ANTHROPIC_API_KEY=sk-ant-xxx go test ./pkg/provider/ -v -run TestAnthropicChatStream

ANTHROPIC_API_KEY=sk-ant-xxx go test ./pkg/provider/ -v -run TestAnthropicChatSync

TestApproval_Approved verifies that when OnApproval returns true, the tool executes normally and its real result is returned to the model.

TestApproval_Denied verifies that when OnApproval returns false, the tool is NOT executed and the model receives a rejection message.

TestApproval_NilCallback verifies that when OnApproval is nil (not set), tools with NeedsApproval=true are auto-approved. This ensures backwards compa

Unit test for the package-private bearerToken helper. Covers the non-obvious cases: empty header, wrong scheme, exact-prefix collision.

Verifies that all hardcoded sections are present in the prompt.

Verifies that when SOUL is empty, the default identity is used.

Verifies that memories are wrapped in context fencing tags.

Verifies that when memories are empty, no memory section appears.

Verifies that when USER is empty, no User Profile section appears.

Verifies that when SOUL is provided, it replaces the default identity.

Verifies that content exceeding character limits is truncated.

Verifies that USER profile is injected under a "User Profile" heading.

Integration test: actually hits DuckDuckGo. Run with: go test ./pkg/tool/websearch/ -run TestCanaryDDG -v This test validates that the regexes stil

Integration test: actually fetches a public URL. Run with: go test ./pkg/tool/webfetch/ -run TestCanaryFetch -v Validates that the HTTP client, SSR

Empty context returns (nil, false), not a panic or zero-Claims pointer.

--- tests -----------------------------------------------------------------

Verify that time range codes map correctly, especially "y" -> "t" (DuckDuckGo's quirk).

TestDisconnectCleansRegistry verifies the handler unregisters the conn from the (real DB) Registry when the client closes.

TestEditFile_AmbiguousMatch verifies that old_text appearing more than once is rejected. This forces the model to provide more surrounding context to

TestEditFile_BadJSON verifies that malformed JSON arguments return an error rather than panicking.

TestEditFile_DeleteText verifies that passing an empty new_text effectively deletes old_text from the file. This is the intended way to remove code —

TestEditFile_FileNotExist verifies that editing a non-existent file returns an error with the file path.

TestEditFile_MultilineReplace verifies that old_text and new_text can span multiple lines. This is the common case for real code edits — the model sen

TestEditFile_NotFound verifies that old_text not present in the file returns an error. The file must remain unchanged.

TestEditFile_SingleReplace verifies the happy path: old_text appears exactly once and is replaced with new_text. The rest of the file content must rem

--------------------------------------------------------------------------- Token estimation tests ---------------------------------------------------

TestEstimateMessageTokens_TextOnly verifies the basic character-to-token formula: tokens = (role + content chars)*2/5 + 12 overhead.

TestEstimateMessageTokens_ToolResult verifies that a tool result message (role="tool") correctly counts Role, Content, and ToolCallID.

TestEstimateMessageTokens_WithToolCalls verifies that ToolCall metadata (ID, function name, arguments) is included in the token estimate. Without this

TestEstimateTokens_EmptySlice verifies that an empty message slice returns 0 tokens.

TestEstimateTokens_MultipleMessages verifies that estimateTokens sums individual message estimates correctly.

TestEstimateToolTokens_Empty verifies that no tools = 0 tokens.

TestEstimateToolTokens_WithTools verifies that tool definition token estimation accounts for name, description, and JSON-serialized parameters. This i

TestEvalArithmetic_BadJSON verifies that malformed JSON arguments (i.e. the arguments string from OpenAI is corrupted or not valid JSON) return an err

TestEvalArithmetic_BasicOps verifies that the four basic arithmetic operators produce correct results. These are the most common expressions the model

TestEvalArithmetic_DivisionByZero verifies that dividing by zero returns an error rather than panicking or returning Inf. In Go constant evaluation, i

TestEvalArithmetic_FloatDivision verifies that float division returns an exact rational fraction. types.Eval uses Go's arbitrary-precision constant ar

TestEvalArithmetic_IntegerDivision documents Go's integer division semantics: 10/3 truncates to 3, not 3.333 or "10/3". This is standard Go behavior (

TestEvalArithmetic_InvalidExpression verifies that a syntactically invalid expression (not valid Go constant syntax) returns an error rather than pani

TestEvalArithmetic_Parentheses verifies that parentheses correctly override operator precedence. Without parentheses, 1+2*3 = 7; with them, (1+2)*3 =

TestExecCommand_BadJSON verifies that malformed JSON arguments return an error rather than panicking.

TestExecCommand_BasicCommand verifies the happy path: a simple command runs successfully and its stdout is returned.

TestExecCommand_FailedCommand verifies that a non-zero exit code does NOT cause Run to return an error. Instead, the exit status is appended to the ou

TestExecCommand_Pipe verifies that shell features like pipes work, confirming that "sh -c" delegation is functioning correctly. Without "sh -c", the p

TestExecCommand_Stderr verifies that stderr output is captured alongside stdout. CombinedOutput merges both streams, so diagnostic messages and error

Verify that output exceeding 50KB is truncated at the byte limit, and that multi-byte UTF-8 characters are not split at the boundary.

Verify that output exceeding 2000 lines is truncated at the line limit.

Verify that UTF-8 multi-byte characters are not split when truncating at the byte limit.

Verify that Execute rejects non-http(s) schemes.

Verify that a basic search passes query, default count, and empty time range to the provider.

Verify that count values over 100 are rejected and the default is used instead.

Verify that the model can override the default result count.

Verify that count=0 falls back to the default (0 is not a valid count).

Verify that when MaxResults is not set (zero value), the default of 5 is used.

Verify that a whitespace-only query returns an error without calling the provider.

Verify that Execute rejects empty URL.

Verify that Execute rejects invalid JSON args.

Verify that malformed JSON args return an error.

Verify that an invalid time_range defaults to no filter instead of returning an error.

Verify that calling Execute with no provider configured returns a helpful error.

Verify that an empty result set returns a human-readable "no results" message.

Verify that provider errors are surfaced in the tool result.

Verify that Execute blocks private/local URLs (SSRF protection).

Verify that a valid time_range is passed through to the provider.

Verify that extractDDGURL correctly unwraps DuckDuckGo's redirect URLs and handles direct URLs, missing trailing params, and original URLs with query

Verify that extractText removes script, style, and noscript blocks, strips tags, and normalizes whitespace into clean readable text.

Verify that extractText handles empty input without panicking.

TestFallbackSummary verifies the crude fallback: each message is truncated to 10% (min 200 chars) and joined with " | ".

TestFindSafeBoundary_BetweenBoundaries verifies that when targetIndex falls between two boundaries, the earlier one (at or before target) is returned.

TestFindSafeBoundary_EmptyHistory verifies that an empty history returns 0 (no boundaries exist, nothing to cut).

TestFindSafeBoundary_ExactMatch verifies that when targetIndex falls exactly on a Turn boundary, that boundary is returned.

TestFindSafeBoundary_FallbackToAfter verifies the defensive edge case: if no boundary exists at or before targetIndex (e.g. history starts with non-us

TestForceCompress_AppendToExistingSummary verifies that if a summary already exists (from a prior summarization), the truncation note is appended rath

--------------------------------------------------------------------------- Force compression tests --------------------------------------------------

TestForceCompress_PreservesToolCallPairs verifies that truncation at Turn boundaries keeps tool_call/tool_result pairs intact. If we cut in the middle

TestForceCompress_SetsSummaryNote verifies that forceCompress records how many messages were dropped, so the model knows context was lost.

TestForceCompress_SingleTurn verifies the edge case where history has only one Turn (one user message + many tool calls). cutIndex would be 0, so forc

TestForceCompress_TooShortToCompress verifies that history with ≤2 messages is left untouched — cutting it would make things worse.

Verify that formatFetchResult includes all metadata fields.

Verify that formatJSON pretty-prints valid JSON.

Verify that formatJSON returns the original string for invalid JSON.

Verify that formatResults includes query, titles, URLs, and snippets in the output.

After a successful login the stored token is attached to subsequent calls; the protected endpoint returns 200 with the decoded user.

Without a token a protected call surfaces as ErrUnauthorized so the CLI can prompt for re-login.

TestGetCurrentTime_ReturnsValidRFC3339 verifies that the tool returns a string that can be parsed back as RFC 3339. We don't assert the exact time (th