Skip to content

CLI reference

harness 2.0.1 exposes 80 commands. Every command accepts --help. Exit codes are documented in exit codes.

Global options go before the command name (harness --no-json run list):

Option Type Required Default Description
--version flag no false Print the harness version and exit
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)
--install-completion flag no Install completion for the current shell.
--show-completion flag no Show completion for the current shell, to copy it or customize the installation.
Command Summary
harness init Create .harness/project.yaml for a repository, using the detected technology profiles, make Git ignore .harness/ (in...
harness inspect Detect the technology profiles of a repository (read-only), with confidence and evidence.
harness doctor Check the local environment (Python and Git required; Node.js and npm reported as NOT_APPLICABLE when absent; Git...
harness do Make a change from TEXT in one command: create the task (the text is its intent and, without --criterion, its...
harness check Run the validators of the gate and the enabled diff checks on the workspace, without recording anything...
harness status Show the full status projection of a run: phases, validations, findings, gate, human decision, event-chain check and...
harness trace Export the trace of a run as Markdown, JSON, JSONL, SARIF or a GitLab Code Quality report (codequality: the findings,...
harness verify Verify the record of a run (or of every run): the event chain, the head of the chain against its anchor outside...
harness export Export a run as a portable evidence bundle: its event chain, records, every artifact they reference and a manifest...
harness verify-approval Check in CI that the range --base..--head is a ChangeSet a person approved: recompute its ChangeSet digest from the...
harness gc Apply the retention settings to runs that ended (closed, cancelled or rejected) longer ago than them:...
harness inbox List the runs of the project that wait for a person, oldest first: a decision in DECISION (gate status, digest,...
harness metrics Local metrics of the governed runs, with zero model calls: tokens (input, output, cache) and cost by agent, model,...
harness registry List the projects whose run registry lives in the state directory (runtime.stateDir: auto, $HARNESS_STATE_DIR or...
harness review-code Review a change with the review panel outside a governed run: reviewers by domain over diff slices, the layered rule...
harness retrospect Derive non-mutating observations and recommendations from a closed run.
harness config validate Validate the project configuration and print the resolved profiles, workflow, validators, capabilities and policies,...
harness config lint Report contradictions between the harness configuration and the agent instruction files (AGENTS.md, CLAUDE.md,...
harness task create Validate a task file (YAML or JSON) and persist it as a versioned task.
harness task list List the tasks persisted in the project.
harness task show Show one persisted task.
harness task questions Show the clarification questions INTENT asked about a task: the open request (asked about the current revision of the...
harness task clarify Answer the open clarification questions of a task and store the revised task.
harness task confirm Confirm the operational contract of a task's current revision, bound to its digest (intake.operationalContract).
harness run start Create a run for a task and execute the normative phases until a stop condition.
harness run continue Resume a run from its persisted state, for example after fixing the environment or after REQUEST_CHANGES.
harness run cancel Cancel a run and terminate its active process group.
harness run cleanup Remove the worktree of a finished isolated run (run start --isolate worktree).
harness run quarantine Keep the changes of a run that stopped without approval as a quarantined patch and restore the workspace to its...
harness run list List the runs of the project, newest first.
harness gate decide Record a human decision bound to the current ChangeSet digest.
harness evidence list List the evidence records of a run with their artifact references and digests.
harness evidence attach Attach evidence.
harness findings list List the structured findings of a run with severity, rule and location.
harness memory add Record a memory entry with its provenance.
harness memory list List the memory records of the project with their status: active, superseded, expired, unapproved, limit or...
harness memory manifest Show the context manifest recorded for a run: the records it applied, their digest and the candidates it excluded...
harness memory approve Approve a proposed record.
harness memory invalidate Invalidate a record without deleting it: it stops entering any context and the actor and the reason stay on record.
harness recommendation list List the retrospective recommendations of a run with the decision recorded for each.
harness recommendation decide Accept, edit or reject a retrospective recommendation.
harness plugins list List the built-in extensions and the external plugin protocol version.
harness benchmark run Run the synthetic microbenchmarks (hashing, transitions, gates, events, artifacts, process launch).
harness benchmark scenarios Compare a direct patch-and-test path with the complete governed path on generated Python and Node.js projects.
harness api serve Serve the local API and web dashboard.
harness artifact show Print an artifact's content after verifying its digest, or its descriptor with --describe.
harness exceptions list List the exceptions of the project (review.exceptions): who granted each, on which run and digest, its scope,...
harness rules health Show how each rule and validator behaved across the runs of the project: how often it fired, blocked a gate, was...
harness outcome record Link something that happened after a run (an incident, a revert, a hotfix, a regression) to it.
harness outcome list List the outcomes linked to the runs of the project, oldest first.
harness plan show Show the decomposition PLANNING proposed for a run (planning.decomposition), its digest, its status and the...
harness plan decide Approve or reject the proposed decomposition, bound to its digest.
harness acceptance show Show the acceptance tests proposed for a run (verification.acceptanceTests), their digest, their status and, once...
harness acceptance decide Approve or reject the proposed acceptance tests, bound to their digest.
harness acceptance dispute Decide a dispute of the frozen acceptance tests (verification.acceptanceTests.disputes): a run whose VERIFICATION...
harness budget show Show the usage of a run and of its task (reported tokens and cost, measured wall time of the agent calls) against the...
harness budget raise Raise a budget limit of a run blocked with budget.exceeded; the raise is recorded on the run's event chain with the...
harness routing calibrate Report the cost per approved task of every routing decision recorded in the project (agentRouting), by provider...
harness agent-contract schema Print the JSON Schema of the request contract (what a parent sends an agent call), of the response contract (what the...
harness agent-contract validate Validate a JSON file against a contract; output is the JSON an agent prints, checked against the limits of its call...
harness agent-contract scripts List the scripts and commands agents may call as black boxes, or show the contract (invocation, inputs, outputs,...
harness agent-contract calls Show the agent calls of a run, the iterations each agent used against its cap and, with --kind (and --iteration),...
harness pr publish Post the decision brief of a run on a pull or merge request (one comment per run, updated when published again) and...
harness pr create Open a pull or merge request from the closure branch of a run (push the branch first), with the repository's pull...
harness pr status Set the commit status governed-harness from a run: success once it closed approved, pending while it waits for a...
harness pr forge Show the forge the workspace resolves to (delivery.forge or the origin remote), its API URL, transport and the...
harness verification show Show the verification plan of a run (the rung each criterion requires, the rungs this environment reaches and why),...
harness verification decide Decide on a run whose preflight is UNAVAILABLE: continue it uncertified for what this environment cannot verify...
harness standards show Show the language standards packs: the detected and effective packs (repository overrides applied), their tools and...
harness architecture show Show the architecture of the project: the configured section, whether the project is new or existing, the cached...
harness architecture decide Decide the architecture, bound to the digest of the proposal: approve or reject the layer rules a survey inferred...
harness architecture refresh Mark the cached architecture survey stale: the next run surveys the project again (one call).
harness project show Show what the harness detects about the project, deterministically and without an agent call: new or existing (and...
harness mcp serve Serve the harness to an agent session over stdio (Model Context Protocol, JSON-RPC 2.0, one message per line).
harness review Show the decision brief of a run: what was asked, what changed, the risks with file and line, what was verified on...
harness review verify Check the review evidence of a commit (refs/harness/review/pass/SHA or pass-warn) without calling any model: the...
harness review variance Run the same review several times without cache and report its variance: verdict stability, how often each finding...
harness review rules show Show the effective review rule catalog: the rules of the three layers (built-in, language packs, project) after...
harness review rules sync Write the generated rules block of every project reviewer (.harness/review/agents) from the catalog, between its...
harness review hook install Install a pre-push hook that runs harness review-code --mode hook: a FAIL or UNKNOWN verdict, or a base that cannot...

harness init

Create .harness/project.yaml for a repository, using the detected technology profiles, make Git ignore .harness/ (in .git/info/exclude, so the tree stays clean) and write an example task. Prints the detected profiles and the next commands.

harness init [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--force flag no false Replace an existing project configuration
--gitignore, --no-gitignore flag no --gitignore Make Git ignore .harness/ (state, artifacts and copies of the code live there)
--ignore-file exclude | gitignore no exclude Where the .harness/ entry goes: exclude writes the repository's .git/info/exclude, so an existing repository's tree stays clean (the .gitignore outside a Git repository); gitignore writes the project's .gitignore
--example-task, --no-example-task flag no --example-task Write an example task to .harness/task.example.yaml
--agent-skills flag no false Also write the skill of the governed flow for Claude Code (.claude/skills/harness/SKILL.md) and Codex (.codex/skills/harness/SKILL.md)
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness inspect

Detect the technology profiles of a repository (read-only), with confidence and evidence.

harness inspect [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness doctor

Check the local environment (Python and Git required; Node.js and npm reported as NOT_APPLICABLE when absent; Git identity reported) and, with --path, the project: its configuration, the agent CLI of the configured provider on PATH, the agent sandbox mechanism, the validators, the baseline commit and whether .harness/ is ignored by Git. Each check that is not PASSED says how to fix it. Exit code 2 when a required check FAILED; a WARNING does not fail.

harness doctor [OPTIONS]
Option Type Required Default Description
--path path no Also validate the project in this directory
--install-hooks flag no false Run the hook installation command the project declares (environment.gitHooks.install) before the checks
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness do

Make a change from TEXT in one command: create the task (the text is its intent and, without --criterion, its acceptance criterion), run it and, on a terminal, answer INTENT's questions and show the decision brief and ask for the decision. No task file is needed. Small, risk-free tasks take the fast lane (friction.fastLane). With --pre-approve the approval you give now is applied only when the gate passed, there is no risk factor and the task is size S. Exit codes as run start: 0 closed, 4 a decision is pending, 6 stopped.

harness do [OPTIONS] {text}
Option Type Required Default Description
text str yes
--criterion str no An acceptance criterion (repeatable; default: the text itself)
--owned str no A workspace path the change may touch (repeatable)
--title str no Task title (default: first line)
--provider str no Agent provider id; defaults to the project agentProvider
--pre-approve flag no false Approve in advance (friction.preAuthorization): applied at DECISION only if the gate passed, the ChangeSet has no risk factor and the task is size S
--pre-approve-hours int range no Validity of the pre-authorised approval
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--interactive, --no-interactive flag no --interactive On a terminal, answer the questions and decide in the same command
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness check

Run the validators of the gate and the enabled diff checks on the workspace, without recording anything (runtime.gateContract). The agent can run it before it finishes; the implement request carries the exact command. Exit 0 when every mandatory validator passes and no enforced check reports a blocking problem, 6 otherwise.

harness check [OPTIONS]
Option Type Required Default Description
--run str no Run whose task and baseline the diff checks use (default: the latest run that wrote a check state)
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness status

Show the full status projection of a run: phases, validations, findings, gate, human decision, event-chain check and metrics.

harness status [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness trace

Export the trace of a run as Markdown, JSON, JSONL, SARIF or a GitLab Code Quality report (codequality: the findings, with the same fingerprints as SARIF). Under governance.verifyRecords: true the run is verified first (as harness verify) and a run that does not verify is not exported (exit code 6).

harness trace [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--format str no markdown markdown, json, jsonl, sarif or codequality
--output path no Write to this file instead of standard output
--path path no Project directory

harness verify

Verify the record of a run (or of every run): the event chain, the head of the chain against its anchor outside .harness (governance.chainAnchor), every record that has an event against that event, the execution's pointers and every referenced artifact against its digest. Prints a report and never repairs anything. With --bundle, verify an evidence bundle of harness export instead: its entries against the manifest, the event chain, the artifacts and the decision records. Exit code 0 when everything verifies, 6 when any check fails (the report's hint then names the next commands for each run that does not verify), 3 for an unknown run or bundle.

harness verify [OPTIONS]
Option Type Required Default Description
--run str no Run (execution) identifier; without it, every run of the workspace
--bundle file no Verify a portable evidence bundle instead (no workspace needed)
--path path no Project directory

harness export

Export a run as a portable evidence bundle: its event chain, records, every artifact they reference and a manifest with the digest of each entry. harness verify --bundle and harness verify-approval --bundle check it without the workspace. A run whose event chain does not verify is not exported (exit code 5; 6 under governance.verifyRecords when the record does not verify).

harness export [OPTIONS]
Option Type Required Default Description
--bundle file yes Write the portable evidence bundle (tar.gz) here
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness verify-approval

Check in CI that the range --base..--head is a ChangeSet a person approved: recompute its ChangeSet digest from the two revisions and require an unexpired APPROVE or APPROVE_EXCEPTION bound to that digest in a bundle (or the workspace record). Any other change in the range fails. Exit code 0 when it passes, 5 when no approval matches.

harness verify-approval [OPTIONS]
Option Type Required Default Description
--base str yes Revision the change starts from (the pull request base)
--head str no HEAD Revision to verify (the merged tree)
--bundle path no Evidence bundle with the approval (repeatable)
--workspace, --no-workspace flag no --workspace Also look for the approval in the workspace's .harness/state.db
--path path no Repository directory

harness gc

Apply the retention settings to runs that ended (closed, cancelled or rejected) longer ago than them: retention.artifactDays deletes the run's artifacts that no kept run uses and records a retention.artifacts.pruned event; retention.eventDays removes the run, its events, records and artifacts. Open runs and memory records are never touched. Without --apply only the report is printed.

harness gc [OPTIONS]
Option Type Required Default Description
--apply flag no false Delete what the report lists; without it nothing is deleted
--path path no Project directory

harness inbox

List the runs of the project that wait for a person, oldest first: a decision in DECISION (gate status, digest, blocking findings, the risk factors an APPROVE must acknowledge), answers to clarification questions in INTENT, and every other wait before DECISION (kind plan, decomposition, acceptance, acceptance-dispute, architecture, contract, deferred or preflight, with the digest it binds to), each with the next command. With --approve, --reject, --request-changes, --decisions or --batch it records several decisions, each bound to its own ChangeSet digest (a batch file may list acknowledgeRisks per decision): a stale digest or a refused decision is reported for that run (the others are recorded) and the command exits 5 when any was refused.

harness inbox [OPTIONS]
Option Type Required Default Description
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)
--approve str no Batch: APPROVE a run as RUN=sha256:DIGEST, the digest the inbox shows (repeatable)
--reject str no Batch: REJECT a run as RUN=sha256:DIGEST (repeatable)
--request-changes str no Batch: REQUEST_CHANGES on a run as RUN=sha256:DIGEST (repeatable)
--decisions file no Batch file (YAML or JSON): a list of run, decision, changeSetDigest and an optional rationale
--batch flag no false On a terminal: decide every pending decision of the inbox one after the other
--rationale str no Batch: the rationale recorded with each decision
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness metrics

Local metrics of the governed runs, with zero model calls: tokens (input, output, cache) and cost by agent, model, task and phase, models used, lines by agent invocation and by person, issues resolved and features delivered, time per task, phase, agent call and human wait, quality, friction against the targets per size and daily and weekly trends. Costs of providers that report only tokens are estimated from the price table and labelled estimated. The HTML file is self-contained (inline CSS and SVG, nothing fetched). No per-person indicator is computed.

harness metrics [OPTIONS]
Option Type Required Default Description
--format str no html, md, json, csv (one row per run) or prometheus; printed, or written to --output. Without it, metrics.json, metrics.md and metrics.html are written to --output-dir
--output path no File for --format
--output-dir path no Directory of the three files written without --format (default .harness/metrics)
--open flag no false Open the HTML report in the default browser
--since str no Only runs created since DATE (YYYY-MM-DD, ISO 8601, 7d, 2w, 12h)
--task str no Only the runs of this task
--model str no Only agent calls of this model
--agent str no Only agent calls of this agent provider id
--all-repos flag no false Every repository of the run registry (runtime.stateDir), not only this one
--prices file no Price table (YAML or JSON, USD per million input, output and cache tokens by model id) for providers that report tokens without a cost; overrides metrics.prices
--narrative flag no false Add a narrative summary: one call to metrics.narrative.command, on demand
--path path no Project directory

harness registry

List the projects whose run registry lives in the state directory (runtime.stateDir: auto, $HARNESS_STATE_DIR or the user's data directory) with their latest runs: the runs of several repositories in one place.

harness registry [OPTIONS]
Option Type Required Default Description
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness review-code

Review a change with the review panel outside a governed run: reviewers by domain over diff slices, the layered rule catalog, a fixed output contract and a verdict the harness recomputes. Rules a tool verifies never reach a model; reviewers run only on their signals; answers are cached. A passing review is recorded as refs/harness/review/pass/SHA (or pass-warn). Exit code 6 when the verdict is FAIL or UNKNOWN (under review.panel.mode enforce) or when hook mode cannot fetch the base.

harness review-code [OPTIONS]
Option Type Required Default Description
--mode str no manual manual: the branch against its base; hook: the same, fetching the base first (a failed fetch aborts); staged: the index against HEAD
--base str no Base branch or commit; default: the pull or merge request base of the CI, the branch convention (review.panel.baseBranches), then the default branch
--head str no HEAD The commit to review (manual and hook)
--provider str no Provider of the reviewers (an agentProviders id or simulated)
--fallback-provider str no Provider of the one retry after an invalid answer
--model str no Force one model for every reviewer
--skip str no A reviewer id to skip (repeatable)
--pull-request int no Pull or merge request number for the comment (default: from the CI environment)
--comment, --no-comment flag no Post one comment per passing result on the pull or merge request (default: review.panel.comment)
--cache, --no-cache flag no --cache Use the review cache
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness retrospect

Derive non-mutating observations and recommendations from a closed run. Nothing is applied automatically.

harness retrospect [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness config validate

Validate the project configuration and print the resolved profiles, workflow, validators, capabilities and policies, the settings that are declared but not applied (declarative) and a warning for each one the project relies on (warnings).

harness config validate [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness config lint

Report contradictions between the harness configuration and the agent instruction files (AGENTS.md, CLAUDE.md, .cursorrules, .cursor/rules, .github/copilot-instructions.md): tool versions, coverage thresholds, statements about the tests and instructions to bypass a control (--no-verify, a forced push, git add -A). Each conflict names the source that wins by instructions.precedence. Exit 6 when there is any issue.

harness config lint [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness task create

Validate a task file (YAML or JSON) and persist it as a versioned task. A task needs at least one acceptance criterion (exit code 2 otherwise), except under intake.criteriaPolicy: enforce: there a task without criteria is stored with criteriaPending: true and INTENT asks for its criteria (rule C0). Under governance.pinTaskRevision: true a task id that has an open run is refused with exit code 5: revise it with task clarify during INTENT or use a new id.

harness task create [OPTIONS]
Option Type Required Default Description
--file file yes Task file (YAML or JSON)
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness task list

List the tasks persisted in the project.

harness task list [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness task show

Show one persisted task.

harness task show [OPTIONS]
Option Type Required Default Description
--task str yes Task identifier (taskId)
--path path no Project directory

harness task questions

Show the clarification questions INTENT asked about a task: the open request (asked about the current revision of the task), every earlier request and the recorded answers.

harness task questions [OPTIONS]
Option Type Required Default Description
--task str yes Task identifier (taskId)
--path path no Project directory

harness task clarify

Answer the open clarification questions of a task and store the revised task. Run harness run continue afterwards to assess the revision in INTENT. An unknown question id or an empty answer exits with 2, no open request with 3, and a task with a run past INTENT or an actor id of an agent, validator or the harness with 5.

harness task clarify [OPTIONS]
Option Type Required Default Description
--task str yes Task identifier (taskId)
--file file yes Answers file (YAML or JSON): answers by question id, optional criteria and requirement changes
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness task confirm

Confirm the operational contract of a task's current revision, bound to its digest (intake.operationalContract). Under enforce INTENT waits for it; then harness run continue. With --pre-approve the same act records an approval in advance bound to the contract digest and the condition gate passed, no risk factor, size S. A stale digest, a non-human actor or --pre-approve without friction.preAuthorization.mode: allow exits 5.

harness task confirm [OPTIONS]
Option Type Required Default Description
--task str yes Task identifier (taskId)
--digest str yes Digest of the operational contract shown by task questions
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--pre-approve flag no false Under friction.preAuthorization: also approve the run in advance, applied at DECISION only if the gate passed, the ChangeSet has no risk factor and the task is size S; otherwise you are asked as usual
--pre-approve-hours int range no Validity of the pre-authorised approval in hours (default friction.preAuthorization.defaultHours, at most maxHours)
--rationale str no `` Why you approve in advance (recorded with --pre-approve)
--path path no Project directory

harness run start

Create a run for a task and execute the normative phases until a stop condition. Exit code 4 means the automated phases finished and a human decision is pending; 6 means a validation, policy or blocking condition stopped the run. With --isolate worktree a branch or worktree directory that already exists stops the run (exit 6); nothing is reset or deleted.

harness run start [OPTIONS]
Option Type Required Default Description
--task str yes Task identifier (taskId)
--provider str no Agent provider id; defaults to the project agentProvider
--isolate str no worktree: run in a Git worktree of its own on a new branch from the updated base branch; none: in place (default: workspace.isolation.mode)
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness run continue

Resume a run from its persisted state, for example after fixing the environment or after REQUEST_CHANGES.

harness run continue [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness run cancel

Cancel a run and terminate its active process group. The cancellation is recorded as an event.

harness run cancel [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness run cleanup

Remove the worktree of a finished isolated run (run start --isolate worktree). Only a worktree the harness created and without uncommitted changes is removed; its branch is kept. Anything else exits 5.

harness run cleanup [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness run quarantine

Keep the changes of a run that stopped without approval as a quarantined patch and restore the workspace to its baseline (governance.stopTheLine). This releases a line blocked by stopTheLine: block. A run that is closed or still open for a decision is refused (exit 5).

harness run quarantine [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness run list

List the runs of the project, newest first.

harness run list [OPTIONS]
Option Type Required Default Description
--path path no Project directory

harness gate decide

Record a human decision bound to the current ChangeSet digest. A decision for a stale digest, or APPROVE over a gate that did not pass, is rejected with exit code 5. On a terminal, omitting --decision, --change-set-digest or --rationale shows the decision brief, asks for the missing values and asks the person to type the start of the digest the decision binds to; a digest that is not confirmed records nothing (exit 5). Without a terminal the three options are required (exit 2). Under governance.confirmDecisionDigest: true the same confirmation is asked on a terminal even when every option is given. An actor id of an agent, validator or the harness is refused with exit 5; without --actor the decider is the Git user under governance.deciderIdentity: git (human.local, with a warning, when Git has no identity).

harness gate decide [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--decision APPROVE | REJECT | REQUEST_CHANGES | APPROVE_EXCEPTION no Human decision; asked interactively when omitted on a terminal
--change-set-digest str no Current ChangeSet digest shown by status or review (sha256:...); confirmed interactively when omitted on a terminal
--rationale str no Justification recorded with the decision; asked interactively when omitted
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--continue, --no-continue flag no --continue Resume the run after recording the decision
--expires-in str no APPROVE_EXCEPTION under review.exceptions: validity such as 14d, 36h or 2w (default review.exceptionDays)
--expires-at str no APPROVE_EXCEPTION: ISO 8601 instant the exception expires
--scope str no APPROVE_EXCEPTION: findings covered, as rule or rule:path (repeatable; default: the blocking findings of the gate, by fingerprint)
--alternative-evidence str no APPROVE_EXCEPTION: evidence that compensates for the blocking findings
--follow-up str no APPROVE_EXCEPTION: issue or task that will remove the exception
--interactive, -i flag no false Show the decision brief and confirm the ChangeSet digest even when every option is given
--acknowledge-risk str no APPROVE or APPROVE_EXCEPTION under verification.riskFactors: a risk factor of the ChangeSet the decider acknowledges (repeatable; required for every factor whose action is acknowledge)
--check str no Under review.manualChecklist: a checklist item the person verified (repeatable); APPROVE needs every item of the run
--change-request str no REQUEST_CHANGES under review.structuredChanges: a blocking item as 'description::condition' where the condition is test:NODE_ID (a pytest node id), absent:REGEX or text (repeatable)
--path path no Project directory

harness evidence list

List the evidence records of a run with their artifact references and digests.

harness evidence list [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness evidence attach

Attach evidence. With --run and --item D-CRITERION, close a deferred verification with a JUnit, SARIF or CI status report (bound to the item's ChangeSet digest and commit; an expired item exits 5). With --manual, store a person's attachment bound to the run's ChangeSet (and to a checklist item with --item), or to a task revision as intake context the agent receives (--task).

harness evidence attach [OPTIONS]
Option Type Required Default Description
--file file yes File to attach
--run str no Run the evidence is about
--task str no Task the attachment is intake context of (with --manual)
--item str no Deferred item (D-CRITERION) to close, or checklist item
--manual flag no false A person's attachment (screenshot, video, log) instead of a test or CI report
--format str no auto auto, junit, sarif or ci-status (deferred items)
--case str no Only the JUnit test cases whose name contains this text
--commit str no Commit the evidence is about (checked against the item's)
--note str no What the attachment shows
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness findings list

List the structured findings of a run with severity, rule and location.

harness findings list [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness memory add

Record a memory entry with its provenance. NORMATIVE, PROJECT and RETROSPECTIVE entries enter a context only once approved.

harness memory add [OPTIONS]
Option Type Required Default Description
--level NORMATIVE | PROJECT | TASK | EPHEMERAL | RETROSPECTIVE yes Scope of the record
--key str yes Stable name of the record within its level
--value str yes JSON object, or plain text stored under the key 'text'
--task str no Task identifier; required for TASK
--run str no Run identifier; required for EPHEMERAL
--valid-until str no ISO 8601 instant after which the record no longer applies
--supersedes str no Identifier of the record this one replaces
--sensitive flag no false Withhold the value from the context manifest and the agent
--approve flag no false Record the entry as approved by the acting person
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness memory list

List the memory records of the project with their status: active, superseded, expired, unapproved, limit or out_of_scope.

harness memory list [OPTIONS]
Option Type Required Default Description
--task str no Evaluate the status for this task
--run str no Evaluate the status for this run
--path path no Project directory

harness memory manifest

Show the context manifest recorded for a run: the records it applied, their digest and the candidates it excluded with the reason.

harness memory manifest [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness memory approve

Approve a proposed record. The approval is a new record that supersedes the proposal; approving a record that needs no approval or is already approved exits with code 5.

harness memory approve [OPTIONS]
Option Type Required Default Description
--memory str yes Memory record identifier
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness memory invalidate

Invalidate a record without deleting it: it stops entering any context and the actor and the reason stay on record.

harness memory invalidate [OPTIONS]
Option Type Required Default Description
--memory str yes Memory record identifier
--reason str yes Why the record no longer applies
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness recommendation list

List the retrospective recommendations of a run with the decision recorded for each.

harness recommendation list [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness recommendation decide

Accept, edit or reject a retrospective recommendation. The decision is kept as retrospective memory: an accepted or edited recommendation enters the context of later runs, a rejected one stays as history. A second decision on the same recommendation exits with code 5; nothing is applied to rules, gates or configuration.

harness recommendation decide [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--recommendation str yes Recommendation identifier shown by retrospect
--decision ACCEPT | EDIT | REJECT yes Decision on the recommendation
--rationale str yes Justification recorded with the decision
--statement str no Edited text of the recommendation; required with EDIT
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness plugins list

List the built-in extensions and the external plugin protocol version.

harness plugins list [OPTIONS]

harness benchmark run

Run the synthetic microbenchmarks (hashing, transitions, gates, events, artifacts, process launch).

harness benchmark run [OPTIONS]
Option Type Required Default Description
--output path no Write to this file instead of standard output
--iterations int range no 1000 Iterations per in-process operation

harness benchmark scenarios

Compare a direct patch-and-test path with the complete governed path on generated Python and Node.js projects.

harness benchmark scenarios [OPTIONS]
Option Type Required Default Description
--output path no Write to this file instead of standard output
--iterations int range no 3 Repetitions per scenario

harness api serve

Serve the local API and web dashboard. Keep it bound to 127.0.0.1.

With the api section that harness init writes, every route requires Authorization: Bearer TOKEN. Your token comes from the variable in api.tokenEnv (HARNESS_API_TOKEN by default); when it is not set, a new token is generated and printed once on standard error. Without the section there is no authentication, as in 1.0.0.

harness api serve [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--host str no 127.0.0.1 Bind address; keep the loopback default
--port int range no 8765 TCP port

harness artifact show

Print an artifact's content after verifying its digest, or its descriptor with --describe. Artifacts are stored redacted; a digest mismatch exits with 1.

harness artifact show [OPTIONS] {reference}
Option Type Required Default Description
reference str yes
--describe flag no false Print the descriptor (digest, size, media type, metadata)
--path path no Project directory

harness exceptions list

List the exceptions of the project (review.exceptions): who granted each, on which run and digest, its scope, alternative evidence and follow-up, when it expires and the runs whose gate relied on it. An expired exception no longer covers any finding.

harness exceptions list [OPTIONS]
Option Type Required Default Description
--status str no all all, active or expired
--expiring-within int range no Only active exceptions expiring within N days
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)
--path path no Project directory

harness rules health

Show how each rule and validator behaved across the runs of the project: how often it fired, blocked a gate, was excepted, fired on a ChangeSet that was later corrected, fired in a rejected run or in a run later linked to an outcome. Read-only; nothing is changed and no figure is per person.

harness rules health [OPTIONS]
Option Type Required Default Description
--since int range no Only runs created in the last N days
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)
--path path no Project directory

harness outcome record

Link something that happened after a run (an incident, a revert, a hotfix, a regression) to it. The outcome feeds harness rules health and the causal retrospective; nothing is applied. An actor id of an agent, validator or the harness exits with 5.

harness outcome record [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--kind str yes INCIDENT, REVERT, HOTFIX, REGRESSION or OTHER
--summary str yes What happened
--reference str no Link or identifier of the incident, revert or fix
--observed-at str no ISO 8601 instant it happened (default: now)
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)
--path path no Project directory

harness outcome list

List the outcomes linked to the runs of the project, oldest first.

harness outcome list [OPTIONS]
Option Type Required Default Description
--run str no Only the outcomes of this run
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)
--path path no Project directory

harness plan show

Show the decomposition PLANNING proposed for a run (planning.decomposition), its digest, its status and the sub-tasks completed so far.

harness plan show [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness plan decide

Approve or reject the proposed decomposition, bound to its digest. APPROVE runs the sub-tasks in order, each with its own verification and gate; REJECT keeps the task whole. A stale digest, a non-human actor or a decision other than APPROVE or REJECT exits 5. An APPROVE with --no-continue exits 0 once it is recorded, as gate decide does.

harness plan decide [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--decision APPROVE | REJECT | REQUEST_CHANGES | APPROVE_EXCEPTION yes APPROVE or REJECT
--digest str yes Digest of the proposed plan shown by plan show
--rationale str yes Justification recorded
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--continue, --no-continue flag no --continue Resume the run after recording the decision
--path path no Project directory

harness acceptance show

Show the acceptance tests proposed for a run (verification.acceptanceTests), their digest, their status and, once approved, the frozen files and their digests.

harness acceptance show [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness acceptance decide

Approve or reject the proposed acceptance tests, bound to their digest. APPROVE writes and freezes them, runs them once on the workspace before the change and resumes the run; every later VERIFICATION checks that they are unchanged and pass. A stale digest or a non-human actor exits 5.

harness acceptance decide [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--decision APPROVE | REJECT | REQUEST_CHANGES | APPROVE_EXCEPTION yes APPROVE or REJECT
--digest str yes Digest shown by acceptance show
--rationale str yes Justification recorded
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--continue, --no-continue flag no --continue Resume the run after recording the decision
--path path no Project directory

harness acceptance dispute

Decide a dispute of the frozen acceptance tests (verification.acceptanceTests.disputes): a run whose VERIFICATION failed only on those tests, after an implementation whose self-report says a test is wrong, waits for a person (kind acceptance-dispute in harness inbox; harness acceptance show lists the disputed tests and the reasons). UPHOLD keeps the tests: the run verifies again and goes on to its correction cycles or fails, as without the dispute. REVISE removes the frozen files and returns the run to SPECIFICATION, where new tests are proposed and wait for harness acceptance decide; once approved and frozen the run verifies the same change again. A stale digest or a non-human actor exits 5.

harness acceptance dispute [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--decision UPHOLD | REVISE yes UPHOLD or REVISE
--digest str yes Digest of the dispute (harness inbox)
--rationale str yes Justification recorded
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--continue, --no-continue flag no --continue Resume the run after recording the decision
--path path no Project directory

harness budget show

Show the usage of a run and of its task (reported tokens and cost, measured wall time of the agent calls) against the budget limits, the limits a person raised and what remains.

harness budget show [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness budget raise

Raise a budget limit of a run blocked with budget.exceeded; the raise is recorded on the run's event chain with the person and the rationale, then harness run continue resumes the run. A non-human actor, a lower limit or a missing rationale exits 5, with a hint that names harness budget show and the command to retry.

harness budget raise [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--scope str yes call, task or run
--metric str yes costUsd, tokens or wallSeconds
--to float yes The new limit; it must be above the current
--rationale str yes Justification recorded
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--path path no Project directory

harness routing calibrate

Report the cost per approved task of every routing decision recorded in the project (agentRouting), by provider family, call kind, task size, model and effort, and suggest the cheapest implement rung per size among groups with at least two approved runs. Nothing is applied: a person edits agentRouting.tables.

harness routing calibrate [OPTIONS]
Option Type Required Default Description
--path path no Project directory

harness agent-contract schema

Print the JSON Schema of the request contract (what a parent sends an agent call), of the response contract (what the parent receives) or of a script contract.

harness agent-contract schema [OPTIONS] {contract}
Option Type Required Default Description
contract str yes

harness agent-contract validate

Validate a JSON file against a contract; output is the JSON an agent prints, checked against the limits of its call kind. Exit 0 when it keeps the contract, 6 when it does not (the violation is printed).

harness agent-contract validate [OPTIONS] {document} {file}
Option Type Required Default Description
document str yes
file path yes
--kind str no Call kind whose limits output is checked against

harness agent-contract scripts

List the scripts and commands agents may call as black boxes, or show the contract (invocation, inputs, outputs, errors) of one.

harness agent-contract scripts [OPTIONS] [script]
Option Type Required Default Description
script str no
--path path no Project directory

harness agent-contract calls

Show the agent calls of a run, the iterations each agent used against its cap and, with --kind (and --iteration), exactly what was sent, what the parent received, and whether the stored request matches its digest and the context the contract allowed.

harness agent-contract calls [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--kind str no Show one call kind
--iteration int no Show one iteration
--path path no Project directory

harness pr publish

Post the decision brief of a run on a pull or merge request (one comment per run, updated when published again) and attach the run's findings: SARIF to GitHub code scanning, a Code Insights report on Bitbucket; GitLab and Azure DevOps read the report from CI artifacts. GitHub is the default; another forge is detected from the origin remote or set with --forge or delivery.forge. Information only: nothing is decided on the forge.

harness pr publish [OPTIONS]
Option Type Required Default Description
--pr int range yes Pull request number
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--repository str no owner/name (default: delivery.publisher or the origin remote)
--transport str no gh (the GitHub CLI) or api (HTTPS with the token in delivery.publisher.tokenEnv, GITHUB_TOKEN by default)
--sarif, --no-sarif flag no Upload the run's findings to code scanning
--commit str no Commit the SARIF report belongs to (default: the PR head)
--forge str no github, gitlab, bitbucket, azure-devops or gitea (default: delivery.forge, else detected from the origin remote)
--path path no Project directory

harness pr create

Open a pull or merge request from the closure branch of a run (push the branch first), with the repository's pull request template and the decision brief as its description and the configured labels. The creation is recorded as a delivery.pull-request.created event. Nothing is approved on the forge: the approval is the run's digest-bound decision.

harness pr create [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--base str no Target branch (default: delivery.forge.baseBranch)
--head str no Source branch (default: the closure branch of the run)
--title str no Title (default: the task title)
--label str no Label, added to delivery.forge.labels (repeatable)
--draft, --ready flag no Open it as a draft (default: delivery.forge.draft)
--forge str no github, gitlab, bitbucket, azure-devops or gitea
--repository str no Repository path on the forge (default: the origin remote)
--transport str no cli (gh or glab) or api (HTTPS, token from the environment)
--path path no Project directory

harness pr status

Set the commit status governed-harness from a run: success once it closed approved, pending while it waits for a human decision, failure otherwise.

harness pr status [OPTIONS]
Option Type Required Default Description
--commit str yes Commit whose status is set
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--target-url str no Link shown with the status (for example the CI job)
--forge str no github, gitlab, bitbucket, azure-devops or gitea
--repository str no Repository path on the forge (default: the origin remote)
--transport str no cli (gh or glab) or api (HTTPS, token from the environment)
--path path no Project directory

harness pr forge

Show the forge the workspace resolves to (delivery.forge or the origin remote), its API URL, transport and the environment variable its token is read from. Calls nothing.

harness pr forge [OPTIONS]
Option Type Required Default Description
--path path no Project directory

harness verification show

Show the verification plan of a run (the rung each criterion requires, the rungs this environment reaches and why), its preflight (READY, PARTIAL or UNAVAILABLE), the certification of its ChangeSet, the deferred items and the manual checklist.

harness verification show [OPTIONS]
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--path path no Project directory

harness verification decide

Decide on a run whose preflight is UNAVAILABLE: continue it uncertified for what this environment cannot verify (those criteria end WAIVED, never certified; the unavailable probes are not run). Recorded on the run with the person and the rationale. A run that does not wait on such a preflight, or a non-human actor, exits 5.

harness verification decide [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--continue-uncertified flag no false Continue without the rungs the preflight found unreachable (required)
--rationale str yes Justification recorded
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--continue, --no-continue flag no --continue Resume the run after recording the decision
--path path no Project directory

harness standards show

Show the language standards packs: the detected and effective packs (repository overrides applied), their tools and the validators added for the tools the repository configures; with --lang every card of one pack; with --file the cards selected for those files. Works without a project configuration. Reads only.

harness standards show [OPTIONS]
Option Type Required Default Description
--lang str no One pack (python, typescript, java, ...) with all its cards
--file str no A workspace path: show the cards an implement call and the review checklist would get for it (repeatable)
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness architecture show

Show the architecture of the project: the configured section, whether the project is new or existing, the cached survey or the ADR, the layer rules in force with their forbidden imports, and the source layout the survey cache is keyed by.

harness architecture show [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness architecture decide

Decide the architecture, bound to the digest of the proposal: approve or reject the layer rules a survey inferred (existing project), or choose one of the options of a new project, which is recorded as an ADR; under architecture.catalogue: experts --style chooses any style of the catalogue instead. The approved rules are enforced as forbidden dependencies in VERIFICATION. A stale digest or a non-human actor exits 5.

harness architecture decide [OPTIONS]
Option Type Required Default Description
--run str yes Run (execution) identifier, latest or a unique prefix such as run_1a2b
--digest str yes Digest shown by the blocked phase
--rationale str yes Justification recorded
--decision APPROVE | REJECT | REQUEST_CHANGES | APPROVE_EXCEPTION no APPROVE or REJECT the layer rules a survey inferred
--option str no The architecture option chosen for a new project
--style str no A style of the catalogue chosen for a new project (architecture.catalogue: experts)
--actor str no Identifier of the person acting (recorded, not authenticated). Defaults to the Git user under governance.deciderIdentity: git, otherwise human.local. Ids of agents, validators and the harness (agent., validator., harness.*) are refused with exit code 5.
--continue, --no-continue flag no --continue Resume the run after recording the decision
--path path no Project directory

harness architecture refresh

Mark the cached architecture survey stale: the next run surveys the project again (one call). Approved rules stay in force until a new survey is decided.

harness architecture refresh [OPTIONS]
Option Type Required Default Description
--path path no Project directory

harness project show

Show what the harness detects about the project, deterministically and without an agent call: new or existing (and why), profiles, standards packs, testing strategy, architecture status, project setup answers and forge.

harness project show [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness mcp serve

Serve the harness to an agent session over stdio (Model Context Protocol, JSON-RPC 2.0, one message per line). Tools: project, status, inbox, task create, questions and clarify (the person's answers, marked as relayed), run start (provider session by default) and continue, check, review and standards. No tool decides for a person: gate, acceptance, plan and architecture decisions stay in the terminal. Register it, for example, with claude mcp add harness -- harness mcp serve.

harness mcp serve [OPTIONS]
Option Type Required Default Description
--path path no Project directory

harness review

Show the decision brief of a run: what was asked, what changed, the risks with file and line, what was verified on which ChangeSet digest and what was not, active exceptions, retries and corrections, what changed since the last decision and the exact decide command. It only reads the record. The subcommands manage the review panel (#57).

harness review [OPTIONS] COMMAND [ARGS]...
Option Type Required Default Description
--run str no latest Run (execution) identifier, latest or a unique prefix such as run_1a2b
--diff flag no false Include the ChangeSet diff (redacted)
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)
--path path no Project directory

harness review verify

Check the review evidence of a commit (refs/harness/review/pass/SHA or pass-warn) without calling any model: the report is bound to the commit, its digest is intact and its diff is the diff between its base and the commit. Exit code 6 when it does not verify.

harness review verify [OPTIONS]
Option Type Required Default Description
--sha str no HEAD The commit whose review evidence to check
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness review variance

Run the same review several times without cache and report its variance: verdict stability, how often each finding appears and the mean pairwise agreement (Jaccard). Every run calls the reviewers again.

harness review variance [OPTIONS]
Option Type Required Default Description
--runs int range no 3 Identical reviews to run
--mode str no manual manual or staged
--base str no Base branch or commit
--provider str no Provider of the reviewers
--model str no Force one model for every reviewer
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness review rules show

Show the effective review rule catalog: the rules of the three layers (built-in, language packs, project) after precedence and supersedes, the inactive rules and why, the reviewers and the ones whose rules block drifted. Reads only.

harness review rules show [OPTIONS]
Option Type Required Default Description
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness review rules sync

Write the generated rules block of every project reviewer (.harness/review/agents) from the catalog, between its BEGIN and END markers; the only writer of that block. With --check nothing is written and a drift exits with 6.

harness review rules sync [OPTIONS]
Option Type Required Default Description
--check flag no false Only report project reviewers whose rules block drifted (exit 6)
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)

harness review hook install

Install a pre-push hook that runs harness review-code --mode hook: a FAIL or UNKNOWN verdict, or a base that cannot be fetched, stops the push.

harness review hook install [OPTIONS]
Option Type Required Default Description
--force flag no false Replace a pre-push hook the harness did not write
--path path no Project directory
--json, --no-json flag no Print JSON, or readable text with --no-json (default: JSON unless standard output is a terminal)