All chapters
  1. Start hereOne change, from diff to production
  2. Before stage 1The team setup: what changes when you are not alone
  3. Stage 1Code change: catch it while the agent is still editing
  4. Stage 2Commit: what must be true before a commit exists
  5. Stage 3Pull request: arrive already reviewed, with evidence
  6. Stage 4Review: who reviews what, at hundreds a day
  7. Stage 5CI/CD: spend scarce capacity like a budget
  8. Deep diveCI/CD agents: where they live and how they run
  9. Deep diveBuild the agents: pinpoint the change, then fix it
  10. Stage 6When the change is an agent
  11. OperateVolume and cost: what breaks at 300 a day
  12. AdoptAdopt it in steps
  13. ReferenceHook events: which one, with a script for each
  14. ReferenceGlossary
  15. AboutAbout this site
Reference

Glossary

Every term this site uses, each with what it is, what it is not, and how it shows up in the running example.

Every term this site uses, in one place. Each card opens the same drill-down as the ? badges on the diagrams and the underlined phrases in the text: one line, why it matters here, what it is, what it is not, and how it shows up in the running example.

Active versionThe one version of a pipeline agent whose claims start the verifier and the revert flow.Bare modeA way to start the agent so that it does not discover project instructions, skills, hooks, agents or tool servers by itself.Behaviour diffThe list of eval and trigger cases whose outcome differs before and after a change.Blast radiusThe size and reach of the harm a wrong change could cause.CalibrationWhether an agent that says 90% is right about 90% of the time.ChangelistOne unit of change that is reviewed and merged as a whole.CI pipelineThe server-side sequence of builds and tests a change must pass.CI triage agentGiven a failing job, it reads the evidence through tools and names a culprit change with a confidence score.Coding agentA model that reads files, edits them and runs commands, step by step, until the task is done.commit-msgGit runs it after the message is written and before the commit is created.Deny ruleA line in agent settings that makes the harness refuse a matching action.End-to-end test environmentA complete running copy of the system, where a test follows a real user flow from start to finish.EvalA fixed set of cases, a way to run the agent on them, and a way to grade the results.Fix-on-request agentA subagent that fixes one review finding on a pull request's branch when a person asks for it in a comment.Frozen worldThe logs, changes and bug records for a case, saved as they were.Gate filesThe files that decide how other changes are judged or what an agent loads and may do: .github/**, CODEOWNERS, CLAUDE.md, .claude/**, .mcp.json, .githooks/**, agents/** and ci/**.Git hookAn executable file, named after a git event, that git runs for anyone who commits or pushes in that clone.Golden setThe labelled cases an eval runs on.HarnessThe program around the model: it runs tools, applies permissions and fires hooks.Headless runOne non-interactive run: a prompt in, a result out, then it exits.Hook eventThe moment a program offers for your script to run, such as before a commit or after an edit.HookA script or call that runs automatically at a defined point in the agent's work.Hybrid replayA tool cache keyed by the request, with live fallback and a count of misses.Judge modelA model given a rubric and a transcript, asked for a grade.lane checkThe required check named lane. A workflow computes green, yellow or red from the changed paths; the policy app posts the check and passes it only when that lane's approval condition is met.LaneGreen, yellow or red: how much review and testing a change needs.MCPA standard way for an agent to discover and call tools that a separate program, a tool server, provides.Merge queueA line of approved changes, each tested against main plus the changes before it.PluginOne folder that carries several agent components, so they can be loaded or shared together.Policy appAn app identity that approves green-lane pull requests and re-evaluates the lane check when a review arrives.post-commitGit runs it after creating the commit. It cannot stop anything.PostToolUseThe harness fires it after a tool call succeeded. The script cannot undo the call; it can tell the model what to fix.pre-commitGit runs it at the start of every commit. A non-zero exit means no commit.pre-pushGit runs it once before sending commits, however many commits the push contains.PreToolUseThe harness fires it after the model asks for a tool call and before the call runs. The script can cancel the call.Prompt cacheThe provider stores the stable start of a prompt so later calls read it cheaply.Pull requestA branch offered for merging, plus the review and the check results attached to it.Red-build fix agentA pipeline agent that proposes a repair on its own branch after the verifier supports a claim and main is green again.Retry budgetA fixed number of automatic re-runs per change, counted across its whole lineage, before a person must approve more.Review receiptA record of the inputs and verdict of a review. Signed by a trusted runner, it lets CI verify instead of repeat.Review refA branch name pattern, such as review/..., whose push starts the trusted review runner.Reviewer agentAn agent with no edit tools, in a fresh context, that reviews a diff and posts findings.SandboxLimits on files and network that apply whatever the command text says.Separation of dutiesWhoever wrote a change, person or agent, is not the one who approves it.Shadow modeThe new version sees real inputs and its answers are recorded, not used.Skill fileA text file with a name, a description and instructions for one kind of task.Smoke setA small part of the golden set that runs on every change.Stand-in tool serverA small server with the same read-only tool interface as production, backed by frozen data.SubagentA separate agent that the main agent hands one task to, or that a whole run can take the role of.Test selectionChoosing tests from the diff instead of running all of them.TraceThe ordered list of tool calls, results and model messages in one run.Trigger testA table of prompts with the skills each one may, and may not, invoke.TurnOne round of the agent loop that starts with a request and ends when the model stops and answers.VerifierA job with no model in it that tests whether the agent's answer holds.Workflow fileA file in the repository that tells the CI system what to run, when, and with which permissions.