Unified Harness Protocol

Changelog

All notable changes to the Unified Harness Protocol.

2026-10-19, addition: an update can name the version it was built from (ETag, If-Match)

PUT /v1/harnesses/{harness_id} replaces the whole configuration, so a client that builds its body from a read (an app applying the copy of the harness it keeps, an upgrade step) silently overwrote whatever changed between its read and its write. A server MAY now answer a harness's read and update with an ETag, the harness's version, and one that does MUST honour If-Match on that harness's update (Harnesses §5.2): a version that is no longer current is refused 412 precondition_failed and nothing is written; * matches the harness as it is; an update without If-Match replaces as before. The precondition is judged after scope (another caller's harness is still 404) and before the body. Under VERSIONING an optional request header and a response header fit within a version: a server that sends no ETag is unchanged, and a client learns support from the ETag it reads rather than from a capability. 412 joins invalid_request_error's statuses, and the code is unprefixed on the precedent of idempotency_key_reused, as a code this specification now covers.

Conformance suite 2026.10.19.post6 (2026-11-04)

B-01…B-06 check the addition below on a server that reports session_branching: the first response's file bytes after three turns, absence of later files, original-session isolation, inspection provenance, the branch_unavailable refusal for a running response, the 400 for a branch with no previous response (and no task run in the session it named), and a continuation in a branch that names no origin. Missing model-created fixtures remain explicit skips; refusing a valid fresh fixture's branch fails a server that reports the capability. A server without the capability skips all six, so no verdict it had changes.

2026-10-19, addition: explicit session branching (capability session_branching, #477)

branch: true beside previous_response_id creates a new session from that response's committed user files and transcript, leaving the original untouched. Responses identify the origin with metadata.branched_from; session inspection identifies the source session too. Only the response that branched carries branched_from; the session's detail keeps the origin, with the turn it started from. Missing or unreachable checkpoints return 409 branch_unavailable. The request is opt-in; a non-boolean branch, true without a previous response, and branch with an input_response are 400 invalid_input, never run some other way. Under VERSIONING a new capability and an optional request field fit within a version: branching is optional at every class, a server without the capability may ignore branch as every server did before, and a client reads the capability before sending it. Earlier versions served from the same implementation share the additive behavior. The unprefixed code is part of this accepted UEP's canonical specification, following the existing input_not_pending precedent, rather than an independently invented vendor error.

Conformance suite 2026.10.19.post5 (2026-11-03)

T-13 and T-14 check the addition below on a server that reports input_requests. T-13 asks the harness, through the ask_user tool §3.2 names, for a choice between two options, and requires the task to end incomplete with reason input_required, the question as an input_request item and in the last message's text; then an answer outside the options is 422 input_response_invalid, an answer to a request the response does not hold is 409 input_not_pending, the answer that fits continues the task to a terminal status that is not failed, and the same answer again is 409. T-14 asks for a secret, answers with a nonce, and requires the nonce in nothing a client reads: the response to the answer, both responses read back, and the task's input_items. A model that never asked measured nothing, so both skip; a server without the capability skips, so no verdict it had changes. The packaged schema is regenerated from the OpenAPI document, which now carries harness_idempotency as the hand-edited copy already did, and incomplete_details is typed.

2026-10-19, addition: a task can end by asking the person (capability input_requests)

A harness that needs one answer — a choice, a confirmation, a value it must not guess, a secret it must not invent — ends its turn on the question (Tasks §3.2). The response ends incomplete with incomplete_details.reason input_required, the question is its last input_request item (kind is choice, confirm, free_text or secret; options[] carry {id, label}; expires_at is optional) and its last message's text, and the answer is an input_response item on the next task, chained on the response that asked — a normal continuation, the way a function_call_output answers a function_call. Nothing waits: the turn is over, and the question stands until it is answered, passed over by another task (named in metadata.voided_input_requests) or expires. Three refusals, one per remedy: 409 input_not_pending, 410 input_request_expired, 422 input_response_invalid, after which the request stays pending. A secret answer reaches the harness outside the conversation and appears nowhere a client reads. A base without a question tool of its own is given one named ask_user. incomplete widens from "a budget stopped it" to "stopped short, with a reason" (Lifecycle §3), and incomplete_details joins the response's field table it was missing from. Under VERSIONING a new capability, a new output item type a client already tolerates, and an input item a server without the capability never sees fit within a version; the three codes are unprefixed on the precedent of idempotency_key_reused, as codes this specification now covers. Issue #476, with #479's typed request.

Conformance suite 2026.10.19.post4 (2026-10-29)

F-10 checks the addition below on a server that reports harness_idempotency: a repeat of a create's key answers with the first harness, a repeat carrying a different request is refused 409 idempotency_key_reused, and once that harness is deleted a repeat creates one. It deletes every harness it made. A server without the capability skips F-10, so no verdict it had changes.

2026-10-19, addition: a create's Idempotency-Key (capability harness_idempotency)

POST /v1/harnesses (Harnesses §5.1) takes an optional Idempotency-Key, as a task does. Without one, a client whose create answer was lost cannot tell whether the harness exists, and its retry makes a second. A server that reports the new capability keeps a create's key for at least 24 hours in the caller's scope. A repeat answers with the harness the first request created, waits for it while it is still being created, and is refused 409 idempotency_key_reused when it carries a different request. Once that harness is deleted the key names none, and a repeat creates one. Under VERSIONING an optional request header and a new capability fit within a version: a server without the capability ignores the header, as every server did before.

2026-10-19, deprecation: the harness models view's fallback

fallback in GET /v1/harnesses/{harness_id}/models (Harnesses §3) is deprecated. It names the model a server would run in place of one it cannot serve, which Tasks §1.3 makes the harness's default, so it only ever repeats default; a server that refuses an unavailable model (422 model_unavailable), as both maintained servers do, runs none in its place. Nothing changes on the wire: the field keeps its type, and servers keep sending it. Under the deprecation rules it is removed no earlier than the second version after this one. A client reads default and each model's available.

Conformance suite 2026.10.19.post3 (2026-10-19)

No change to the protocol, the schema or any verdict a conformant server gets.

T-12 measures the step budget (Tasks §1.1). A task sent with max_step: 1 whose work needs three tool calls must end incomplete, never completed or failed. Nothing in the suite sent max_step before this, so a server that ignored the budget scored as well as one that honoured it. The fixture gains a tool, uhp_steps, whose answer asks for the next step, and the check counts the rounds from the response's own items: the tool call and the answer to it are two items the server produced, so a model that only claims a call cannot make the count rise. A tool named with a server's own MCP prefix (mcp__<server>__uhp_steps) counts as the fixture's tool.

  • A server that ran the chain past the budget and reported completed or failed fails.
  • A server that names max_step in metadata.ignored_fields, the answer §1.1 gives a server that cannot bound a base, passes.
  • incomplete_details.reason is reported in the result and never failed: §1.1 requires the status and not the reason, and a UEP is what would make the reason a rule.
  • A chain that stayed inside the budget, or never began, says nothing about the rule and is skipped.

T-12 needs the fixture's uhp_steps, so the public copy of the fixture is rebuilt from this commit; until it is, T-12 skips against the default --plugin-mcp-url.

Conformance suite 2026.10.19.post2 (2026-10-19)

The packaged copy of the 2026-10-19 schema marks HarnessModels.fallback deprecated. No check or verdict changes.

Conformance suite 2026.10.19.post1 (2026-10-19)

No change to the protocol, the schema or any verdict a conformant server gets.

T-11 reads a background task back while it is still running and requires the lifecycle table's name for the state, in_progress. No check looked at a task before it reached a terminal status, so a server could answer every background POST and every poll with a name of its own and stay green. HarnessRouter answered running, which the table does not define, while its own stream emitted in_progress for the same turn: a client written to the specification waited for in_progress, or treated the unknown status as an error. The check skips rather than fails when the task finishes before it can be read mid-turn, since there is then nothing left to observe.

2026-10-19, addition: when each turn began and ended

Additive. Each item of a session's turns (Sessions §3) SHOULD carry created_at, when the turn began, and completed_at, when it ended, both in Unix seconds as every other time in the protocol. completed_at is null while the turn is running and when the server does not know when it ended. An imported turn carries the time it was imported for both. A client no longer reads each turn's response to learn its time. TurnItem gains both fields.

2026-10-19, addition: earlier turns (capability input_history)

Additive. A request without previous_response_id may carry the conversation so far as user and assistant messages before the last user message (Tasks §2.1). A server that reports the new input_history capability gives them to the harness as its history, treats only the last user message as the task's own, keeps them as the session's first turns marked imported: true in /turns (completed, no model ran, no cost). Beyond its bounds it imports none of them and reads the request as without the capability, never refusing it. A server without the capability, or one that reports it false, is unchanged.

2026-10-19

Against 2026-10-04, one rule changes and the rest is additive.

Retiring a harness may retire its sessions with it, under a stated restore window (Harnesses §5.3). Every earlier version said a server MUST NOT delete the sessions or responses that used a deleted harness. That stays the rule for a server that reports the new harness_retirement capability false, or says nothing. A server that reports it true MAY remove them with the harness, under three conditions: - it states the window when the harness is retired: the delete response carries sessions: "retired" and restorableUntil (Unix seconds), and its own interface tells the person the window's length before they confirm; - within the window, a restore brings back the harness and its sessions together; - only after the window may it remove them for good. The one exception is a person's own request to erase their data, which may remove their retired sessions sooner; - it never retires a harness mid-task: it cancels the task first, or refuses the delete with 409 and the new code harness_busy, saying why.

A server may withhold them from lists and reads while the window runs. The delete response gains sessions (kept or retired; absent means kept) and restorableUntil. This is the one change that is not additive. Versioning records what it means for a server that also serves an earlier version.

Harness writes refuse what they cannot store, now stated in Harnesses §5.1: a field the harness does not have and a default_model its base does not run are 422, the latter with model_not_available, now in the error table beside the task-time model_unavailable. Both maintained servers have done this since 2026-10-10.

Conformance suite 2026.10.19. New check F-09 (class Full) creates a harness, runs a task on it and deletes the harness. A server whose harness_retirement is false or absent passes when the session still reads afterwards and the response says nothing else. One that reports it true passes when the delete response says sessions: "retired" with a future restorableUntil.

Conformance suite 2026.10.4.post11 (2026-10-15)

No change to the protocol, the schema or any verdict a conformant server gets. ME-01 assumed the memory the suite creates is a root of the server's tree and asked its child for exactly one ancestor. A server whose memories are its own workspaces places a new memory inside the caller's, which the chapter allows, and would have failed for that alone. The child's ancestors are now expected to be the memory's own ancestors and then the memory, root first.

Conformance suite 2026.10.4.post9 (2026-10-06)

No change to the protocol, the schema or any verdict a conformant server gets.

A check that quotes the server's answer in its failure message could not read it. Eighteen messages, reached by eight of the nine Environments checks and all ten Memories checks (in the check itself or in a helper they share), were built from r.text, and a client result had body, json and header() but no text. A message is evaluated only when its assertion fails, so nothing showed against a conformant server. Against a server that really refused the call, the check raised AttributeError, the suite filed it as ERROR, "the check itself broke", and the server's own sentence about why it refused never reached the report: a real non-conformance read as a defect in the suite. A result now has text, the body decoded and never raising, and those checks FAIL with the status and the server's words. Reported and fixed by @Jiaqiiii523 (#411), with a test that reads the checks and refuses any attribute a result does not have; it now also follows an answer that comes back from a helper and one read straight off the call, which between them are a third of the answers the checks read.

A report names the suite that produced it. suite_version was a second copy of the version and had not moved since 2026.9.12.post4, so every report written by 2026.10.4 through 2026.10.4.post8 carried that older name. The version is now written once, in the package, and the build reads it from there. No report published under reports/ was written in that window.

2026-10-04 (patched the same day)

Additive to 2026-09-28: every request and object valid under the previous version is valid here.

Patch of 2026-10-04, same version. A record's content is an ordered list of parts, always, on every read (Memories §4.3). The chapter as first published that morning said content was text or one file reference, which could not hold a picture beside what it shows, or a conversation turn with a file in it. Two kinds of part are defined, text and file; a modality is a media type of a file part, not a kind of part, so a new one needs no change here, and any other kind is x.-prefixed and carried unchanged. A file part may carry the words that stand for it (text, with text_source), which is what recall by words finds; a part may carry the role that said it, so a turn is observed in this same shape. A string remains shorthand for one text part on a write. A file part's bytes are read at GET /v1/memories/{id}/records/{rid}/content/{index} (was /content). A provider's capability document says what it keeps in content (media, bytes, describes), replacing the files flag, and a write of a media type the provider does not keep is refused with memory_unsupported, never stored without its file. Schema: MemoryContentPart; MemoryRecord.content is an array. Suite 2026.10.4.post1: ME-02 expects the list, ME-09 exercises parts, order, refusals, and a file part kept with its description and read back byte for byte, or refused, as the provider declares. No implementation had shipped against the morning's text.

Second patch of 2026-10-04, same version. A memory's id is opaque. The schema had required the hmem_ prefix, which a server whose memories are a tree it already keeps (its own workspaces, its own folders) could meet only with a second id space and a lookup between the two. hmem_ remains the prefix of a server that mints its own ids. Schema: the pattern is gone from Memory.id, the path parameter and a harness's entry. Suite 2026.10.4.post2: ME-01 no longer asserts the prefix.

Third patch of 2026-10-04, same version. recall searches the memory and everything below it the caller may read; the morning's text had it act on one memory. A search is how a caller finds where something is kept, so each result now carries memory (id, name): the place to walk from. It never searches an ancestor. depth narrows it (0 is the memory alone; it used to widen, from a default of one memory). A server MAY cap the memories one question covers and says so in degraded. list, get, history and the queries still act on one memory. The provider capability recall.max_depth is gone: the subtree is the server's to compose. Schema: MemoryRecall.results[].memory is required. Suite 2026.10.4.post3: ME-03 rewritten.

Fourth patch of 2026-10-04, same version. One vocabulary for who acts on a memory. A grant's principal was "an opaque typed identifier" whose kinds each server chose, and a record's written_by named its kinds only by example (harness, consolidator); two implementations had already stamped a person as member and as user. Now four kinds serve both places: member, harness, group (holds grants, never writes) and provider (writes, never holds a grant); a server's own kind is x.-prefixed. consolidator is gone: a consolidation's records are written by the provider, with consolidation_id beside it, and a record derived from an observation carries on_behalf_of. A credential is not a kind. Schema: MemoryWriter, and a pattern on a grant's principal. Suite 2026.10.4.post4: ME-02 checks the writer's kind.

Fifth patch of 2026-10-04, same version. A person and an agent are the same kind, and access has one mechanism. The fourth patch listed member (a person) and harness (an agent) as two kinds; they are one, member, and which of the two a member is, is its type (human, agent), a property no rule depends on. With that, a harness no longer carries a list of attached memories with an access that narrowed a grant: an agent reaches what it was granted, on the memory, as a person does. /v1/harnesses/{id}/memories reads the agent's principal and what it was granted, and sets two things: default_memory_id and observe. An agent joins a group only when someone adds it. Kinds are now member, group and, for a writer only, provider. Schema: HarnessMemories reshaped, HarnessMemorySettings added, HarnessMemoryEntry removed. Suite 2026.10.4.post5.

Sixth patch of 2026-10-04, same version. One graph. Entities and the relationships between them are not a second structure beside records: an entity is a record (new core type entity), a reference is an edge, and a relationship with something to say is a fact that references its subject and its object, so it has a writer, a time, a history and a source like any record. Five reference names are defined (derived_from, part_of, about, subject, object). One read, POST /v1/memories/{id}/graph, returns records as nodes and references as edges around a record or for a memory, the same for every provider; a provider declares where its entities come from in graph.entities (derived, stated, none). Also: an agent's membership of a group follows the rule a person's does (the fifth patch forbade a server to place an agent in a group on its own, which is not equal standing where every new person joins one). Schema: MemoryGraph, MemoryGraphRequest, MemoryRecordRef. Suite 2026.10.4.post6: ME-10.

The chapter's opening, §1, the list of an agent's tools (memory_graph) and §14 were then brought in line with the six patches above; no rule changed.

Seventh patch of 2026-10-04, same version. A record MAY carry a title: one line that names it, apart from what it says (a document's heading, an entity's name, a fact's statement). It is optional, at most 300 characters, matched by a search together with the content, and revised like any other field. A record with a title MAY have empty content. An entity's name is its title. Schema: MemoryRecord.title. Suite 2026.10.4.post7: ME-02 checks that a title reads back apart from the content and may stand alone.

The graph read without around starts from the records of the memory and of what is below it, the scope a search has; the sixth patch's text said the memory alone, which returned nothing for a memory whose records are all kept in its children.

A record names at most one subject and at most one object; a second of either is refused with memory_invalid (422). Stated because two implementations had to agree on it: the reference gateway kept such a record and a graph engine refused it.

Two things the first server built from this chapter by another team read differently from the reference gateway, both the chapter's fault, both now said once: supersedes is the NUMBER of the version a version replaced (the example showed a record id, and the schema allowed either); and a reference the reader may not follow is { record_id, available: false } with no memory_id (the chapter returned the memory's id, which tells the reader that the memory exists).

Records that follow a source (§5.4), found by the second server. A server that shows a document as one record per section cannot let a section be revised or forgotten through the memory: the next save of the document would undo it. The chapter had no word for that, so the server's refusal was undefined and a kept file was caught by the same refusal. Now: such a record carries follows (kind, id, name of its source), revise and forget on it are refused with memory_unsupported in words that say where the source is changed, and a client offers neither. A file someone kept is NOT such a record: its title and its line are its writer's statement, and it is revised and forgotten like any other; forgetting it takes the file out of reads, erasing it removes the bytes. Schema: MemoryRecord.follows, optional.

What a caller was given (GET /v1/memories?granted=true), found by the second server. The listing without parent says where a caller enters the tree, and nothing said where below that it may do more. A server that ran an agent whose memories another server kept built the agent's tools from the entry listing, and an agent that read a memory and wrote one of its children was offered no tool to write with. granted=true lists every memory the caller may read whose parent it cannot see, and every memory below those on which what it may do differs from the parent, with its privileges on each. It is stated in what the caller may do, not in how access is kept, so a server that lets a member in by a role answers it as one that keeps a grant per memory does. §9 names it as what an agent's tools are built from. Conformance: ME-01 asks it.

Facts that change (§4.4), standing records and observed conversations, found while preparing three public memory benchmarks for the second server. A memory that keeps facts has to tell an update ("I moved to Porto") from an addition ("I got a fourth bike") and from a disagreement ("I have never written a route", after describing one), keep the time something was said apart from the time it became true, and give an agent its standing instructions without being asked. The chapter had a word for none of these, so every writer and every server would have invented its own. Now:

  • A fact's attributes have defined keys: predicate, value, holds (one or many), stated_at, as_said, precision, and standing (on any record). Without time.valid_from, a fact became true when it was said. A server that declares holds places a one fact among the others on its subject and predicate by their world time: what was true at T ends at T (a version of the server's own, on_behalf_of the writer), a different value from the same instant is linked with contradicts and both stay, and a fact about the past ends where the next value began. The answer to the write says what it did, in closed and conflicts, never stored.
  • contradicts is a defined rel, kept on both records. A recall result on one side brings the other, after the ranked results, with score null and why ["contradicts"], outside limit.
  • GET /v1/memories/{id}/prime reads a memory's standing records, newest first, within a token budget, from that memory alone unless depth widens it; a server priming an agent reads each memory it was granted at depth 0. The capability prime says whether a provider answers it.
  • An observed episode carries time.valid_from, a content part may carry name beside role, and a conversation is observed one session to an episode, in order. A job says queued, running, completed or failed (pending stays for a provider that cannot tell the first two apart), with what it wrote, closed, conflicts and usage.
  • A revision no longer changes the world time of the version it replaces: that version gets only its belief-time stamp, so as_of reads what the memory believed then. Stated because the second server's store did otherwise; its existing histories needed no repair.

Additive: every request and object valid before is valid now. Schema: MemoryContentPart.name, the episode's time, MemoryJob (memory_id, closed, conflicts, usage, times; queued and running), MemoryRecord.closed and .conflicts on a write's answer, MemoryProvider.prime and .holds, MemoryPrime and its path. Suite 2026.10.4.post10: ME-11 (five facts on one subject: ended, linked, placed in the past, placed inside an earlier value), ME-12 (both sides of a disagreement from a recall of one), ME-13 (the prime read: only standing records, newest first, a budget, the memory alone).

  • Memories (Memories), the Harness Memories sub-protocol, optional at every class behind the memories capability. A memory is a named node in a tree: it holds records and may have child memories. A grant gives a principal privileges (read, write, create, delete) on a node and the node's descendants inherit it, up to a restricted node; what a caller may not read is answered exactly as what does not exist.
  • Records are written two ways (observe: raw episodes, the provider decides what to derive, and may answer 202 with a job; remember: one record as stated) and removed two ways (forget: closed, trace kept; erase: content gone, with a list of what could not be reached). A change appends a version; history, as_of and named snapshots read the past.
  • recall acts on one memory, by meaning, words and fields, and answers with the parent and children the caller may read, so an agent walks the tree itself, up or down. What a provider did not do is said in degraded; abstain says that nothing returned answers the question. Named queries, and free queries where a provider offers them, run in the provider's own language, confined to what the caller may read by a means the statement cannot undo.
  • A server keeps no memory of its own: records are kept by a provider the server is connected to, behind one contract, and each provider says what it does in a capability document (GET /v1/memories/providers): isolation, derivation, signals, history, erase, queries, types.
  • A harness is attached to memories at PUT /v1/harnesses/{id}/memories; attaching is granting. A task names one more in metadata.memory. A session reaches memory at four moments: priming, the agent's tools, observation after a turn, and consolidation, whose runs are visible, bounded and reversible.
  • Extension record types carry their own schema and operations, for resources that are more than text.
  • New objects in the schema: Memory, MemoryRecord, MemoryRecall, MemoryGrant, MemoryProvider, MemoryConsolidation and their requests; 23 paths under /v1/memories and /v1/harnesses/{id}/memories. New error codes: memory_not_found, memory_forbidden, memory_record_not_found, memories_not_attached, memory_invalid, memory_unsupported, memory_busy, memory_unavailable.
  • Conformance suite 2026.10.4: ME-01 to ME-08 drive the tree, records, recall and the provider's capability document on a server that reports the capability and has a provider connected; a server without the capability skips them. They pass, 8 of 8, on a build of the reference server with a hosted memory service as the provider. The suite runs with one credential, so isolation between principals is not yet checked by it.

2026-09-28 (patched 2026-09-29)

Additive to 2026-09-12: every request and object valid under the previous version is valid here.

Patch of 2026-09-29, same version. The task-side reference to an environment is metadata.environment, not a top-level request field: the task surface is a Responses request and metadata is its extension point, where harness_id already travels (Tasks §1.2, Environments §5). The response reports metadata.environment beside session_id. The harness and session objects keep environment as an ordinary field. The published chapter had carried the field at the top level for one day; the schema, the conformance suite (2026.9.28.post1: EN-05 sends it in metadata, EN-07 asserts the response echo; post2: EN-02 compares the returned bytes, which post1 could not read) and the reference server (0.26.7) changed together. The reference server passes the full class, 84 of 84, on post2.

Patch of 2026-09-29, second, same version. A build's record says which step is running (stage) and carries its log while it runs, and a server MUST write it as steps end, so a client can show progress (Environments §4). An optional GET /v1/environments/packages/check answers what the manager's registry says about a spec before any build (EnvironmentPackageCheck), so a client can refuse a typo where it is typed. The environment's status is building while a build runs, whichever version is active; a task keeps reading the active version meanwhile. Suite 2026.9.28.post3: EN-09 exercises the check on a server that offers it; reference server 0.26.10.

  • Environments (Environments), the Environments sub-protocol, optional at every class behind the environments capability. An environment is a project's files and installed dependencies, built once and mounted read-only at a fixed path in every session that names it, beside the session's own writable working directory. The chapter defines the object (henv_), files by path and whole-project import, builds and versions with an active pointer and rollback, the environment field on the harness object and metadata.environment on a task, what a session sees (the mount, the variables, the instruction, the record), and the errors.
  • The session object carries environment (Sessions §3).
  • New error codes: environment_not_found, environment_not_ready, environment_busy, environment_exists, environment_invalid, environment_unavailable.
  • Conformance suite 2026.9.28: checks EN-01 to EN-08 (EN-07 runs one task that reads the environment and tries to write it), skipped on a server without the capability.

Conformance suite 2026.9.12.post4 (2026-09-27)

  • The public fixture answers at https://uhp-fixture.harnessrouter.ai/mcp, the suite's new default; post3 pointed at the container's own address.

Conformance suite 2026.9.12.post3 (2026-09-27)

The protocol is unchanged; the suite that measures it is.

  • The plugin checks point a plugin at a server that resolves and answers. A host may refuse an MCP server it cannot reach at configuration time and record that in skipped; with the unresolvable placeholder the checks used before, such a host failed P-02, P-04, P-06 and P-07 for refusing the placeholder rather than for anything about plugins (the hosted HarnessRouter, 2026-09-27). The suite ships the server (uhp-conformance-fixture, pip install "uhp-conformance[fixture]"), defaults to a public copy of it, and takes --plugin-mcp-url for a copy the host under test can reach.
  • P-11, new: a plugin's MCP server answers the agent's tool call. The agent is asked to call the fixture's tool with a nonce and the fixture's answer, a function of the nonce, must come back. It is the only plugin check that runs a task, and the only one that shows an agent reaching a plugin's server rather than a server deriving and refusing correctly.
  • Measured implementations are re-measured under this suite before their badges are shown again.

2026-09-12

Additive to 2026-08-11: every request and object valid under the previous version is valid here, and a client written against it keeps working against a server that serves this one. The previous version stays published at its own address and a server SHOULD keep serving it.

Added

  • Plugins (Plugins), the Harness Plugins sub-protocol. A plugin is an Agent Plugins 1.0.0 package, carried as files and installed into a harness through a new plugins array on the harness object. The server derives manifest, mcpServers, skills and skipped from the package on every write; the package's servers and skills join the harness's own for every turn. UHP defines no package format of its own: as Agent Plugins wraps Agent Skills by saying where a skill lives and deferring the skill's format, this chapter wraps Agent Plugins by saying how a package travels, binds and runs, and defers the package's contents.

  • The composition rule (Plugins §4). A harness's own mcpServers and skills keep exactly the meaning they had: what a client wrote there, never a plugin's components copied in. This is what makes the version additive. A harness configured before this version is byte-for-byte a valid harness now, with plugins empty, and a client that reads a harness and writes it back cannot install a plugin's servers twice. Component names are unique within a harness across the direct lists and every enabled plugin; a collision is refused at configuration time with plugin_conflict rather than resolved silently at run time.

  • Export (Plugins §5): GET /v1/harnesses/{id}/plugin returns the harness's own tools and skills as an Agent Plugins package, credentials omitted and each omission recorded, that installs into another harness or any other Agent Plugins client unchanged. The direct fields were always the contents of a plugin without a name; this makes that an operation.

  • stdio MCP servers, inside plugins (Plugins §2.2): a plugin's mcp.json may declare a process (command, args, env, cwd), reported on the plugin as a PluginMcpServer. The harness's own mcpServers list stays remote-only and byte-for-byte what it was, because a process needs the root and data directory only a package provides. A server whose base cannot run a transport refuses the plugin with unsupported_transport.

  • Discovery: the plugins capability, optional at every class, and plugin_schemas, the Agent Plugins manifest schemas the server installs (Lifecycle §2).

  • Error codes: plugin_invalid (422), unsupported_plugin_schema (422), unsupported_transport (422), plugin_conflict (409), plugin_not_found (404) (Errors §3).

  • Security §9, plugins are third-party code (Security): stdio servers run inside the agent's sandbox and nowhere else, only the two placeholders are ever expanded, plugins are installed by whoever manages the harness and never per request.

  • Schema: Plugin, PluginManifest, PluginSkill, PluginSkipped, FileList; the skill files endpoint, which the prose named and the OpenAPI document omitted, is now in both.

Conformance

  • P-01 to P-10, class Full, gated on the plugins capability: the schema list, package round trip and derivation, refusal of a package without a manifest, refusal of a name collision, survival of an unrelated edit, export that installs again without credentials, skipped components recorded, enabled: false preserved, refusal of an unsupported manifest schema, refusal of duplicate plugin names. Each is proven against a defect stub carrying one mistake at a time, the way the R-series is. That made the suite 74 checks.

  • X-09, class Extended: a file uploaded through POST /v1/files is accepted and can be sent as task input by id (Files §1.2). X-05 sends its file inline and never reaches the upload endpoint, so a server whose uploads all failed passed the chapter (HarnessRouter CE 0.17.3, #198). Proven against a defect stub: an upload that fails outright, one that answers without an id, and one that truncates each fail X-09 alone while X-05 stays green. The suite is now 75 checks.

2026-08-11, additive clarifications

These were published as additive changes to 2026-08-11 before 2026-09-12 and are carried into it unchanged.

Clarified

  • Session deletion is named (Sessions §6): DELETE /v1/sessions/{id} is the protocol's session delete, with the semantics the old path always had (transcript, trace and working folder go; the session stops counting toward any allowance; a later GET is 404). DELETE /v1/traces/{id} stays as an alias a server MAY keep, which is what the reference implementation does with one handler behind both paths.

  • Session sharing is codified (Sessions §5), closing the four gaps of #44: the share object is schema'd (SessionShare: id and url required, url may be base-relative), DELETE /v1/sessions/{id}/share is the named revocation endpoint (other forms MAY be kept), a bodyless POST publishes, and revocation MUST reach every link minted. Sharing itself remains a MAY. These are the shapes the reference implementation already demonstrates and the R-series already measures; the suite now asserts them instead of probing for them.

  • Turn items have a shape (Sessions §3): each item of GET /v1/sessions/{id}/turns MUST carry id (the response id) and status, and SHOULD carry user / assistant / tools / files. Previously additionalProperties: true and nothing else, which held X-04 at "the endpoint answered 200".

Conformance

  • R-01/R-02 validate the share object against the schema and FAIL (no longer skip) when it carries no url; R-06 asserts the named DELETE revocation endpoint and FAILs (no longer skips) when it does not work; X-04 validates every turn item. New defect stub mode: a server whose revocation exists only as a toggle now fails R-06.

  • R-08: a bodyless POST publishes. The sentence was written into §5 above and left unenforced, named as a known gap rather than hidden. The rest of the R-series mints through a helper that retries a 400/422 with {"enabled": true} — the retry is what lets the series measure a toggle-dialect server at all, and it is also what let this sentence go untested: such a server passes R-01…R-07 on a request §5 does not require anyone to accept, while refusing the one it does. R-08 is the one check that does not retry. New defect stub mode, bodyless_rejected, is what the reference implementation was before this series existed; the matrix proves R-08 is the only check that reddens on it. No specification or schema change: this enforces a sentence that is already written.

  • tools and include are reserved and ignored (Tasks §1.4). Both arrived with the OpenAI Responses wire shape this version stays compatible with, and neither ever had defined semantics in UHP. A server accepts them and does not act on them.

tools cannot mean what it means in the Responses API, where the client executes tools and returns the result as input: UHP puts both the call and its result in output, so there is no input path for the return leg and the loop the field implies cannot be completed. The other plausible reading — per-request MCP servers — is already covered by Harnesses §4.1, and would be an escalation primitive besides: it would let any caller point the agent at an endpoint of their choosing, executed with the harness owner's credentials and workspace. The rule underneath, stated so it can be applied again: narrowing is safe, widening is escalation.

  • metadata.ignored_fields is specified (Tasks §1.1), and required when a request carries a reserved field. Ignoring has to be observable, for the same reason model substitution is reported in §1.3: a silently dropped field is indistinguishable from an honoured one.

  • GOVERNANCE.md records "a declined field is not a pending one" — the general rule these two fields are the worked example of. Due to @aenawi in #42.

Nothing is removed and no client breaks: a request sending either field was already accepted and already had no effect. Removal is a question for the next version.

Conformance

Three checks at class Core, in both directions, where the suite previously sent neither field: T-08 a request carrying them is accepted, T-09 they are reported in metadata.ignored_fields, T-10 a request that sent neither is not told one was ignored. Core goes from 37 checks to 40.

2026-08-11 — first published version

The initial specification, extracted from the shipping HarnessRouter implementation rather than designed in the abstract. Every endpoint and event described here was already running in production before it was specified; the work was to write down the contract precisely, close the gaps that writing it down revealed, and make the result testable.

Defined

  • Architecture — client / server / harness roles, three conformance classes (Core, Extended, Full), the six-object model, transport and authentication.
  • Lifecycle — UHP-Version negotiation, the GET /v1/uhp discovery document, task states (in_progress, completed, failed, incomplete, cancelled), session lifecycle, concurrency rules.
  • Harnesses — discovery, the harness object, model catalogues, computed available, and harness management at class Full.
  • Tasks — POST /v1/responses, the response object, harness selection via metadata.harness_id, model substitution reporting, idempotency.
  • Streaming — the SSE event vocabulary, sequence_number ordering guarantees, reconnection.
  • Sessions — continuation via previous_response_id, listing, inspection, cancellation, sharing.
  • Files — inline and uploaded input, artifact annotations, download, preview, retention.
  • Errors — a single error envelope, a closed set of codes, retry rules.

Gaps this closed in the reference implementation

Writing the specification exposed three places where the implementation had no defined behaviour:

  • No capability discovery. A client had to guess what the server supported, or discover it from a 404. Added GET /v1/uhp.
  • No protocol version on the wire. Nothing identified which contract a response was written to. Added the UHP-Version header on every response.
  • Unstructured errors. Failures returned a bare human-readable string, so a client had to match on prose to decide whether to retry. Added the structured error envelope, with the previous string retained as a deprecated alias so existing clients keep working.

Known compromises

Recorded in VERSIONING.md: mixed field casing between the task and harness surfaces, and session deletion living at /v1/traces/{id}. Both are kept for compatibility and marked for a future major version.

Unified Harness Protocol 2026-10-19 · Apache-2.0 · defined and maintained in the HarnessRouter open-source repository. The standard can be implemented independently of any hosted service.