* fix(ai/gemini): coerce numeric/boolean enum values to strings
Gemini Function Calling only accepts string enum values (see
https://ai.google.dev/gemini-api/docs/function-calling). Tool schemas
containing enum: [1,2,3] or enum: [true,false] were passed through
unchanged, causing Gemini to reject the entire request with 400
TYPE_STRING errors at 'function_declarations[X].parameters...enum[N]'.
Extend cleanSchemaForGeminiWithDefs to coerce enum and const values
via a shared helper (number/boolean/bigint -> string, drop null and
non-coercible entries, de-duplicate), and retype the containing schema
as 'string' since Gemini's validator only accepts enum on string-typed
schemas.
Adds seven vitest cases covering numeric/boolean enum, const, nested
schemas, and dedup/empty-after-coercion edge cases.
Fixes#104566
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(ai/gemini): make enum retyping property-order independent
Address @clawsweeper review on #104567:
The initial patch set cleaned.type = 'string' inline while visiting the
'enum' or 'const' key. For schemas where 'type' appears AFTER 'enum' /
'const' in the input object — a valid ordering since JSON Schema property
order is not semantic — a later 'type' pass would overwrite the coercion
via the tail 'cleaned[key] = value' branch, and Gemini would still receive
an invalid declaration.
Move the type override out of the loop: track whether any enum/const was
coerced via a boolean flag, and force cleaned.type = 'string' after the
loop finishes. This makes the fix independent of input property order.
Also add curly braces around single-line ifs to satisfy check-lint
(eslint(curly)) which failed on the previous commit.
New reversed-order regression tests added.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(ai): preserve Gemini enum schema types
---------
Co-authored-by: leonme <leonme@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* feat(tooling): extend strict-ratchet lane to all remaining leaf packages
* fix(packages): burn down indexed-access debt in ai and agent-core
All 196 noUncheckedIndexedAccess errors fixed behavior-identically
(iteration, .at()/slices, DataView byte math, guarded queue peeks) plus
cleanup finds: a sparse-tool diagnostic bug, deduplicated Responses
call/item ID parsing with an honest string|undefined splitter, and a
clearer local failure for redacted-thinking blocks missing signatures.
Removes all 17 remaining non-null assertions across ai, gateway-client,
and llm-core so the assertion ban aligns 1:1 with the ratchet lane.
speech-core joins memory-host-sdk as structurally deferred (imports
src/** via plugin-sdk path mappings).
* feat(agents): derive a provider-declared default utility model when unset
When agents.defaults.utilityModel is not set, utility tasks (titles, progress
narration) now use the primary provider's declared small model
(modelCatalog.providers.<id>.defaultUtilityModel: OpenAI -> gpt-5.6-luna,
Anthropic -> claude-haiku-4-5). Auth is inherent because derivation follows
the agent's primary provider. Setting utilityModel to an empty string
disables utility routing entirely; narration turns on automatically when a
default resolves and stays off otherwise.
Formatting/lint/tests verified on Testbox (oxfmt, oxlint, plugins:inventory,
docs:map, import-cycles, check:test-types, 4 test files).
* fix(ai): drop the temperature parameter for models that reject it
The GPT-5.6 family 400s on temperature via the Responses API (live-verified:
gpt-5.6-luna/-terra reject it; gpt-5.5 and gpt-5.4-mini/nano accept it).
supportsOpenAITemperature gates all three OpenAI payload builders, with a
catalog compat override (compat.supportsTemperature) declared for the 5.6
family in the openai manifest. Without this, utility tasks (titles,
narration) would fail once gpt-5.6-luna becomes the derived default.
Gates on Testbox: oxfmt, oxlint, plugins:inventory, import-cycles,
check:test-types, 6 test files. Live proof on Testbox: gpt-5.6-luna and
claude-haiku-4-5 one-shot completions succeed with temperature requested.
* fix(agents): carry the primary model's auth profile onto the derived utility default
A primary like openai/gpt-5.5@work previously reached utility tasks via the
profiled fallback; the derived default shares the provider, so its trailing
auth profile carries over instead of silently switching to default
credentials (Codex review). Gates on Testbox: oxfmt, check:test-types, tests.
* docs: realign the manifest provider-fields table after adding defaultUtilityModel
* fix(ai): cache and case-insensitively match Azure deployment map
resolveAzureDeploymentNameFromMap re-parsed the deployment-map string into
a new Map on every call (a hot path: streams and lifecycle hooks) and
looked model ids up case-sensitively. A request for `GPT-4o` against a
`gpt-4o=deployment-gpt-4o` map therefore fell back to the raw model id and
404'd on Azure ("deployment not found").
Cache the parsed lookup map keyed by the raw deployment-map string (bounded
so memory stays flat) and normalize keys to lowercase so lookups are
case-insensitive. Deployment names (the values) stay verbatim because Azure
requires the exact deployment name, and the fallback still returns the
original-cased model id.
Fixes#102936🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(ai): prefer exact-case match, case-insensitive only as fallback
Addresses review: lowercasing every key regressed configs whose deployment
map distinguishes keys by case (e.g. `GPT-4o=prod-a,gpt-4o=prod-b`), where an
exact `GPT-4o` request previously resolved to `prod-a` but would now collapse
to the last lowercased entry.
Cache an exact-case lookup alongside the lowercased one and resolve exact
first, using the case-insensitive map only as a fallback. Every previously
working exact mapping is preserved; the case-insensitive path only rescues
ids that would otherwise 404. Adds a regression test for exact-case
precedence.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* test(ai): prove Azure deployment map resolves on the real provider path
Adds an integration test that drives the real streamSimpleAzureOpenAIResponses
path against a loopback server (via the AI transport host fetch) and asserts the
deployment name resolved from AZURE_OPENAI_DEPLOYMENT_NAME_MAP is what lands in
the outgoing request `model` field — the value Azure routes on, and the source
of the reported 404 when it fell back to a mixed-case model id.
Covers the mixed-case fix (GPT-4o -> deployment-gpt-4o on the wire) and the
exact-case no-regression case (GPT-4o -> prod-a).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* test(ai): use block-body promise executors to satisfy oxlint
no-promise-executor-return flagged the arrow executors around server
listen/close; use statement bodies so nothing is returned from the executor.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* refactor(ai): simplify Azure deployment lookup
* refactor(ai): simplify Azure deployment lookup
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <peter@steipete.me>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(mistral): use truncateUtf16Safe for error text truncation
Replace naive .slice(0, maxChars) with truncateUtf16Safe() in
truncateErrorText() to prevent surrogate pair splitting in
Mistral provider error messages shown to users.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(ai): keep Mistral errors UTF-16 safe
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* feat(secrets): resolve SecretRef model credentials at egress via process-local sentinels
SecretRef-managed model-provider credentials now travel as opaque
oc-sent-v1 sentinels through auth storage, stream options, and SDK
config; the guarded model fetch injects real values into headers and
URLs immediately before the SSRF-guarded send and fails closed on
unknown sentinels. packages/ai adapters converge on the host guarded
fetch where the SDK supports custom fetch and unwrap at construction
where it does not. Resolved values (and their percent-encoded forms)
register for exact-value log redaction. Kill switch:
OPENCLAW_SECRET_SENTINELS=off. Also fixes a pre-existing unhandled
rejection race in capNonOkResponseBodyLazily (pipeThrough writer leak).
* test(plugin-sdk): update public surface budget
* fix(anthropic): resolve thinking as disabled when legacy budget is below 1024
When adjustMaxTokensForThinking collapses the thinking budget below the
Anthropic minimum (1024), option resolution now sets thinkingEnabled to
false instead of always forcing it to true. This keeps every downstream
consumer (payload, replay, temperature, tool-choice) consistent — they
all see the same disabled state instead of an enabled flag with a
missing or API-rejected thinking block.
|| → ?? in both builders is defensive: the resolution layer already
prevents invalid budgets from reaching the builder through the normal
path, but ?? preserves an explicit zero when the builder is called
directly.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix(anthropic): guard raw streamAnthropic builder path against sub-minimum budgets
Add budget guards in both Anthropic payload builders so direct
streamAnthropic and bundled-plugin callers (e.g. Mantle) that bypass
option resolution also get the disabled-state rule instead of producing
API-rejected { type: "enabled", budget_tokens: < 1024 } requests.
Add proof-anthropic-thinking-budget.mts driving real production
functions with terminal output and negative control.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix(anthropic): normalize legacy thinking budgets
Co-authored-by: Pick-cat <huang.ting3@xydigit.com>
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
The active user turn was decorated at request-build time with an inbound
metadata block (Conversation info / reply-target / sender / forwarded /
chat-history), but the session store persists the bare text and the LLM boundary
strips that block from historical replay. So a user message's serialized bytes
changed retroactively — decorated as the active turn, bare as history — on every
turn, invalidating any prefix/exact-match provider prompt cache from that point
onward, every turn. (#90811 fixed the timestamp half of this same asymmetry.)
Rework of the earlier "strip the active turn" approach (which lost model-visible
context — ClawSweeper's P1) into one that keeps the context AND recovers the
cache cost:
- Route current-turn inbound metadata out of the user prompt and into the hidden
runtime-context custom message (the mechanism from openclaw#89428), and
relocate that carrier to the ABSOLUTE TAIL of the wire request (after the
active user turn and any tool-call scaffolding). The user turn stays bare and
byte-identical in both the active and historical positions; the volatile
carrier vanishes next turn exactly where the assistant reply begins anyway, so
the request is an append-only prefix-extension through the active user turn.
- A durable marker (UserMessage.runtimeContextCarrier, set by convertToLlm from
the carrier's details) lets the Anthropic SDK transport, the managed Anthropic
transport (anthropic-transport-stream / anthropic-payload-policy), and
OpenAI-completions skip the carrier when selecting the deepest cache_control
breakpoint, keeping the anchor on the last stable user turn.
Storage stays BARE (invariant preserved); runtime-only turns (room events) keep
their existing inline behavior, which is byte-stable because room-event/system
context is not strip-eligible; legacy bare-stored sessions are unchanged.
Reviewed with adversarial gpt-5.5/codex sweeps (correctness, byte-identity,
compatibility, test adequacy): found + fixed the managed Anthropic transport
cache-anchor path and added the missing OpenAI-completions/managed-transport
breakpoint tests; a theoretical runtime-only mutation was verified unreachable
and pinned with a test. Final sweep clean.
Co-authored-by: openclaw#89428 (runtime-context handoff approach)
The ChatGPT backend routes requests by session_id (codex-cli sends it); without it each request lands on an arbitrary machine and the prompt cache misses. Measured on live traffic: 56.0% -> 81.8% TRUE cache hit rate (12-turn A/B, identical prompts).
* fix: preserve Bedrock live API providers
* fix: derive Bedrock smoke region from AWS config
* fix: honor Bedrock discovery region in smoke
* fix: keep Bedrock live smoke on Bedrock runtime
---------
Co-authored-by: openclaw-clownfish[bot] <280122609+openclaw-clownfish[bot]@users.noreply.github.com>
* fix(mistral): correct model id typo in usesReasoningEffort helper
* chore: re-trigger CI
* test(mistral): cover medium reasoning effort
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* refactor: extract reusable AI runtime package
* refactor: complete AI provider relocation
* refactor: keep llm core internal
* refactor(ai): make @openclaw/ai self-contained with host policy ports
Move pure transport helpers (tool projections, strict-schema normalization,
prompt-cache boundary, stream guards, anthropic/openai compat, request
activity) from src into packages/ai; move utf16-slice into
normalization-core. Inject host policy (guarded fetch, redaction,
strict-tool defaults, diagnostics logging) through AiTransportHost with
inert library defaults installed by src/llm/stream.ts. Narrow the public
barrel to instance-scoped createApiRegistry/createLlmRuntime; the
process-default runtime moves behind internal/ and
registerBuiltInApiProviders takes an explicit registry. Delete the
src/llm/api-registry re-export facade.
* fix(ai): teach node, jiti, and vite resolvers the @openclaw/ai and utf16-slice subpaths
The workspace alias tables in root-alias.cjs, plugin-sdk-native-resolver,
sdk-alias, the shared vitest config, and the Control UI vite config only
knew @openclaw/llm-core; Node-side plugin loading resolved @openclaw/ai
through the pnpm symlink to the unbuilt dist (checks-node-compact CI
failures), and the Control UI build broke on the new
normalization-core/utf16-slice subpath.
* chore(ui): drop leftover service-worker debug logging
* build(release): ship @openclaw/ai with its own shrinkwrap and honest dependency set
packages/ai declares only its six real runtime deps (kysely, chalk, json5,
tslog, zod, fs-safe, and proxyline were never imported); orphaned root deps
removed. generate-npm-shrinkwrap now treats publishable packages/* like
publishable plugins so the AI tarball pins its transitive tree even though
workspace deps are omitted from the root shrinkwrap. knip learns the
package entry points; the tsdown dts neverBundle option moves to its
documented deps.dts home; the README documents the no-semver internal/*
contract and host ports.
* docs(ai): add minimal external-consumer example app
examples/ai-chat consumes only the public @openclaw/ai surface (built dist
via the workspace link): isolated runtime, built-in provider registration,
one streamed completion. Supports Anthropic/OpenAI via env keys and a
keyless local Ollama target; live-verified against Ollama.
* docs(ai): document the @openclaw/ai package and workspace shrinkwrap boundary
* chore(check): include examples/ in duplicate-scan targets
* fix: emit normalization package subpaths
* fix: complete AI package boundary artifacts
* fix: align AI package boundary contracts
* fix(ci): stabilize package release contracts
* test: align documentation contract checks
* test: keep cron docs guard aligned
* test: align restored docs contract guards
* test: follow upstream docs contracts
* docs: drop superseded talk wording
