@evolvingmachines/evolve 0.0.55-project-sable.20260916.db0bc0bb

package/dist/cli/index.d.cts

@@ -0,0 +1,270 @@
+#!/usr/bin/env node
+import { T as Trial, a as TrialAnalysis, A as AgentInput, J as JobCreate, P as PublishDatasetInput, D as DatasetVersion, b as DatasetImport, C as Check, c as JobEvent, d as DatasetFailedTask, e as DatasetImportProgress, R as Rubric, f as JobSecretInline, g as JobSecretRef, h as TraceEvent } from '../types-DlpTxdR_.cjs';
+
+/**
+ * evolve — the CLI for Evolve hosted datasets & jobs.
+ *
+ * Noun-verb grammar over the hosted client: `evolve <noun> <verb>`, plus
+ * first-class top-level commands that need no noun — `run` (which takes
+ * `job start`'s flags), `analyze`, and `upload`, each spelled, helped and
+ * dispatched as a command in its own right (Harbor registers all three the
+ * same way, as top-level commands: their cli/main.py). Singular
+ * nouns are canonical; `job`/`trial`/`analysis`/`dataset` also answer to
+ * their plurals as hidden aliases, but `agents` does NOT — that word is
+ * reserved for the managed-agents CLI and refuses with the reason — and
+ * `skills` is not the plural of `skill`: it is the one local group, serving
+ * the bundled skill files a coding agent reads as its manual (skills.ts),
+ * with no API call. `session` is the managed-agents lane's first noun here —
+ * list and inspect the sessions your SDK runs recorded, headless. Every
+ * other command speaks ONLY through the SDK clients (datasets() / agents() /
+ * jobs() / trials() / analyses() / skills() / auth() / sessions()) — no raw
+ * HTTP lives here.
+ *
+ * Output: human tables on a TTY, tab-separated rows when piped, --json for
+ * the rendered machine shape (NDJSON for --watch event streams), -q for
+ * ids-only lists. Exit codes: 0 success (watch: job COMPLETED / publish
+ * COMPLETED / every analysis completed), 1 runtime/API failure (watch:
+ * FAILED or CANCELLED / any analysis failed), 2 usage error.
+ */
+
+/** Usage-level error: bad command line, not a runtime failure. Exit code 2. */
+declare class CliUsageError extends Error {
+    constructor(message: string);
+}
+declare const USAGE: string;
+interface Invocation {
+    /** Canonical "<group> <verb>", or "help" / "version". */
+    command: string;
+    positionals: string[];
+    flags: Record<string, string | number | boolean | string[]>;
+}
+declare function parseArgs(argv: string[]): Invocation;
+/**
+ * -c reads real YAML through the standard `yaml` package — the hand-rolled
+ * subset reader is retired. PyYAML's reading is the contract, so the schema is
+ * YAML 1.1 (`yes`/`on` are booleans, `012` is octal, a flow mapping is a whole
+ * sequence item) with two carve-outs: the 1.1 spec's bare `y`/`n` booleans,
+ * which PyYAML never adopted — `n: 3` keeps its key — and the 1.1 spec's
+ * number shapes PyYAML never adopted either, where a float needs a dot and a
+ * signed exponent and an integer may not be zero-padded, so `e3`, `1e3`, `08`
+ * and `0:0` stay the strings a caller wrote. The schema is PINNED, not
+ * merely defaulted: PyYAML's resolver has no mode but 1.1 and reads a file the
+ * same way whatever its `%YAML` directive says, while this library lets an
+ * explicit `%YAML 1.2` swap in the 1.2 core schema. One file, one reading.
+ * Three refusals sit on top of the library: a second document, an unresolvable
+ * tag (both PyYAML refusals too), and a duplicate key — PyYAML silently keeps
+ * the last value, and that silence is exactly the corruption a config file
+ * cannot afford. Every refusal names its `source` and carries `:line` wherever
+ * the parser places the problem — the library's alias-bomb guard fires at
+ * resolution, after parsing, and has no position to give. An empty or
+ * comment-only file is an empty object.
+ */
+declare function parseYamlConfig(text: string, source: string): unknown;
+/** A path into the parsed config: object keys and list indices, root first. */
+type ConfigPath = (string | number)[];
+/**
+ * `parseYamlConfig` plus a `locate` that answers "what line does this path
+ * start on", so a schema refusal can point at the file the way a parse
+ * refusal already does. JSON configs have no locator — JSON.parse keeps no
+ * positions — and every message treats the line as optional for that reason.
+ */
+declare function parseYamlConfigLocated(text: string, source: string): {
+    value: unknown;
+    locate: (path: ConfigPath) => number | undefined;
+};
+/**
+ * Parse repeatable --ak key=value pairs into the wire's agents[].kwargs
+ * object. The `config` key gets the channel's one client-side mechanic: a
+ * string value is a LOCAL settings file (the server never reads a client
+ * path), resolved here to its parsed content — JSON or TOML by extension,
+ * with a both-ways attempt for anything else. Every other key rides verbatim;
+ * acceptance is the server's.
+ */
+declare function parseAgentKwargs(pairs: string[], read?: (path: string) => string): Record<string, unknown>;
+/**
+ * Read a local prompt file for `analyze -p` / `run --analyze-prompt` — Harbor's
+ * `-p/--prompt` (their cli/analyze.py:252-255), whose TEXT replaces the platform's
+ * default body as the analyzer's instruction template (analyzer.py:130-134
+ * `prompt_path.read_text()`). Read verbatim, no parsing: the tokens
+ * (`{trial_path}`, `{task_section}`, `{criteria_guidance}`) are rendered
+ * server-side. Ruled here: the file must be readable and non-empty; the
+ * server stores it whole at any length (no invented number) and refuses
+ * only an empty or NUL-bearing one typed (`invalid_input` naming
+ * `analyze.prompt`).
+ */
+declare function loadPromptFile(path: string, read?: (path: string) => string): string;
+/**
+ * Read and parse a local rubric file for `analyze -r` / `run --analyze-rubric`
+ * into the spec's Rubric shape — Harbor's own loader law (their
+ * cli/quality_checker/models.py load_rubric): TOML, YAML, or JSON by
+ * extension, anything else refused by name. Ruled here: the shape, unknown
+ * fields refused naming them, and the two spec MINIMUMS — a non-empty
+ * criteria list, and each criterion's three fields present as non-empty
+ * strings. The server owns the rest of the bounds (criteria count, field
+ * lengths, the snake_case name grammar) and refuses them typed
+ * (`invalid_rubric`).
+ */
+declare function loadRubricFile(path: string, read?: (path: string) => string): Rubric;
+/**
+ * Parse repeatable --secret NAME[@LABEL][=ENVNAME] references into the wire's
+ * secrets[] objects. The grammar splits on the FIRST '=' (everything after it
+ * is the in-sandbox env name) and then the FIRST '@' (everything after it is
+ * the label), so neither delimiter is legal inside the parts — which matches
+ * the server's vocabularies (env-var-shaped names, [A-Za-z0-9._-] labels).
+ * Only the shape is ruled here; name/label semantics (reserved names, the
+ * 'default' fallback, the ambiguity refusal) are the server's.
+ */
+declare function parseSecretRefs(values: string[]): JobSecretRef[];
+/**
+ * Parse repeatable --secret-inline NAME[@LABEL]:DELIVERY=VALUE entries into
+ * the wire's inline secrets[] objects. DELIVERY sits BEFORE the '=' so the
+ * value — everything after the FIRST '=' — is passed through byte-for-byte
+ * and may contain '=', ':' and '@' unescaped. The head splits on its FIRST
+ * '@' (label) and LAST ':' (delivery), matching the label grammar
+ * ([A-Za-z0-9._-], no ':' possible). Only the shape is ruled here;
+ * name/label semantics and the collision refusal are the server's.
+ */
+declare function parseInlineSecrets(values: string[]): JobSecretInline[];
+/** Parse repeatable KEY=VALUE pairs into an env map. */
+declare function parseEnvPairs(pairs: string[], flag: string): Record<string, string>;
+/**
+ * Build the POST /api/jobs body from a parsed `job start` / `run` invocation:
+ * -c loads a base config in the spec's own vocabulary, then every explicitly
+ * passed flag overrides the corresponding field. -i/-x/-l are stamped onto
+ * EVERY dataset selector — per-selector filters with one flag grammar, so a
+ * glob that matches nothing in one dataset simply filters nothing there.
+ */
+declare function buildJobInput(inv: Invocation, read?: (path: string) => string): JobCreate;
+/**
+ * Build the datasets().publish() input from a parsed `dataset publish`
+ * invocation. `--name`/`--version` are optional with `--dir` (a corpus
+ * carrying a dataset.toml manifest supplies them server-side — Harbor's
+ * dataset layout — and the SDK refuses before uploading when neither the
+ * flags nor a manifest exist) and with `--from hub:…` (the server defaults
+ * them from the resolved package). A git or `--from <url>` source requires
+ * both — its corpus is only fetched after the server accepts the publish.
+ *
+ * `--from` takes the FETCHABLE sources, one flag for both spellings: a plain
+ * https URL is a public tarball (`archive_url` on the wire), and the `hub:`
+ * prefix marks a Harbor hub package whose reference part is exactly Harbor's
+ * own grammar (`org/name[@ref]` — their CLI takes the same reference as a
+ * bare positional; the prefix exists only because this one flag also accepts
+ * URLs).
+ */
+declare function buildPublishInput(inv: Invocation): PublishDatasetInput;
+/**
+ * Build the agents().create() input from a parsed `agent add` invocation.
+ * `--install-script` names a FILE; its contents are what the SDK uploads.
+ */
+declare function buildAgentInput(inv: Invocation, readScript?: (path: string) => string): AgentInput;
+interface CliIO {
+    out(line: string): void;
+    err(line: string): void;
+    /**
+     * True when stdout is an interactive terminal. Drives the table-vs-TSV
+     * split on list commands; TERM=dumb counts as non-interactive. Defaults to
+     * false when absent so redirected output always gets machine-safe rows.
+     */
+    tty?: boolean;
+    /**
+     * Ask the human a yes/no question — the destructive-verb prompt. Present
+     * only when stdin can actually answer (an interactive terminal); absent,
+     * a destructive verb must be told --yes, exactly Harbor's non-TTY
+     * posture. The question rides stderr so stdout stays machine-clean.
+     */
+    confirm?(question: string): Promise<boolean>;
+}
+interface ListColumn<T> {
+    key: string;
+    header: string;
+    cell(row: T): string;
+}
+/** Exported for its test, like trialDetailLines above it. */
+declare const TRIAL_COLUMNS: ListColumn<Trial>[];
+/**
+ * Full-detail rendering of one trial — evolve trial show. Exported for
+ * tests, like the other line renderers.
+ */
+declare function trialDetailLines(run: Trial): string[];
+/**
+ * Full-detail rendering of one analysis run — evolve analysis show. The same
+ * facts `trial show` folds into its `analysis` rows, standing on their own:
+ * status and model, one row per criterion verdict, the summary, the
+ * analyzer's own metered spend (four decimals like every analyzer-spend
+ * figure — an analysis costs cents), and the typed failure when there is
+ * one. Exported for tests, like the other line renderers.
+ */
+declare function analysisDetailLines(analysis: TrialAnalysis): string[];
+/**
+ * One trace event line — evolve trial download --stream trace-parsed, and
+ * every other verb that prints a parsed transcript. Null for a line the
+ * rendering HIDES: a harness's own `usage` event (owner's law 2026-09-10 —
+ * tokens and money shown anywhere come from the gateway meter, never from
+ * what a harness prints; the raw event still rides `--json`). The gateway
+ * meter's usage event (`gatewayUsageOf`) renders as a money line: model,
+ * tokens in / out (cached), cost as the gateway priced it, the call's start
+ * instant.
+ */
+declare function traceEventLine(event: TraceEvent): string | null;
+/**
+ * Compact one-line rendering of one live-progress change for --watch — the
+ * poll-line adaptation of Harbor's publish progress display (their overall
+ * bar is spinner + "M of N" + description + elapsed/remaining, rich Live at
+ * 10 fps: REFERENCES/Harbor src/harbor/cli/publish.py:231-238; ours is a
+ * line per observed server write, so it keeps the M-of-N and elapsed columns
+ * and drops the animation). During the copy phase the line states "N of M
+ * images already banked" — Harbor's "skipped (exists)" idea (publish.py:388)
+ * in this platform's banked vocabulary.
+ */
+declare function importProgressLine(progress: DatasetImportProgress, nowMs?: number): string;
+/**
+ * The settled progress record for the final --watch block: wall-clock per
+ * phase, the publish's image economics (built / mirrored / banked), and the
+ * CodeBuild copy-minutes meter — the shape of Harbor publish's settle
+ * summary (per-item Build/Upload timing table + "Published N, skipped M
+ * task(s) in X.XXs": REFERENCES/Harbor src/harbor/cli/publish.py:288-315).
+ */
+declare function progressSettleLines(progress: DatasetImportProgress): string[];
+/** Compact one-line rendering of one publish status change for --watch. */
+declare function importStatusLine(job: DatasetImport): string;
+/**
+ * Compact one-line rendering of one version state change for --watch's
+ * settle phase: the version's own walk (BUILDING, then READY/ARCHIVED or
+ * FAILED).
+ */
+declare function versionStatusLine(version: DatasetVersion): string;
+/** Compact one-line rendering of one SSE event for --watch. */
+declare function eventLine(event: JobEvent): string;
+/**
+ * The settled wave, one row per analyzed trial — Harbor's "Trial Analyses"
+ * table carried over: the criterion verdicts, the analyzer's own cost, and a
+ * summary excerpt. A failed analysis renders its typed failure in place of
+ * verdicts and repeats it in full below the table — never a silent absence.
+ * Exported for its test, like the other line renderers.
+ */
+declare function analysisResultLines(runs: Trial[]): string[];
+/**
+ * Harbor's check renderers (cli/analyze.py:29-81), as lines: ONE task
+ * prints its checks table (Check | Outcome | Explanation, the criterion
+ * titled) and the agent cost; several print the summary table (Task | Pass
+ * | Fail | N/A | Cost ($)) with an errored task's row dashed, then one
+ * `❌ task: reason` line per failed task and the total agent cost — their
+ * `_render_checks_table` / `_render_check_summary`. A task that is still
+ * queued or running has no verdict yet and prints its status in the
+ * outcome column.
+ */
+declare function checkResultLines(check: Check): string[];
+/** The record head for `check show` and the verb's return: id, status, source, policy, then Harbor's tables. */
+declare function checkDetailLines(check: Check): string[];
+/**
+ * The partial-publish model's honest ending for `dataset publish --watch`:
+ * "built N of M tasks — K failed to build", plus where the reasons live and
+ * the one fix (a re-publish — versions are immutable). Empty on a fully
+ * built version, so the common case keeps its exact output.
+ */
+declare function buildSettleSummaryLines(version: DatasetVersion, job: DatasetImport): string[];
+/** One live line per failed task outcome during `dataset publish --watch`. */
+declare function failedTaskOutcomeLine(entry: DatasetFailedTask): string;
+declare function runCli(argv: string[], io?: CliIO): Promise<number>;
+
+export { type CliIO, CliUsageError, type Invocation, TRIAL_COLUMNS, USAGE, analysisDetailLines, analysisResultLines, buildAgentInput, buildJobInput, buildPublishInput, buildSettleSummaryLines, checkDetailLines, checkResultLines, eventLine, failedTaskOutcomeLine, importProgressLine, importStatusLine, loadPromptFile, loadRubricFile, parseAgentKwargs, parseArgs, parseEnvPairs, parseInlineSecrets, parseSecretRefs, parseYamlConfig, parseYamlConfigLocated, progressSettleLines, runCli, traceEventLine, trialDetailLines, versionStatusLine };

package/dist/cli/index.d.ts

@@ -0,0 +1,270 @@
+#!/usr/bin/env node
+import { T as Trial, a as TrialAnalysis, A as AgentInput, J as JobCreate, P as PublishDatasetInput, D as DatasetVersion, b as DatasetImport, C as Check, c as JobEvent, d as DatasetFailedTask, e as DatasetImportProgress, R as Rubric, f as JobSecretInline, g as JobSecretRef, h as TraceEvent } from '../types-DlpTxdR_.js';
+
+/**
+ * evolve — the CLI for Evolve hosted datasets & jobs.
+ *
+ * Noun-verb grammar over the hosted client: `evolve <noun> <verb>`, plus
+ * first-class top-level commands that need no noun — `run` (which takes
+ * `job start`'s flags), `analyze`, and `upload`, each spelled, helped and
+ * dispatched as a command in its own right (Harbor registers all three the
+ * same way, as top-level commands: their cli/main.py). Singular
+ * nouns are canonical; `job`/`trial`/`analysis`/`dataset` also answer to
+ * their plurals as hidden aliases, but `agents` does NOT — that word is
+ * reserved for the managed-agents CLI and refuses with the reason — and
+ * `skills` is not the plural of `skill`: it is the one local group, serving
+ * the bundled skill files a coding agent reads as its manual (skills.ts),
+ * with no API call. `session` is the managed-agents lane's first noun here —
+ * list and inspect the sessions your SDK runs recorded, headless. Every
+ * other command speaks ONLY through the SDK clients (datasets() / agents() /
+ * jobs() / trials() / analyses() / skills() / auth() / sessions()) — no raw
+ * HTTP lives here.
+ *
+ * Output: human tables on a TTY, tab-separated rows when piped, --json for
+ * the rendered machine shape (NDJSON for --watch event streams), -q for
+ * ids-only lists. Exit codes: 0 success (watch: job COMPLETED / publish
+ * COMPLETED / every analysis completed), 1 runtime/API failure (watch:
+ * FAILED or CANCELLED / any analysis failed), 2 usage error.
+ */
+
+/** Usage-level error: bad command line, not a runtime failure. Exit code 2. */
+declare class CliUsageError extends Error {
+    constructor(message: string);
+}
+declare const USAGE: string;
+interface Invocation {
+    /** Canonical "<group> <verb>", or "help" / "version". */
+    command: string;
+    positionals: string[];
+    flags: Record<string, string | number | boolean | string[]>;
+}
+declare function parseArgs(argv: string[]): Invocation;
+/**
+ * -c reads real YAML through the standard `yaml` package — the hand-rolled
+ * subset reader is retired. PyYAML's reading is the contract, so the schema is
+ * YAML 1.1 (`yes`/`on` are booleans, `012` is octal, a flow mapping is a whole
+ * sequence item) with two carve-outs: the 1.1 spec's bare `y`/`n` booleans,
+ * which PyYAML never adopted — `n: 3` keeps its key — and the 1.1 spec's
+ * number shapes PyYAML never adopted either, where a float needs a dot and a
+ * signed exponent and an integer may not be zero-padded, so `e3`, `1e3`, `08`
+ * and `0:0` stay the strings a caller wrote. The schema is PINNED, not
+ * merely defaulted: PyYAML's resolver has no mode but 1.1 and reads a file the
+ * same way whatever its `%YAML` directive says, while this library lets an
+ * explicit `%YAML 1.2` swap in the 1.2 core schema. One file, one reading.
+ * Three refusals sit on top of the library: a second document, an unresolvable
+ * tag (both PyYAML refusals too), and a duplicate key — PyYAML silently keeps
+ * the last value, and that silence is exactly the corruption a config file
+ * cannot afford. Every refusal names its `source` and carries `:line` wherever
+ * the parser places the problem — the library's alias-bomb guard fires at
+ * resolution, after parsing, and has no position to give. An empty or
+ * comment-only file is an empty object.
+ */
+declare function parseYamlConfig(text: string, source: string): unknown;
+/** A path into the parsed config: object keys and list indices, root first. */
+type ConfigPath = (string | number)[];
+/**
+ * `parseYamlConfig` plus a `locate` that answers "what line does this path
+ * start on", so a schema refusal can point at the file the way a parse
+ * refusal already does. JSON configs have no locator — JSON.parse keeps no
+ * positions — and every message treats the line as optional for that reason.
+ */
+declare function parseYamlConfigLocated(text: string, source: string): {
+    value: unknown;
+    locate: (path: ConfigPath) => number | undefined;
+};
+/**
+ * Parse repeatable --ak key=value pairs into the wire's agents[].kwargs
+ * object. The `config` key gets the channel's one client-side mechanic: a
+ * string value is a LOCAL settings file (the server never reads a client
+ * path), resolved here to its parsed content — JSON or TOML by extension,
+ * with a both-ways attempt for anything else. Every other key rides verbatim;
+ * acceptance is the server's.
+ */
+declare function parseAgentKwargs(pairs: string[], read?: (path: string) => string): Record<string, unknown>;
+/**
+ * Read a local prompt file for `analyze -p` / `run --analyze-prompt` — Harbor's
+ * `-p/--prompt` (their cli/analyze.py:252-255), whose TEXT replaces the platform's
+ * default body as the analyzer's instruction template (analyzer.py:130-134
+ * `prompt_path.read_text()`). Read verbatim, no parsing: the tokens
+ * (`{trial_path}`, `{task_section}`, `{criteria_guidance}`) are rendered
+ * server-side. Ruled here: the file must be readable and non-empty; the
+ * server stores it whole at any length (no invented number) and refuses
+ * only an empty or NUL-bearing one typed (`invalid_input` naming
+ * `analyze.prompt`).
+ */
+declare function loadPromptFile(path: string, read?: (path: string) => string): string;
+/**
+ * Read and parse a local rubric file for `analyze -r` / `run --analyze-rubric`
+ * into the spec's Rubric shape — Harbor's own loader law (their
+ * cli/quality_checker/models.py load_rubric): TOML, YAML, or JSON by
+ * extension, anything else refused by name. Ruled here: the shape, unknown
+ * fields refused naming them, and the two spec MINIMUMS — a non-empty
+ * criteria list, and each criterion's three fields present as non-empty
+ * strings. The server owns the rest of the bounds (criteria count, field
+ * lengths, the snake_case name grammar) and refuses them typed
+ * (`invalid_rubric`).
+ */
+declare function loadRubricFile(path: string, read?: (path: string) => string): Rubric;
+/**
+ * Parse repeatable --secret NAME[@LABEL][=ENVNAME] references into the wire's
+ * secrets[] objects. The grammar splits on the FIRST '=' (everything after it
+ * is the in-sandbox env name) and then the FIRST '@' (everything after it is
+ * the label), so neither delimiter is legal inside the parts — which matches
+ * the server's vocabularies (env-var-shaped names, [A-Za-z0-9._-] labels).
+ * Only the shape is ruled here; name/label semantics (reserved names, the
+ * 'default' fallback, the ambiguity refusal) are the server's.
+ */
+declare function parseSecretRefs(values: string[]): JobSecretRef[];
+/**
+ * Parse repeatable --secret-inline NAME[@LABEL]:DELIVERY=VALUE entries into
+ * the wire's inline secrets[] objects. DELIVERY sits BEFORE the '=' so the
+ * value — everything after the FIRST '=' — is passed through byte-for-byte
+ * and may contain '=', ':' and '@' unescaped. The head splits on its FIRST
+ * '@' (label) and LAST ':' (delivery), matching the label grammar
+ * ([A-Za-z0-9._-], no ':' possible). Only the shape is ruled here;
+ * name/label semantics and the collision refusal are the server's.
+ */
+declare function parseInlineSecrets(values: string[]): JobSecretInline[];
+/** Parse repeatable KEY=VALUE pairs into an env map. */
+declare function parseEnvPairs(pairs: string[], flag: string): Record<string, string>;
+/**
+ * Build the POST /api/jobs body from a parsed `job start` / `run` invocation:
+ * -c loads a base config in the spec's own vocabulary, then every explicitly
+ * passed flag overrides the corresponding field. -i/-x/-l are stamped onto
+ * EVERY dataset selector — per-selector filters with one flag grammar, so a
+ * glob that matches nothing in one dataset simply filters nothing there.
+ */
+declare function buildJobInput(inv: Invocation, read?: (path: string) => string): JobCreate;
+/**
+ * Build the datasets().publish() input from a parsed `dataset publish`
+ * invocation. `--name`/`--version` are optional with `--dir` (a corpus
+ * carrying a dataset.toml manifest supplies them server-side — Harbor's
+ * dataset layout — and the SDK refuses before uploading when neither the
+ * flags nor a manifest exist) and with `--from hub:…` (the server defaults
+ * them from the resolved package). A git or `--from <url>` source requires
+ * both — its corpus is only fetched after the server accepts the publish.
+ *
+ * `--from` takes the FETCHABLE sources, one flag for both spellings: a plain
+ * https URL is a public tarball (`archive_url` on the wire), and the `hub:`
+ * prefix marks a Harbor hub package whose reference part is exactly Harbor's
+ * own grammar (`org/name[@ref]` — their CLI takes the same reference as a
+ * bare positional; the prefix exists only because this one flag also accepts
+ * URLs).
+ */
+declare function buildPublishInput(inv: Invocation): PublishDatasetInput;
+/**
+ * Build the agents().create() input from a parsed `agent add` invocation.
+ * `--install-script` names a FILE; its contents are what the SDK uploads.
+ */
+declare function buildAgentInput(inv: Invocation, readScript?: (path: string) => string): AgentInput;
+interface CliIO {
+    out(line: string): void;
+    err(line: string): void;
+    /**
+     * True when stdout is an interactive terminal. Drives the table-vs-TSV
+     * split on list commands; TERM=dumb counts as non-interactive. Defaults to
+     * false when absent so redirected output always gets machine-safe rows.
+     */
+    tty?: boolean;
+    /**
+     * Ask the human a yes/no question — the destructive-verb prompt. Present
+     * only when stdin can actually answer (an interactive terminal); absent,
+     * a destructive verb must be told --yes, exactly Harbor's non-TTY
+     * posture. The question rides stderr so stdout stays machine-clean.
+     */
+    confirm?(question: string): Promise<boolean>;
+}
+interface ListColumn<T> {
+    key: string;
+    header: string;
+    cell(row: T): string;
+}
+/** Exported for its test, like trialDetailLines above it. */
+declare const TRIAL_COLUMNS: ListColumn<Trial>[];
+/**
+ * Full-detail rendering of one trial — evolve trial show. Exported for
+ * tests, like the other line renderers.
+ */
+declare function trialDetailLines(run: Trial): string[];
+/**
+ * Full-detail rendering of one analysis run — evolve analysis show. The same
+ * facts `trial show` folds into its `analysis` rows, standing on their own:
+ * status and model, one row per criterion verdict, the summary, the
+ * analyzer's own metered spend (four decimals like every analyzer-spend
+ * figure — an analysis costs cents), and the typed failure when there is
+ * one. Exported for tests, like the other line renderers.
+ */
+declare function analysisDetailLines(analysis: TrialAnalysis): string[];
+/**
+ * One trace event line — evolve trial download --stream trace-parsed, and
+ * every other verb that prints a parsed transcript. Null for a line the
+ * rendering HIDES: a harness's own `usage` event (owner's law 2026-09-10 —
+ * tokens and money shown anywhere come from the gateway meter, never from
+ * what a harness prints; the raw event still rides `--json`). The gateway
+ * meter's usage event (`gatewayUsageOf`) renders as a money line: model,
+ * tokens in / out (cached), cost as the gateway priced it, the call's start
+ * instant.
+ */
+declare function traceEventLine(event: TraceEvent): string | null;
+/**
+ * Compact one-line rendering of one live-progress change for --watch — the
+ * poll-line adaptation of Harbor's publish progress display (their overall
+ * bar is spinner + "M of N" + description + elapsed/remaining, rich Live at
+ * 10 fps: REFERENCES/Harbor src/harbor/cli/publish.py:231-238; ours is a
+ * line per observed server write, so it keeps the M-of-N and elapsed columns
+ * and drops the animation). During the copy phase the line states "N of M
+ * images already banked" — Harbor's "skipped (exists)" idea (publish.py:388)
+ * in this platform's banked vocabulary.
+ */
+declare function importProgressLine(progress: DatasetImportProgress, nowMs?: number): string;
+/**
+ * The settled progress record for the final --watch block: wall-clock per
+ * phase, the publish's image economics (built / mirrored / banked), and the
+ * CodeBuild copy-minutes meter — the shape of Harbor publish's settle
+ * summary (per-item Build/Upload timing table + "Published N, skipped M
+ * task(s) in X.XXs": REFERENCES/Harbor src/harbor/cli/publish.py:288-315).
+ */
+declare function progressSettleLines(progress: DatasetImportProgress): string[];
+/** Compact one-line rendering of one publish status change for --watch. */
+declare function importStatusLine(job: DatasetImport): string;
+/**
+ * Compact one-line rendering of one version state change for --watch's
+ * settle phase: the version's own walk (BUILDING, then READY/ARCHIVED or
+ * FAILED).
+ */
+declare function versionStatusLine(version: DatasetVersion): string;
+/** Compact one-line rendering of one SSE event for --watch. */
+declare function eventLine(event: JobEvent): string;
+/**
+ * The settled wave, one row per analyzed trial — Harbor's "Trial Analyses"
+ * table carried over: the criterion verdicts, the analyzer's own cost, and a
+ * summary excerpt. A failed analysis renders its typed failure in place of
+ * verdicts and repeats it in full below the table — never a silent absence.
+ * Exported for its test, like the other line renderers.
+ */
+declare function analysisResultLines(runs: Trial[]): string[];
+/**
+ * Harbor's check renderers (cli/analyze.py:29-81), as lines: ONE task
+ * prints its checks table (Check | Outcome | Explanation, the criterion
+ * titled) and the agent cost; several print the summary table (Task | Pass
+ * | Fail | N/A | Cost ($)) with an errored task's row dashed, then one
+ * `❌ task: reason` line per failed task and the total agent cost — their
+ * `_render_checks_table` / `_render_check_summary`. A task that is still
+ * queued or running has no verdict yet and prints its status in the
+ * outcome column.
+ */
+declare function checkResultLines(check: Check): string[];
+/** The record head for `check show` and the verb's return: id, status, source, policy, then Harbor's tables. */
+declare function checkDetailLines(check: Check): string[];
+/**
+ * The partial-publish model's honest ending for `dataset publish --watch`:
+ * "built N of M tasks — K failed to build", plus where the reasons live and
+ * the one fix (a re-publish — versions are immutable). Empty on a fully
+ * built version, so the common case keeps its exact output.
+ */
+declare function buildSettleSummaryLines(version: DatasetVersion, job: DatasetImport): string[];
+/** One live line per failed task outcome during `dataset publish --watch`. */
+declare function failedTaskOutcomeLine(entry: DatasetFailedTask): string;
+declare function runCli(argv: string[], io?: CliIO): Promise<number>;
+
+export { type CliIO, CliUsageError, type Invocation, TRIAL_COLUMNS, USAGE, analysisDetailLines, analysisResultLines, buildAgentInput, buildJobInput, buildPublishInput, buildSettleSummaryLines, checkDetailLines, checkResultLines, eventLine, failedTaskOutcomeLine, importProgressLine, importStatusLine, loadPromptFile, loadRubricFile, parseAgentKwargs, parseArgs, parseEnvPairs, parseInlineSecrets, parseSecretRefs, parseYamlConfig, parseYamlConfigLocated, progressSettleLines, runCli, traceEventLine, trialDetailLines, versionStatusLine };