dsh-doctor
DeepSeek Harness 的确定性诊断与恢复
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:astra3294/dsh-doctor#82818eb9b01f660f5cb6c27831e308e818965c08Doctor 以 dsh-doctor 发布，当前版本为 0.4.3。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Deterministic diagnostics and recovery for DeepSeek Harness.
DSH Doctor is the safety net for the two ways a Harness usually breaks itself: it cannot converse (the Web UI opens but the loop is broken), or it cannot start (boot fails after a config, dependency, or plugin change). When the Web UI still opens, the Doctor button calls a loopback-only Host recovery service. When Harness cannot start, the same engine runs as a standalone CLI — including a boot probe that captures the failure and a one-click reset to the last healthy checkpoint, so plugin developers can experiment and always get back to a working state.
DSH Doctor 0.2.0 targets
@deepseek-ai/dsh 0.1.0-rc.6. DeepSeek Harness is a developer preview; unknown versions are scanned in conservative read-only mode.
Add Doctor to the Web profile:
dsh plugin --profile web add dsh-doctor
dsh --profile web
This installs one package with both Host and browser halves. It contributes:
If the Web UI cannot start:
npx dsh-doctor recover --profile web
# Scan one profile (offline and path-anonymized by default)
npx dsh-doctor scan
# Probe whether the Harness can boot and diagnose the failure from its output
npx dsh-doctor boot
# One-click recovery: reset configs to the healthy checkpoint, reconcile
# dependencies, then verify the boot again
npx dsh-doctor recover --profile web
npx dsh-doctor recover --profile web --yes # also apply confirmed actions
# Snapshot the current profile as a healthy recovery baseline
npx dsh-doctor checkpoint --profile web
# Start the Harness; diagnose automatically when it fails to boot
npx dsh-doctor launch --auto-recover
# Restore a checkpoint
npx dsh-doctor rollback <checkpoint-id>
recover runs the safe actions with zero confirmation, then verifies the result by booting the Harness when it was down. The exit status is 1 when errors remain (or boot verification fails), warnings return 0 unless --strict is used, and invalid arguments or internal failures return 2. JSON output is available with --json.
cordis.yml, cordis.patch.yml, and pnpm workspace syntax and permissions;session.jsonl.zstd) and storage caches for corruption;Every mutation follows the same transaction:
Safe one-click actions prepare Doctor state and reset the profile configuration to the latest healthy checkpoint. On a cold start with no healthy checkpoint, Doctor can synthesize minimal valid config files — after explicit confirmation, with the broken originals preserved in the pre-repair checkpoint. Dependency installation, disabling a boot-crashing plugin, quarantining a corrupt session file, and re-enabling a Doctor-disabled plugin require explicit confirmation. Doctor refuses to disable the connection, layout, settings, conversation, runtime, module-loader, or Doctor rescue entries.
Checkpoints are stored locally under $DSH_HOME/doctor/checkpoints/, with ten retained per profile and the newest healthy checkpoint always kept. They may contain exact copies of local configuration so that rollback is byte-accurate; they are never included in reports or uploaded. API keys are never printed, changed, or sent by Doctor.
See Recovery matrix and Security policy.
Requires Node 24 and pnpm.
pnpm install
pnpm check
npm pack --dry-run
MIT