An intent-gated SDLC for AI-generated code. Every gate verdict is computed by a small auditable CLI from an evidence ledger, never asserted by a model, and a red gate physically blocks git commit.
/plugin marketplace add galando/temper /plugin install temper
That’s the whole install. Your first /temper:temper "…" sets the project up: config, scaffold, and the commit gate.
The intent comes first. Then each stage hands its work to the next: criteria become scenarios and tasks, Haiku agents build each task test first, Review sends real bugs back, and Check traces every criterion to evidence before the commit.
INTENT → PLAN → DESIGN? → BUILD → REVIEW → CHECK → COMMIT
You approve the Problem and success criteria before any exploration or architecture spends tokens. A wrong intent multiplies into wrong everything — correcting it here costs words; after Plan it costs the plan.
Each stage’s PASS/FAIL is computed by temper gate from an evidence ledger of real commands and exit codes. A red gate blocks git commit via a native pre-commit hook; a human override is recorded with its reason, never erased.
BDD scenarios are derived from a measured blast radius, so every planned file traces to a behavior — the structural defense against over-engineering.
temper bands watches metric history with control bands (pure arithmetic, no model); a breach is drafted as the next intent and rides the same gates. Fixes write a committed lessons file every future investigation reads first.
Versioned quality packs (security, TDD, quality, performance…) enforce policy during build and review, with deterministic hook backstops. Works with any CI — the automation surface is commands and exit codes, no platform files. Temper vs the AI-native SDLC playbook →
On Claude Code 2.1.287 or later, Temper also draws a stage bar and the run’s pane, and enforces every stage at the tool layer. Older versions load the plugin as before.
An edit outside the current phase is refused, and so is git commit until the gates pass or you override them. Every refusal names the next step.
A stage bar shows each stage with its result, the time, and each loop back with its reason. A pane shows the Build tasks, the findings, the checks and the criteria. You decide each gate in one question; nothing on the bar decides.
uiMode is full, minimal or off. enforcement is on or off. The CLI state stays the source of truth, and the mod blocks nothing when it cannot tell where the run is.
Early access. Runs in the CLI and the desktop Code tab. Not a sandbox: it guards the editing tools and git commit; the Bash check is best effort. Try it on your machine.
Most AI tools only answer “does the code work?” Temper answers all three questions that matter.
Did we solve the problem? Success criteria with structured validation types, accepted by a human at the intent gate before anything is built.
Does it do the right things? Scenarios derived before architecture. Every file traces to a behavior.
Does the code work? Tests written from scenarios, RED → GREEN evidence the build gate requires. The coverage gate ensures nothing is missed.
All three layers live in a single committed artifact — problem, scenarios, coverage checklist — and the chain of commits is the audit trail.
/temper:temper runs and routes everything else.
The whole pipeline, intent gate to commit — or arm autonomy at the plan gate and it runs unattended, parking before commit
Root cause → failing test (write-protected) → minimal fix → validated
Capture an idea as a committed draft — from anyone — and build it later
Granular control: :plan
:design
:build
:review
:check
:status
:pack
— full reference →
/plugin marketplace add galando/temper /plugin install temper
Then just use it: the first /temper:temper in a project bootstraps config, scaffold, and the native commit gate. Auto-detects Spring Boot, React+TS, Node/Express, FastAPI, Go, and Rust. Optional MCP-powered analysis: Recommended Setup.