Developing
lnk-harness is Link Harness, the harness a Link agent runs by
default: a small router in Rust that answers you, keeps the thread, and
calls the tools Link runs for it. How to use it, and why it works the
way it does: Link Harness.
Its code
This package, modules/agents/harnesses/link/, is its code. It speaks Link's adapter contract
(link_harness_contract::harness), and Link's release builds it and releases it as
one of Link's modules, at the core's version.
Execution Evidence
Native conversation events use the shared
link-harness-wire evidence vocabulary.
An assistant's proposed calls retain original provider IDs separately
from execution references, which use the successfully appended proposal
sequence and call position. A dispatch marker is committed after local
checks and session startup, immediately before a tool request may leave.
Failure to commit it prevents sending. Each result is committed when it
finishes, while model history restores answers to their call order.
Captured Replies
Matching MCP results preserve structured fields alongside the model's text. Captured reply JSON is bounded to 1 MiB, with whole-field omission digests and reasons. Model-text clipping has separate byte counts and a digest. Binary contents use existing stored part references, and failed decoding or storage remains explicit. Unrestricted reply metadata is omitted, and request authorization tokens never enter these records.
Unknown Outcomes
A matching JSON-RPC error is distinct from a local refusal. Transport loss after possible dispatch and incomplete work after a crash remain unknown. Recovery records no invented monotonic duration and never repeats an uncertain effect. Summary-time proposals and outcomes are appended even when no final summary survives. A non-error MCP result is still not proof that a task's postconditions hold.
Model Attempts
Each selected model attempt records its final canonical request after fallback selection and system substitution. Text and argument digests, source event references, and transformed source spans identify included content without copying prompts or opaque thinking. Source attribution follows tool answers through pairing, and synthesized unanswered replies remain synthetic. Media selection is canonical only, since a provider's conversion can omit a selected image or document.
Saved Configuration
The private harness configuration uses the shared saved-file channel,
declared beside Config, and names schema 1. The shared no-schema policy
reads existing unversioned configuration as the current shape. Missing
read registrations stay unverified. Reads retain the file-identity cache,
and writes retain the atomic private writer with mode 0600. Private
endpoint bindings stay in configuration and never enter execution evidence.
Skill Delivery
Skill receipts must match the actual returned text at a host-registered
read boundary. That text is the call's content, read through its
untrusted-content fence (link_untrusted::unfenced): File's is the text
itself, and Memory's is the document's JSON, its text field checked and
its other fields matched against the structured content. Neither tool
puts the text in its structured content, which holds the receipt alone.
Skill Credit
A whole source receipt cannot grant full credit after model-text clipping. Explicit invocation preserves original input apart from the expanded body. Source loading, inclusion in a recorded model attempt, and following the instructions are separate assertions.
Build and test
From the repository root, with the programs its tests run already built:
cargo build --locked -p link-harness
cargo test --locked -p link-harness
cargo test --locked -p link-harness --test harness turns::cargo fmt --checkcargo clippy --all-targets --locked -- -D warningsIts fake model
Every test's fake model is tests/common's fake_model: it keeps each
request and streams the events the test makes of it. A test that orders
a step against an answer holds each answer after its first piece with
Hold::Until and lets it go with LetGo, so the order is what the test
let go and never how long anything took.
Conformance and the sandbox
tests/conformance.rs runs the conformance kit, the checks every
harness passes. tests/sandbox.rs calls Link Memory as a tool. The tests also run the core's lnk and Link Memory's lnk-memory-tool from the workspace's target/debug. The sandboxed tests skip without bubblewrap unless LNK_REQUIRE_SANDBOX=1 makes that a failure, as CI sets it (Test Environments).