Tdd Workflow

TDD Workflow — Red-Green-Refactor Protocol

This skill teaches the protocol Claude follows AFTER a test watcher is running. For watcher setup (starting the watcher pane and managing long-running processes), see terminal:terminal-interaction. For framework-specific pass/fail/running/idle signals, see terminal:framework-signals.

The TDD loop requires a dedicated pane topology: Claude edits code in one pane while a persistent test watcher runs in a second pane. Claude reads the watcher with watch-pane and capture-pane but never restarts it.

Overview

Pane Topology

┌──────────────────────┬────────────────────┐
│  Claude work pane    │  Test Watcher      │
│  (Bash / code edit)  │  (helper pane)     │
└──────────────────────┴────────────────────┘

The left (work) pane is where Claude runs file edits and any auxiliary commands. The right (watcher) pane runs the test framework in watch mode for the entire session. Claude only reads the watcher pane — it never sends keystrokes to it except at teardown.

State Machine

The TDD protocol is a 5-state machine. Always know which state you are in before taking an action.

RED    → parse failure (file:line:test name from watch-pane output)
         → edit code → save → enter WAITING

WAITING → watch-pane fires "pattern:RUNS|Running" → change detected
         → if watch-pane times out: re-save with touch, watch-pane again
         → proceed to poll for RED or GREEN

GREEN  → report pass count → leave watcher running → ask user to continue

COMPILE_ERROR → fix compile error first → re-enter WAITING

IDLE   → Ctrl+C watcher pane → report summary: N tests fixed

State Transitions

Priority Rule

COMPILE_ERROR takes priority over RED. If the watcher output shows both error[E and a test failure, fix the compile error first — the test failure may disappear once the code compiles.

Critical Timing Rules

After saving a file, wait for the "change detected" signal before reading results.

• Do not read the watcher output immediately after a save. The watcher has not yet noticed the change.
• Use watch-pane with a trigger pattern for the framework's "running" or "change detected" signal first (e.g., RUNS , Running, Ding!).
• Only after that signal fires is it safe to watch for the final RED or GREEN outcome.
• If the "change detected" signal does not appear within the timeout, re-trigger with touch on the saved file. The file system event may have been missed.

Violating this rule produces false RED readings — Claude thinks the test is still failing when the watcher simply has not rerun yet.

Failure Extraction Regexes

Use these regexes against watch-pane output (or capture-pane snapshot content) to extract the file, line, and test name from a RED state. Extract before editing so you know exactly where to make the change.

Jest/Vitest:  /at Object\.<anonymous> \((.+):(\d+):(\d+)\)/
              /FAIL (src\/[^\s]+\.test\.[tj]s)/
Cargo:        /panicked at '[^']+', (.+):(\d+):\d+/
Pytest:       /FAILED (tests\/[^\s]+)::([^\s]+)/
              /File "([^"]+)", line (\d+)/

Match group 1 is the file path, match group 2 is the line number. For Jest/Vitest's FAIL line, match group 1 is the test file path. For Pytest's FAILED line, match group 1 is the module path and group 2 is the test name.

Watcher Management

Never restart the watcher between iterations. One watcher per project.

• The watcher pane is created once at the start of the TDD session and left running until the user is done.
• mcp__plugin_terminal_mux__watch-pane observes non-destructively. The watcher keeps running between watch-pane calls.
• Do not send Ctrl+C to the watcher pane mid-session. This destroys watch mode and requires re-setup.
• If two Claude agents are working on the same project, they share the single existing watcher — neither creates a second one. Check for an existing watcher pane before creating a new one.
• At teardown (IDLE state), call mcp__plugin_terminal_mux__close-pane({ slot: 1 }) — it interrupts the watcher and then kills the pane, or releases it if it was one the user had left open.

Worked Example — Jest/Vitest End-to-End

This example shows the full loop from watcher start through a single RED-to-GREEN cycle.

Step 1: Start the Watcher (enter WAITING/IDLE)

mcp__plugin_terminal_mux__start-and-watch({
  slot: 1,
  command: "bun test --watch",
  pattern: "press a to rerun|Waiting for file changes|Waiting\\.\\.\\.",
  mode: "medium",
  timeout: 30
}) → WatchResult  // confirms watcher is initialized and idle

Step 2: Read Failure (enter RED)

Once the user reports a failing test or the watcher shows a fail signal, read the current state:

mcp__plugin_terminal_mux__capture-pane({ slot: 1 })
→ snapshot contains:
    FAIL src/utils/format.test.ts
    ● formatDate › returns ISO string
      Expected: "2026-01-15"
      Received: "01/15/2026"
      at Object.<anonymous> (src/utils/format.test.ts:12:5)

Apply the Jest/Vitest regex: file = src/utils/format.test.ts, line = 12.

Step 3: Extract and Edit

Read the failing file at the extracted location:

Read({ file_path: "/absolute/path/src/utils/format.ts" })

Locate the formatDate function and correct the date format. Use Edit to apply the fix. Save (the file write is the save).

Step 4: Enter WAITING — watch-pane for Change Detection

Do not read the watcher yet. Use watch-pane to block until the watcher notices the change:

mcp__plugin_terminal_mux__watch-pane({
  slot: 1,
  triggers: "pattern:RUNS |pattern:Running|error",
  mode: "line",   // notify on every new line — fast response
  timeout: 15
}) → WatchResult
// If event == "timeout": re-trigger with Bash touch, then watch-pane again
// If event == "pattern:...": change detected, watcher is re-running

If watch-pane times out (no change detected), re-trigger:

Bash({ command: "touch /absolute/path/src/utils/format.ts" })
// Then call watch-pane again

Step 5: watch-pane for Final Result (RED or GREEN)

mcp__plugin_terminal_mux__watch-pane({
  slot: 1,
  triggers: "pattern:Tests: \\d+ passed|pattern:Tests: \\d+ failed|pattern:FAIL |pattern:error\\[E|exit|error",
  mode: "bunch",   // notify every 10 lines — balanced for test output
  timeout: 60
}) → WatchResult
// WatchResult.output contains the full test run output
// WatchResult.event tells you: "pattern:Tests: N passed", "pattern:FAIL", etc.

Step 6: Report and Continue

Report to the user:

"All 14 tests passing. formatDate now returns ISO strings. Watcher is still running.
Would you like to continue with another failing test, or stop here?"

Leave the watcher pane running. If the user is done, send Ctrl+C to the watcher pane and kill it:

mcp__plugin_terminal_mux__close-pane({ slot: 1 })   // interrupts the watcher, then releases or kills the pane

Report summary: "Session complete. 1 test fixed across 1 file."

Process Liveness Check

If you need to verify the watcher is still running before calling watch-pane:

mcp__plugin_terminal_mux__pane-state({ slot: 1 })
→ { isAlive: true, foregroundCmd: "bun", waitingForInput: false }
// isAlive: false → watcher crashed; recreate it
// waitingForInput: true → watcher may be paused; check output