PRODUCT 02 — EARLY ACCESS THE OPERATION: ROUTINE ENGINEERING WORK NAME PROVISIONAL

From ticket to pull request, without leaving your repository.

Agents that plan a change, write it, test it and open the PR — on your branches, through your CI, inside your review rules. Your engineers keep the merge button.

The tickets nobody wants are the ones a system should be doing.

Bump the dependency. Add the rate limit. Backfill the tests. Real work, well understood, and a poor use of the people who understand it. This product takes those and hands back a reviewable pull request. Never a merge.

01 — A RUN, AS IT ACTUALLY READS

$ svakrit run [ENG-2841]
  “Rate-limit the public search endpoint”

[01] read     4 files · middleware chain, config, tests
[01] plan     token-bucket limiter, reuse request-context
                redis client. no new deps.
     ———— awaiting plan approval ————
[02] approved [reviewer] · “use the shared config key”
[02] applied  comment folded into plan

[03] write    api/middleware/ratelimit.py     [+124]
              api/middleware/__init__.py      [+3  ]
              api/config/defaults.py          [+7  ]
              tests/test_ratelimit.py         [+52 ]

[04] refused  add dependency [ratelimiter@2.1]
              rule [ENG-DEP-00] · no new third-party deps
              wrote against existing client instead
              tradeoff noted in PR body

[05] ci       your pipeline, not ours
              run 1  FAIL  [test_search_paging] (flake)
              run 2  PASS  reported as flaky, not retried away

[06] pr       opened [#1204] → base main
              merge blocked: policy requires named approval

  the agent will not merge, and structurally cannot.
ILLUSTRATIVE · TICKET, FILES, TIMES AND RULE IDS ARE EXAMPLES
POLICY IN FORCE
[ENG-PLAN-01]Plan approved before any edit
[ENG-PATH-02]Never touch infra/ or .github/
[ENG-DEP-00]No new third-party dependencies
[ENG-NET-03]No secrets read, no outbound calls
[ENG-MERGE-01]Named human approval to merge
RUN SUMMARY · ILLUSTRATIVE
Wall clock[18 min]
Human time[2 min] plan review
Refusals1
Silent skips0
THE POINT OF LINE [04]

A refusal names the rule, what the agent wanted, and what it did instead. An agent that quietly worked around the policy would be the failure — not the one that stopped and said so.

02 — THE LOOP, IN A REPOSITORY

The same four steps CortexVigil runs on a camera feed.

That is not a coincidence or a marketing frame. It is the same runtime and the same governance layer, pointed at a git history instead of a control room.

01

Observe

The repository, the ticket, the CI history and the conventions this team already follows. Read, not guessed.

02

Decide

A written plan: what would satisfy this ticket, and would this team accept it in review. You approve or redirect it.

03

Act

Write the change, run your CI, open the pull request on your branch, with the tests that prove it. Never merge.

04

Account

Every tool call, file read, refusal and retry against the run. Reviewable at the next audit, not only while it is live.

03 — WHAT IT DOES

Not a chat window. A colleague with a checklist.

This product is in development and this page describes where it is going. We would rather write that plainly than let you discover it in a trial.

Plans before it edits

Every run opens with a written plan against the real codebase. You approve or redirect it before a single line changes.

Lives in your repository

Your branch naming, your commit conventions, your CI, your PR template. Nothing new for the team to learn or log into.

Writes the tests too

A change arrives with the cases that prove it, and a note on what it deliberately did not cover, so review knows where to look.

Reports failure honestly

A flaky test is named as flaky, not retried until it passes. A ticket it cannot do comes back with a reason rather than a plausible wrong answer.

Leaves a trail

Every tool call, file read, refusal and retry is logged against the run. Reviewable months later, not only while you are watching.

Stays inside your walls

Self-hosted in your VPC, with the model provider you have already approved. Your source never becomes anyone else’s training set.

04 — GUARDRAILS

The interesting part is what it refuses to do.

An agent with commit access is a security boundary before it is a productivity feature. The policy layer is written by you, enforced outside the model where no clever input can argue with it, and logged whether it fires or not.

“The agent decided to” should never be the end of an incident review.

Path and scope limits Directories, file types and services it may never touch — enforced before the edit, not caught in review afterwards.
Named human gates Plan approval, review approval, merge approval, each a specific person rather than a checkbox the agent can tick.
Secret and network isolation No production credentials, no unlisted outbound calls. The sandbox is the default, not the paid tier.
Full audit trail Every action and every refusal, attributed and timestamped, exportable to whatever you already keep logs in.

We are taking a small number of early-access teams.

We would rather work closely with a few teams on real repositories than open a waitlist we cannot serve. If you have a codebase and a backlog you would trust an agent with under supervision, tell us about it.

[hello@svakrit.ai]