Harness Wire

link-harness-wire holds the words a harness's messages and skills are written in: Link's own connection's threads, what came of a message, and a skill as the Agent Skills standard writes one. Its runtime, how a host calls a harness in its own process, is in Harness Prompts. Its code is in link-harness-wire.

Execution Evidence

The evidence vocabulary records what a runtime observed about a tool request. A proposal identifies the calls, their original provider IDs, the turn or summary that proposed them, and the admitted work. A call's identity is its successfully appended proposal's sequence and position within that proposal, scoped to the agent and conversation. A proposed sequence is not an identity until the store accepts its append.

Persistence Scope

Committed records are successful appends acknowledged by the owning Store before dependent work proceeds. Recovery covers process interruption while the host retains its storage and treats unfinished effects as unknown. Host failure and power loss remain unverified, since append acknowledgements establish no disk-flush guarantee. Snapshot completeness describes retained records at capture and supplies no stronger persistence guarantee; partial or missing tails remain explicit gaps.

Dispatch Markers

A dispatch marker is persisted after local validation and session startup, immediately before a request may leave the process. Failure to persist that marker prevents sending. The marker proves possible dispatch, and a crash between the marker and sending remains uncertain. Monotonic offsets from the proposal's clock describe the request's interval, independently of the order parallel requests finish.

Matching Replies

A completed request received a usable MCP result with a matching JSON-RPC ID. Both regular JSON and streamed responses must match and contain exactly one usable result or error. A matching JSON-RPC error is a remote error, a local refusal happened before dispatch, and loss of a response after possible dispatch is unknown. A completed result's isError and structured outcome remain separate from task success.

Dispatch Parameters

A dispatch records the actual server label, operation, and bounded arguments sent from the harness to MCP. Original model intent remains in the proposal, so normalizing absent arguments to an empty object is visible. Transport metadata, authorization, and endpoint URLs are excluded. Later namespace routing belongs to its own implementation boundary.

Captured Contents

Reply fields are retained whole or explicitly omitted with their byte count, digest, and reason. Content blocks retain their original indexes. Binary media uses the conversation's existing part references, with decoding, unsupported contents, and storage omissions recorded. These records describe retained contents and omissions, not complete raw JSON. The model-facing text has its own digest, size, and clipping counts.

Work Context

Work context uses the admitted WorkRef. Authorization tokens and unrestricted request metadata are never execution evidence. Evidence versions and unexpected fields are refused rather than inferred.

Validating a Conversation

evidence::index joins records from one native conversation by appended proposal and call position. It refuses orphaned, duplicate, out-of-order, and impossible transitions, including normalized call IDs that collide within a proposal. Missing transitions remain missing. Recovery can omit a finish offset when a new process cannot reconstruct the original monotonic clock.

Skill Loading

Skill records attach either to their containing owner event or to a tool execution. They retain the source read receipt, source and supplied content identities, and the containing model view. A managed read source is selected by the host's checked startup binding. An unregistered service's receipt remains service-asserted. Neither a tool-side whole read nor a familiar path proves that the whole skill reached the model.

Exported Sessions

A local export's versioned manifest classifies every portable update as native or imported. Native records retain allowlisted metadata and the runtime's evidence, including internal events with no visible update. Imported updates remain archival even when their metadata claims native execution. Importing a manifest never establishes local execution.

Internal Media

Internal execution and summary events generate no portable message updates. Their native records retain media part references and report InternalMedia omissions because this export cannot represent their bytes as neutral Agent Client Protocol (ACP) contents. A complete stored snapshot therefore does not establish media fidelity or delivery; assertions requiring the omitted bytes remain unknown.

Snapshot Completeness

Snapshot completeness describes the stored records that were retained. Partial log tails, incomplete compact frames, and a missing prefix are explicit gaps. A store without a completeness guarantee reports unknown. A complete snapshot can still lack evidence or hold unfinished calls. Manifest validation refuses repeated sequences, overlapping or missing update classifications, out-of-bounds ranges, and unsupported versions.

link is Link's own connection, lnk agent chat send's, with no account and no bot, from the agent's owner alone. Its threads are link:<name>, and main unless a chat names one.

A thread's name

link_thread accepts a name of 1 to 64 ASCII letters, digits, -, _, and ., starting with a letter or digit. A single ~ may separate it from a copy suffix with the same character and starting rules, as in main~hermes; the complete name still fits within 64 bytes. Other names are refused.

Checking a thread

is_link_thread checks the link: connection prefix alone. check_link_thread additionally validates the thread name through link_thread, refusing another connection or an invalid name.

What came of a message

Asked is what came of a message sent to the harness: its answer, whole, or that it joined a turn already running, whose answer, on the message that started it, answers it too.

A skill

A skill is a folder named for it, holding SKILL.md, whose front matter says its name and description above its instructions. Link reads only those two: a harness lists them in its prompt, and its model reads the rest when a task needs it.

Its name

check_name takes the standard's names: 1 to 64 lowercase letters, digits and -, with no - first, last or twice in a row.

Reading one

parse reads a SKILL.md into a Skill, its name and its description on one line, and refuses one with no front matter between two lines of ---, or whose name or description is missing or not as the standard says. A description is at most 1,024 characters.

Front Matter Values

The reader accepts plain text, quoted scalars, and indented block values for name and description. It normalizes whitespace into one line and cleans display-control characters before returning the metadata. A list, map, or other unsupported value cannot supply either required field.

Parser Scope

This is a focused metadata reader, rather than a general YAML document loader. Other front matter fields and the instruction body remain outside Skill; the harness owns loading the body when the skill is invoked.