mirror of
https://github.com/deepseek-ai/deepseek-harness
synced 2026-08-15 21:04:50 +00:00
669771097d
Per-session policy overrides (sandbox/mode, approval/policy) never crossed the delegation boundary: a spawn child of a read-only-switched parent ran under the wider deployment default, and a fork child missed any switch made after its seed boundary — delegation was a bypass channel for a user's tightening. The in-process driver now snapshots the delegating parent's override chain and stamps it onto the child through the canonical write paths (SandboxPolicyService.inheritOverride / ApprovalService.inheritOverride), anchored inside the child's first turn via a one-shot agent/prompt-submit listener: turn-enclosed (durable), ahead of the first request (an inherited 'never' reaches the child's first system prompt), and positioned after any stale fork-seed switch so the ordinary last-event-wins fold resolves it. Only overrides are copied — an unswitched parent stamps nothing and the child follows the live deployment default; both services are consumed opportunistically, so compositions without them delegate unchanged. Nesting composes by construction (each stamp folds the already-stamped parent log). Evidence: inheritance.spec.ts drives scripted-model children into the real dsh-fs-sandbox fence through the real write tool (disk-state + denial-marker assertions; spawn, stale-seed fork, grandchild, escalation fail-closed, and no-stamp guards), inheritOverride contract tests in both service suites, and the recorded subagent-sandbox-inheritance ACP snapshot (read-only preset → delegate → child denied, replayed keylessly). See .agents/notes/implemented/feature/2026-07-25-subagent-policy-inheritance.md.
Packages
Packages use the @deepseek-ai/dsh-* scope. Each is a Cordis Service subclass or function plugin; contributions use ctx.effect(), ctx.on(), or ctx.waterfall(). Authoring rules: package and root.
Hierarchy
Packages live at packages/<group>/<pkg>/; groups are containers, while names remain @deepseek-ai/dsh-<pkg>. Each group README is the canonical package/ctx-key map.
| Group | Role | Release expectation |
|---|---|---|
core/ |
Product API spine: sessions, prompts, tools, agent services, and the concrete loop | Product — stable surface |
goal/ |
Persisted same-session goal state and lifecycle | Product — stable surface |
llm/ |
LLM capability family: the abstract service + provider adapters | Product — stable surface |
bash/ |
Bash capability family: executor seam, local impl, model-facing tool | Product — stable surface |
pty/ |
Persistent PTY capability family: owner-scoped sessions, local implementation, and model-facing tools | Product — stable surface |
code-runtime/ |
Code-execution capability family: the runtime seam for model-written programs + a worker-thread backend | Product — stable surface |
sandbox/ |
Process-confinement seam; bwrap/Landlock/Seatbelt backends | Product — stable surface |
fs/ |
Filesystem capability family: seam, local impl, model-facing file tools, bash-backed discovery tools | Product — stable surface |
lsp/ |
LSP capability family: seam, generic stdio provider, and the lsp tool |
Product — stable surface |
skill/ |
Skill capability family: the provider registry, local provider, and model-facing catalog/loader | Product — stable surface |
compact/ |
Compaction capability family: the abstract seam + a basic backend (tool deferred) | Product — stable surface |
context/ |
Model-visible request context, including workspace instructions and time context | Product — stable surface |
subagent/ |
Subagent capability family: the provider-registry seam and the model-facing delegation tool | Product — stable surface |
tasks/ |
Generic background-task runtime and model-facing task_* control tools |
Product — stable surface |
workflow/ |
Workflow capability family: the script-engine seam, worker-thread engine, and model-facing workflow and fresh-agent ralph tools |
Product — stable surface |
web/ |
Web capability family: seam, search/fetch provider impls, and the model-facing web tools | Product — stable surface |
spill/ |
Spill capability family: storage seam, local impl, tool-result spill policy | Product — stable surface |
todo/ |
Todo/planning family: the model-facing todo_write tool |
Product — stable surface |
plan/ |
Plan collaboration state with a direct entry command and reviewed exit | Product — stable surface |
timeout/ |
Tool-call timeout policy: the tools/execute deadline enforcer |
Product — stable surface |
guard/ |
Loop-hygiene guards: advisory repeat-call reminders | Product — stable surface |
cordis/ |
Self-referential runtime toolset: inspect the live runtime's plugins and services, mount/unmount model-written plugins (design) | Product — stable surface |
hooks/ |
Hook bridges + the shared Claude Code / Codex wire-protocol library | Product — stable surface |
session-persistence/ |
Persistence capability family: the seam + JSONL/SQLite backends | Product — stable surface |
session-query/ |
Session retrieval: logical corpus, bounded reads, lineage, and event relationships | Product — stable surface |
session-title/ |
Log-backed session titles: fallback service, shared LLM policy, and opt-in providers | Product — stable surface |
sdk/ |
Project SDK tooling | Product — stable surface |
ui/ |
Editor/client integration surfaces: ACP bridge, JSON-RPC SDK server, user-approval/user-interaction seams, ask-user tool | Product — stable surface |
examples/ |
Demo bundles (agent-spine + TUI/one-shot CLI/ACP/JSON-RPC bins) the leaves load | Support — example infra |
support/ |
Support infrastructure (testkits, invariants, replay, Loader smokes) | Support — lower compatibility expectations |
util/ |
Low-level zero-dependency utilities shared across groups (Branded<B>, Harness home/path helpers, timeout, retention) |
Support — small, stable, harness-dep-free |
Groups distinguish product API from support infrastructure. New packages join an existing group; a new group updates its README and this table.
Dependencies
The dependency graph is generated: docs/module-graph.md (pnpm run gen-module-graph, freshness-gated in CI).
Extension plugins depend on interfaces, never the concrete loop. dsh-agent-loop is swappable; UI, hook, and tool plugins use dsh-agent. Composition bundles, including dsh-agent-spine-demo, may depend on spine plugins. Capabilities split into interface / implementation / consumer packages; see capability seams.
Package READMEs cover purpose, APIs, extension points, and Model Experience unless on the model-agnostic omission allowlist. They also carry ## Known Limitations and Deferred Work or use its allowlist.