Run executes the full-scan pipeline: enumerate → filter → token-filter → dispatch one subtask per file → collect comments.

New creates a new SessionHistory with the given repo directory.

String renders a one-line human-readable estimate. Money is intentionally omitted — pricing varies per provider/model and we don't want to imply a pre

ResolveEndpoint resolves an endpoint without per-run overrides.

--- Provider-based resolution tests ---

NewCommentCollector creates an empty collector.

NewRegistry creates an empty, mutable registry.

Add appends a comment to the collector.

New creates a new Agent from the given arguments.

NewRetryCollector returns an empty collector.

Execute runs the tool with the given arguments and returns the result string.

Finalize marks the session as complete, sets the end time, and persists the final summary record. When a frozen manifest was stored via SetFinalManife

One OCR comment as ``ocr review --format json`` emits it.

Get returns the provider registered under name.

NewRunner returns a Runner bound to the given dependencies.

NewTextMessage creates a message with simple string content.

Writer returns the current stdout writer (real stdout or discard).

NewResolver builds a Resolver with the following priority: 1. Custom rule file specified via --rule flag (first match wins) 2. Project-local .opencode

emitRunResult is the post-LLM-run finalization shared by `ocr review` and `ocr scan`: resolves comment line numbers, records telemetry, restores stdou

Run executes the full review pipeline: parse diffs -> plan per file -> LLM tool-loop -> collect comments.

Freeze prevents further mutations. Call once setup is complete.

NewAgent creates a scan Agent from the given args. The Session is auto-created (review_mode = full_scan) when not supplied.

Register adds a tool provider. Panics if the registry is frozen.

New creates a Runner that allows at most maxConcurrent simultaneous git subprocesses. If maxConcurrent <= 0 the default (16) is used.

GetOrCreateFileSession returns the FileSession for the given file path, creating one if it doesn't exist yet.

Freeze builds the immutable report. It must be called once, after the run is done, and runID is the review session ID. The three return shapes are di

LookupProvider returns the preset provider by name. The returned Provider has its own copy of the Models slice.

ResolveLineNumbers populates StartLine/EndLine on each comment by matching the ExistingCode against the corresponding file's diff hunks (primary), or

newBuilderWith registers a set of selected items on a fresh builder and sets a valid input mode so Finalize passes the mandatory-mode check.

MarkCompleted records that the item's subtask completed. Whether it produced comments does not affect completion.

mustFinalize finalizes and fails the test on an unexpected validation error.

Err returns the first error encountered while creating or writing the underlying file, or nil if none occurred. Text-mode rendering drops the per-writ

NewOpenAIClient creates a new OpenAI-compatible LLM client. ExtraHeaders are applied per request (not baked into the SDK client) so SessionKeyTemplate

dispatchSubtasks runs the Plan + Main phases for each changed file concurrently.

parseReviewFlags provides test compatibility: parses args through a fresh cobra command instance and returns the resulting reviewOptions.

setTestHome redirects the user home resolved by os.UserHomeDir() to dir for the duration of the test. Setting HOME alone is NOT enough on Windows: os.

Comments returns all collected comments.

--- Delete custom provider tests ---

setTestHome redirects the user home resolved by os.UserHomeDir() to dir for the duration of the test. Setting HOME alone is NOT enough on Windows: os.

setTestHome redirects the user home resolved by os.UserHomeDir() to dir for the duration of the test. Setting HOME alone is NOT enough on Windows: os.

metaCtx attaches identity the way the review path will in P4.

Content extracts the text content from the first choice, falling back to reasoning content.

RecordReviewItemDone persists the file-level checkpoint used by resume.

RunPerFile drives the main LLM conversation loop for a single file. It sends messages with the configured tool definitions, executes any tool calls re

StartSpan creates a new span from the given context. When telemetry is not enabled, it returns a no-op span so callers can safely defer .End().

CompletionsWithCtx sends a chat completion request with context support for cancellation and timeout. The deferred finalizeRequest is the client boun

ExtractText returns the concatenated text content from a Message's Content field. Handles both plain string and content block array formats.

Item returns a copy of the checkpoint for fingerprint.

NewOpenAIResponsesClient creates a client for the OpenAI Responses API. URL normalization mirrors NewOpenAIClient: cfg.URL is forced to end in respons

Helper: run publish() with a Recorder and no-op sleep.

ListProviders returns all built-in providers sorted by provider name. Each returned Provider has its own copy of the Models slice in registry order.

Drive a poster method over a sequence of fake urlopen outcomes.

One OCR comment as `ocr review --format json` emits it.

parseScanFlags provides test compatibility: parses args through a fresh cobra command instance and returns the resulting scanOptions.

recordUntimed records an attempt with zero timestamps. These tests are about numbering, outcome and aggregation, where the derived durations are nois

stripANSI removes ANSI escape sequences from a string so tests can assert against plain text content.

NewAnthropicClient creates a new Anthropic Messages API client. The Anthropic API manages prompt-cache affinity server-side, so unlike the OpenAI clie

Since returns a copy of all comments stored at index ≥ start. Returns nil when no new comments have been added since the snapshot.

Every test below reaches Freeze without finalizing anything itself: the client boundary in CompletionsWithCtx does it. That is deliberate — Freeze ref

resolveRuleEntries scans each entry's Rule field. When the value looks like a file path, it reads the file content and replaces the Rule. Absolute pat

AppendTaskRecord adds a new task record to the file session for the given file path and task type. It auto-assigns the RequestNo based on existing rec

CompletedCount returns the number of file-level checkpoints replay recovered. Review reuse is narrower — see ReusableItem — while scan, having no mani

Finalize decides the request-level outcome for m exactly once. The three inputs are the recorded attempts, the error the logical request returned to

ModelListContains reports whether target matches any entry in models (trimmed).

ReadLines returns a window of lines from the file plus the total line count. startLine is 1-based; maxLines is the maximum number of lines to collect.