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).
Three findings from the PR-4 convergence round:
(A) A root-undefined binding argument passed normalization untouched, so
the sub-call DISPATCHED and only then failed the tool/code-dispatch append
(Session.append rejects undefined event data) — a sub-call executed with
no log record, violating the nothing-executes-unlogged contract. And the
tool received the SAME object later handed to the append, so a tool
mutating its args desynced the logged record from what was dispatched (or
re-poisoned the append). jsonNormalizeArgs now rejects undefined up front
with a model-correctable message and returns TWO independent parses of the
canonical JSON text: the tool gets one, the event logs the sibling —
identical by construction, mutation-proof.
(B) The bridge built its bindings record with plain-object assignment, so
a registered tool named __proto__ hit the prototype setter and silently
vanished (the runtime host resolves binding names as own properties). The
record is now null-prototype with defineProperty, mirroring the
worker-side namespace build.
(B) The header-pin sanity assertions ran only inside NON-pinning
scenarios, so a class consisting solely of its pinning scenario (the two
Code Mode classes) would accept a re-recorded pin carrying several headers
or a header-delta. A fixtures meta-test now asserts every pinning fixture
directly.
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 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 Date.now()/Math.random()/argless-new-Date() bans existed solely to
keep scripts resume-compatible for the deferred journaling/resume
feature. Pre-support for an unimplemented feature is speculative cost:
scripts may now read the clock freely; implementing resume reintroduces
the bans as a script-contract tightening. The RFC's shipped-state
description is updated in place, the tool DESCRIPTION drops the
constraint sentence (the pinned text-turn header follows), and the
engine README's trust-premise paragraph now leans on absent globals
alone.
The tool-order feature canonicalizes the model-facing list (alphabetical
absent a configured toolOrder), so the header-pinning text-turn fixture is
re-recorded on the stacked tree — every other fixture stores the header as
scrubbed tokens and needed nothing. AGENTS.md condensed back under its
ceiling after the merge union.
Residuals from the Codex re-review:
1. A system delta's insert was flattened to one token, so deltas differing
only in inserted-line count compared equal. Now one {{system}} token per
inserted line — position AND extent survive, content does not.
2. The live uniformity guard folded only request/header snapshots, so a
mid-run header CHANGE (request/header-delta) could diverge from the pin
invisibly. Non-pinning runs now assert zero header-delta events: a
scenario that legitimately changes its header mid-run exists to show
that change, so it must pin (fail-loud until it does).
Codex review findings on the pinned-header change:
1. scrubRequestHeaders flattened a request/header-delta's whole
system/tools payload to one token, so two meaningfully different deltas
compared equal. Now the structural facts survive — keepStart/keepEnd
line positions, added/removed/changed tool NAMES — and only the bulk
(inserted prompt lines, schema bodies) is tokenized.
2. The one-pin design rested on an unasserted premise (all sessions
compose the same header). Every non-pinning scenario now asserts, live,
that each request/header its run produces equals the pinned fixture's
header (both sides normalized against their own volatile values), so a
session-dependent header fails loud until it gets its own pin.
Verified the guard bites: perturbing the pinned fixture's prompt fails
a non-pinned scenario with the intended message.
3. RFC de-slopped per docs/AGENTS.md: no PR reference, no SHOULD
spec-speak; Decision/Verification/Consequences updated for 1 and 2.
Restack on the carved-out foundation (#192), per review feedback on #170.
The seam files resolve to the carve-out's revision — its prompt-order
neutrality fix (backends no longer inject 'tools'; the structured runtime
gates its own capture-tool registration) restores the subagent tools to
master's front position, so every recorded fixture is re-recorded on the
stacked tree and the authored error-finish/cancel headers re-patched to the
stacked tool list ([subagent, subagent_fork, workflow, todo_write, ...]).
Every session.jsonl fixture embedded the full composed system prompt and
complete tool-schema list in its request/header event (~8 KB on one line,
identical across the suite), so any prompt or tool-schema edit forced a
re-record or hand-edit of every fixture — see the dynamic-workflows PR for
the churn pattern this removes.
Now exactly one scenario (text-turn, flagged pinsHeader) commits and
compares that content verbatim; every other fixture stores and compares it
as {{system}}/{{tools}} tokens via the new pure scrubRequestHeaders
normalizer (applied to both compare sides and to record-mode writes, so a
re-record cannot reintroduce the content). request/header-delta payloads
are scrubbed the same way; config/reason stay verbatim — a model swap
SHOULD churn every fixture, a prompt edit should not. Replay is unaffected:
script derivation reads only assistant/chunk events.
Fixture meta-guards enforce the split: non-pinning fixtures must be fixed
points of the scrub, the pinning fixture must not be, and exactly one
scenario pins. Committed fixtures migrated through the same function.
Docs: pinned-header RFC (implemented/testing), base snapshot RFC + testing
policy + llm-replay module doc/README updated.
Master's reconstructable-requests overhaul (#179) meets the workflow tool:
- subagent-inprocess structured-output nudge becomes a system-prompt section
plus logged context (the injected-request waterfall shape is gone upstream)
- snapshot fixtures re-recorded on the merged tree so every request/header
carries the workflow tool; authored error-finish/cancel headers patched to
the merged tool list and system text
- architecture.md condensed back under its word ceiling; module graph regenerated
One re-record after the header events landed: recorded scenarios
re-harvested against the live API; the three fs-writing fixtures are
relativized (this recording's model happened to emit absolute
file_path arguments, which only round-trip through replay when the
path is cwd-independent) and, with the two never-re-recorded authored
scenarios (error-finish, cancel), rewritten in the normalized
authored-fixture form the harness documents — each now carrying the
request/header snapshot the loop logs before its first dispatch.
Keyless replay verified green across all 35 scenarios.
Beyond the mechanical conflicts (provider capability lines vs master's new
inheritsParentContext field; generated catalogs regenerated rather than
hand-merged; knip/lockfile), three master-side reworks required semantic
adaptation of this branch:
- The persona rework removed AgentOptions.systemPrompt, which was the
structured-output instruction's channel. The instruction now rides the
SAME final-request enforcement listener that injects the schema'd tool:
appended per request to final.system (per-request wire state, not agent
prompt state). Tests assert the wire request (adapter.requests) instead
of child.options; the bare-direct-dispatch test pins the no-system arm.
- Tool guidance moved out of deployment prompts into per-tool prompt
sections; the examples' workflow paragraph became a tool:<toolName>
section contributed by dsh-tool-workflow (explicit-ask-only policy),
and both example personas resolve to master's minimal identity+behavior
form. tool-workflow gains inject: systemPrompt (+ peer dep, tsconfig
ref); the export-shape guard updated.
- The uniform-RFC-format gate: the dynamic-workflows RFC restructured to
the implemented/ skeleton (bare Status line; Proposal -> Decision;
What-was-rejected -> Alternatives considered; new Consequences), and
the overall-run-timeout deferral is now recorded in the RFC's Deferred
list. The doc-graphs atlas classification gains the workflows seam
(workflow-vm implementation, tool-workflow consumer).
Master's harness-identity section made "empty assembled prompt" states
unreachable through the loop, so the instruction-append is a plain
undefined-ternary and the structured tests assert append-not-replace.
All snapshot goldens (including workflow-run) replay unchanged. Full
local CI-equivalent gate sequence green on the merged tree.
Review round 2 (tianyicui inline comments):
- dsh-system-prompt itself registers the harness:identity (-100) and
deployment:persona (0) sections — they must survive a swapped loop
plugin, so they leave dsh-agent-loop; the persona text is the plugin's
own validated 'persona' config. The model/cwd variables STAY on the
loop: runtime facts of the agents it drives.
- AgentOptions.systemPrompt is deleted with all its forwarding plumbing:
the app configs' systemPrompt keys become 'persona' routed through
dsh-agent-core (schema = z.intersect of the owners'), the ACP bridge
and tool-subagent stop carrying persona configuration, and subagent
children now render the deployment persona like every other agent.
- Example personas drop transport/interface trivia (ACP, CLI) — facts
irrelevant to the model.
- Root CONTEXT.md removed (not idiomatic); its persona definition was
wrong under the new ownership anyway.
- Docs, READMEs, the prompt-variables RFC, and generated catalogs
updated; new loop test pins the assemble-waterfall escape valve
(an emptied assembly sends NO system field).
Reconciliations beyond textual conflicts:
- product rename (DeepSeek Code -> DeepSeek Harness SDK) applied to the
PR-added assertion in system-prompt.spec.ts that master's rename
commit could not reach
- architecture.md: master's rewrite kept; this PR's prompt-assembly
semantics re-added in the new doc's voice (Turn Flow footnote +
service-spine row), within the 1630-word ceiling
- cordis catalog regenerated into master's split events.md/services.md
(events-and-services.md deleted); module graph and doc graphs
regenerated to pick up this PR's new events and dependency edges
A new capability family at packages/workflow/ in the bash seam shape,
modeled on Claude Code's dynamic workflows: the model writes a JavaScript
orchestration script (export const meta = {...} + plain-JS body), a runtime
executes it, and the script — not the conversation — holds the loop, the
branching, and the intermediate results.
- dsh-workflow (ctx.workflows): abstract WorkflowService + run vocabulary
(WorkflowRun whose result NEVER rejects) + observe-only workflow/* events
carrying data snapshots (id + meta, never the live run), per-listener
contained like subagent/*.
- dsh-workflow-vm: in-process node:vm engine. Meta extraction via a
string/comment-aware scanner (template interpolation rejected; literal
evaluated alone in an empty timed context; statement blanked line-
preservingly so stacks keep script line numbers). Hooks: agent(prompt,
{label, phase, schema, model}) over ctx.subagents, parallel(), pipeline()
(no cross-stage barrier), phase(), log(), args. Fatal-vs-null discipline:
hook misuse (unknown/deferred options, bad arguments, unsupported
schemas, tripped caps, seam start failures, cancellation) throws fatal
WorkflowErrors the combinators RE-THROW — never dissolved into the
per-item null reserved for child failures. Realm boundary: inbound values
materialized by descriptor walks that never invoke accessors (defineProperty
copies, __proto__-safe); outbound values rebuilt in-realm via the
context's own JSON.parse. Determinism bans (Date.now/Math.random/argless
new Date) kept so future resume support cannot break scripts. Caps and
timeouts are validated Config. Every hook promise carries a no-op
rejection consumer (app-boot exits on unhandled rejections).
- dsh-tool-workflow: the model-facing workflow tool, synchronous like
dsh-tool-subagent (start → await → try/finally dispose; abort bridged;
non-completed → isError). Generic render card titled by a textual
meta.name sniff. The tool description carries the authoring contract.
Wired into examples/{coding-agent,acp-agent} with explicit-ask-only
guidance. Coverage at every tier: unit (meta scanner, materializer incl.
counting-getter and __proto__ regressions, combinator semantics,
concurrency ceiling, caps, cancellation, no-unhandled-rejection abandon),
integration over the real spawn stack, with-key e2e (real two-phase run +
the tool through the registry pipeline), and a recorded ACP snapshot
scenario (workflow-run, 1 child session). RFC:
docs/rfc/implemented/feature/2026-07-05-dynamic-workflows.md (deferred
work explicitly listed). AGENTS.md budget 1575 → 1590 for the new group's
layout line.
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
Codex review, with its own probes, showed the RFC overclaimed: the
disable patch carried no name assertion despite the text crediting one,
and an id rename is not fail-loud — the skipped patch's warning needs a
logger the replay app deliberately lacks, and the resulting keyless
adapter entry fails inside its fiber without reaching the
unhandled-rejection guard (verified by a subprocess probe of the real
installFailLoud + boot composition). The overlay now asserts
name: dsh-llm-deepseek on the patch (a reused id can never disable the
wrong plugin), and the RFC records the honest residual: an id rename
degrades to config rot with replay output still correct (llm-replay
owns the stream short-circuit), plus the insert-collision last-wins
fact.
examples/acp-agent/cordis.snapshot.yml is a 26-line declarative
overlay: one entry mounts @cordisjs/plugin-include on ./cordis.yml with
patches that disable the llm-deepseek entry by id and insert
llm-replay. Every other entry is the live tree loaded through the
include, so the replay tier exercises exactly what ships and an
app-shape change lands once — the silent-drift class the hand-mirrored
125-line twin invited is structurally gone. The bin is untouched;
recording still boots cordis.yml; assertEntriesLoaded tolerates the
disabled entry by design. All snapshot scenarios pass unchanged,
byte-identical goldens included; the include applies patches at load
time only, which a one-shot replay boot is exactly.
Implements docs/rfc/implemented/testing/2026-07-04-single-source-acp-replay-config.md
(moved from proposed/ and amended); the acp-snapshot-tests and
hook-snapshot-matrix RFCs' replay-config facts are amended in the same
change.
Two pieces of dsh-acp surface were unreachable from any shipped config:
- AcpConfig.agentName/agentVersion: the app package hands the bridge only
{ model, systemPrompt }, so no leaf cordis.yml could set them; they were
settable only by direct-mounting the bridge (a unit test). Hardcode
agentInfo at the initialize site and delete the fields, their schema
defaults, the ?? fallbacks, and the TODO(double-default) whose subject
vanishes. The handshake wire value is unchanged (all snapshot initialize
lines byte-identical).
- The toolKindFor name heuristic special-cased bash*/read*/write/edit*
names in the generic-fallback path, violating the bridge's own design
rule ("the bridge never special-cases tool names"). Every first-party
tool ships its kind via presentCall; the fallback now renders the
neutral kind 'other'. The fallback is reachable when a presentCall
throws OR when model args fail the tool schema (defineTool's presentCall
wrapper returns undefined on violations) — the latter shows up in one
committed golden (hook-codex-posttool-block: three bash calls missing
the required description), whose kind cells flip execute->other. That
3-line golden refresh is the whole transcript delta.
The empty-arguments branch of parseToolArguments lost its only exercise
with the deleted heuristic test; it is live behavior (JSON.parse('')
throws, so the guard is what renders a zero-arg call as rawInput {}), so
it gets a dedicated pin instead of deletion.
RFC moved to docs/rfc/implemented/simplification/ and amended to shipped
reality: fallback reachability includes schema-invalid args, and the
golden churn is exactly the three kind cells (the original zero-churn
claim held only for the branding half).
Load both hook bridges in the ACP example (dsh-hooks-claude → ./hooks.json,
dsh-hooks-codex → ./codex-hooks.json) so the full-transcript snapshot tier can
exercise each dialect against the real app. An absent config file is a silent
no-op, so a scenario carries only the file it needs and the other bridge
vanishes — verified byte-identical against every pre-existing snapshot.
Add a scenario per hook point × its headline Decision outcome, both dialects:
UserPromptSubmit block (authored, keyless) + context-fold, PreToolUse deny/ask,
PostToolUse block/context, Stop force-continue. The mid-turn scenarios are
recorded against the real API with the hook active, so the model's reaction to
a denied/blocked/force-continued turn is part of the replayed transcript.
SessionStart and SubagentStart are deliberately excluded (detached best-effort
inject races the log position — a recorded golden fails 10/10 on its own
replay), as is SubagentStop (observe-only, zero transcript footprint — a golden
could never be proven to fail). Both stay on the bridges' unit coverage. See
docs/rfc/implemented/testing/2026-07-04-hook-snapshot-matrix.md.
A Write CREATE rendered its completed tool_call_update as the model-facing
result TEXT (`<path>…</path>…Created file`), which — because an ACP
tool_call_update.content REPLACES the call's content — clobbered the
new-file diff the pending call installed. So Zed showed the diff, then
replaced it with raw XML-ish text; only overwrite/edit looked right
(their result re-sends a diff).
write's presentResult now ALWAYS returns a diff card for a successful
write: the applied contextual hunk from `meta` when there is one
(overwrite), else an args-derived whole-file diff (`oldText: null`) for a
create or an unchanged-content overwrite. This matches claude-agent-acp,
where the create diff rides on the update and no result text replaces it.
An error still falls through to generic rendering so its message shows.
edit is unchanged (it always has a hunk; no whole-file fallback).
Re-recorded fs-write / fs-write-overwrite goldens; the create's completed
update is now a {type:'diff'} block, not the XML result text.
fs write/edit now emit a result-time contextual-diff tool_call_update
(the applied hunk with ±3 context lines, one hunk per replace_all site),
matching what claude-agent-acp sends and what makes an editor render the
change in place. The call-time snippet diff stays; the result hunk
supersedes it (ACP content-replace).
Mechanism:
- A persisted tool-private `meta` channel: execute may return
`{ content, meta }`; `meta` (JsonValue) rides on the tool/result event
and is handed back to presentResult, so the diff reproduces on replay
(event-sourced). JsonValue is now exported from dsh-session.
- The backend returns raw before/after text (storage facts) on
FsWriteOutcome/FsEditOutcome; the tool computes the hunk via the npm
`diff` package's structuredPatch. A create has no before → no result
diff; a failed/aborted mutation carries no meta.
- ToolResultView gains a DiffResultView; the bridge's result-side switch
renders it as {type:'diff'} content blocks.
RFC: docs/rfc/implemented/architecture/2026-07-02-result-time-applied-hunk-diffs.md
(justifies the npm `diff` runtime dep over vendoring and the meta channel);
the render-intent-union RFC's Non-goal is updated to record this shipped.
All fs snapshot goldens re-recorded; edit/overwrite gain the contextual
result diff, create/read/policy-reject unchanged in structure.
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
The fs-policy gate throws FS_NOT_OBSERVED when the model edits a file it
never read; that rejection surfaces as a failed tool_call_update, but no
snapshot pinned it — a regression that dropped or mis-rendered the failed
card would pass every gate. Record a scenario that edits a seeded file
without a preceding read: the edit is vetoed, the file stays unchanged on
disk, and the transcript shows the pending edit card followed by a
status:'failed' update carrying the policy error.
Tianyi Cui
Yichen Jiang
kingwl
imccyu
Hypatia May