Execution and limits
Reviewers and executors
A reviewer is a named construction (Bravo or Charlie) at a version, run on an executor that says whose credentials each stage uses.
On GitHub, checks and reviews name only the reviewer and version: "Sigma Charlie v6". They do not name the executor, the models or any cost. The executor is recorded on the run and shown on Sigma's run page and repository settings. Checks and reviews published before public format v2 keep the name they were published with, such as "Sigma Charlie v6 (logins)".
Lineages and versions
Bravo and Charlie are sibling constructions, not a ranking and not release stages. In the "crossed" construction Bravo's producer is Codex and its verifier is Claude; Charlie's producer is Claude and its verifier is Codex.
A version identifies the prompts, the result schema and the attempt rules. Versions 4, 5, 6 and 7 are built in the checked source. Version 7 uses the Claude Agent SDK on the gateway executor only; it does not upgrade historical runs or add a subscription path. From version 6 the verifier returns one verdict per candidate instead of repeating the candidate's text, and Sigma reassembles the frozen texts itself. Running the same version on a different executor does not make a new version. Older Alpha/Beta names and versions keep their original identity in history.
Read the version on the repository's configured reviewer or on the run itself. A newer construction in Sigma's source changes nothing in your repository until an operator binds it there.
Stages: producer and verifier
The producer reads the change and proposes candidates; Sigma records them as suspicions. The verifier, in a fresh context, checks each candidate and confirms, rejects or reclassifies it. Only confirmed candidates are verified findings. For hosted reviews in Comments mode, when the producer proposes nothing, the verifier stage is skipped with the recorded reason "No candidate findings to verify." Gateway reviews contributing to Rules still require verification, even with no candidates. Reviewers never see each other's findings while they work.
Executors
The paths below describe Sigma's configured hosted reviewers. Personal runner self-service on your own computer is not yet available on the hosted service.
An executor records how the two stages ran and whose credentials they used. It says nothing about who owns a Mnemoverse account, who pays an invoice, or whether a configured login still works.
Wide tables: scroll horizontally to see more columns.
| Execution path | Producer | Verifier | Prerequisites |
|---|---|---|---|
Sigma's API keys (gateway) |
API through Sigma's credential gateway | API through the gateway | a configured reviewer and available API budget |
Claude subscription + Sigma API (owner-login) |
the owner's Claude Code subscription login | API through the gateway | a Charlie reviewer, the login enabled on the launcher, and the repository within the scope the operator allowed |
Claude + ChatGPT subscriptions (logins) |
the owner's Claude Code subscription login | the owner's Codex ChatGPT login | a Charlie reviewer, both logins enabled, and the repository on the operator's explicit list of private repositories |
Subscription paths exist for Charlie only, because Bravo's producer is Codex. The Codex login is used only on private repositories the operator listed; Sigma checks the repository's visibility at admission and again right before execution, and treats an unreadable repository as public, so the review is refused. A refused subscription review posts a neutral Review refused check that says "this reviewer is not enabled for this repository" and does not name the login. No model work happens. The exact cause stays in Sigma.
A configured path is not proof of a working login. Runs keep the executor they actually used; changing a repository's configuration does not rewrite history.
For maintainers. Newly connected owner repositories use the installation's configured default path at version 6; existing repository configurations are preserved. Invited repositories use the Bravo and Charlie version 7 gateway roster. Changing the roster is an operator action, not a setting on the page. The modes (Automatic, On request, Off) change schedules only; they cannot add or remove an executor.
For operators. Versions are bound per repository with
--reviewer-version {4,5,6,7}at rebind time; version 7 requires gateway execution. The Codex login path needsCOMPOSE_PROFILES=codex-login, the login volume andSIGMA_OWNER_CODEX_LOGIN_REPOSITORY_IDS; see runbook R4. Each login path has its own egress proxy with a fixed allowlist, described in Where it runs.
Mnemoverse sign-in, GitHub access and model-provider authorization are separate connections. Signing into your profile does not authorize Claude or ChatGPT subscription use, and provider credentials are never accepted in PR comments.
What each path costs you
Reviews and checks in the current format (public format v2) show no cost on GitHub. Cost is recorded on Sigma's run page, for the account that connected the repository. Reviews and checks published before public format v2 can still show a Cost section or cost lines; they keep that wording.
Sigma records money in two different ways and never adds them together:
- API stages are metered by Sigma. The API ledger records known (settled) and held (reserved, unsettled) amounts, derived from tokens. They are estimates, not invoices.
- Subscription stages are observed, not billed. Sigma records the provider tool's own usage report and values it at API list price for comparison. That number is not charged by Sigma and is not counted against Sigma's API spending cap. It says nothing about the provider's subscription price or the remaining plan allowance.
On the run page each stage and attempt has its own line: a producer that needed two attempts has two. Subscription usage is outside Sigma's API cap. A missing line means the information is unavailable, not that the stage was free. See Cost lines and the run page.
Provider constraints
Technical support for a login does not grant permission to offer it commercially. The following provider documentation was checked on 1 October 2026:
- Anthropic: end users can sign into an unmodified Claude Code binary through Anthropic's own flow under the conditions for hosting that binary. This does not permit a third-party app to collect Claude credentials, intermediate subscription usage, or offer its own Claude.ai login. Anthropic legal and compliance.
- OpenAI: Codex supports API-key and ChatGPT sign-in; API usage is billed separately from a ChatGPT subscription. Its authentication guide recommends API keys for programmatic CI/CD. Codex authentication.
- Sign in with ChatGPT is a separate identity and token-sharing integration, with explicit eligibility for locally hosted or open-source and selected commercial applications. It is not a synonym for using a personal Codex login, and Sigma has not implemented it. Token sharing, integration quickstart.
These links describe provider requirements. This guide does not claim that Sigma's current hosted or paid subscription arrangement has been approved.
Usage windows and waiting
A subscription stage can hit its provider's usage window. When the provider reports a reset time, the job waits for it instead of failing:
- Sigma waits for a reset up to 6 hours away (the operator can change this), plus a 60-second margin after the reported time.
- A job waits at most three times in a row; the fourth refusal ends the run as unfinished, with the refused attempts kept in history.
- While waiting, the check reads Review waiting and names the reset time in UTC, without naming the provider. A scheduled resumption is not a promise that the provider will then accept the work.
A reset that is unknown or further away than the waiting window leaves the run unfinished, with recovery guidance on the check. Read that job's status rather than inferring a plan-wide limit from one refusal; Limits and budget lists what the maintainer can do.
Choosing a reviewer for one request
Any repository writer can ask for one exact reviewer on a pull request.
- Prerequisites: the reviewer is configured for this repository, or is a registered version with the same lineage and executor as a configured reviewer. Both the selected reviewer and its configured authority must be enabled, and the repository must remain within the connected account's scope.
- Action: comment exactly
/sigma review <reviewer>@<version>/<executor>, for example/sigma review charlie@6/logins. The App-mention form@<app slug> review <selector>is equivalent. - Expected result: a neutral check "Sigma Charlie v6" appears on the head commit with the title Review queued, then Review in progress, then Review published.
- Check: the first line of the published review names the reviewer, the version, the head and the base you expected. GitHub does not show the executor; the run page does.
- Recovery: an unknown, disabled or unavailable selector posts Reviewer
selection refused without model work; comment
/sigma reviewwithout a selector to run the repository's configured reviewers, or name an enabled one exactly.
For coding agents. The selector grammar is
[a-z][a-z0-9-]*@<version>/(gateway|owner-login|logins); versions 1, 2, 4, 5, 6 and 7 are registered; version 7 requires gateway execution. A selector cannot bypass Off, a private-repository gate or an absent executor. Extra text after the selector is refused.