The DeepSeek search and Search/Fetch API providers now report readiness
from credential presence alone, on any model route: the account token
first on the deepseek-account route, then a literal apiKey or a
configured apiKeyEnv reference. A conversation on another provider's
model keeps web_search and web_fetch while a DeepSeek key is configured,
as on master, so the route refusal WEB_PROVIDER_ROUTE_DISABLED and its
upgrade guide are removed.
The pi-ai-conversation-updates and deepseek-messages-system-prompt
snapshots replay their master recordings again, the headless keyless
smoke points its search key reference at an unset variable, and the
readiness Agent Note records the rejected route and endpoint rules.
A user who gives DeepSeek search a key of its own has chosen what that key
pays for, so the route check now applies only to the shared
DEEPSEEK_API_KEY. A literal apiKey, or an apiKeyEnv other than
DEEPSEEK_API_KEY, opens every model route for Messages search and for the
Search and Fetch APIs; a configured endpoint does not, since the search
still sends the shared key. The Web search settings page saves its key
under the section's apiKeyEnv, DEEPSEEK_API_KEY by default, so a key saved
there keeps serving only the DeepSeek routes.
The route refusal now names the settings that give search a key of its
own, and the READMEs, the readiness Agent Note, and the upgrade guide
describe the exception.
Review found two faults in the endpoint and configuration checks. The
providers took the official chat route's endpoint from the first active
dsh-llm-deepseek-api-key Loader row, so a composition that isolates several
instances answered for an unrelated one: a correct route lost search, or a
gateway's key reached DeepSeek. And a key saved on the Web search settings
page lands under the default DEEPSEEK_API_KEY reference, so the inference
that search had its own configuration never saw it.
Both DeepSeek web packages now serve only the deepseek-account and
deepseek-official routes, whatever key they are configured with, and no
longer compare the chat endpoint with their own. ready() checks the route
and then credential presence as before; the API-key resolver repeats the
route check and refuses with WEB_PROVIDER_ROUTE_DISABLED. The packages no
longer read Loader rows or depend on dsh-llm-deepseek-api-key, so the
lockfile, the module graph, and the keyless smoke expectations that the
endpoint check had changed return to master's.
An upgrade guide covers conversations on other providers' models, which
lose DeepSeek web search and, with the Search and Fetch APIs selected,
web_fetch.
The Web search switch selects the DeepSeek Search and Fetch APIs together,
and both reuse the DeepSeek chat key by default. Neither provider answered
readiness, so a step without a usable key was offered web_search and
web_fetch whose every call failed, and a conversation on another provider's
model or behind a gateway sent the chat key to DeepSeek.
dsh-web gives WebFetchProvider an optional ready(llmProvider) and adds
ctx.web.fetchReady(), which asks only the provider fetch() selects, as
searchReady() does for search. tool-web registers web_fetch with a step
activation condition on fetchReady(), so a step that cannot fetch leaves the
tool out of declarations, guidance, and PTC bindings and refuses a call to
it as UNKNOWN_TOOL; without web_fetch the search guidance uses its
search-only text. The shipped HTTP fetch provider has no ready() and stays
admitted.
web-deepseek-api applies the API-key rules of Messages search to both APIs:
without a literal apiKey, an apiKeyEnv other than DEEPSEEK_API_KEY, or a
baseURL other than the default, the chat key goes only to the DeepSeek
routes and only to an API base on the official chat route's origin. The
refusal sits in the key's only resolver, and both providers' ready()
answers presence in the order an operation resolves credentials. The
transport repeats these checks so the package still does not depend on
dsh-web-search-deepseek.
Master's step-scoped conditional activation resolves each tool's condition
before a model step assembles, and both the declarations and dispatch read
that admitted set. Prompt assembly no longer carries tool declarations, so
tool-web's system-prompt/assemble listener could not remove web_search any
more, and removing a declaration there never stopped a repeated call from
running.
tool-web now registers web_search with a when condition that asks
ctx.web.searchReady() about the route of the step's model selection, and its
guidance section renders from the step's admitted set. A step that does not
admit the tool leaves it out of declarations, guidance, and PTC program
bindings, and refuses a call to it as UNKNOWN_TOOL without reaching a
provider. Tests that execute web_search directly now run inside a resolved
step, as the executor requires for conditional registrations.
Docs, the readiness Agent Note, and the event producer/consumer table follow
the new mechanism; the tools README gains an explicit conditional-tools
anchor for cross-locale links.
The DeepSeek Search API returns a short `snippet` for display and a longer
`summary` written for model reading. Carry `summary` on `WebSearchSource`,
map it from the Search API, and render it in place of the snippet in
`web_search` output while the search card keeps the snippet.
`tool-web` collapses summary whitespace and caps each summary at the new
`searchSummaryMaxChars` setting (800 characters by default, the upper end of
the guide's typical range; live summaries measured 1,000-2,300). The cut keeps
surrogate pairs intact and ends in an ellipsis. The PTC tool output type gains
the optional field, so the affected system-prompt sidecars are refreshed, and
the web search round now exercises a multi-line summary.
The HTTP fetch provider resolved each hostname, rejected any non-public
answer, and pinned the connection. That refused every page behind
TUN-mode proxies with fake-IP DNS and did not confine a model that has
shell network access. Requests now go through the installed proxy
dispatcher and the system resolver with no address restriction.
HttpFetchResolver and the HttpFetchProvider resolver parameter are
removed; the upgrade guide covers the migration.
Co-authored-by: turtle1999 <[email protected]>
Replace the `routes` allowlist with the user's own configuration. A search
whose key or endpoint the user configured (a literal apiKey, an apiKeyEnv
other than DEEPSEEK_API_KEY, baseURL, or $DEEPSEEK_SEARCH_BASE_URL) serves
every model route without the route or endpoint check. Without such
configuration the key is the DeepSeek chat key: only the DeepSeek account and
official routes may use it, and only toward a search endpoint on the official
chat route's origin. The account token rule stays as before.
Both refusals now name the settings that give search its own setup.
DeepSeek native search is served by DeepSeek's endpoint, so two checks now
gate the API key, in ready() and again in the key resolver of every search:
- Route: a model request on a route other than deepseek-account,
deepseek-official, and the new `routes` setting is refused with
WEB_PROVIDER_ROUTE_DISABLED, so another provider's conversation does not
bill DeepSeek unless a deployment lists the route.
- Endpoint: the official chat route's own key reference goes only to a
search endpoint on that route's chat origin, read from the active
dsh-llm-deepseek-api-key Loader row with that package's resolver;
otherwise WEB_PROVIDER_ENDPOINT_MISMATCH. A literal key, a separate key
reference, or a search endpoint on the chat origin passes.
The provider passes a resolver's own WebError through unchanged. A real
Loader composition pins the endpoint check. The pi-ai-conversation-updates
scenario gets its own tool sidecar without web_search, and the legacy
deepseek-messages route loses web_search in deepseek-messages-system-prompt.
Delete @deepseek-ai/dsh-invariants, every package's ./invariant companion
and its tests, the invariant gates and Vitest setup, the scoped-event
resolver generator, and the INVARIANT-coded listener rethrow paths.
Update docs, READMEs, Agent Notes, and add an upgrade guide.
On the account route a search reads the account grant before any API key
and fails with WEB_PROVIDER_ERROR when that read throws; it never falls back
to the key. ready() swallowed the failed read and then accepted a literal,
stored, or environment key, so it answered true for a search that fails.
ready() now checks the account token first on the account route, answers
false when that read fails, and only then checks the literal key and the
credential reference. Tests pair the readiness answer with the real search
failure for a literal key and an environment key.
Move per-request web_search withholding out of the DeepSeek provider:
- dsh-web: add the optional WebSearchProvider.ready(llmProvider) and
ctx.web.searchReady(llmProvider?), which asks only the provider that
search() would select; a provider without ready() and a selection
failure answer true.
- dsh-web-search-deepseek: implement ready() with the existing credential
check and drop the assemble listener together with its system-prompt,
tools, and tool-web dependencies.
- dsh-tool-web: remove web_search and its tool:web_search section in a
prepended system-prompt/assemble listener when searchReady() answers
false for the route model selection assigned to the request.
A composition that selects another search provider no longer loses
web_search when DeepSeek has no credential. The Agent Note is renamed to
2026-09-27-web-search-offered-by-provider-readiness and records the split.
The DeepSeek search provider now removes dsh-tool-web's web_search tool and
its prompt section from each model request whose search would have no
credential: no account token for the request's deepseek-account route and
search endpoint, no literal apiKey, and no configured apiKeyEnv credential.
A prepended system-prompt/assemble listener reads the route model selection
writes for the request, so a model switch or a newly stored key applies to
the next request through the existing dynamic tool updates. PTC
presentations and execution-time authentication are unchanged.
Tool guidance sections keep only what the tool definitions do not say.
Apply review: reword interrupt_agent and list_agents, carry "a rejected
escalation stays denied" in the rejection error, pin
sandboxPermissionsDescription, and fix stale README and Agent Note
statements. Record the tool-definition rules in agent-experience.
Compaction threshold snapshots shift their context windows by the
370-token system-prompt estimate reduction; escalation-rejected moves to
a v4 fixture for the new rejection text.
git mv timeout-policy beside repeat-tool-guard (both are loop-hygiene
policies on the tool-execution pipeline, and the timeout/ group name
collided with util/timeout) and tool-cordis into self-modification/
(naming the role the framework name obscured). Merged/renamed group
README triplets, tsconfig globs, generator sources, hierarchy tables,
catalogs, and the timeout-policy design note's group references
follow. Adds the fifth FIXME marker (dsh-timeout-guard, recorded as a
suggestion to settle at resolution time).
guard + self-modification suites: 197 passed.
Address the code-review bot findings on the web result card:
- web_fetch's card truncated now derives from the shared renderFetchOutput
helper, matching the effective truncation the model-facing text reflects
(provider cap, source cut, or output cap), instead of the provider-only flag.
- Drop the redundant content copy from both web result views; a UI without the
web capability falls back to the raw tool/result content. Narrow the TUI
transcript view.content access accordingly.
- Set the result-state title from the call args (query/url) so a window-
truncated replay keeps a title.
- Project meta from the seam result types rather than hand-rolled value types.
- Sync the card vocabulary across core tools README, docs/core-data-structures,
the adding-a-tool cookbook, and the tool-web package README (both languages,
re-recorded pairings); regenerate the cordis api-catalog and cordis-inspect
snapshot; revise the Agent Note.
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.
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.
Yichen Jiang