Two review findings on the turndown swap, both verified empirically:
- Unclosed-tag nesting makes the synchronous turndown/domino walk
superlinear (measured: depth 512 ~0.15s, 2k ~2s, 20k ~5s), during
which the cooperative fetchTimeoutMs timer cannot fire. renderBody
now preflights nesting depth with a linear tag scan and passes
bodies past 512 levels through raw; the try/catch stays for markup
the scan cannot see (comment-hidden tags), simulated in tests via a
converter throw.
- Markdown escaping can expand converted HTML ~2x (100k underscores
render as 200k chars), so provider body caps no longer bounded the
model-visible result. formatFetchOutput now caps the complete output
(header + body + footer) under new fetchMaxOutputChars config
(default 200000 = 2x the local provider's default body cap), reusing
the truncation notice.
README EN+ZH, config catalog, Agent Note EN+ZH updated; the new
web-fetch fixture is migrated to the packed layout master now
requires; tool-web coverage stays 100% per-file.
The simplified request path no longer anchors an unchanged request/header on
resume (agent.ts logs a header only when it differs from the folded baseline),
so resume-turn logs drop that event and later seqs shift down. Refresh the
keyless session-log fixtures to match; no model scripts changed.
Implements the turndown Agent Note from the NIH dependency audit (full
variant, not the minimal entities-only fallback): dsh-tool-web's fetch
rendering now converts HTML through turndown + @joplin/turndown-plugin-gfm
(atx headings, fenced code, dash bullets, GFM tables/strikethrough) over
the real domino DOM, with script/style/noscript removed wholesale. The
hand-rolled ~86-line regex converter html.ts and its entity tables are
deleted; renderBody wraps the conversion in try/catch falling back to
the raw HTML body, because turndown's recursive DOM walk overflows with
a RangeError on pathological nesting (measured: 4k levels on the main
thread, 8k in a worker) where the regex version could never throw.
Closure weight, measured: tool-web IS in the single-exe runtime closure,
and the exe asset globs would pack ~7.9 MB of the three new packages —
but ~6 MB of that is domino's test corpus, with runtime lib/ at ~550 KB
against a ~174 MB artifact (<0.5% either way), so the swap wins.
Per testing policy the previously-missing keyless web_fetch snapshot
ships in the same change: the acp-agent `web-fetch` scenario boots a new
web.cordis.yml overlay (web seam + real dsh-web-fetch-local provider +
tool-web fetch-only + a loopback HTTP fixture server on a fixed port
serving deterministic HTML with entities, a GFM table, and nesting), so
recording and keyless replay both drive the real HTTP fetch and real
conversion end to end; the scenario pins the new `web` header class.
The Agent Note moves proposed -> implemented and is rewritten per the
lifecycle contract (Decision/Consequences/Testing, closure verdict and
alternatives recorded); tool-web and acp-agent READMEs updated in both
languages and pairs re-recorded.
Conflict resolution: scripts/type-equiv.manifest.json takes master's new
paired-derivative format (one primary entry per pair) and re-adds this
stack's CodeDispatchLog entry in that format. zh README pairs brought
along for the dispatch-log arm (spill-policy behavior/limitations bullets,
tools limitation bullet now pointing at the shipped bounding).
The scenario inspects the tools service API; the round-2 content-contract
JSDoc change shifted its rendered output. Keyless DSH_SNAPSHOT=refresh —
the resulting fixture is byte-identical to the one the shiki branch
already carries (the downstream trees were green for this reason).
Responding to ds-review-bot round 2 on #661:
- logWork is bounded: past maxParallelSubCalls pending shaped-append tasks
the ordered commit lane holds (Promise.race drains one), so a slow spill
backend backpressures the run instead of accumulating unbounded pending
I/O and retained results. Tasks self-remove on settlement; run
settlement still drains every task inside the open turn. New spill test
drives three oversized reads against a hung backend at cap 1 and proves
the third dispatch cannot start until a save drains.
- shapeDispatchLog's catch uses errorMessage() (total), so a thrown value
with a throwing toString cannot escape the containment and lose the
settle event.
- CodeDispatchLog.content documented as the RENDERED result projection
(native tool/result vocabulary), not what the program received — the
program gets the structured value; doc pair + type-equiv re-synced.
- New RECORDED tui-agent snapshot scenario code-mode-dispatch-spill: the
real Loader-visible composition (worker runtime + spill-local + policy)
drives an oversized bash sub-call end-to-end; replay proves the durable
dispatch copy is bounded to preview + locator while the program value
stays whole (the outer result carries just the line count).
Agent Note updated (both languages).
The automation bridge inherited two master-era assumptions the message
machine no longer honors. A prompt blocked at pre-turn admission opens no
turn, so no turn/end could ever settle it — the bridge now watches
whenIdle() and reports a turnless slot as cancelled (the disposed-agent
guard moved to a registry identity check before send). A failed turn no
longer rejects at its turn/end either: agent.retry() closes the failed
turn and opens a successor on the same history, so the bridge holds the
terminal error and lets a retry-triggered turn/start adopt the prompt,
rejecting only at quiescence with no successor.
Also: refresh the empty-response-retry fixture for retry-as-turn logging,
adapt master-side tests to the unified send()/UserMessageData API and
registry-fact disposal, resync the doc pairs both sides touched, trim
architecture.md back under its word ceiling, and regenerate the event and
persistence catalogs.
The previous hygiene fix re-serialized with json.dumps defaults (spaced
separators), leaving the line byte-identical; explicit compact
separators make the header-scrub guard pass.
The re-records for the SDK-prompt change had harvested live v4-pro
headers while every replay overlay pins v4-flash, so keyless replay
diverged on provenance; keyless refresh reconciles the affected
scenarios (dispatch-start pairs preserved).
This branch intentionally appends plugin-sourced context (plan-mode
notices, the skill catalog) after the prompt, so the request's last message
is no longer the user's text. The scripted adapter keyed every scenario off
messages.at(-1) only, missed its triggers, and fell through to the
ask_user_question default — the frozen 'Waiting for the first token' PTY
timeouts. It now scans all text since the last assistant message.
The scripted (non-recorded) fixtures' run_code calls predate the required
description parameter, so replay rejected them at validation before any
dispatch: patch the scripted programs' args (tool/call, message blocks,
and chunk deltas together) and refresh goldens keylessly. Also picks up
the v4-pro re-records of the code-mode scenario pair whose live model
drifted from the overlay pin, and drops tmp-path churn.
The bridge replaces its serialization queue with a pool that reuses the
native concurrency contract: submissions classify through
registry.executionMode (fail-closed isConcurrencySafe), start strictly in
submission order, overlap up to the validated maxParallelSubCalls config
(default 10; 1 restores serial), and exclusive calls drain the pool, run
alone, and bar later calls. Each started sub-call logs a
tool/code-dispatch-start event at pool entry; the existing
tool/code-dispatch settles the pair (started ⇔ settles exactly once;
abandoned queued calls log neither). SDK prompt guidance now states the
true Promise.all contract — re-recorded across every code/both-mode
snapshot (plus the stale cordis-dynamic-toolchain fixture gaining the
required description arg).
Client: CodeSubCall widens to RunningToolCall | ToolResultNode — starts
land the running shape (rows wear the native running ring), settles
replace in place preserving start order, callTime pairs to the start
time. Fixture emits start/settle pairs; jsdom pins the running sub-row;
runtime specs pin in-place settlement and out-of-order completion.
The tasks/ family now matches the capability-seam shape: @deepseek-ai/dsh-tasks
keeps the abstract TaskService (ctx.tasks contract, vocabulary types, snapshot
invariant companion) and the new @deepseek-ai/dsh-tasks-local carries the
process-local registry (LocalTaskService: in-memory store, settlement,
owner-cleanup effects, teardown, TASK_WAIT_TIMEOUT). Compositions and test
harnesses now load dsh-tasks-local; producers, TaskKindMap merges, and
dsh-tool-tasks keep importing the seam only.
Producer misconfiguration diagnostics name dsh-tasks-local because loading the
implementation is the fix. The registry behavior suite moves to tasks-local;
the seam keeps a stub-subclass registration test and the probe-based invariant
suite.
The origin/master merge added the session-query-spill and
escalation-approved acp scenarios, whose pinned tool-schemas still
carried additionalProperties: true on the todo_write item schema. This
PR tightens that to false (model-visible via the request header), so
re-record it in the two new expected outputs. session-sandbox-root,
escalation-rejected, and fs-escalation-approved compare against the
escalation-approved pinned header and pass once it is fixed.
