English | 한국어
Give your coding agent one local brief to start a pull request review from, so it explores the repository less.
QAMap reads the branch on your machine and prints one size-limited brief: the numbered
diff, the tests and callers that name each changed function, the history behind removed
lines, and easy-to-miss changes such as a removed throw or a field that is read
but never assigned. Claude Code or Codex reviews from it. QAMap itself makes no
LLM call and uploads nothing.
42 public pull requests from 12 open-source projects, chosen by a rule fixed before the runs:
| Claude Code CLI 2.1.292, one fixed model | Alone | With QAMap |
|---|---|---|
| Tokens, all 42 pull requests | 86.1M | 48.6M (-44%) |
| Pull requests where QAMap used fewer tokens | - | 35 of 42 |
| Blind LLM grade, 36 pull requests: preferred (4 ties) / valid findings | 18 / 19 | 14 / 16 |
| Runs that started a test runner or package install in a static review | 5 of 48 | 0 of 48 |
| Runs that found a regression the project fixed later (6 regressions, 2 runs each), found + partly found | 3 + 0 | 5 + 5 |
With QAMap is the build that became 0.5.2, measured before one later fix that changes a flow line in 4 of the 42 briefs and lowers one brief's confidence label. Its changes were designed after reading the reviews and briefs of these 42 pull requests, so no row is independent evidence. The last row's change signals target those six regressions, and this is the first of three measurements in which QAMap found more of them. The preference leans to Claude Code alone and is not statistically significant (p = 0.60). Codex was not measured. Tokens are the host's usage receipt, mostly cached input; the list-price estimate fell 34%. Single runs vary: one pair of byte-identical briefs used 0.9M and 2.0M tokens. Protocol, every run and limits.
With Node.js 20 or newer, install the CLI and set up the repository once for Claude Code and Codex. Commit the files it writes on the default branch, outside the reviewed diff:
npm install -g @ivorycanvas/qamap
qamap init --agentOn the branch to review, ask your agent: "Use QAMap to review this PR." The benchmark
used qamap init --agent --review-mode report, which lets agents run QAMap in this
repository without asking first. To read the brief yourself, run qamap qa brief.
init --agent writes an AGENTS.md section, skill files and qamap.config.json.
qa brief only reads the repository and saves its report under ~/QAMap-reports/.
The adoption guide covers other package managers and uncommitted changes.
The host needs access to the local repository and terminal.
/plugin marketplace add IvoryCanvas/QAMap
/plugin install qamap@ivorycanvasThe plugin calls the QAMap CLI, installed separately with
npm install -g @ivorycanvas/qamap. Then ask "Use QAMap to review this PR."
Installing the plugin is not consent to analysis or test execution. The
Claude Code setup guide also covers the directory listing.
Token boundary: QAMap's local analysis makes no model calls. An agent still uses model tokens to invoke QAMap and interpret the report; savings are not guaranteed.
| Brief section | What the reviewer gets |
|---|---|
| Changes | Numbered hunks, change signals, and lines a hunk leaves out. |
| References | Tests that name each changed function, its callers and their tests. |
| What to verify | Pattern checks to turn into action and expected result, or dismiss. |
| Unknowns, Not fully shown | What the brief could not settle, and the git diff to read next. |
For human reviewers, qamap qa prints a QA plan, qamap qa run executes a
selected repository command, and qamap e2e draft . --dry-run previews optional
automation. Nothing runs unless you ask. See the brief guide and
the command reference.
The human QA plan, qamap qa, for a public fixture that changes a subscription renewal flow:
Open the exact CLI output and first-run walkthrough.
QAMap reads the diff, history and tests, finds the changed declarations with their tests and callers, adds change signals and QA focus, and fits them into one brief. It follows direct change evidence before broad guesses, labels name matches as name matches, and lists what it could not trace or fit. When evidence is insufficient, it says so instead of inventing a contract or a passing result. Coverage and limits.
| Goal | Guide |
|---|---|
| Review one branch | First-run walkthrough |
| Adopt QAMap in a team | Adoption guide |
| Use QAMap from an agent | Agent integration |
| Set up the Claude Code plugin | Claude Code setup |
| Review every command | Command reference |
| Inspect benchmark evidence | Benchmarking |
QAMap is pre-1.0. A brief or QA plan is evidence for a reviewer, not a product
specification or proof: a person decides whether a behavior is intended or broken.
False positives, missed risks, and unusable drafts are especially useful. Start with CONTRIBUTING.md, and never publish private repository, customer, or credential data.