Two independent multipliers were pushing bare `pnpm run lint` past node's
default heap:
1. parserOptions.project expanded to every package tsconfig plus the root
one — each matched config built its OWN ts.Program, and the dev paths
map pulls sibling package SOURCES (plus the vendored Cordis stack) into
every such program, so resident memory grew as sum-of-closures, not
repo size: ~4.6 GB peak for 425 repo files. projectService shares one
tsserver-style graph: ~2.0 GB peak, ~28 s → ~14 s wall.
2. `eslint .` traversed .claude/ harness-local state — stale worktree
checkouts there carry tens of thousands of additional .ts files
(whole-repo copies), roughly tripling the work again even under the
project service. Other checkouts are not this one's sources; ignore
them like node_modules. (#169 carries the identical ignore line inside
its chain; the hunks dedupe on its next rebase.)
Type-aware rules verified live under the service: a floating-promise
probe still trips no-floating-promises.
Review finding, valid — the previous SDK fix covered only the
non-default branch: in the default mode under Code Mode the wire filter
dropped exit_plan_mode but the registry-rendered tools:sdk section
still advertised its binding, offering default-mode agents a call that
can only error and breaking the byte-identical claim (a no-dsh-mode
deployment's registry never saw the tool, so its SDK never listed it).
The SDK re-render extracts to one helper both branches share: the
non-default branch passes the mode's visibility rule, the default
branch hides exactly the exit binding. The pinning test now compares
the default-mode SDK byte-for-byte against a bare deployment without
dsh-mode — the strongest form of the invariant the RFC states.
A review finding claimed plan mode under the registry's 'both'
presentation leaves non-allowlisted native schemas on the wire. The
claim misreads the filter predicate — 'keep visible plus run_code' IS
'drop everything else' — and the SDK re-render shares the same visible()
predicate, which is precisely the remedy the finding requests. This
test refutes it empirically on unchanged code (first run green) and
stays as the regression pin: plan + both = [exit_plan_mode, read,
run_code] on the wire, read documented and write absent in the SDK.
Review follow-up on the residual the previous commit accepted — and the
acceptance was wrong, because the fix is clean: in Code Mode the SDK
section IS the soft surface (the wire carries only run_code), section
text resolves in assemble's base, and renderToolsSdk is an exported
pure renderer. The outermost wrapper therefore re-renders tools:sdk
from the same visibility predicate the wire filter applies (allowlist,
exit-IFF-plan, minus run_code mirroring the registry's own exclusion):
a plan-mode program is documented exactly the callable bindings — read
and the exit, never the denied write. The default mode leaves the
section untouched (absence of policy), both pinned by tests.
The soft layer's promise — the model is never encouraged toward a tool
the gate denies — now holds in Code Mode too; the only remaining
prompt-honesty residual is a prepend-after-load assemble listener,
where the gate still covers execution.
Review follow-up: the blanket 10s default correctly targets worker-thread
cold-start races (starting, first-script-line, async child-registration
messages — genuinely CPU-bound under CI contention), but four waits assert
something different — that the HOST reacted PROMPTLY to an event that
already happened (a settled result, an observed worker death). Those had no
cold-start left to wait on, so the generous default just widened the window
a real regression could hide in.
Verified by injecting a 6s delay into the settle-reap's abort call: the
un-overridden helper's test still passed in ~6s. The same mutation now fails
in ~1s with the explicit 1000ms override restored on all four sites (the
abort-on-settle test's own assertion, the two worker-death cancel/dispose
reap checks, and the dispose-ack-race check). The other 12 waits keep the
10s default — they run BEFORE a result is awaited, waiting on the worker to
actually start rather than on a host reaction. Doc comment corrected to
describe the split instead of claiming every wait is a cold-start race.
Review finding, valid: under the registry's Code Mode the assembly's
only wire tool is run_code, which the plan allowlist filtered out —
leaving the model with NO tools at all, the exit review included. The
composition exists today (the acp-agent example ships a code-mode
overlay), so plan mode bricked it outright.
run_code is a transport, not a capability: every bridged sub-call is
serialized back through ToolRegistry.execute() carrying the same agent,
so tools/pre-execute judges each capability individually — exactly like
native calls. Both layers now exempt it by name: the filter keeps it
visible (tests pin plan-mode Code Mode assembly = ['run_code']) and the
gate passes the wrapper while the same run's write sub-call still
denies with the plan-mode reason.
Documented residual, same class as the prepend-after-load one: the SDK
section renders from the registry's store, so a plan-mode program may
be offered bindings whose dispatch the gate then denies — nothing runs
that a native call could not.
Review finding at the seam of two surfaces this branch added to the
same stdin: with an ask_user_question (or plan-review) prompt active,
the line handler dispatched every line as the answer first, so
'/mode plan' typed mid-question was recorded as free-text feedback —
model-visible in the tool result — and the mode never changed. Command
handling now runs before answer dispatch: the command executes, the
question stays pending and still owns the next non-command line. A
literal '/mode…' free-text answer is the trade-off deliberately spent —
a swallowed command that becomes review feedback costs far more than
that contrived answer shape.
Review finding with a real in-repo instance: the structured runtime's
per-spawn final-assembly wrapper (prepend, post-next) re-injects
structured_output OUTSIDE the mode filter, so a structured child in
plan mode would see a tool the gate then denies — the soft policy and
the hard gate telling different stories. The suggested fix (make the
mode filter outermost) cannot beat that instance: prepend unshifts, so
the per-spawn listener always registers later and wraps outer.
Two-part resolution instead. Semantically, structured_output enters the
shipped plan allowlist — it is a child's pure result channel, the same
ask/report class as ask_user_question and exit_plan_mode, so the
filter, the re-injection, and the gate now agree wherever a structured
child runs in plan mode. Mechanically, the filter registers with
prepend anyway: it now wraps outside every append-registered listener
regardless of load order (regression test pins a pre-registered
post-next mutator being filtered), narrowing the documented cosmetic
residual to prepend-after-load listeners only, where the gate still
covers execution. Severity note: no execution breach existed — the gate
held throughout; this closes the prompt-honesty gap.
The spec's 16 vi.waitFor sites used the 1s default timeout to wait for
worker-thread startup and child registration — CPU-bound work that blows
past 1s on a contended runner. The CI coverage lane (4 vitest workers plus
suites that spawn real subprocesses) hit this 3 times across 4 recent PR
runs, each a different subset of the cancellation/worker-death tests, each
green on rerun.
Every wait now goes through a shared helper with a 10s bound, and the file
sets a 30s test timeout to make room for it. The one deliberately tight
wait keeps its 800ms bound through the helper's override — it proves the
host (not the wedged worker's later loop turn) delivered the cancel, so a
generous bound would erase what it tests. No behavior under test changed.
spec.onError is a caller-supplied callback boundary, but the flattening
catch invoked it unguarded: a throwing sink rejected the whole async result,
breaking the seam's "result never rejects on a child-level failure"
contract (and docs/defensive-patterns.md's contain-callback-exceptions
rule). The sink's own throw is now swallowed — named as such — while the
original child failure still settles as stopReason 'error'. Regression test
drives a spawn failure through a throwing sink and asserts result resolves.
Same defect as the codex backend's, fixed there on PR #240; this is the
symmetric fix for the already-merged ACP backend.
Review hardening (the finding's ordering premise did not hold — see the
PR thread — but its failure-path kernel did): onBoundary cleared the
pending intent BEFORE appending the mode/set, so a backend rejecting
that one write lost the switch forever while the picker kept showing it
optimistically. The intent is now cleared only after the append lands;
a failed flush stays parked and the next healthy boundary converges the
log with the picker. The containment test extends to pin the re-park
and the retry.
The bridge's re-notify keeps deriving from the logged event's value —
now documented in place: the service holds ONE coalesced pending slot
(every flush reads the latest selection, so a stale flush cannot
exist), and for any other writer the logged value is the truth the
picker should track, in log order.
The three wedged-child scenarios busy-spin their worker for 1.5s while the
host waits for the start RPC; under a loaded 2-core CI runner (this branch
adds several parallel suites) the RPC lands after vi.waitFor default 1s,
failing the coverage lane three runs in a row at the same three sites. An
explicit 10s waitFor timeout (well inside each test own 15s budget) makes
the assertions load-tolerant without weakening them. Belongs upstream with
dsh-workflow-workerthread; carried here because it gates this PR.
The three-entry cordis.yml (dsh-sandbox-local + dsh-bash-sandbox at a
read-only default + dsh-approval) served over ACP: the first live approval
composition. Recorded snapshot scenarios pin the wire end to end —
config-options advertisement, the mode-switching arc as the suite pinned
header (both switches, the prompt-section delta, one changed-by-the-user
notice per knob, a confined write landing under the switched mode), and
both escalation branches over scripted permissionAnswers (a grant runs
confined under workspace-write; a rejection executes nothing and pins the
fail-closed text). The with-key escalation e2e drives a real model +
real runner + the real bridge answerer, world-verified; ci.yml snapshot
lane and e2e.yml install bubblewrap so the confined replays actually
execute.
Both RFCs move to implemented/ (Decision/Consequences form, deferred
phases tracked in their own sections), with every cross-reference flipped.
effective(session) = findLast(the session own knob events)?.value ?? the
composition-config default. One log-only event per knob, owned by its
domain (bash/sandbox-mode in dsh-bash, approval/policy in dsh-approval),
each exporting the same three-piece kit: the event declaration, a pure
fold, and THE write path — a switch IS its event; no owner service, no
facts map. Restart immunity and multi-session isolation fall out of the
log replay by construction.
Execution follows the fold on both sides: the bash tool stamps
escalation grant > session override > executor default, and the approval
seam prepends the never-gate that auto-rejects before any interactive
answerer. Visibility is two layers per knob: a per-agent prompt section
states the effective value on every request (logged through
request/header*, so what-the-model-was-told replays from the log), and an
agent/pre-step narrator injects at most one coalesced delta notice with
positional attribution (user switch vs operator/config drift). The ACP
bridge advertises one capability-gated select per composable knob with
currentValue folded per session, validates set_config_option against the
closed vocabularies, and anchors idle switches at the next turn
prompt-submit under the turn-enclosure contract.
The tool gate advertises sandbox_permissions (an enum of exactly the modes
STRICTLY WIDER than the mounted executor default — the schema makes a
non-widening request inexpressible) plus a required justification, exactly
when ctx.bash.sandboxMode reports a confining mode at registration:
composition truth, never a dead lever. An escalating call resolves
ctx.approval BEFORE anything executes with the audit-self-contained reason
"escalate sandbox to <mode>: <justification>"; allowed-once stamps the
granted mode onto that one bash request (the seam-level per-call override),
while rejected / cancelled / unavailable and the no-service / no-agent
paths each fail closed with their own error text and execute nothing. The
description teaches the flow only when the fields exist: retry the SAME
command once after a real denial, never preemptively; a rejected
escalation is final. No new session events: the attempt is an ordinary
tool/call, the decision is the approval audit pair, the outcome an
ordinary tool/result whose facts name the mode it ran under.
dsh-bash grows the per-call policy carrier: BashExecRequest.sandboxMode
(request-optional, spec required-but-nullable — the owner pattern; resolve()
is the one explicit defaulting step) and the BashExecutor.sandboxMode
capability fact (undefined in the base class — composition truth the tool
layer can read). dsh-bash-local carries the field verbatim and confines
nothing.
dsh-bash-sandbox extends LocalBashExecutor and hands ctx.sandbox the exact
argv it is about to spawn. A denial is a RESULT FACT (the command RAN;
result.sandbox.denied is orthogonal to exitCode/signal), classified
conservatively against the wrap own dialect; a RUNNER failure outranks
denial — foreground re-throws the structured SANDBOX_UNAVAILABLE, a settled
background task stamps sandbox.runnerFailed — so a broken sandbox never
reads as a failing command and the command never runs unconfined.
dsh-tool-bash renders the markers and teaches the model not to retry around
a policy denial; escalation and per-session switching are staged follow-ups.
ctx.sandbox (dsh-sandbox): confine(argv, policy) returns the argv to spawn
instead — wrapped so the process and its children run confined — plus the
enforcement completeness and the backend denial/runner-failure dialects;
no usable backend throws the fail-closed SANDBOX_UNAVAILABLE. Policy rides
per call. dsh-sandbox-local selects by platform and caches the verdict:
multi-candidate chains probe FUNCTIONALLY in preference order (Linux:
bwrap → the registry-installed node-addon-landlock-run launcher), a sole
candidate is selected unprobed (darwin: sandbox-exec/Seatbelt) and fails
closed at execution via runnerFailureSignatures; win32 is a reserved empty
chain. Profile parity is honest per backend (documented temp-area and ABI
differences; enforcement full|partial is a structured result fact).
CI: the sandbox-e2e matrix proves real-kernel confinement per rung (bwrap,
Landlock per architecture through the registry-installed launcher,
Seatbelt), failing on a silent all-skip; the packed-install rehearsal
installs the launcher family from the registry and asserts the binary
executable apart from kernel enforcement.
The ACP bridge registers the first real approval answerer: an ask for an
agent it owns becomes session/request_permission attached to the already-
streamed tool call (one-shot allow_once/reject_once only), outcomes map
conservatively (unknown optionId never grants, client cancel → cancelled),
and foreign or call-less requests delegate down the waterfall. The snapshot
harness accepts scripted permissionAnswers (FIFO; an unscripted prompt
answers cancelled, fail closed) so recorded scenarios can drive the wire
keylessly.
ctx.approval (dsh-approval): request() dispatches the approval/request
waterfall and always resolves a closed outcome — allowed-once / rejected /
cancelled / unavailable — never rejects; zero listeners fall through to
fail-closed unavailable; abort settles cancelled and discards late answers;
throwing or rogue answerers are contained as unavailable; every ask lands
the log-only approval/asked / approval/decided audit pair. dsh-tools routes
a pre-execute ask through the seam opportunistically (ctx.get) with three
distinct deny reasons, keeping the historical ask→deny degrade when the
seam is absent.
The per-session policy tier, the ACP bridge answerer, and the sandbox
escalation asker are staged follow-ups of the approval-seam RFC.
Live failure: deepseek-v4-flash answered a greeting entirely in the
reasoning channel — no text block, no tool calls. The serializer's
null-content fallback produced an assistant message with neither
content nor tool_calls, which the API 400s ('Invalid assistant message:
content or tool_calls must be set'). Because that message sits durably
in the session log, every later turn of the session re-derived the same
history and failed identically — one all-reasoning response bricked the
session permanently (log: turns 2 and 3 failing byte-identically).
content is now always the flattened text ('' when there is none); the
passback rule still keeps reasoning_content off plain turns. The old
null shape was pinned by a test whose comment claimed the wire accepts
it — live-falsified, updated together with the code, plus a regression
test for the reasoning-only shape. Existing bricked logs resume cleanly
under the fix (the poisoned message now serializes as '').
Live-session feedback (a real Zed elicitation round-trip): the model
presented its finished plan as a plain reply and asked the USER to
switch modes — the exact reversal the roadmap warns about — because the
shipped section's 'present it with the exit_plan_mode tool' read as a
suggestion. The section now says a finished plan is delivered by
calling exit_plan_mode, preferred over pasting it as a plain reply or
asking the user to switch modes — firmer, without imperatives.
ask_user_question enters the shipped plan allowlist (asking is
read-only-safe), and the section points a blocked decision at it. The
plan-acp-agent example composes the bash family (default mode only —
plan's allowlist keeps excluding it, so the two modes now demo a real
difference) plus tool-ask-user; both recorded scenarios re-recorded:
the pin now shows plan = [ask_user_question, exit_plan_mode, read,
todo_write] and post-exit default = the full eight-tool surface.
Hypatia May
Tianyi Cui
Yichen Jiang
kingwl
pku-xht
Ziya