Repository navigation
Releases
v1.8.0-rc.8 — DeepSeek Harness 0.2.2-rc.3 Compatibility
codex-dsh-native-team v1.8.0-rc.8
Pre-release — released 2026-11-16
External GPT leadership for native DeepSeek Harness Agent Teams — powered by an MCP bridge.
Keep the native team. Replace the reasoning Lead.
This is the second public release of the project and a release candidate
(1.8.0-rc.8), not a stable 1.0. It adds compatibility with DeepSeek Harness
Desktop / Host 0.2.2-rc.3 and a safer offline maintenance upgrade path. Read
the current limitations before installing.
Compatibility warning.
1.8.0-rc.7and earlier adapters do not work on
DSH0.2.2-rc.3: the runtime removed the legacyctx.agentTeamsAPI they
depended on, so the old adapter fails at initialization with
TEAM_API_UNSUPPORTED. Use1.8.0-rc.8with0.2.2-rc.3.
What this project is
codex-dsh-native-team (the DSH MCP Bridge) is an MCP server — registered as
deepseek_harness — that lets an external GPT agent, primarily OpenAI Codex,
act as the only reasoning Lead of a native DeepSeek Harness Agent Team.
The bridge exposes exactly three MCP tools (dsh_delegate_task, dsh_team,
dsh_team_wait) and routes every operation to the official DSH native Team and
Subagent services. DSH keeps its native Root Session, workers, teams, shared task
board, messaging and lifecycle. The DSH Root performs zero model inference:
planning, decisions, coordination and final integration are done by the external
GPT. Workers run in their own native DSH environments with real tools, Skills
and peer communication, and can escalate genuine blockers to the external Lead.
Single-worker and multi-worker delegations both use a real native team.
"Replace the reasoning Lead" means the external GPT takes over the Lead's
decision duties; the DSH Root Session and the official Native Team internals are
kept, not deleted or reimplemented.
Architecture
External GPT / Codex Lead
│ MCP (stdio), 3 tools
▼
MCP Bridge (deepseek_harness)
│ official DSH native Subagent + Team services (0.2.2-rc.3)
▼
DSH Native Team Root (0 model inference)
│
▼
Native DSH Workers ── native tools, global Skills, peer messaging
│
└── task completion / blockers / events ──► wait wakes the external Lead
- The DSH Root Session is real and official, created through
agents.create
and bound withWorkspace.attachSession; root model requests are forbidden by
guards. - Workers keep the official parent/child relations and use the official
list_agents/send_message/wait_agenttools. - The external Lead waits for events instead of polling, decides, and performs
final integration on the exact accepted task revisions.
See the architecture documentation
for the full component and protocol description.
Main changes in 1.8.0-rc.8
Compatibility — official Native Subagent API
DSH 0.2.2-rc.3 removed the legacy ctx.agentTeams methods
(membership, tryMembership, listMembers, spawnTeammate, sendMessage,
waitForChange, interrupt) that 1.8.0-rc.7 required. The RC.8 adapter is
migrated to the official Native Subagent API of 0.2.2-rc.3:
- root/member identity through
subagents.rootOf, membership through
subagents.listDescendants; - Worker creation through
subagents.startActivation(providerspawn,
deliverycaller), using the returned child Session ID as the Worker identity; - messaging through
subagents.sendMessage, interruption through
subagents.interrupt; - the adapter keeps its own event-driven
WaitCoordinator— it is not replaced
by the nativewait_agenttool, and no polling loop was added.
The three MCP tools, the external-GPT-only Lead, the inert DSH Root, the shared
native task board, blocker escalation and the 25-minute default wait are
unchanged.
Team Mode is explicit and persisted
New roots are switched into native Team Mode explicitly (team/mode set to
enabled) and the bridge waits for the official session-persistence flush barrier
before reporting team creation as complete. Historical roots without an explicit
mode are treated as enabled by the official projection and are not silently
mutated; read-only team views expose teamModeEnabled. A closed historical root
is not auto-enabled, and an empty view from a closed team is not treated as
"no tasks".
Worker Session IDs, messaging and task ownership
Native Worker Session IDs are the real identity. Task ownership requests keep
the owner key but carry the Worker Session ID, and the persisted/returned field
remains ownerId. Historical Worker names are kept as read-only external
aliases for existing teams; new Workers cannot reuse an alias, and ambiguous
names require the Session ID.
Native message relay still observes the official agent-message source with the
sender's Session ID and the new body prefix (Agent <name/path/id> sent a message:), while historical Team message from ...: prefixes remain parseable
for replay. Blocking reports are validated against the real sender, current task
ownership and revision, persisted with dedupe and cursor commit before the wait
wakes the Lead, and directed answers pin both the questionId and the Worker
Session ID.
Reliability
- Cold restore now discovers readiness once per root cohort and validates the
whole cohort with one official session list after every root flushes; a
missing root still rejects the whole publication. Production cold restart with
46 historical teams completed in about 49 seconds. - Host/adapter shutdown preserves real disposal errors, keeps the error journal,
waits for persistence and no longer releases the transport or control lock
when close fails. - Control ownership orphaned by the offline maintenance flow is reclaimed only
after the previous process is proven gone. - Completion notifications still come only from a completed current task
revision on the native task board; idle state, settlement or ordinary
messages are not treated as completion.
Safe upgrade
RC.8 adds an operator-driven, fail-closed offline maintenance procedure for
moving an existing installation to a new build on the same 0.2.2-rc.3
runtime:
- an exclusive maintenance window: DSH Desktop/Host must be fully exited through
the official quit path, with an officialsessionPersistence.flush()receipt
and real operating-system exit codes; - kernel single-writer Session leases plus profile/control/storage locks held
during the cutover, so no concurrent writer exists; - a full consistency backup generated and verified at byte level, with the
official readers re-checking sessions, teams, task revisions, event cursors
and acknowledgements before any file is swapped; drift after the backup aborts
the cutover; - program-level forward recovery: recovery restores only the bridge/adapter
program files of the same verified RC.8 build and keeps the latest official
sessions, teams, task revisions, cursors and user configuration. The scenario
was rehearsed during the RC.8 production validation.
There is no automatic rollback, and cross-version downgrade is not
supported. In particular, RC.7 cannot initialize on 0.2.2-rc.3 and is refused
before any write; old business snapshots are audit material and must never be
restored over newer data. The online maintenance entry points remain read-only
preflight/refusal by design.
Verification
- 220/220 automated tests passed, 0 failed, 0 skipped, 0 cancelled
(20 test files, re-run on the public release tree). node scripts/check-architecture.mjs: 24 maintained source modules, 3 MCP
tools, 0 legacy backend dependencies, 0 unreachable modules.- Reproducible adapter build: rebuilding
native-team-plugin/dist/host.mjs
from this release tree produced a byte-identical bundle
(12ca4e5d…) and the same build fingerprint (633ddb…0c34) as the
production-verified adapter. - Real native-team acceptance on DSH Desktop/Host
0.2.2-rc.3on Windows with
real model Workers: single-Worker task, dual-Worker native bidirectional
messaging, a real[codex:blocking]question answered throughquestionId
and resumed to completion, a Global Skill loaded inside a Worker, and the
default wait contract (omittedtimeoutMs→ 1,500,000 ms) with early wakes. - The DSH Root recorded
rootModelRequests: 0across all roots and root events
in acceptance. - Historical data was preserved through the production maintenance window:
45 historical teams and 2,648 original events unchanged, 428 session files and
38 workspaces preserved, and a cold restart kept team/task revisions, cursors
and session files without prompt replay.
Requirements
- Windows (primary and only tested platform; see limitations).
- Node.js >= 24.
- DSH Desktop / Host
0.2.2-rc.3exactly — the officially tested runtime.
RC.8 requires the native Team/Subagent services present in that runtime. - An external MCP-capable client logged in — primarily OpenAI Codex — with
the bridge registered asdeepseek_harness(stdio,node src/server.mjs,
managed tool timeout 3900 s). - An absolute workspace directory the Team will operate in. Installation is
read-only by default; workspace writes require the explicit-Writeopt-in.
Full installation and upgrade instructions:
INSTALLATION.md.
Upgrading from 1.8.0-rc.7
scripts/install.ps1 only performs a fresh install into an empty install
directory. Existing installations must use the documented upgrade flow, not a
fresh install into the old directory:
- Read INSTALLATION.md
and the maintenance prerequisites. - Take a verified backup of the current in...
v1.8.0-rc.7 — Native DSH Agent Teams led by an external GPT
codex-dsh-native-team v1.8.0-rc.7
Pre-release — released 2026-11-12
External GPT leadership for native DeepSeek Harness Agent Teams — powered by an MCP bridge.
Keep the native team. Replace the reasoning Lead.
This is the first public open-source release of the project. It is a release
candidate (1.8.0-rc.7), not a stable 1.0: read the
current limitations before installing.
What this project is
codex-dsh-native-team (the DSH MCP Bridge) is an MCP server — registered as
deepseek_harness — that lets an external GPT agent, primarily OpenAI Codex,
act as the only reasoning Lead of a native DeepSeek Harness Agent Team.
The bridge exposes exactly three MCP tools (dsh_delegate_task, dsh_team,
dsh_team_wait) and routes every operation to DSH's official experimental
Native Agent Team Service and Subagent Runtime. DSH keeps its native Root
Session, workers, teams, task board, messaging and lifecycle. The DSH Root
performs zero model inference: planning, decisions, coordination and final
integration are done by the external GPT. Workers run in their own native DSH
environments with real tools, Skills and peer communication, and can escalate
genuine blockers to the external Lead. Single-worker and multi-worker
delegations both use a real native team.
"Replace the reasoning Lead" means the external GPT takes over the Lead's
decision duties; the DSH Root Session and the official Native Team internals are
kept, not deleted or reimplemented.
Architecture
External GPT / Codex Lead
│ MCP (stdio), 3 tools
▼
MCP Bridge (deepseek_harness)
│ official Native Agent Team Service + Subagent Runtime
▼
DSH Native Team Root (0 model inference)
│
▼
Native DSH Workers ── native tools, global Skills, peer messaging
│
└── task completion / blockers / events ──► wait wakes the external Lead
- The DSH Root Session is real and official, created through
agents.create
and bound withWorkspace.attachSession; root model requests are forbidden by
guards. - Workers keep the official parent/child relations and use the official
list_agents/send_message/wait_agenttools. - The external Lead waits for events instead of polling, decides, and performs
final integration on the exact accepted task revisions.
See the architecture documentation
for the full component and protocol description.
Main changes in 1.8.0-rc.7
Global Skills integration
The native Worker preset now mounts the official DSH skill filesystem and
skill-tool components at the same scope as the Worker. New workers see the same
Skill catalog DSH makes available natively (21 native catalog entries were
observed in acceptance), project .dsh/skills and .agents/skills entries take
precedence, same-name overrides and disable-model-invocation are handled by
the official registry, and Skill directory resources can be read read-only
inside the existing permission boundary. Skill recommendations use directory
metadata only and add 0 model calls; Skill bodies are loaded on demand.
Skills installed only under Codex .codex/skills or Codex plugin directories are
not automatically added to the DSH search roots — see
current limitations.
Default 25-minute wait
When timeoutMs is omitted, dsh_team_wait waits up to 25 minutes
(1,500,000 ms). Waits wake early on task revision completion, real blockers,
important authorizations, serious errors, and integratable state; ordinary
progress is buffered and does not produce heartbeats. requested_timeout_ms
(null means omitted), effective_timeout_ms and recommended_next_wait_ms are
distinct fields, and short waits remain available for explicit diagnostics.
Proactive blocker escalation
A Worker that hits a genuine blocker (missing required input, missing
permission/tool dependency, a verification loop that stops converging) reports
it proactively through the existing send_message path with a first line of the
form [codex:blocking id=<stable-key> task=<task-N>]. The bridge persists the
problem per Worker/task/revision, wakes the external Lead through the existing
wait, and the Lead answers with the existing send operation (questionId). No
fourth MCP tool and no parallel communication service was added.
Also in this release
- Normal synchronous waits omit
timeoutMsinstead of using polling-style
short timeouts; short waits are for explicit diagnostics only. - Worker guidance covers Skills, blocker reporting and low-yield verification
loops; the Worker policy grew while the MCP instructions were trimmed. - Risk-based review only: no default reviewer, no mutual approval, no per-file
reviewer. Independent review is used only for security/permission/
data-corruption risk, core algorithm or interface complexity, design
conflicts, or an explicit review request. - Final integration is performed directly by the external Lead;
integration_ready
is reported when all necessary tasks are complete. - Persistent native Team sessions; teams are archived after 90 days of
completed, quiescent and acknowledged state.
Verification
1.8.0-rc.7baseline: 185 automated tests passed, 0 failed;
node scripts/check-architecture.mjsreports 24 maintained source modules,
3 MCP tools and 0 legacy backend dependencies.- Measured against DSH Desktop / Host
0.2.2-rc.1on Windows with real Codex CLI
sessions using the default wait: e.g. 115,901 ms wake onworker_completed,
19,238 ms / 48,975 ms wakes onlead_decision_needed, 37,651 ms on
team_ready_for_review, worker completion → wait wake 44/40 ms, blocker wake
38 ms. Existing 44 native teams and 551 unrelated workspace files were
preserved during the measured deployment. - The DSH Root recorded
rootModelRequests: 0(no root request headers and no
root model responses) in acceptance.
Requirements
- Windows (primary and only tested platform; see limitations).
- Node.js >= 24.
- DSH Desktop / Host
0.2.2-rc.1(the tested version) with the official
experimental Agent Team Profile and Auto Review enabled. - An external MCP-capable client logged in — primarily OpenAI Codex — with
the bridge registered asdeepseek_harness(stdio,node src/server.mjs). - An absolute workspace directory the Team will operate in. Installation of the
adapter is read-only by default; workspace writes require the explicit
-Writeopt-in.
Full installation instructions:
INSTALLATION.md.
Release assets
| Asset | Description | SHA-256 |
|---|---|---|
dsh-mcp-bridge-1.8.0-rc.7.zip |
Full source and build tree for this release: MCP bridge source, scripts, tests, public docs, the native adapter plugin (source, dist/host.mjs, third-party licenses) and the prebuilt adapter package under packages/. Use it to install the bridge with npm ci --ignore-scripts and to build/install the adapter. |
Listed in SHA256SUMS-1.8.0-rc.7.txt (exact values also in the release description) |
dsh-codex-team-adapter-1.8.0-rc.7.tgz |
Prebuilt native DSH plugin package (dist/host.mjs, dist/THIRD_PARTY_LICENSES.txt, cordis.patch.yml). Installed with scripts/install.ps1 -Workspace <absolute-workspace> -Archive <absolute-path-to-tgz> [-Write]. |
Listed in SHA256SUMS-1.8.0-rc.7.txt (exact values also in the release description) |
SHA256SUMS-1.8.0-rc.7.txt |
SHA-256 checksums for the two assets above. | — |
Verify the downloaded files before installing:
Get-FileHash .\dsh-mcp-bridge-1.8.0-rc.7.zip -Algorithm SHA256
Get-FileHash .\dsh-codex-team-adapter-1.8.0-rc.7.tgz -Algorithm SHA256The digests must match the values published in SHA256SUMS-1.8.0-rc.7.txt and in the release description on this page. GitHub's automatically generated "Source code
(zip)" and "Source code (tar.gz)" archives contain the repository at the
v1.8.0-rc.7 tag; the ZIP release asset adds the prebuilt adapter package under
packages/.
Quick start
# 1. Install the bridge (from an unpacked release ZIP)
npm ci --ignore-scripts
npm test
# 2. Install the native adapter into a DSH workspace (read-only by default)
.\scripts\install.ps1 -Workspace C:\path\to\your\workspace `
-Archive C:\path\to\dsh-codex-team-adapter-1.8.0-rc.7.tgz
# 3. Probe the native Host from the bridge directory
node src\cli.mjs doctorThe installer registers the bridge as the managed stdio MCP server
deepseek_harness in your client config automatically (it writes the
~/.codex/config.toml block; never commit your real client config). Start a new
client session, or reconnect the MCP connection, so the v1.8.0-rc.7 tool schema
loads. A first delegation can be as short as:
{"requestId": "work-1", "prompt": "Implement and verify ...", "permission": "workspace-write"}More examples, the multi-worker flow and troubleshooting are in the
README
and INSTALLATION.md.
Current limitations
- Release candidate quality — not a stable
1.0; APIs and behavior may
change between release candidates. - Compatibility is only what was tested: DSH Desktop/Host
0.2.2-rc.1,
Node.js >= 24, Windows. No claim of universal compatibility with other DSH or
Codex versions or with macOS/Linux. - The outer Codex client/wrapper may still incur its own model-continuation cost
during long waits. - The DSH Root doing zero inference does not mean total token cost is zero:
workers and the external GPT still consume tokens. - Auto-archive after 90 days is supported; **auto-delete (15 days after
ar...