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.
One principle: every fact in the assembled prompt has exactly one owner.
- dsh-system-prompt: merge-extensible AssembleContext on assemble();
a variable(name, provider) registry; {{name}} interpolation in
renderPrompt, strict (unknown/valueless/malformed references throw);
duplicate section and variable names rejected; assembly carries
resolved section text + variables through the assemble waterfall.
- dsh-agent declares AssembleContext.agent; dsh-agent-loop registers
the agent:persona section (order 0 - identity renders before tool
guidance) and the model/cwd variables, and drops its string join:
renderPrompt(assembly) IS the full prompt.
- Tool guidance moves to its owners: descriptions carry per-tool
semantics; sections only cross-call habits (tool:bash exit-code
habit at order 105; read's not-shell nudge). todo/subagent need no
section - their descriptions already carry the contract.
- SubagentProvider.inheritsParentContext (spawn/acp false, fork true);
dsh-tool-subagent derives truthful per-provider wording and resolves
the provider at load (backend must be listed first).
- Example personas shrink to identity + behavior with {{model}} (and
{{cwd}} in the ACP tree); the welcome banner stops enumerating tools.
RFC: docs/rfc/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md
Replace the "bag of optional fields" tool-presentation types
(ToolCallPresentation / ToolResultPresentation / ToolTerminal) with a
card-tagged discriminated union — the standing FIXME(tool-presentation).
A tool declares one render intent per call/result and the ACP bridge
switches on `card`:
ToolCallView = generic | terminal | diff
ToolResultView = generic | terminal
The `diff` card is new: fs write/edit now emit an ACP {type:'diff'}
content block (an editor's inline diff), which the old shapes could not
express. The bridge also relativizes a file card's title against the
session cwd (mirroring claude-agent-acp's toDisplayPath) while keeping
locations/diff paths raw, and derives the no-capability fenced console
fallback from a terminal result's output. read gains the window-in-title
(`Read foo.txt (5 - 8)`) and an always-set location line, matching the
reference adapter field-for-field.
Migrates all three producer families (tool-fs, tool-bash, tool-todo) and
the sole consumer (the ACP bridge) together — the source does not compile
piecewise. Adds snapshot coverage for the terminal _meta path (a new
capability-advertising scenario) and re-records the fs goldens to show the
diff cards. Applied-hunk (result-time, context-line) diffs need a new
result/event shape and are a follow-up.
RFC: docs/rfc/implemented/architecture/2026-07-02-tool-render-intent-union.md
Bring the interception-seams branch onto current master (via A→B). The
substantive reconciliation is master's compaction `agent/pre-step` serial seam
meeting C's interception seams:
- types.ts: keep BOTH master's `agent/pre-step` AND C's new interception events
(`agent/prompt-submit`, `agent/session-start`, `agent/turn-continuation`→
`ContinuationDecision`); drop the turn-mirror declarations (removed on A).
- loop.ts: the merged per-turn order is `turn/start` → per queued msg
`agent/prompt-submit` (rewrite/inject/block) → (fully-blocked ⇒ zero-step
`rejected`) → per step: drain steering → assemble system prompt →
`agent/pre-step` (compaction, OUTSIDE the step) → `step/start` → single
`deriveMessages()` → model → tools/pre-execute·dispatch·post-execute. No
turn-mirror emits; `closeTurn()` is the A-simplified single-call form.
- Docs (architecture, core.md, agent/agent-loop READMEs, catalog) reconciled to
show C's interception seams alongside `agent/pre-step`, no turn/step mirrors.
- rfc/README: dropped the stale `proposed/` compaction row (master moved that RFC
to implemented/); kept C's new `pre-tool-input-rewrite` proposed row.
- interception.spec.ts: migrated its two `agent/turn-end` reason collectors to
the `turn/end` session event, and ADDED a cross-test proving a
`prompt-submit` rewrite + additionalContext is VISIBLE to an `agent/pre-step`
listener on the same turn — pinning the merged seam ordering (compaction sees
the post-prompt-submit surface, not stale history).
Address review: the "trusted-plugin surface" framing overstated the security
story. A model driving the `bash` tool already has equivalent power to set env
vars and feed stdin through ordinary shell syntax (`FOO=bar cmd`, heredocs), so
the `env`/`stdin` seam fields grant it no new capability — and they cannot
exfiltrate the harness's ambient credentials, because the credential SCRUB in
dsh-bash-local (which strips *KEY*/*SECRET*/*TOKEN* from process.env before the
child sees it) is the actual control, and it works regardless of these fields
(tool-call args are static JSON, never shell-evaluated).
So drop the "dangerous / trusted-plugin boundary" language across the RFC, the
three bash-package READMEs, the bash/src/types.ts JSDoc, and docs/bash.md (both
the type-equiv blocks — kept 1:1 with source — and the prose). The reality that
remains: the `bash` tool doesn't EXPOSE env/stdin as parameters because they'd
be redundant with shell syntax; the fields exist for in-process plugins (the
hooks bridges) to pass a JSON payload + CLAUDE_* vars cleanly. The guard test is
kept but reframed: it catches a future `...args` spread that would silently
forward model input into the post-scrub env merge, NOT a trust wall. No code or
behavior change.
Codex's PR-C review found two (A) blockers:
- tools/post-execute could corrupt the protected outcome. postExecute passed the
mutable `result` to listeners and then read result.callId / spread result on the
return paths, so a listener mutating the reference (flipping isError, rewriting
callId, injecting an error) escaped the decision channel. Now the authoritative
callId/isError/error are SNAPSHOT before the waterfall and the return value is
rebuilt from the snapshot + the typed PostToolDecision — the decision is the only
sanctioned way to change the outcome, and callId is always exec.callId. Added a
regression test that mutates the result reference and asserts it has no effect;
proven to fail red on the unfixed code.
- Public docs/JSDoc still advertised the removed `tools/execute` waterfall after the
split. Swept every current-state reference to tools/pre-execute + tools/post-execute:
the ToolRegistry class JSDoc (and the regenerated catalog), loop.ts's ASCII flow
(also added the prompt-submit/session-start steps it was missing), the package-map
READMEs (packages, core, agent-core), core-data-structures core.md/tools.md, the
bash + acp + invariants src/READMEs (the deferred permission gate is the
tools/pre-execute deny/ask seam now), the cookbook, and the implemented RFCs whose
factual seam catalog drifted. codec.ts's totality prose now lists `rejected`.
Proposed-RFC references are left as-is (frozen proposals, validated when built).
The hooks subsystem runs external hook commands the Claude Code / Codex way:
JSON payload on stdin, context in CLAUDE_PROJECT_DIR / CLAUDE_PLUGIN_ROOT env.
Reusing the ctx.bash seam for that needs two new inputs — but stdin and arbitrary
env are exactly what dsh-bash-local's credential scrub exists to keep away from
model-driven commands. So this adds them as a TRUSTED-PLUGIN surface:
- BashExecRequest + BashExecSpec gain optional `stdin` and `env`. They are plain
optionals on the resolved spec (not required-but-nullable like `owner`): a
missing one means "none", the safe default, not a security footgun.
- dsh-bash-local threads them through resolve/run/start. `env` merges AFTER the
credential scrub, so a trusted caller's explicit entry wins even on a
credential-shaped name — the scrub guards the harness's OWN ambient creds from
model-driven commands, not a trusted plugin. stdin is always a pipe, closed
immediately (with bytes when supplied, empty otherwise — EOF as before); an
EPIPE from a child that exits without reading is swallowed.
- The model-facing dsh-tool-bash NEVER forwards model input into stdin/env (its
request is command/workdir/timeoutMs/signal/owner only). A regression guard
drives the real tool with adversarial args and asserts the request carries
neither field — proven to go red if the consumer ever forwards them.
Configurable scrub (in an earlier sketch) is dropped as speculative: the explicit
`env` field already gives a trusted caller full control, and no caller needs to
broaden the ambient scrub. Documented in a new architecture RFC, the bash.md
type-equiv blocks, and the three bash READMEs.
Merge brought in the RFC-classification reorg and two new doc gates;
rewrite every drifted packages/<name> cross-link (Markdown link targets,
moved-README relative depths, and .ts comment paths) to the grouped paths.
Add two doc-sync/hygiene gates so the manual checks this restructure
needed become automated:
- verify-package-paths.ts: flags a packages/<path> reference (in Markdown
or a .ts comment/string) that does not resolve AND names a real package
in a segment — i.e. a stale path to a MOVED package. A path naming a
non-existent package (a forward-looking proposal) is left alone, so it
applies uniformly across proposed/implemented/rejected.
- check-workspace-constraints: assert the packages/<group>/<pkg> depth-2
shape (group dirs carry no package.json; no flat or over-nested
packages). Group names stay open; only the shape is fixed.
Move the 18 flat packages/<name> packages into role-grouped dirs:
core/, llm/, bash/, session-persistence/, ui/, support/. Group dirs are
pure containers; each package keeps its @deepseek-ai/dsh-* name.
Collapse the per-package tsconfig paths maps (base + typecheck) into one
@deepseek-ai/dsh-* wildcard with a candidate per group, and derive the
publint list from the hierarchy. Update all depth-coupled globs/configs
(workspace, tsdown, vitest, eslint, knip, tsconfig includes/refs,
per-package tsconfigs, generators, doc-script scopes, type-equiv manifest)
and the cross-package/script relative imports in tests.
Fix doc-typecheck's workspacePaths() to parse tsconfig JSONC via the
TypeScript API instead of a regex comment-strip, which corrupted the
new wildcard `/*/` path candidates.
WIP: doc cross-links and package/RFC docs still to update.
Tianyi Cui
kingwl