143 docs
Reference

Agent tools CLI

API reference for the `143-tools` commands available to coding agents.

Coding agents call 143-tools from inside the sandbox to query connected integrations without handling provider credentials directly. The available commands depend on which integrations are configured for the session, so 143-tools --help is the source of truth at runtime.

Coding agent in sandbox
143-tools CLI
143 integration broker
Sentry, Linear, logs, Slack, GitHub, previews
Agents get scoped tool access without seeing provider credentials.

Use this page as a lookup table. For task flows, start with the guide for the integration or preview workflow you are configuring.

CLI contract

Agent-facing commands use a hierarchical shape:

143-tools <namespace> <action> [--flag value ...]
143-tools <namespace> --help
143-tools <namespace> <action> --help
143-tools --help

Results are printed to stdout, usually as JSON. Flags map directly to the tool input schema. Array flags use comma-separated values, for example --states triage,in_progress. Boolean flags accept true or false.

Flat command names such as sentry_list_errors, linear_get_task, log_query, and create_pr are no longer supported. Error messages include the new command shape and the help command the agent should run next.

Use jq when scanning result sets:

143-tools linear list_tasks --team ENG --limit 25 | jq '.[].identifier'

Namespaces

Only configured namespaces appear in 143-tools --help.

NamespaceCommands
sentrylist_errors, get_error, get_error_trend, find_related_errors
linearlist_tasks, get_task, find_related_tasks, update_task, create_task
notionsearch_documents, get_document
githublist_recent_prs, get_pr_reviews
circlecilist_flaky_tests, get_job_test_results, get_recent_test_failures
logsquery, context, fields, stats
previewcreate, ensure, status, list, stop, restart, update, observe, act, control, request_handoff, screenshot, console, inspect, interact, multi_viewport, visual_diff, assert
slacksearch_messages, get_thread, send
issuecreate
prcreate
projectpropose
evaladd
code-review-historylist, get, policy, update_policy, prior-rounds, prior-round, prior-messages

Error tracking

These commands are exposed for configured error trackers such as Sentry. The namespace uses the provider name.

sentry list_errors

List unresolved errors with severity, occurrence count, and affected-user summaries.

FlagTypeRequiredDescription
--projectstringNoProject slug to filter by.
--severitycritical | high | medium | lowNoSeverity filter.
--sincestringNoISO 8601 lower bound for last seen time.
--limitnumberNoMax results. Defaults to 25.

Common use: start a production bug investigation with high-impact unresolved errors.

143-tools sentry list_errors --severity critical --limit 20

sentry get_error

Get full details for one error, including stack trace, tags, and error type.

FlagTypeRequiredDescription
--error_idstringYesError or issue ID from the provider.

Common use: inspect the stack and tags before searching the codebase.

143-tools sentry get_error --error_id 12345

sentry get_error_trend

Get occurrence trend data over a time window.

FlagTypeRequiredDescription
--error_idstringYesError or issue ID from the provider.
--periodstringNoDuration such as 24h, 7d, or 14d. Defaults to 14d.

Common use: decide whether a suspected fix should be urgent because the error is spiking.

143-tools sentry get_error_trend --error_id 12345 --period 24h

Find errors that likely share a root cause, such as a matching stack prefix or culprit.

FlagTypeRequiredDescription
--error_idstringYesError or issue ID from the provider.

Common use: avoid fixing one symptom while missing duplicate production failures.

143-tools sentry find_related_errors --error_id 12345

Tasks

These commands are exposed for configured task managers such as Linear. The namespace uses the provider name.

linear list_tasks

List tasks matching filters with state, priority, and assignee summaries.

FlagTypeRequiredDescription
--teamstringNoTeam key, such as ENG.
--statescomma-separated stringsNoState names such as triage,backlog,in_progress.
--priorityurgent | high | medium | lowNoPriority filter.
--limitnumberNoMax results. Defaults to 25.

Common use: find the highest-priority engineering work related to the current repo.

143-tools linear list_tasks --team ENG --priority urgent --limit 25

linear get_task

Get full task details, including description, comments, and linked issues.

FlagTypeRequiredDescription
--task_idstringYesTask ID or identifier, such as ENG-123.

Common use: read the complete ticket before implementing a fix.

143-tools linear get_task --task_id ENG-123

Find linked tasks, sub-issues, or tasks in the same project.

FlagTypeRequiredDescription
--task_idstringYesTask ID or identifier.

Common use: check whether another task already covers the same root cause.

143-tools linear find_related_tasks --task_id ENG-123

linear update_task

Update task metadata or add a comment.

FlagTypeRequiredDescription
--task_idstringYesTask ID or identifier.
--priorityurgent | high | medium | lowNoNew priority.
--statestringNoTarget state name.
--commentstringNoComment to add.

Common use: leave an implementation note or move a task after completing work.

143-tools linear update_task --task_id ENG-123 --comment "Opened a PR with the fix."

linear create_task

Create a new task.

FlagTypeRequiredDescription
--titlestringYesTask title.
--team_keystringYesTeam key, such as ENG.
--descriptionstringNoMarkdown description.
--priorityurgent | high | medium | lowNoPriority.
--labelscomma-separated stringsNoLabels to apply.

Common use: create follow-up work discovered during a code change.

143-tools linear create_task --team_key ENG --title "Add regression coverage for webhook retries" --priority high

Documents

These commands are exposed for configured document stores such as Notion.

notion search_documents

Search documents by text query.

FlagTypeRequiredDescription
--querystringYesSearch query text.
--workspacestringNoWorkspace or space to search within.
--limitnumberNoMax results. Defaults to 10.

Common use: find design docs, RFCs, postmortems, or launch plans before changing behavior.

143-tools notion search_documents --query "webhook retry policy" --limit 10

notion get_document

Fetch a document's full content.

FlagTypeRequiredDescription
--doc_idstringYesDocument ID from search results.

Common use: read the authoritative product or architecture context for a change.

143-tools notion get_document --doc_id abc-123

Code review

These commands are exposed for configured code review sources such as GitHub.

github list_recent_prs

List recent pull requests with titles, authors, review status, and change size.

FlagTypeRequiredDescription
--statemerged | open | closedNoPR state. Defaults to merged.
--limitnumberNoMax results. Defaults to 20.

Common use: learn recent local patterns before editing the same area.

143-tools github list_recent_prs --state merged --limit 20

github get_pr_reviews

Get reviews and inline review comments for a PR.

FlagTypeRequiredDescription
--pr_numbernumberYesPull request number.

Common use: inspect prior reviewer feedback so a new change follows accepted conventions.

143-tools github get_pr_reviews --pr_number 1234

CI test insights

These commands are exposed for configured CI providers such as CircleCI.

circleci list_flaky_tests

List flaky tests detected by CircleCI.

FlagTypeRequiredDescription
--branchstringNoRestrict to flakes seen on this branch.
--workflow_namestringNoRestrict to a workflow.
--limitnumberNoMax results. Defaults to the provider's full list.

Common use: choose a flaky test to investigate when CI is unstable.

143-tools circleci list_flaky_tests --limit 25

circleci get_job_test_results

Fetch individual test results and failure messages for one CI job.

FlagTypeRequiredDescription
--job_numbernumberYesCI job number.

Common use: read the exact failure output for a flaky occurrence.

143-tools circleci get_job_test_results --job_number 123456

circleci get_recent_test_failures

Get recent failure occurrences for a test, including failure messages.

FlagTypeRequiredDescription
--test_namestringYesTest function or case name.
--classnamestringNoClass or file grouping, recommended for disambiguation.
--limitnumberNoMax occurrences. Defaults to 5.

Common use: compare failure messages across runs before changing a test.

143-tools circleci get_recent_test_failures --test_name "TestWebhookRetries" --limit 5

Logs

Log commands use a shared logs namespace with a --provider flag. Configured providers can include victorialogs and mezmo.

logs query

Run a read-only provider-native log query over a bounded time range.

FlagTypeRequiredDescription
--querystringYesProvider-native query text.
--providerstringNoLog provider to use when more than one is configured.
--sincestringConditionalDuration such as 15m, 1h, or 7d. Required unless --start_time/--end_time are provided.
--start_timestringNoRFC3339 lower bound.
--end_timestringNoRFC3339 upper bound.
--limitnumberNoMax results. Defaults to 100, max 1000.
--directiondesc | ascNoResult order. Defaults to desc.
--fieldscomma-separated stringsNoField names to include.
--include_rawbooleanNoRequest redacted raw provider payloads when authorized.

Common use: investigate recent API errors with a bounded production log search.

143-tools logs query --provider victorialogs --query 'service:api AND level:error' --since 1h --limit 100

logs context

Fetch neighboring logs around an event anchor. At least one of --id, --cursor, or --timestamp is required. Using --timestamp also requires --query.

FlagTypeRequiredDescription
--idstringConditionalStable event or log ID anchor.
--cursorstringConditionalOpaque event cursor anchor.
--timestampstringConditionalRFC3339 timestamp anchor. Requires --query.
--querystringNoProvider-native query text.
--providerstringNoLog provider to use.
--sincestringNoBounded lookback window.
--start_timestringNoRFC3339 lower bound.
--end_timestringNoRFC3339 upper bound.
--beforenumberNoLogs before the target. Defaults to 20, max 100.
--afternumberNoLogs after the target. Defaults to 20, max 100.
--fieldscomma-separated stringsNoField names to include.
--include_rawbooleanNoRequest redacted raw provider payloads when authorized.

Common use: expand from one failed request to the surrounding request lifecycle.

143-tools logs context --provider victorialogs --query 'request_id:"req_123"' --timestamp 2026-06-05T12:00:00Z --since 15m --before 25 --after 25

logs fields

List common indexed or queryable fields for a provider.

FlagTypeRequiredDescription
--providerstringNoLog provider to use.
--querystringNoOptional provider-native query.
--sincestringNoLookback window. Defaults to 24h, max 7d.
--limitnumberNoMax field names or sampled records. Defaults to 100.

Common use: discover available fields before composing a precise log query.

143-tools logs fields --provider victorialogs --since 24h --limit 100

logs stats

Run lightweight provider-native aggregate log stats. This command appears only for providers that support stats.

FlagTypeRequiredDescription
--querystringYesProvider-native query text.
--providerstringNoLog provider to use.
--sincestringConditionalDuration such as 15m, 1h, or 7d. Required unless --start_time/--end_time are provided.
--start_timestringNoRFC3339 lower bound.
--end_timestringNoRFC3339 upper bound.
--group_bycomma-separated stringsNoField names to group by.
--intervalstringNoTime bucket interval, such as 5m or 1h.
--limitnumberNoMax grouped rows. Defaults to 100.

Common use: group errors by service, org, or endpoint before drilling into raw logs.

143-tools logs stats --provider victorialogs --query 'level:error' --since 1h --group_by service --limit 20

Previews

Preview commands use 143's platform-controlled preview lifecycle and browser inspector. Inside a coding-agent sandbox, session-scoped commands default to CODING_SESSION_ID, so screenshots and browser checks run against the live session workspace before the branch is published. Browser diagnostics also accept an explicit --session-id or --preview-id.

preview create

Create or reuse a preview.

FlagTypeRequiredDescription
--session-idstringConditionalSession UUID for a session preview. Recommended for agent iteration.
--repositorystringConditionalRepository full name or unique short name for a branch preview.
--branchstringConditionalPushed branch name for a branch preview.
--waitbooleanNoWait until the preview is ready or fails.
143-tools preview create --session-id "$CODING_SESSION_ID" --wait
143-tools preview create --repository example-org/example-app --branch feature/foo --wait

preview ensure, preview observe, and preview act

Coding-agent sessions can use the native verification loop without passing a preview or session ID:

143-tools preview ensure --wait
143-tools preview observe --path /
143-tools preview act --steps '[{"action":"click","role":"button","name":"Save"}]'
143-tools preview update --wait

observe returns screenshot reference metadata, URL, title, viewport, capture time, a bounded accessibility tree, optional bounded DOM, readiness and browser restoration status, and error-level console messages after the supplied cursor. Pass --output path/to/image.png to write the image into the workspace without keeping base64 in command output. act accepts structured actions such as navigate, click, fill/type, select, check/uncheck, press, hover, scroll, wait for selector/URL/text/readiness/network idle, coordinate click, and viewport changes, then returns the action result and a fresh observation. Navigation is restricted to the active preview origin and the repository's browser.allowed_paths policy.

Use preview control to inspect whether the agent, a human, or a pending handoff owns browser input. When login, MFA, CAPTCHA, or human judgment is required, call:

143-tools preview request_handoff --reason "MFA approval required"

This pauses agent actions until an authorized user takes control in the session preview panel and returns it. The browser URL, cookies, and local storage remain on the same session-owned context throughout the handoff.

preview status

Read status, URL, freshness, and the recommended update mode.

FlagTypeRequiredDescription
--session-idstringConditionalSession UUID for the active session preview.
--preview-idstringConditionalPreview UUID for a branch preview.
143-tools preview status --session-id "$CODING_SESSION_ID"

preview update

Make a session preview reflect recent workspace edits. The platform selects the fastest safe path: browser reload, soft service restart, full recycle, cold relaunch, or no-op. Config overrides escalate to full recycle.

FlagTypeRequiredDescription
--session-idstringNo in a coding-agent sandboxSession UUID. Defaults to CODING_SESSION_ID in a coding-agent sandbox.
--pathstringNoPath to reload/check. Defaults to /.
--waitbooleanNoWait when a restart is started.
--force-modestringNoDiagnostic override: browser_reload, soft_service_restart, full_recycle, cold_relaunch, or noop_current.
--reload-browserbooleanNoReload the browser context when possible. Defaults to true.
--configJSON stringNoOptional preview config override.
143-tools preview update --session-id "$CODING_SESSION_ID" --wait

Browser Inspection

Use these after creating or updating a preview. Pass either --session-id or --preview-id.

CommandRequired flagsPurpose
preview screenshottargetCapture a screenshot. Optional flags: --path, --viewport-w, --viewport-h, --full-page, --delay-ms, --inline-base64. Responses include capture.url, repeated under the former artifact key during the current rollout.
preview consoletargetRead browser console messages. Optional --level error filters output.
preview inspecttarget plus --selector or --x/--yInspect DOM metadata for an element.
preview interacttarget, --stepsExecute browser actions from a JSON step array. Step timeouts accept timeout_ms.
preview multi_viewporttargetCapture mobile, tablet, and desktop screenshots, or pass --viewports JSON. Each capture includes stored-reference metadata.
preview visual_difftarget, --before-snapshot-id, --after-snapshot-idCompare two stored preview snapshots.
preview asserttarget, --assertionsRun browser assertions from a JSON assertion array.
143-tools preview screenshot --session-id "$CODING_SESSION_ID" --path / --viewport-w 1280 --viewport-h 720 --inline-base64 false
143-tools preview inspect --session-id "$CODING_SESSION_ID" --selector '[data-testid=save]'
143-tools preview interact --session-id "$CODING_SESSION_ID" --steps '[{"action":"click","selector":"[data-testid=save]","screenshot":true}]'
143-tools preview multi_viewport --session-id "$CODING_SESSION_ID" --path /
143-tools preview console --session-id "$CODING_SESSION_ID" --level error

Messaging

These commands are exposed for configured message sources such as Slack.

slack search_messages

Search messages by text query.

FlagTypeRequiredDescription
--querystringYesSearch query text.
--channelstringNoChannel name or ID.
--limitnumberNoMax results. Defaults to 10.

Common use: find user reports, incident discussion, or rollout context.

143-tools slack search_messages --query "checkout timeout" --limit 10

slack get_thread

Get a full conversation thread.

FlagTypeRequiredDescription
--message_idstringYesMessage ID of the thread root.

Common use: read the full discussion around a bug report before coding.

143-tools slack get_thread --message_id msg-456

slack send

Send a plain-text Slack message through 143's platform-managed Slack connection. This command appears only when the session or automation has the Slack notification capability.

FlagTypeRequiredDescription
--channel-idstringYesSlack channel ID to send to, such as C123.
--textstringYesPlain-text message body.

Common use: post automation completion or status updates to a channel chosen by the automation owner.

143-tools slack send --channel-id C123 --text "Automation completed successfully."

The command returns delivery state and Slack message coordinates:

{"status":"sent","channel_id":"C123","message_ts":"1700000000.000100"}

143 workflow tools

These commands are exposed by 143 itself when the session has the matching capability.

issue create

Create a new engineering issue and return its UUID.

FlagTypeRequiredDescription
--titlestringYesConcise issue title.
--descriptionstringYesDetailed context and evidence.
--severityinfo | warning | error | criticalNoIssue severity. Defaults to info.
--tagscomma-separated stringsNoTags to categorize the issue.

Common use: file follow-up work from an agent investigation.

143-tools issue create --title "Webhook retries drop idempotency key" --description "Observed while fixing ENG-123." --severity warning

pr create

Queue first-class 143 pull request creation for the current session.

FlagTypeRequiredDescription
--session_idstringNoCompatibility override. Normally omit it: the backend derives the current session from the signed tool token. If supplied, it must match that token.
--draftbooleanNoWhether to create a draft PR. Omit to use the repo default.
--author_modeauto | app | userNoPR author mode. Omit to use the default.
--team-reviewerscomma-separated stringsNoGitHub team slugs to request after the PR is ready. Qualified owner/team names must match the repository owner.
--titlestringNoAutomations with a pull_request publish policy only: the exact single-line PR title, at most 120 bytes.
--bodystringNoAutomations with a pull_request publish policy only: the exact Markdown description, at most 60,000 characters. Mutually exclusive with --body-file.
--body-filepathNoRead the staged Markdown description from a file inside the sandbox. Mutually exclusive with --body.

Common use: open a PR through the same workflow as the app after tests pass.

143-tools pr create --draft false

For a manual session, include a team explicitly requested by the task:

143-tools pr create --draft false --team-reviewers example-org/code-reviewers

Team requests require a ready PR. An explicit draft request, or an inherited draft default without --draft false, is rejected before publication. The backend's draft-first review workflow still requests teams after marking the PR ready.

This command queues the durable 143 publication workflow; it does not call the GitHub PR API from the sandbox. Sandbox GitHub API access is read-only for pull requests, while git push uses a separate repository-bound contents token.

The command returns the workflow's actual asynchronous status: review_in_progress, pr_queued, already_published, manual_publication_required, or blocked. A running review or queued publication is not a created pull request; report the returned state and wait for the session Overview to converge.

For an automation with a frozen pull_request publish policy, the command can register its requested teams before the agent finishes. It returns awaiting_automatic_publication; this stores metadata without queuing early publication. The normal completion and review gates still determine when the PR becomes ready. This path requires the existing publishing capability and does not change the automation's identity or grant new permissions. Registration is rejected if the configured publication will remain a draft. If draft policy changes after registration, normal publication follows that policy and reports that the incompatible team requests were omitted.

Automation metadata calls accept team, title, and description options; draft and author choices remain controlled by the configured publication policy:

143-tools pr create --team-reviewers example-org/code-reviewers

An automation whose goal depends on specific PR text, such as a title prefix or a line a later run searches for, can stage that text the same way:

143-tools pr create --title "docs: Weekly dependency report" --body-file /tmp/pr-description.md

The automatically opened PR uses the staged title and description instead of generated ones. 143 trims surrounding whitespace, inserts linked Linear issue keys such as [ENG-123] at the start of the title (after a conventional-commit type such as feat:, shortening a long title to fit), and appends its standard footer and hidden publication marker to the description. Each call replaces only the options it passes; the last staged value for each option wins until completion freezes the metadata. Other sessions reject --title and --body on pr create; use pr update once their PR exists.

Confirmed team requests are recorded durably. Publication retries do not request the same team again after it has finished reviewing and removed its pending request. An uncertain request outcome is reconciled before any retry. Failure to request one team does not prevent requests to the remaining teams or undo an opened PR. Incomplete requests remain publication-job errors; automatic merge waits for the requested review handoff.

pr update

Update the title and/or Markdown description of the current session's existing primary Pull Request.

FlagTypeRequiredDescription
--session-idstringNoCompatibility override. Normally omit it; if supplied, it must match the signed current-session token.
--titlestringNoReplacement Pull Request title.
--bodystringNoReplacement Markdown description. Mutually exclusive with --body-file.
--body-filepathNoRead the replacement Markdown description from a file inside the sandbox. Mutually exclusive with --body.

At least one of --title, --body, or --body-file is required. Prefer a file for substantial descriptions:

143-tools pr update --body-file /tmp/pr-description.md

The server verifies that the Pull Request belongs to the token-scoped session and repository, updates GitHub with the server-held App installation, preserves 143's hidden publication marker and preview footer, refreshes the local PR mirror, and records an agent audit event. The sandbox's raw GitHub token remains read-only for Pull Requests, so gh pr edit is intentionally not a substitute for this command.

automation execute-action and automation action-status

Available to automation runs with the admin-configured Resumable automation actions capability (automation_actions), including scheduled, manual, event-triggered, and continuous target runs. Each call executes one configured action. There is no required bundle or provider order.

143-tools automation action-status --operation_key daily-report:2026-09-22
143-tools automation execute-action --file /tmp/notification.json
143-tools automation execute-action --resume --operation_key daily-report:2026-09-22 --action_key notify

For a Slack-only workflow, the request file can contain:

{
  "operation_key": "daily-report:2026-09-22",
  "action_key": "notify",
  "kind": "slack_notification",
  "text": "The daily report is ready."
}
KindRequest contentAdmin-configured destination
github_labelpr_number, head_shaRepository and label
github_team_reviewpr_number, head_shaRepository and team slug
github_issue_commentpr_number, head_sha, textRepository
github_issue_comment_updatepr_number, head_sha, text, source_action_idRepository
github_pull_request_reviewpr_number, head_sha, comments, optional textRepository
notion_tracking_rowproperties: property names mapped to string valuesData source and allowed property names/types
slack_notificationtextChannel ID and optional user group ID

GitHub actions require an open PR at the supplied full lowercase head SHA in the session's configured repository. Slack and Notion can omit PR/head preconditions in every run mode, allowing pending steps to resume across pushes. Supply a precondition when the content must remain valid for a specific commit; it is always enforced on resume. In continuous target sessions, any supplied PR must match that target.

The Update this automation’s GitHub PR comments grant authorizes edits separately from comment creation. Supply the UUID of the original successful github_issue_comment receipt as source_action_id. You can find the UUID in the original receipt or in the prior comment's raw body in the server marker <!-- 143-automation-action:UUID -->. The server matches the receipt to the current organization, automation, repository, PR, and action scope, then verifies the comment's ownership and expected body. The server updates that comment in place, preserves the original marker, and adds a marker for the current update. Send the new body in text without the reserved 143 server markers. The update has no GitHub comment ID field or fallback to comment creation.

Keep the same original source_action_id across intentional revisions and give each revision its own action key within the operation. If a revision is interrupted, inspect its status and use the original keys and frozen request when resuming. An uncertain edit requires reconciliation.

The Post an advisory GitHub PR review with inline comments grant publishes one GitHub COMMENT review. Supply one to five comments with path, line, and body. Each line must be on the right side of the current PR diff. The server mentions the PR author in the review summary and records the review ID and permalink. Optional summary text and each inline body are limited to 4 KiB. Leave mentions and reserved 143 delivery markers to the server. Read back published reviews and inline threads using the review ID.

For an unknown inline review send, run automation execute-action --resume with the original keys. 143 checks up to 200 GitHub reviews for the original action marker, commit, and app identity. It records a matching receipt without resubmitting the review. An inconclusive lookup leaves the step in needs_attention.

Notion supports title, rich_text, url, date (YYYY-MM-DD), and select properties. Include the configured title property; select options must already exist. No schema changes are made. Slack text is escaped to prevent mention/link markup injection. The server renders [PR #N](https://github.com/OWNER/REPO/pull/N#pullrequestreview-ID) as a labeled link when its repository matches the session and both PR numbers match. Other link and mention markup remains escaped. An admin can configure slack_user_group_id with a Slack user group ID to have the server prefix each notification with that group's mention. Keep mention markup out of request text. Deploy API and worker support before configuring this field. Before rolling back to older binaries, disable affected automation actions, reconcile unfinished sends, and remove the field from their grants.

For continuous PR targets, an admin can also set slack_pr_author_map_path to a relative JSON file path. The server reads that file from the repository's default branch and matches the live PR author against github_name in its array of entries. Exactly one match with a valid slack_id adds that author's mention. Missing, ambiguous, or invalid user mappings produce a warning and leave the group alert available. Lookup and decoding failures stop before a Slack send and can be retried with the original keys. Deploy API, worker, and frontend support before configuring the path. Before rolling back to binaries without this field, disable affected actions, reconcile unfinished sends, and remove the path from their grants.

Keys contain 1–160 letters, digits, periods, underscores, colons, slashes, or hyphens, beginning with a letter or digit. Reuse an operation key for the same logical workflow and an action key for the same step. Distinct steps can use the same kind. Separate scheduled occurrences should use separate operation keys. Per-target sessions isolate keys by target; other runs isolate keys by automation and repository. Keys survive job retries, later runs, and session reconstruction.

An operation has at most 100 recorded steps. Slack text is limited to 4 KiB, GitHub comments to 32 KiB, and Notion content to 32 KiB across at most 20 properties (8 KiB each). The whole request is limited to 64 KiB. --file cannot be combined with --resume or --request. MCP clients supply the same object as request; resume supplies resume: true, operation_key, and action_key.

Responses include data.status and per-action keys, IDs, states, destinations, and provider receipts. Status is not_started, partial, in_progress, needs_attention, or delivered. Execution returns the selected step; status returns all recorded steps for the operation. delivered describes those recorded steps, not completion of the entire automation goal. Execution prints JSON and exits nonzero for incomplete delivery or an error.

After interruption, inspect status and resume with the original keys. Successful actions return their stored receipts with reused: true, without sending again. Receipts include their creation/completion timestamps and originating/last sending run IDs. Use a distinct operation key for each new occurrence, such as a dated daily report. Pending actions and definite provider rejections can be retried with frozen content. Changed content or destinations under an existing key are rejected. A changed PR head blocks unfinished work with the old precondition.

needs_attention means an uncertain send may have succeeded without a receipt. Report its action ID for operator reconciliation. Never retry using direct provider APIs or new keys. Other independent steps remain callable. Continuous sessions keep their normal read allowlist plus these explicitly granted commands; general provider write tools remain unavailable.

eval add

Add a candidate eval task from a session launched by the eval settings bootstrap flow. This namespace is not exposed to ordinary coding sessions.

FlagTypeRequiredDescription
--pr_numbernumberYesSource pull request number.
--pr_titlestringYesSource pull request title.
--base_commit_shastringYesCommit SHA before the fix.
--solution_commit_shastringYesCommit SHA containing the fix.
--solution_diffstringYesDiff that solved the issue.
--issue_descriptionstringYesReproducible eval task prompt.
--scoring_criteriastringYesJSON array of scoring criteria.
--complexitytrivial | simple | moderate | complexYesCandidate task complexity.
--fitness_scorenumberYesCandidate quality score from 0 to 1.
--fitness_reasoningstringYesWhy the candidate is useful for regression protection.
--evidencestringNoOptional JSON evidence gathered while selecting the candidate.
--warningscomma-separated stringsNoReviewer warnings about ambiguity or setup risk.

Common use: while bootstrapping evals, record a PR-derived coding task that should become a reviewed eval candidate.

143-tools eval add --pr_number 42 --pr_title "Fix checkout timeout" --base_commit_sha abc123 --solution_commit_sha def456 --solution_diff "$(git show --format= --patch def456)" --issue_description "Reproduce and fix the checkout timeout." --scoring_criteria '[{"name":"fixes-timeout","grader_type":"llm_judge","description":"The checkout path no longer times out."}]' --complexity moderate --fitness_score 0.9 --fitness_reasoning "Real regression with clear before/after behavior."

143 code review history

Read access to past automated code reviews for the session's repository, mirroring the session-history namespace. These commands exist so agents can audit how the review policy behaved on real pull requests — which reviews approved, blocked, or escalated, and on what evidence — and then propose policy adjustments. Requires the review_feedback capability.

code-review-history list

List past code reviews for the current repository, newest first. Rows are compact summaries: PR context, decision, status, risk verdict, and the policy_id of the policy version that governed the review. The full review body is only returned by get.

FlagTypeRequiredDescription
--decisionapproved | comment_only | needs_human_review | blockedNoFilter by review decision.
--statusqueued | running | completed | failed | stale | cancelledNoFilter by review run status.
--outcomeautomatically_approved | completed_not_approvedNoFilter by posted outcome.
--acceptablebooleanNotrue for reviews judged acceptable risk, false for reviews flagged for humans.
--searchstringNoMatch PR title, repo name, session title, or PR number.
--created_afterstringNoOnly reviews created after this RFC3339 timestamp.
--created_beforestringNoOnly reviews created before this RFC3339 timestamp.
--cursorstringNoPagination cursor: the id of the last row from the previous page (also returned as meta.next_cursor).
--limitnumberNoMax results. Defaults to 20, capped at 50.

Common use: sample recent decisions before judging whether the policy is behaving as intended.

143-tools code-review-history list --decision blocked --created_after 2026-06-01T00:00:00Z --limit 20

code-review-history get

Get one review in full: the posted review body, every finding (severity, confidence, file and line range, and whether it was posted as an inline comment), and each reviewer agent's structured verdict. Look up by the session_id field from a list row.

FlagTypeRequiredDescription
--session_idstringYesCode review session ID from code-review-history list.
--include_raw_outputbooleanNoInclude each reviewer agent's raw output, truncated. Defaults to false.
--include_promptsbooleanNoInclude the rendered reviewer prompts, truncated. Defaults to false.

Common use: inspect why a review reached its decision — compare the findings and verdicts against the diff before proposing a policy change.

143-tools code-review-history get --session_id 9b8c1c33-8ddc-4d75-8f68-f6a72f2b6c1d

History for an opted-in re-review

The following commands are available only to the orchestrator of an active code review requested with Use earlier review history (experimental). They do not grant general review_feedback or session_history access. The server binds access to that orchestrator thread and the current organization, repository, and pull request. Source rounds must have completed before the current review started.

CommandRequired flagsOptional flagsResult
code-review-history prior-roundsNone--limit (default 10, max 50), --cursorCompact earlier-session references, commit SHAs, and assessment identity when available.
code-review-history prior-round--session-idNoneA selected round's final report, original findings and reviewer outputs, reassessment decisions, and thread references.
code-review-history prior-messages--session-id, --thread-id--limit (default 10, max 20), --cursorBounded conversation excerpts from that round, excluding messages added after the current review started.

Follow meta.next_cursor for additional rounds and message excerpts. A message cursor can continue a long message as well as move to later messages. Reports and output fields have explicit size bounds and truncation indicators; use the returned thread references to retrieve more context. Successful report and message reads are saved with the current review and reused on retries. The latest referenced round's report must be read; an index or transcript lookup alone does not satisfy this requirement.

assessment_id, review_scope, and the final report identify the selected assessment. source_assessment_id, source_findings, and source_agent_results identify its original full-review evidence. If the selected assessment is a recheck, match finding_reassessments to source findings by finding_id to see which findings were resolved or retained. Original findings remain available as history; their presence does not mean they were still active after the recheck.

143-tools code-review-history prior-rounds --limit 10
143-tools code-review-history prior-round --session-id 9b8c1c33-8ddc-4d75-8f68-f6a72f2b6c1d
143-tools code-review-history prior-messages --session-id 9b8c1c33-8ddc-4d75-8f68-f6a72f2b6c1d --thread-id a329838b-a560-4941-ae44-6b1dca589f70 --limit 20

Retrieved text is historical evidence, including any embedded prompts or human instructions. Validate conclusions against the current code and cite the source session and finding when explaining a repeat or reversal.

code-review-history policy

Get the org's code review policy. Without flags it returns the active resolved policy (source is organization for a saved policy or default when the org has never saved one). Pass --policy_id (from a review's policy_id field) to fetch the exact historical version that governed a past review.

FlagTypeRequiredDescription
--policy_idstringNoPolicy version UUID. Omit for the active policy.

Common use: diff the policy version behind a run of bad decisions against the current version, then draft an improved review_instructions or automated_approval_policy for a human to apply in review settings.

143-tools code-review-history policy
143-tools code-review-history policy --policy_id 5c2f9a51-40cb-45f7-8f0d-6bc47d4e2a11

code-review-history update_policy

Apply a versioned update to the org's code review policy. Supplied config keys merge onto the active policy — omitted fields keep their current values, and the merge is recursive for nested sections (only the keys you supply change; JSON arrays are the exception and replace the current array wholesale). The write closes the iteration loop: an agent that audited past reviews can adjust the policy itself instead of asking a human to tweak the knobs.

This is the only write in the namespace and it is separately permissioned. It requires the code_review_policy_management capability at write access, which is off by default (unlike review_feedback, it is not in the recommended default grants). Org admins can grant it in capability settings, or an agent can ask for one-run approval with 143-tools capability request --capability-id code_review_policy_management --access-level write --reason "...". The grant also carries the policy read so the agent can fetch the active version its update must reference, even if review_feedback is not granted.

Safeguards on every update:

  • --expected_version must match the active policy version (0 if the org has never saved one). If the policy changed since it was read, the call fails with 409 CODE_REVIEW_POLICY_VERSION_CONFLICT and the current version — re-read with policy and retry.
  • --reason is required and recorded in the org audit log along with the session that made the change.
  • Policies are insert-only versions; a human can review the audit trail and restore any earlier version from review settings.
FlagTypeRequiredDescription
--configstringYesJSON object of policy fields to change, same shape as the policy output's config (e.g. {"review_instructions":"..."}).
--expected_versionnumberYesActive policy version this change is based on. 0 when the org has never saved a policy.
--reasonstringYesWhy the policy is changing; stored in the audit log. Max 2000 characters.

Common use: after list/get show a pattern of wrong decisions, apply the corrected instructions directly.

143-tools code-review-history update_policy \
  --expected_version 7 \
  --reason "Blocked three doc-only PRs this week; policy now exempts docs-only diffs" \
  --config '{"automated_approval_policy":"Approve documentation-only changes that touch no executable code."}'

Sandbox GitHub CLI transport

The sandbox's installed gh wrapper invokes 143-tools gh <arguments...>. Arguments, stdin, stdout, ordinary stderr, pagination, and exit status follow the GitHub CLI. Its normal transport uses a private Unix socket to forward requests to api.github.com or uploads.github.com over TLS. The host records normalized routes, status, duration, and quota headers with the session's resolved identity. Credentials and HTTP bodies are excluded from telemetry. The CLI cache remains enabled; cached responses never count as remote requests.

Explicit GH_TOKEN, GITHUB_TOKEN, nondefault GH_HOST, and existing Unix transport settings preserve their original transport and bypass observation. Configuration, auth, alias, extension, release, and workflow-run commands also bypass the transport. Direct CLI binaries, arbitrary HTTP clients, and Git fetches are outside this API measurement. API authorization-header overrides are counted with an unknown principal rather than attributed to the session's installation. Observed token-cache hits and controller deferrals are recorded separately.

The standard 143-tools github integration source also reports actual API requests through the credential socket. Its token is resolved once per source, and observation reports reuse that identity without another credential lookup. Static tokens and custom API hosts remain outside installation attribution.

Token minting is attributed to sandbox_token; subsequent API observations use sandbox_gh or sandbox_tools. Tool processes reuse a binding for the same resolved credential and session. Binding capacity failures produce a warning without blocking credential issuance. Cache-hit and local-deferral details are debug logs; their bounded activity counters are recorded regardless of log level.

Observed wrappers disable GH_DEBUG to keep HTTP bodies and headers out of the diagnostic stream. A child terminated by a signal returns status 128 + signal. Archive, artifact, asset-download, and workflow-log API endpoints use the native CLI transport so their redirects retain the CLI's host and credential handling. The observed transport rejects other cross-host redirect targets.

Sandbox reports carry github_observation_trusted=false. The host binds their identity to the credential session, but the reported quota values are sandbox observations and do not control cooldowns or quota admission.

On this page