Anthropic's tool search compatibility table lists Haiku 4.5 as supported,
but the capability gate denied all Haiku. Sonnet 5 and Mythos 5 were also
missing from the tool's model picker list.
Flip the defaults for two settings that have been validated behind an
opt-in flag:
- `chat.agentHost.copilot.toolSearch.enabled`
- `github.copilot.chat.preferLongContext.enabled`
Both settings are defined in more than one place, so each pair is
updated together:
- Tool search: the workbench setting registration in
`chat.shared.contribution.ts` and the agent-host schema default in
`copilotCliConfig.ts` that the renderer forwards into.
- Prefer long context: the contributed default in the Copilot extension's
`package.json` and `ConfigKey.PreferLongContext`, which throws a
`BugIndicatingError` at load if the two drift, plus the forwarded
`preferLongContextEnabled` root-config default in `agentHostSchema.ts`.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
nes: fix: preserve ghost-text state for speculative reuse
Keep speculative next-edit results linked to their stable cache entry so suggestions shown as ghost text remain suppressible when reused in another view kind. Add coverage for speculative reuse, rebasing, and cache replacement.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 93fb295d-2415-4079-b354-784417ad06f7
* Coalesce content exclusion fetches to stop exhausting the GitHub API rate limit
Every caller that discovered a new repository triggered a refresh of the
content exclusion rules for *every* known repository, so request volume grew
quadratically with repository count. In a workspace with many git repos (an
AOSP checkout, in the reported case) this produced ~16k requests to
api.github.com in 30 minutes, exhausting the account's 5k/hour REST budget and
starving everything sharing it, including Copilot token refresh.
The endpoint is https://api.github.com/copilot_internal/content_exclusion, so
it draws on the user's ordinary REST quota rather than a CAPI budget.
- Coalesce per repository using shared DeferredPromises, a short batching
window and a bounded-concurrency Limiter, so each repo is fetched at most
once per TTL and each caller only waits on the repos it asked for. A
regression test measures 75 requests -> 3 for 26 repositories.
- Only cache rules on a successful response. Empty placeholder rules were
written on discovery and left in place on failure, making a failed fetch
indistinguishable from "this repo has no exclusions" and preventing a retry
for 30 minutes, so exclusions silently stopped applying while rate limited.
- Only memoise a negative verdict once the relevant rules actually loaded,
which otherwise left files checked during an outage permanently allowed.
- Add a shared rateLimitBackoffMiddleware covering 429 and quota-exhausted 403
responses, honouring Retry-After and x-ratelimit-reset, and move both
RemoteContentExclusion and CloudSessionApiClient onto it. This replaces a
third hand-rolled copy of the same backoff logic.
- Precompile glob patterns, track the regex rule count directly, and only
invalidate memoised results when rules that could change an outcome arrive.
Fixes#322275
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 32f94728-158b-4639-b789-a5dd483f043f
* Address review feedback on cache invalidation and request lifecycle
Five correctness issues raised in review, each with a test that fails against
the previous implementation.
- rateLimitBackoffMiddleware: a response that was already in flight could clear
a block established by a concurrent rate-limited request, letting later calls
reach the server during the window the server asked us to wait out. The
backoff is now only reset once the active block has elapsed.
- isIgnored returned a memoised verdict before reaching the staleness check, so
a URI that had been evaluated once never triggered a refresh and could miss
newly added exclusions indefinitely. Verdicts are now tagged with a rule
generation and are only trusted while the rules behind them are unchanged and
unexpired.
- applyRules only invalidated verdicts when the incoming rules were non-empty,
so a refresh that removed the last rule left files excluded permanently.
Incoming rules are now compared against the previous set, which also avoids
invalidating on an unchanged refresh.
- drainPendingRepos cleared the pending map before its batches completed, so a
lookup arriving while a request was slow queued a duplicate fetch every
batching window. Entries now stay registered until their request settles, and
are removed only if still owned by that attempt.
- dispose only settled repos still queued. Limiter.dispose drops queued
factories without running them, so with more than five batches the callers
awaiting them never resolved. Pending entries are now settled before the
limiter is disposed, and enqueues after disposal resolve immediately.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 32f94728-158b-4639-b789-a5dd483f043f
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 32f94728-158b-4639-b789-a5dd483f043f
Normalize unsupported non-string Gemini enums and restore thought signatures when replaying historical function calls.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Agent Host changes for osortega/agents/checkout-contribution-check-v2-backend
* Address review feedback and fix chat input fixtures
Require non-empty owner/name before building a pull request URI, so
metadata with empty strings no longer yields `https://github.com///pull/42`
and is no longer reported as an available pull request. Cover the case in
the unit test.
Implement `getSession` on the fixture `IAgentSessionsService` stub. The
chat input session toolbar reads it when a session has file changes, which
broke the chatInput FileChanges fixtures.
Condense the context key and chat input comments, and correct the cloud
sessions comment: the guard is reachable whenever `chatSessionPullRequest`
is unknown, not just on older clients.
Keep assistant text complete, group reasoning chunks by ID, and round-trip provider-specific continuation metadata through Responses encrypted content.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Replace the Agent Host BYOK model-options sentinel and private MIME payload with a typed provider request option while retaining encrypted thinking metadata only for opted-in requests.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* chat: enable agent host defaults
Flip the Agent Host migration settings to their rolled-out defaults and enable remote Agent Host support for web clients attached to a remote extension host. Serverless web retains extension-host fallbacks. (Written by Copilot)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test: cover web agent host disablement
Add web-remote coverage for configuration and AI disablement, and clarify the serverless web requirement in the setting description. (Written by Copilot)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test: pin smoke session providers
Pre-seed extension-host smoke profiles before startup and update Agent Host picker labels so the suites exercise their intended providers after the default flip. (Written by Copilot)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* chat: restore agent host setting description
Keep the existing concise setting and policy description while retaining the runtime availability behavior. (Written by Copilot)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Carry encrypted reasoning through a private data part only for Agent Host BYOK requests, preserving existing vscode.lm behavior for other consumers.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Preserve structured reasoning, tool calls, continuation metadata, and usage across the Agent Host renderer bridge while replacing the Chat Completions proxy contract.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* nes: feat: generate samples from workspace recordings
Parse stateful local workspace recordings, select deterministic user-edit and cursor pivots, materialize privacy-safe replay slices, and support bounded parallel datagen without splitting raw timelines.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 62cca64b-0581-410e-ab89-034e20f02ca7
* nes: fix: include cursor boundaries in sample deduplication
Hash the complete post-pivot label so identical prompts with different cursor destinations are rejected as conflicting samples.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 62cca64b-0581-410e-ab89-034e20f02ca7
* nes: fix: consolidate workspace recording imports
Use inline type specifiers so the Copilot extension lint job accepts the new workspace-recording modules.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 62cca64b-0581-410e-ab89-034e20f02ca7
---------
Copilot-Session: 62cca64b-0581-410e-ab89-034e20f02ca7
xtab: fix: clamp tagged patch range to document
Tagged prompt content can add lines beyond the source document. Clamp the CustomDiffPatch pseudo-window to the source range to avoid an unexpected failure.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 62cca64b-0581-410e-ab89-034e20f02ca7
The Gemini BYOK function-declaration converter assumed a JSON Schema
`type` was always a single string. When a tool parameter used a nullable
union such as `{ "type": ["string", "null"] }`, the array was stringified
into `"string,null"` and rejected with `Unsupported type: string,null`.
Through the Agent Host this surfaced as an HTTP 502 that the Copilot SDK
retried five times before failing the turn. Because the failure lives in
the shared converter, regular editor BYOK is susceptible to the same
schema shape.
Normalize nullable unions to Gemini's `nullable: true`, map multi-type
unions to `anyOf`, and handle nullable array items and anyOf/oneOf that
contain `null`. Add regression tests for these cases.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix: send chat image attachments when signed out of GitHub
Image attachments were silently dropped for signed-out users even when a
vision-capable BYOK/local model was configured. The panel Image prompt
element gated image inclusion on
`copilotToken?.isEditorPreviewFeaturesEnabled()`, which is `undefined`
when there is no Copilot token (i.e. not signed in), so the negation
evaluated truthy and the image was omitted before the request was sent.
`isEditorPreviewFeaturesEnabled()` represents an org policy that is only
meaningful when a Copilot token exists; a missing token should not block
the feature. Default the check to enabled when the token is absent, using
the same `?? true` pattern already used in
claudeChatSessionContentProvider. This preserves the org-policy behavior
(explicit `editor_preview_features=0` still omits images) while letting
signed-out / BYOK users send images to vision-capable models.
Applied to both HistoricalImage and Image render paths, and added unit
tests covering signed-out, signed-in, org-policy-disabled, and
non-vision-model cases.
Fixes#323854
Co-authored-by: SamirSaji <samirsaji13@gmail.com>
* fix: report accurate omission reason for policy-blocked images
Address review feedback: when a vision-capable model has images omitted
because org policy disables editor preview features, the omitted-reference
status previously said the model "does not support images", which is
misleading. Set the status description based on the actual cause
(lack of vision vs. org policy). The catch/error path keeps the generic
message.
Co-authored-by: SamirSaji <samirsaji13@gmail.com>
---------
Co-authored-by: SamirSaji <samirsaji13@gmail.com>
Co-authored-by: Dmitriy Vasyura <dmitriv@microsoft.com>
* nes: fix: share in-flight speculative requests across concurrent calls
A speculative request is fired when a suggestion is shown, betting on the
document state that results if the user accepts it. `consumePending()` nulled
the pending slot as soon as the first `getNextEdit` caller claimed the request,
so a second overlapping caller for the same post-edit state could not discover
it and issued a duplicate model request — losing the latency win, the
`ReusedRequestKind.Speculative` telemetry, and the `liveDependentants` sharing
that keeps a joined request alive for multiple dependants.
Rename it to `claimPending()` and keep claimed requests in a map until their
`result` settles, so concurrent callers join the in-flight request instead.
A map rather than a single slot: several speculatives can be claimed at once
(e.g. one per document), and evicting one on the arrival of another would
reintroduce the same bug.
Claimed requests deliberately survive the "nobody will want this anymore"
cancellation reasons (Replaced, Superseded, Rejected, IgnoredDismissed,
trajectory divergence) since they have a live consumer; `invalidateClaimed()`
still cancels them for reasons that invalidate the result itself (CacheCleared,
DocumentClosed, Disposed).
Also settle `firstEdit`/`result` in `_runSpeculativeProviderCall`'s outer catch.
Previously a throw at the first `editStream.next()` left them unsettled forever,
hanging any caller that reused the request — a pre-existing bug that the claimed
map would otherwise widen into a permanent map entry.
Behavior is unchanged when `inlineEdits.speculativeRequests` is off: the claimed
map can only be populated via `_triggerSpeculativeRequest`, which is gated on the
setting, so every new lookup degenerates to an empty-map no-op.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: b9cc340f-f2af-491d-a411-6e1c2c19583e
* nes: test: cover speculative that throws before its first streamed edit
The outer catch in `_runSpeculativeProviderCall` settles `firstEdit`/`result`
so a caller reusing the request cannot hang, but nothing exercised it: every
existing speculative test yields at least one edit or is cancelled.
Add a `throwBeforeFirstYield` harness behavior and a regression test that claims
such a speculative. `getNextEdit` rethrows the underlying error for
`NoNextEditReason.Unexpected`, so the joiner rejects with the provider error
instead of awaiting forever, and the released claim lets the following request
recover with a fresh provider call.
Verified to fail without the fix: the joining call times out after 8s.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: b9cc340f-f2af-491d-a411-6e1c2c19583e
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: b9cc340f-f2af-491d-a411-6e1c2c19583e
Stop hidden editor Copilot CLI session monitoring
Use the editor-window Extension Host visibility setting to control the shared session-state watcher while preserving the existing Agents-window behavior.\n\n(Written by Copilot)\n\nCo-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Reuses the existing Kimi history normalizer for custom Chat Completions endpoints so switched-model tool calls remain valid.
- Keeps BYOK request construction isolated from Copilot-specific model customizations, preserving configured sampling and endpoint behavior.
- Adds regression coverage for call/result pairing and unchanged BYOK sampling parameters.
Fixes#327794
(Commit message generated by Copilot)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix Copilot CLI session monitor startup race
Make the shared session-file watcher follow effective Agent Host ownership settings after startup so delayed experiment defaults can tear down a prematurely created watcher. Add regression coverage for disabling and re-enabling monitoring.\n\n(Written by Copilot)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address session monitor review feedback
Preserve error stack information in watcher setup logs and explicitly verify final watcher cleanup when the session service is disposed.\n\n(Written by Copilot)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
After summarization, `SummarizedConversationHistory` carried the last
summarized thinking block onto the first tool call round for every Anthropic
model, and `setSummaryAndThinking` cleared `round.thinking` whenever it was
called without one.
Only Anthropic's manual thinking mode (`type: "enabled"`) requires an assistant
message to open with a thinking block. Adaptive thinking accepts history as-is,
so on those models the carry-over is an unnecessary edit to a signed assistant
message the API handed back to us. Gate it behind `!supportsAdaptiveThinking`,
and make the thinking assignment overwrite-only so summarization can't strip a
signature off a round that is still replayed.
Endpoints that don't report `supportsAdaptiveThinking` keep the old behaviour.
Refs #327646
Vijay UpadyaandGitHub
Bhavya UandGitHub
Justin ChenandGitHub
Vikram NitinandGitHub
Dileep YavanmandhaandGitHub
vritant24andCopilot
Vritant BhardwajandGitHub
Osvaldo OrtegaandGitHub
Raymond ZhaoandGitHub
Benjamin SteenhoekandGitHub
SrirajandGitHub
Arthur CnopsandGitHub
Chris WestraandGitHub
Ladislau SzomoruandGitHub