The sandbox docs overclaimed a containment contract the design never makes:
"capability access is routed through cordis services, never Node built-ins,
so everything a mounted plugin does stays inspectable and disposable". The
host-realm helpers on the sandbox global (harness, console, btoa) are
reachable functions, so mount code that goes looking can reach the host realm
through one of them — accepted under the trust stance, because the ctx a
mount ultimately receives is fully privileged anyway. Reword the sandbox
module doc, the README trust stance, and the RFC sandbox-semantics section to
say exactly that: the traps and small global surface STEER honest code onto
the cordis services; they are not a security boundary.
Two review findings (#220) on the sandbox context façade:
- Undeclared services were reachable: the façade resolved any live global via
ctx.get(name), so ctx.bash worked without inject: ['bash']. A cross-mount
consumer could then depend on a provider cordis never saw — unmounting the
provider would neither park the consumer nor unwind its registered tools,
leaving a model-visible tool that fails only at execution. The façade now
reads ctx.fiber.inject and refuses any service the mount did not declare
(with a teaching error naming the inject fix), so the dependency is always
visible to cordis and its activation/unload semantics bind.
- ctx.tools.get returned the live ToolDefinition, including execute — mount
code could call another tool directly and bypass ToolRegistry.execute and
its pre/post-execute hooks and accounting. get now returns the same
read-only name/description/parameters view as schemas(), never an invocable.
Adds inject-gate and schema-view regression cases to sandbox-context.spec.ts
(undeclared property/get denied, declared allowed, the cross-mount zombie-tool
scenario refused at call time, get exposes no execute). Package stays at
per-file 100% coverage. RFC, mount description, and tool-catalog updated.
Review finding (#220): the guarded proxy only special-cased ctx.tools, so
mount code could reach an UNGUARDED context through ctx.root, ctx.extend(), or
a service instance's .ctx, then ctx.root.tools.register({…}) to bypass the
marker check and host-realm normalization — a raw vm-realm result would later
error a real agent turn at the session-log plainness check.
The sandbox ctx is now a whitelist façade, not a pass-through proxy: it exposes
only what a mount needs — tools.register (marker-guarded), on/once, provide, the
timer helpers, and injected services resolved through a guarded get — and denies
every framework-plumbing member (root, parent, fiber, reflect, registry, extend,
isolate, intercept, plugin, set, mixin, …) with a teaching error. Injected
services are wrapped so a method returning a Context is rejected on the way back
(the .ctx escape), closing the one indirect leak. There is no context-valued
member left to reach; cross-mount provide/inject is untouched (the plugin's own
inject and the fiber's pending/active gating are unchanged). ctx.plugin (child
plugins) and ctx.set are denied by design; ctx.effect is deferred (FIXME).
Adds tests/sandbox-context.spec.ts covering the escape class (root/extend/fiber/
plugin/set/… denied, the classic root.tools.register bypass, the .ctx escape,
read-only writes) plus the async-service and symbol/in-operator paths for 100%
coverage. RFC/README/tool-catalog/config-catalog updated; api-catalog.ts
regenerated (also picks up the codeRuntime service that entered on the master
merge and was left stale).
The tree SHAPE was the wrong surface for the model: what it needs from
cordis_inspect is what services, plugins, and capabilities are loaded, not the
fiber hierarchy. The plugins section is now a flat name + lifecycle-state list
from ctx.registry (deterministically sorted, one line per instance); the ASCII
tree renderer, the parent→child rebuild, and the dyn-id tree annotation are
deleted — dynamic mounts keep their own richer dynamic section (id, state,
provides, waits). Net -49 lines; RFC and READMEs state the flat-list contract.
Field sessions showed models writing tool schemas in the JSON-Schema dialect
by strong prior — type: 'integer', required: false, then the full
{ type:'object', properties, required: [...] } wrapper — and the rejection
text itself pushed a nearly-correct DSL attempt BACK to raw JSON Schema: one
stats tool cost three consecutive schema errors before mounting. The boundary
now normalizes wherever the input has exactly one meaning (wrapper unwrapped
with the required array becoming per-property flags at any nesting level,
integer → number, required: false → optional, all rebuilt as fresh host-realm
objects) and rejects only genuinely meaningless input, enumerating the valid
vocabulary in the error. Re-running the failing session mounts first-try.
The mount description documents both accepted forms.
The design record for tool-cordis: the three-tool contract, the vm sandbox
trust stance and boundary mechanisms, the dynamic-group lifecycle, cross-mount
provide/inject composition, the generated runtime API catalog, and the
alternatives weighed (per-capability registration tools, hand-maintained API
tables, a mount provenance event, a hardened sandbox).
Both demo:code-mode UIs now share one mechanism: the base example plus a
same-shaped code-mode.cordis.yml include overlay (insert the worker
runtime, flip tools.mode). Previously the REPL side was a hand-forked
example (examples/code-agent) that had also silently diverged — it
dropped compaction and the subagent stack — so the demo's UI argument
switched agents, not just surfaces. The fork is retired: coding-agent
gains the overlay, a Code Mode README section absorbing code-agent's,
and both of its tests (the keyless boot guard, retargeted at the
overlay; the with-key RFC proof, which hand-mounts its own harness and
moves untouched). The RFC's composed-surface and e2e-tier lines, the
examples index, the AGENTS.md smoke table, and the dsh-tools README
link now describe the overlay shape.
Verified live: overlay keyless smoke, with-key code-mode e2e from its
new home, demo:code-mode banner + EOF exit, and the acp handshake.
Code Mode is the point; the UI is just the surface it happens to wear.
demo:code and demo:acp-code collapse into one dispatcher
(scripts/demo-code-mode.mjs): `pnpm run demo:code-mode [repl|acp]` —
repl (default) boots the stdio REPL over examples/code-agent, acp
serves examples/acp-agent's code-mode overlay; each UI runs the exact
node invocation its standalone script ran, and an unknown argument
fails loud with usage. All nine references across READMEs, the RFC,
the overlay header, and the keyless-smoke comment renamed. Smoked all
three paths: usage exit 2, ACP initialize handshake, REPL boot + EOF.
The agent-scope-contexts RFC (implemented) records the decision tree:
the dsh-scope primitive over cordis extend/Context.filter/no-op fibers,
two-level flat scope with shadowing, restriction/grant semantics, the
scoped-dispatch rule with fused helpers, the setup window, and the
alternatives (explicit scope params, isolate, event-filtering-only,
vendored support) with why each lost. CONTEXT.md pins the glossary.
architecture.md gains the Agent Scope section, the dsh-scope spine row,
the scoped turn-flow line, and an extension-table row (ceiling 1640→1790:
the two-layer registration model is a new architectural axis; additions
are condensed to pointers). READMEs of every touched package re-state
their scoped facts; the stale structured-runtime README section is
replaced by the scoped-registration description.
The seam shipped without its own RFC — the reconstructable-requests RFC
was amended with the mechanics, but the decision record (why a
compose-once frozen prefix, and what the per-request before/after shape,
a system-prompt section, a durable history opener, per-turn composition,
and a dedicated session event each lost to) had no home. Implemented
lifecycle, feature class, dated to the first commit of the work.
ds-review-bot critical (follow-up): on the first step of a resumed or
seeded/forked instance, auto-compaction ran before runStep composed
this instance's prefix, so the gate read the PREVIOUS instance's logged
prefix from the header fold — a contributor that grew across
resume/fork (skills added, AGENTS.md grown: exactly the
environment-dependent case) could under-gate and ship an over-window
first request.
The loop now composes agent/session-prefix before the instance's first
agent/pre-step (still once per instance; runStep just reads the cache),
and agent/pre-step carries the composed prefix to its listeners.
CompactService.compactIfNeeded gains the sessionPrefix parameter;
BasicCompactService.estimatePressure gates on the handed value — the
header-fold read is gone, so the estimate is exact at every step
including a resumed/forked instance's first. Composition moving before
the boundary snapshot also means a composing listener's session append
now joins the CURRENT request (documented on the seam).
New coverage: composition precedes pre-step and the seam receives the
composed prefix; cancel and disposal landing inside the composition
window drop the step cleanly; the compact gate test hands the prefix
directly.
Brings in the refreshed base (master merged through the stack after #203
and #205 landed), including the acp-snapshot extraction (#204), and
re-ports this PR's snapshot-suite extensions onto the extracted package:
- dsh-acp-snapshot's Scenario gains headerClass and configPath; the suite
factory pins the request header PER CLASS (construction rejects a
missing or duplicated class pin), forwards a scenario's configPath to
the harness (RunOptions.configPath overrides AgentUnderTest.configPath),
and a new fixtures meta-test asserts every pinning fixture carries
exactly one request/header and no deltas.
- The acp-agent example's thin scenario table re-registers code-mode-turn
and both-mode-turn with their overlay configs and per-class pins; the
committed fixtures replay unchanged.
- The package's synthetic suites cover the new surface (explicit
headerClass on one suite, the default on the other, a configPath
override through the fake bin, and the two construction throws).
Review discussion converged on the industry shape (Claude Code caches
user context per conversation; Codex separates initial context from
diffs; Kimi appends at continuation boundaries to protect prompt
caching): stable openers belong in a compose-once prefix, mid-session
changes belong in append-only history — not in a per-request slot.
agent/session-prefix fires ONCE per loop instance, lazily on its first
request-building step: the composed Message[] is deep-frozen, cached on
the transmission bookkeeping, recorded as EpochHeader.messagePrefix on
the anchoring 'initial'/'resume' snapshot, and reused verbatim for
every request the instance sends — prefix stability is structural, not
a producer discipline, and a resume recomposes with attributable drift.
The request is messagePrefix + boundary snapshot.
The per-step RequestAdvice/RequestAdviceContext surface and the
messageSuffix header field are dropped: the tail slot had no consumer,
and every current update pattern (new AGENTS.md discovered, memory
update, skills change) routes through the existing append-only history
channels — inject(), tools/post-execute additionalContext,
prompt-submit additionalContext — each paid once and prefix-cached
thereafter. The messagePrefix delta arm stays for codec totality; the
loop never produces one in practice.
The RFC's deployment-policy decision is unchanged; state the current
mechanism in place — the per-tool budget is declared on ToolDefinition
(timeoutMs, set by the owning tool plugin from its config) and the
enforcer is zero-config, so a mistyped tool name is impossible.
Regenerate config-catalog (timeout-policy -> no-config; tool-web gains
fetch/searchTimeoutMs), the event graph (tools/change loses its
timeout-policy consumer), the ToolDefinition type-equiv block, and a
source-line drift in the cordis services catalog.
The dsh-tools half of the Code Mode RFC (its fourth, final change): the
registry gains its first config — mode: native | code | both — and OWNS how
its tools reach the model. 'code' contributes exactly one wire tool,
run_code, plus a lazy tools:sdk prompt section declaring every other tool
as a generated TypeScript API (jsonSchemaToTs: total over the defineTool
subset, unknown degradation, lexicographic byte-identical rendering);
'both' ships both representations; 'native' is byte-for-byte the old
behavior. Non-native modes fail every assembly loudly without a
typescript-language ctx.codeRuntime.
run_code's dispatch bridge: JSON-normalizes each binding argument before
dispatch (what dispatches is what the tool/code-dispatch event logs — the
append can never fail on payload shape; BigInt/circulars reject that one
call), serializes all program tool calls through a per-run queue (even
Promise.all — no concurrency-safety metadata yet), routes every sub-call
through tools/pre-execute → tools/post-execute (a deny rejects the
program-side promise), drops sub-call additionalContext (no safe outlet
mid-run; pinned), owns a run-scoped abort that follows the outer signal in
and fires on settlement (in-flight sub-dispatch aborted, queued abandoned,
queue drained before returning), and converts a failed run into
CodeRunFailedError → a structured isError carrying kind + captured logs.
tool/code-dispatch joins SessionEventMap by declaration merging (log-only;
deriveMessages ignores it).
The composed surface: the tools config forwards through agent-core and
both app packages; examples/code-agent + demo:code run the worker runtime
under mode code (keyless boot smoke + a with-key e2e proving the collapsed
[run_code] header, the dispatch events, and the file the program wrote);
two new snapshot scenarios (code-mode-turn, both-mode-turn) record the SDK
section, collapsed header, dispatch events, and result card — each its own
header-pinning class (the harness gains per-scenario config overlays and
per-class pins). Catalogs, graphs, cookbook, hooks-bridge notes, and the
RFC (moved to implemented/, restructured to decision-era headings) updated
in the same change.
The seam's structured-clone boundary admits values JSON does not (BigInt,
Map, circulars), while tool/code-dispatch events must be JSON-appendable —
left unhandled, a sub-call could execute and then fail at logging time.
The bridge now JSON-normalizes binding arguments BEFORE dispatch (a value
that does not survive rejects that one call), so the dispatched form and
the logged form are the same JSON value by construction.
Model-facing tool-call budgets were tangled into each capability's schema
(bash timeoutMs, web_fetch timeout_ms) with no shared home. Add a
tools/execute around-dispatch waterfall to dsh-tools whose base next() is
the dispatch-with-normalization thunk, and a new @deepseek-ai/dsh-timeout-policy
plugin (packages/timeout/) that arms a per-tool deadline on exec.signal and
returns a structured TOOL_TIMEOUT when it wins. Migrate web_fetch (drop the
model-facing timeout_ms) and web_search onto it; the fetch provider keeps its
timeout only as a resource backstop for direct callers. bash and hook command
execution keep BASH_TIMEOUT unchanged.
Named the plugin timeout-policy (not the RFC's tool-timeout) so it does not
trip the gen-tool-catalog packages/*/tool-* completeness guard, and replace
exec.signal by in-place mutation before next() since cordis waterfall next()
ignores passed arguments. RFC moved to implemented/ recording both deviations.
tianyicui's review: the seam name did not say what the event or its
types do. 'advice' reads both ways — advisory content for the model,
and AOP before/after advice woven around a join point (here the
derived history) without modifying it — so RequestAdvice.before/after
are self-describing. Types follow: RequestAdvice / RequestAdviceContext;
the logged EpochHeader fields keep their positional names
(messagePrefix/messageSuffix).
Also sharpens the core.md wording the review flagged as ambiguous:
before-advice sits in front of the ENTIRE derived history, directly
after the system slot (the conventional home for session-stable openers
— an AGENTS.md digest, a skills catalog), after-advice follows the
history's last message. Catalogs and doc graphs regenerated.
A client-callback throw only becomes a JSON-RPC error RESPONSE to the
agent's session/request_permission — runScenario itself kept going, so a
tolerant agent could treat the error as a denial and the scenario would
pass, or worse, record: the impossible click baked into fixture and
golden, green on every replay. The mismatch is now captured as a harness
error while the agent is answered plain cancelled (a well-defined path
it cannot reinterpret), and the step loop rejects the run on it as soon
as the in-flight step settles. The spec asserts the rejection instead of
the agent-side error echo.
The package, coverage, and permission scripting all shipped on this
branch, so the RFC moves to implemented/ with the lifecycle rewrite:
Proposal becomes a present-tense Decision, Acceptance criteria and Risks
fold into Testing/Consequences with what actually pinned each one (the
zero-byte extraction parity, the 100% per-file coverage via the fake
bin, the vitest-in-src caveat, the per-suite pin cost).
The snapshot tier's machinery leaves examples/acp-agent/tests for
packages/support/acp-snapshot (@deepseek-ai/dsh-acp-snapshot), where the
coverage gate measures it and a second example can consume it instead of
forking it: harness.ts (runScenario, parameterized by an AgentUnderTest
{binScript, configPath, tsconfigPath} instead of module constants),
normalize.ts (moved verbatim), and suite.ts (defineAcpSnapshotSuite — the
per-scenario golden/log compares, record write-back, per-suite header pin
with its uniformity guard, and the fixture guard block, lifted from
acp.snapshot.ts). The example file collapses to its scenario table plus
one factory call; env reading (DSH_SNAPSHOT) stays at that edge.
The exactly-one-pin meta-test generalizes from the hardcoded text-turn
name to "exactly one per suite" — which scenario pins is the scenario
table's reviewable choice (per-suite pinning per the proposal RFC).
Extraction parity: pnpm run test:snapshot is 36 passed + fs-policy-reject
failing BEFORE AND AFTER (BSD-sed environment failure, reproduced at the
base commit in a clean worktree — the recorded golden's sed -i syntax is
GNU-only), with zero byte changes under examples/acp-agent/tests/snapshots/.
Coverage for the new src files lands in the next commit.
The harness, normalizers, and suite/guard logic live inside
examples/acp-agent/tests, outside the coverage gate and copyable-only
for a second suite. Propose @deepseek-ai/dsh-acp-snapshot under
packages/support: parameterized runScenario, verbatim normalizers, a
defineAcpSnapshotSuite factory with per-suite header pinning, and
scripted permissionAnswers so an approval round-trip is expressible at
the snapshot tier — the sandbox composition is the immediate consumer.
New group packages/code-runtime/ with the interface package
@deepseek-ai/dsh-code-runtime, per the Code Mode RFC: abstract CodeRuntime
service (run() resolves program failures as an error field, rejects only
for seam misuse), the CodeRunRequest/CodeBindingNamespace/CodeRunResult/
CodeLogEntry/CodeRunFailure vocabulary, and readonly language/isolation
backend descriptors. Registered in the tsconfig maps, packages/README,
architecture service map, and the doc-graph service-role classification;
catalogs regenerated.
The RFC's one forward path token to the worker package becomes an npm-name
mention until PR3 creates that directory (verify-package-paths is
drift-scoped: the now-existing group made the token checkable).
docs/architecture.md ceiling 1630 -> 1640: the doc gained a genuinely new
capability-service row; the row itself is already minimal.
Budget expiry terminated the worker but nothing cancelled an in-flight
host-side sub-dispatch, and a late dispatch could append events after
run_code returned. The bridge now owns a run-scoped AbortController
(follows exec.signal; fired on any run settlement), sub-dispatches get
the run signal, and run_code returns only after the dispatch queue
drains — no post-settlement appends, per dispose-to-quiescence.
(A1) Scope the wire-collapse guarantee honestly: systemPrompt.tools() is
a public multi-provider API, so the mode governs the registry's
contribution (the only shipped source); deliberate extra providers own
what they add, and the shipped-configuration invariant is test-pinned.
(A2) Replace pause-on-pending-RPC timeout with two independent budgets:
computeMs metered by worker.performance.eventLoopUtilization() busy time
(unfoolable by an un-awaited decoy dispatch; probe-verified) plus a
never-pausing maxWallMs ceiling.
(A3) Specify sub-call additionalContext as deliberately suppressed in
the MVP (immediate inject would break call/result adjacency; the plural
channel is named follow-up work).
(B) Orphan-process caveat vs bash-local's group kill; null-prototype
binding namespaces (__proto__/constructor names); per-PR doc artifacts
(packages/README row, architecture service map in PR2, config/tool/
persistence catalogs per owning PR); engines range corrected to
^22.19.0 || >=24.0.0.
