Disclose the exit tool progressively instead of following every plan-mode
transition. The plan projection records whether a logged plan/mode ever
activated plan mode (stateVersion 4), and the tool's condition admits it
for a step whose selected state is active or once the Session has entered
plan mode. Leaving plan mode now changes only the guidance section, so an
addition-only route no longer drops a declaration from its cached prefix.
A call while plan mode is not committed still fails as unavailable outside
plan mode, and a Session that never entered plan mode does not list the
tool. The Web plan-review fixtures return to master's recording because the
request tool lists no longer change on exit.
ctx.tools.register(definition, { when }) adds a synchronous or asynchronous
condition. AgentLoop resolves conditions once per proposed step with the
agent, the model selection that step preparation captured, and the turn
signal, before prompt assembly and agent/pre-step. The resulting
ResolvedToolSet governs tool declarations, Native dispatch, PTC bindings,
and nested calls for that step and its retries, and the turn disposes it
when the proposal is rejected or fails, when no step opens, and after the
step. A false condition removes the tool everywhere; a thrown condition
ends the turn with TOOL_ACTIVATION_FAILED. Calls whose registration was
replaced or removed during the step return UNKNOWN_TOOL.
system-prompt/assemble now transforms PromptPresentation without a tool
array, so tool declarations stay provider-owned. exit_plan_mode is declared
only for steps whose selected plan state is active, the same state that
renders plan guidance; it still rejects a call before the selection is
logged.
Model selection stays as master defines it; this change adds no model
selection event or Agent field.
Store one node table keyed by the 16-hex digest prefix of each node's
canonical subgraph instead of per-root graphs with numeric indices, so a
type change rewrites only the changed nodes and their ancestors.
Refs #6597
`imageLimits` is a boot constant every Session shares: the unit stores no
per-Session state (`stateSchema: z.null()`, `apply: state => state`) and its
view returns `ctx.attachments.imageLimits`. A list row therefore repeated one
host-wide capability fact on every row. It now carries the summary mark like
the three log-derived units, so a summary read skips it while an opened Session
still receives it through `session/follow`, the control baseline, and
`session.projections`.
No other logic changes: the registration, its schema, its constant view, and
the host-side prompt validation keep reading `ctx.attachments` directly.
`docs/persistence-schema.json` records the declaring source location of every
persisted type. Adding the summary-exclusion import and map-entry JSDoc to
`packages/todo/tool-todo/src/types.ts` moved two declarations down by two
lines, so the generated catalog is regenerated. No persisted type, field, or
history record changes: `verify-persistence-changes` reports 65 roots matching
16 history records, and the release/format catalogs are unchanged.
* fix(agent-instructions): budget the user-global and project chains separately
The user-global and project instruction files shared one maxBytes and the user-global files were the broadest, so a large project chain omitted them first. Each chain now renders under maxBytes on its own and the results share one system-reminder frame: unchanged output when everything fits or one chain is empty, at most twice maxBytes otherwise. The project chain keeps its omit-broadest-first, truncate-last rule; reconciliation batches follow the same split; the maxBytes key and the baseline identity are unchanged.
Tests cover the fast paths, both chains fitting only separately, each chain over its own budget, the 2 × maxBytes boundary at exact and tiny limits, multibyte content, reconciliation batches, and a resumed single-budget baseline receiving the user-global file as one transition. The agent-instructions-budget snapshot is refreshed and agent-instructions-budget-chains is added. README, Config.maxBytes JSDoc, the configuration catalog, a new Agent Note, and an upgrade guide state the per-chain bound.
* fix(agent-instructions): keep loaded order in merged frames and note the unshared chain budget
A merged frame now lists sections, omissions and truncations in the loaded order, so a reconciliation batch renders its items in one order whether or not it fits a single budget. The guard comment states only the invariant and points to the Agent Note; the intro selection states its rule. The README names the unshared budget between chains.
---------
Co-authored-by: ZiyaZhang <[email protected]>
Add @deepseek-ai/dsh-agent-controller, composed in the base bundle, as the
entry through which a sender delivers a message from outside an Agent into
its inbox. deliver() runs the sender's content preparation in the Agent's
delivery order, asks the agent-controller/admission waterfall for a
decision, rechecks cancellation and registration, and appends the message
with followup() or steer(). dsh-agent and agent-loop are unchanged.
The Web session/prompt Remote delivers through it: the image model check,
upload receipt resolution, and attachment storage run as prepare, the
receipt binding runs as beforeInsert, and refusals map to
session/prompt-rejected. The Web client keeps its existing echo.
This replaces the Host half of deepseek-harness/deepseek-harness#6178,
whose admission stage lived in dsh-agent.
* fix(web-search-deepseek): disable thinking for auxiliary search requests
The provider discards the generated text of the auxiliary Messages turn,
so reasoning tokens have no consumer. DeepSeek enables thinking by default,
which adds latency and output tokens to every search.
Fixes#6099.
* feat(web): add opt-in traditional Search and Fetch APIs
* test(web): verify API settings persistence and refresh Session evidence
* fix(web): require an explicit Search API base with its route prefix
* docs(web): document direct Search and Fetch API decisions
* fix(web): keep Messages search default behind the new Web Search API switch
- Rename useTraditionalApi/traditionalBaseURL to useSearchApi/searchApiBaseURL and
label the switch "Use the new Web Search API"; Messages search stays the default.
- Default the Search / Fetch API base to https://api.deepseek.com/v1/tools.
- Authenticate the Search API with an API key only: DeepSeek Account tokens
returned an empty HTTP 404 on /v1/tools/* when probed on 2026-10-16.
- Remove the count setting, which tool-web's maxResults always overrides, and
the undocumented DeepSeek-API-Version header.
- Name the active API's base URL field and the current settings path in
endpoint guidance; re-record the endpoint guidance Session.
- Keep WebFetchResult.statusCode required: DeepSeek Fetch reports HTTP 200 and
drops the page title, so the fetch-metadata widening and its upgrade guide
are removed.
* fix(web): accept account sign-in on the live Search API
The public Search and Fetch routes went live on 2026-10-16 and accept the
DeepSeek Account token, so account-route Sessions send it to the Search API
as they do for Messages search, and the settings copy no longer says account
sign-in does not apply.
Record the live behavior in the README and Agent Note: summaries of
1,000-2,300 characters, an HTTP 500 for an unresolvable Fetch target, and
migration notes for gateways without /v1/tools routes. Test doubles now use
the live response fields and RFC 3339 offsets.
* feat(web): configure Search API result count and send the triggering question
- Add web-search-deepseek.count (1-50, default 10, the service default) and
show it as "Results per search" in Search API mode, where "Max searches per
request" applies only to Messages; a smaller request maxResults still wins.
- Send the initiating Session's latest human prompt as ori_question, with
whitespace collapsed and cut to the API's 128-character limit; blank or
absent prompts omit it, and sendOriginalQuestion: false disables it. The
settings hint states that the question is sent.
- Match the web search double on the request's query field, since
ori_question now carries every query in the prompt.
* fix(client): unify web search endpoint override controls
* feat(web): configure registered search and fetch providers through shared settings
* feat(web): include Exa and Perplexity in provider settings
* test(web): cover provider settings branches and refresh plugin snapshots
* revert(web): drop the provider dropdown, schema-driven form, and Exa/Perplexity settings
Revert aed8c5cb609, d88c47dcd80, and 036a5a337d3. The Web search settings
page returns to the DeepSeek toggle; the DeepSeek Search and Fetch API split
is reintroduced as one plugin in the following commit.
* feat(web): move the DeepSeek Search and Fetch APIs into web-deepseek-api
The new web-deepseek-api plugin registers deepseek-search-api and
deepseek-fetch-api over one credential and API base, and exports the
authentication and redirect-rejecting transport that web-search-deepseek now
reuses for Messages search only. The web service reads searchProvider and
fetchProvider for each operation, so the Web search settings switch selects
both APIs together on save and restores Messages search with local HTTP fetch
when turned off. The switch hint links to the DeepSeek search and fetch guide.
* refactor(client): keep the web search endpoint fields on the shared override controls
Drop the always-visible reset and the inherited-address placeholder from the
Web search page, restoring ui-primitives to its base. Each endpoint field shows
the selected section's effective value and offers Reset to default only for an
override, like every other settings field.
* refactor(web): keep Messages search independent of web-deepseek-api
Restore dsh-web-search-deepseek to its own credential and transport code,
keeping only the disabled-thinking request and the current settings guidance,
so the Messages plugin no longer depends on the Search and Fetch API package.
dsh-web-deepseek-api keeps its transport internal; the regions that repeat the
Messages helpers carry jscpd:ignore markers, with the rationale recorded in the
Search and Fetch API Agent Note.
* refactor(web): keep the web-deepseek-api request type module-private
---------
Co-authored-by: LegGasai <[email protected]>
Add @deepseek-ai/dsh-agent-model. It provides ctx.agentModel
(resolveForNextStep, select, applyConfiguredRoute) and owns the
modelSelection projection, which the Agent registry no longer registers.
dsh-agent-loop now injects agentModel; the base and SDK-minimal bundles
and the Python SDK runtime include it.
The provider keeps the projection registered until consumers that are
still unloading finish their cleanup. The model-switch notice becomes an
opt-in per-Agent plugin at @deepseek-ai/dsh-agent-model/notice, and
agent/request now also rejects reasoning-effort changes after step
preparation.
Mount the provider in every composition that loads dsh-agent-loop, and
update consumers, the switch-notice snapshot driver, docs, and the
upgrade guide.
Yichen Jiang