the controls. Linux x64, Node 24.18.0, DeepSeek Harness 0.2.0-rc.2, real gpt-6-sol through its documented openai-codex adapter. Private HOME/XDG/cache directories, one CPU and 3 GiB process-tree memory ceiling. Authentication used access-only transport values; the original OAuth store was unchanged. ### Ownership and follow-up The reduced controls point at client/model-adapter lifecycle ownership, not a REA cleanup defect. Keep this as a compatibility record; investigate upstream disposal/connection ownership before considering a REA change. No REA production workaround, forced exit or timeout-policy change is proposed. ### Verification limits The real DeepSeek API model provider was **not** tested. These tests do not establish behavior for other models/providers, native engines or Windows/macOS. Native deterministic-loopback REA integration verifiers exit successfully and independently validate all 170 schemas and complete named-schema Evidence. Headless JSON tool-result previews are truncated; do not parse them as complete producer JSON. A follow-up inspection reused the installed client's frame scanner to decode every complete frame of the canonical session logs (no torn tails). All 18 REA call inputs and full outputs passed their named contracts, and final cited Evidence IDs matched returned Evidence. No transport credential was found in the complete decoded logs. The process-exit failures remain unchanged. Refs #1587, #1627.

bsakaguchi added the

question

area: mcp

P2: medium

labels

2026-10-15 18:26:07 +00:00

bsakaguchi referenced this issue

2026-10-15 18:28:29 +00:00

[GH-ISSUE #2383] [Tracker] Audit analysis and debugging coverage for the TIOBE top 100 languages (October 2026) #794

bsakaguchi referenced this issue

2026-10-16 21:14:14 +00:00

[PR #597] [MERGED] fix!: repair boundary contracts across CLI and MCP #2097