/**
* std/llm/dialects — the tool-call dialect table, as DATA.
*
* Import with: import { TAGGED_DIALECT, PROVIDER_MARKERS, NAME_ALIASES }
* from "std/llm/dialects"
*
* WHY THIS MODULE EXISTS. "Which tag, marker, or spelling means a tool call"
* is not a parsing question, it is a vocabulary question, and the answer
* changes every time a provider ships a new chat template. Kept as Rust
* constants it was scattered across a dozen files — the tag aliases in the
* tagged scanner, the fence info-strings in the fenced-JSON chunker, the
* name aliases in three separate JSON extractors, the semantic aliases in a
* thousand-line compat module — so adding a dialect meant editing code in
* several places and rebuilding, and no single place answered "what does Harn
* currently recognize?".
*
* Everything here is inert data. There is no scanning, no matching, and no
* policy in this file; it is read by `std/llm/tool_parse` (which owns the
* composition) and handed to the Rust structural scanner (which owns
* delimiting, and needs the vocabulary only to know where a unit ends). That
* split is what keeps Rust proportional to BYTES and Harn proportional to
* CANDIDATES.
*
* ADDING A DIALECT is an edit to this file and nothing else. If a change here
* requires a matching change in Rust, the cut line has been violated — the
* Rust side must consume these lists, never restate them.
*/
/**
* A response tag and every spelling models emit for it.
*
* `canonical` is the name the rest of the runtime uses; `spellings` are the
* forms accepted on the wire. The tagless-underscore variants are not
* theoretical: chat templates strip underscores often enough that rejecting
* `<toolcall>` loses real calls.
*/
pub type TagDialect = {canonical: string, spellings: list<string>}
/** A chat-template markup opener and the close tag it pairs with. */
pub type MarkupDialect = {
opener: string,
close: string,
/** Human-facing label used in parse feedback, e.g. "`<function=...>`". */
label: string,
}
// ── The tagged text protocol ────────────────────────────────────────────────
/**
* Top-level response blocks. A response is a sequence of these and nothing
* else; anything outside them is stray text, which is a violation or salvage
* decision `std/llm/tool_parse` owns.
*/
pub const TAGGED_DIALECT: list<TagDialect> = [
{canonical: "tool_call", spellings: ["tool_call", "toolcall"]},
{canonical: "assistant_prose", spellings: ["assistant_prose", "assistantprose"]},
{canonical: "user_response", spellings: ["user_response", "userresponse"]},
{canonical: "done", spellings: ["done"]},
]
/**
* Tags a weak model wraps around NARRATION *inside* a `<tool_call>` block
* while it is only thinking out loud.
*
* Deliberately tiny and allowlisted. A narration tag is special precisely
* because it is NOT an attempted invocation, so this list must never widen to
* cover unknown tags that look like calls — `<frobnicate>{...}</frobnicate>`
* has to stay a rejected unknown tool, not become silent prose.
*/
pub const NARRATION_TAGS: list<string> = [
"assistant_prose",
"assistantprose",
"thinking",
"reasoning",
]
/**
* Wrapper tags that carry no content of their own. Some chat templates emit
* these around `<invoke ...>` markup; swallowing them silently avoids raising
* two "unknown top-level tag" violations around an otherwise-recovered call.
*/
pub const CONTENTLESS_WRAPPER_TAGS: list<string> = ["function_calls"]
/** Chat-template function markup emitted as plain assistant text. */
pub const MARKUP_DIALECTS: list<MarkupDialect> = [
{opener: "<function=", close: "</function>", label: "`<function=...>`"},
{opener: "<invoke name=", close: "</invoke>", label: "`<invoke name=...>`"},
]
/**
* Line-leading labels models prepend to a bare call. `tool_code:` is Gemma's
* native spelling; the language tags are what a model adds when it thinks the
* runtime wants a code block.
*
* Stripping is safe only at a line-leading position — see `std/llm/tool_parse`
* for the near-miss diagnostic that fires when an UNLISTED label prefixes a
* known tool, which is how a new entry for this list gets discovered.
*/
pub const BARE_CALL_LINE_PREFIXES: list<string> = [
"tool_code:",
"tool_call:",
"tool_output:",
"call:",
"tool:",
"use:",
"python:",
"javascript:",
"typescript:",
"shell:",
"bash:",
]
/**
* Pseudo-tools the agent runtime handles itself. They are callable but never
* appear in a user-declared registry, so name resolution has to add them.
*/
pub const IMPLICIT_TOOL_NAMES: list<string> = ["ledger", "load_skill"]
// ── Provider markers ────────────────────────────────────────────────────────
/**
* Mistral's chat template renders calls as `[TOOL_CALLS]name[ARGS]{...}`, or
* as `[TOOL_CALLS]` followed by a JSON array/object payload.
*/
pub const MISTRAL_CALL_MARKER: string = "[TOOL_CALLS]"
pub const MISTRAL_ARGS_MARKER: string = "[ARGS]"
/**
* DeepSeek's DSML markup. The delimiters are FULLWIDTH vertical bars
* (U+FF5C), not ASCII `|` — a detail worth stating because the two are
* visually identical in most fonts and an ASCII spelling silently matches
* nothing.
*/
pub const DSML_MARKER: string = "<｜DSML｜"
/**
* OpenAI Harmony frame tokens that leak into the visible text channel on some
* gpt-oss routes.
*
* `header` markers introduce a role/channel header whose tail runs to the next
* frame, newline, or wrapper open; `standalone` markers consume only
* themselves. Keeping them apart is what stops a later literal `<|message|>`
* in prose or in a tool argument from skipping intervening `<tool_call>`
* blocks.
*/
pub const HARMONY_HEADER_MARKERS: list<string> = ["start", "channel", "constrain"]
pub const HARMONY_STANDALONE_MARKERS: list<string> = ["message", "end", "call"]
pub const HARMONY_FRAME_PREFIX: string = "<|"
pub const HARMONY_FRAME_SUFFIX: string = "|>"
pub const HARMONY_MESSAGE_MARKER: string = "<|message|>"
pub const HARMONY_TOOL_CALL_HEADER_PREFIX: string = "tool_call to="
/** Harn wrapper opens that a Harmony frame token corrupted mid-tag. */
pub const HARMONY_CORRUPTED_OPENERS: list<string> = [
"<tool_call<|",
"</tool_call<|",
"<assistant<|",
]
/**
* One bracket short of the `[[CALL]]` reserved-token wire opener. A model
* that truncates its opener this way is not writing prose — the stub is
* normalized to a canonical tool-call opener so the recover-or-truncated
* ladder runs instead of the text landing in assistant prose (harn#4486).
*/
pub const RESERVED_MALFORMED_CALL_OPENER: string = "[[CALL]"
// ── Fenced-JSON protocol ────────────────────────────────────────────────────
pub const BACKTICK_FENCE: string = "```"
pub const TILDE_FENCE: string = "~~~"
/** The exact info string that opens a tool block canonically. */
pub const FENCE_OPEN_INFO: string = "tool"
/**
* Info strings that open a tool block WITH a protocol warning. Recognizable
* drift is accepted so the turn progresses; telemetry still sees it.
*/
pub const FENCE_DRIFT_INFOS: list<string> = ["json", "tool_code", "tool_call", "function_call"]
/**
* Separators that may follow a bare `tool` info string and still open a tool
* block (`tool python`, `tool_code`, `tool-call`).
*/
pub const FENCE_INFO_SUFFIX_SEPARATORS: list<string> = [" ", "\t", "_", "-"]
/** Chat-template envelopes that wrap tool calls when a model drops the fence. */
pub const ENVELOPE_OPENERS: list<TagDialect> = [
{canonical: "tool_calls", spellings: ["tool_calls"]},
{canonical: "tool_code", spellings: ["tool_code"]},
{canonical: "tool_call", spellings: ["tool_call"]},
]
/** Tag names that are structural markers, never the name of a tool. */
pub const STRUCTURAL_MARKER_TAGS: list<string> = ["tool", "tool_call", "tool_calls", "tool_code"]
// ── JSON envelope key aliases ───────────────────────────────────────────────
/**
* Keys that carry the tool NAME, in precedence order. Canonical wins when
* several are present.
*/
pub const NAME_ALIASES: list<string> = ["name", "tool_name", "tool"]
/** Keys that carry the ARGUMENTS object, in precedence order. */
pub const ARGUMENT_ALIASES: list<string> = ["args", "arguments", "parameters"]
// ── Name normalization (the compat vocabulary) ──────────────────────────────
/**
* Provider protocol tokens appended to a function name, e.g.
* `run<|channel|>commentary`. Everything from the marker onward is residue.
*/
pub const NAME_CHANNEL_MARKERS: list<string> = [
"<|channel|>",
"<|message|>",
"<|recipient|>",
"<|end|>",
]
/**
* Namespace prefixes cheap hosts prepend to otherwise-bare names
* (`tool.look`, `functions.search`). Stripped ONLY for names that are not
* generic wrappers, so `tool.call` / `tool.exec` / `function.call` keep their
* unwrapping path instead of collapsing to `call` / `exec`.
*/
pub const NAME_NAMESPACE_PREFIXES: list<string> = ["tool.", "tools.", "functions.", "function."]
/** Wrapper names whose ARGUMENTS carry the real `{name, args}` call. */
pub const GENERIC_WRAPPER_NAMES: list<string> = [
"tool",
"tool.call",
"tool.exec",
"tool_call",
"function",
"function.call",
"call",
]
/**
* File-explorer namespaces borrowed from other harnesses (Codex
* `repo_browser.*` and its obvious siblings).
*/
pub const BROWSER_NAMESPACE_PREFIXES: list<string> = [
"repo_browser.",
"repository_browser.",
"workspace_browser.",
"file_browser.",
]
/**
* Browser verbs that map onto a canonical Harn tool. A verb NOT listed here
* keeps its prefix-stripped bare name deliberately: rewriting an unknown verb
* to a tool that may not be registered just trips the ceiling again, and the
* denial feedback guides the model better than a wrong rename.
*/
pub const BROWSER_VERB_ALIASES: dict<string, string> = {
open_file: "look",
view_file: "look",
cat_file: "look",
read_file: "look",
print_tree: "look",
list_tree: "look",
list_dir: "look",
list_directory: "look",
ls: "look",
search: "search",
find: "search",
grep: "search",
}
/** Shell/exec synonyms that mean `run`. */
pub const SHELL_ALIAS_NAMES: list<string> = [
"container.exec",
"container_exec",
"exec",
"sh",
"shell",
"bash",
]
/**
* Argument keys that carry a command, in precedence order. Remapped onto
* `command`/`argv` so a renamed call does not then fail arg validation.
*/
pub const COMMAND_ARG_ALIASES: list<string> = ["command", "script", "cmd"]
/**
* Shell binaries and their inline-script flags. An argv of exactly
* `[binary, flag, script]` is really a command STRING, not an argv vector.
*/
pub const SHELL_ARGV_BINARIES: list<string> = ["bash", "sh", "/bin/bash", "/bin/sh"]
pub const SHELL_ARGV_SCRIPT_FLAGS: list<string> = ["-lc", "-c", "lc", "c"]
/**
* Verbs that exist ONLY as `edit({ action: <verb> })` enum values, never as
* advertised standalone tool names — so folding a top-level call of one into
* `edit` cannot shadow a real tool.
*
* `replace_symbol` and `remove_symbol` are deliberately ABSENT: `replace_symbol`
* is hard-kept in Harn's default agent tool surface, so a top-level call is
* legitimate and rewriting it would shadow a real tool and lose the
* symbol-level semantics. `write_file` / `delete_file` / `patch_file` are
* absent too — they are semantically lossy against `edit`, which has no
* raw-write or whole-file-create action, so a silent rename would just fail
* arg validation one step later.
*/
pub const EDIT_ACTION_VERBS: list<string> = [
"create",
"replace_range",
"replace_body",
"insert_after",
"insert_function",
"delete_range",
"exact_patch",
"add_import",
]
/** `intent` values that identify a tool when a wrapper name carried none. */
pub const INTENT_TOOL_ALIASES: dict<string, string> = {
read: "look",
open: "look",
look: "look",
view: "look",
list: "look",
ls: "look",
search: "search",
grep: "search",
find: "search",
}
/** Argument keys whose mere presence implies a `run` call. */
pub const COMMAND_SHAPE_KEYS: list<string> = ["command", "commands", "cmd"]
// ── HTML character references ───────────────────────────────────────────────
/**
* References a text-format model emits when it escapes its own markup
* delimiters inside a JSON string argument, shipping `if (a <= b)` or
* `xs.map(x => x)` as file content. Left encoded, that source cannot
* compile.
*
* Kept to the operators a model escapes to protect markup framing plus the
* two quote forms. An unlisted reference is emitted verbatim rather than
* guessed at, and decoding runs EXACTLY ONCE per value — a second pass is not
* idempotent for double-escaped input (`&lt;` must decode to `<`, the
* intended literal, not collapse to `<`).
*/
pub const HTML_ENTITIES: dict<string, string> = {
"amp;": "&",
"lt;": "<",
"gt;": ">",
"quot;": "\"",
"apos;": "'",
"#39;": "'",
"#34;": "\"",
}
/** The default completion sentinel a `<done>` block is expected to carry. */
pub const DEFAULT_DONE_SENTINEL: string = "##DONE##"