prreviewbuddy 0.88.1 → 0.91.1

package/CHANGELOG.md

@@ -1,5 +1,87 @@
 # Changelog
 
+## 0.91.1
+
+**Issues lead with what they mean.** Each issue from a new analysis now opens with its consequence:
+a short title saying what can happen, a sentence or two on when and how, and, when there is a
+specific one, the step that would settle it, marked **To verify** or **To decide**. The technical
+title, the reasoning, the code location and the history are under **Technical details** on the
+card, one click (or Enter) away, with nothing to wait for.
+
+Outstanding issues are grouped by the decision they ask of you: **Failures**, **Possible risks**,
+**Intent to confirm** and **Maintenance**, each ordered by severity. The group comes from the
+evidence the analysis read, not from its severity or how confident it is; both are still shown on
+every card. The pane opens with one sentence on what the change touches, a count of what is
+outstanding, and buttons to jump to each group. Folding a group is remembered for the next review.
+
+Resolved and dismissed issues stay in the closed list at the bottom. An issue the new commits fixed
+leads with what the code now does, when the update could see it in the code; one settled by a
+comment never claims a fix. Updates keep each issue's consequence true: they are shown it, and
+restate or withdraw it when the new commits change what it describes.
+
+Reviews made before this are shown as they were, and are not rewritten. Beside the new groups their
+issues sit under **Other issues**; on their own they need no heading.
+
+For coding agents, `list_findings` and `get_finding` (and `prreviewbuddy findings --json` and
+`finding --json`) carry each issue's `consequence` and, on an issue the code resolved,
+`resolved_behaviour`, so a list of findings can be triaged without fetching each one.
+
+## 0.90.0
+
+**Connect your coding agent from Settings.** A new **AI integrations** section in Settings shows,
+for Claude Code and Codex, whether each can use PR Review Buddy, and connects it in one step. You
+see the exact command before it runs, and it is the agent's own `mcp add`, so nothing edits the
+agent's files directly. **Verify connection** starts the registered server and lists its tools;
+**Disconnect** removes it. Verify checks that what answers is PR Review Buddy with all of its
+tools, started with the variables the agent's config gives it. A Codex registration switched off
+in its config shows as **Turned off**. A registration that points at another copy of PR Review Buddy, often an
+older install, shows as such, and **Replace** points it at this one, putting the old registration
+back if the agent refuses the new one. **Copy command** gives the same command to run yourself.
+
+What is registered is the full path of the `prreviewbuddy` that runs this workspace, checked to be
+this version first, because an agent started from a desktop app may not see your shell's `PATH`
+and could otherwise start an older copy. Connect manages each agent's default configuration; a
+Claude Code profile kept in its own directory gets a command to run for it instead. Copy command
+pins the config the row names, so a shell that selects another profile does not write there.
+
+After a finished review, once the guided tour has been answered, the review page mentions AI
+integrations once, and only when an agent is installed and none is connected yet. **Not now**
+puts it away for good. Help has a new **Coding agents** topic covering what an agent can and cannot
+do, connecting, verifying and troubleshooting. The manual commands in the docs and in
+`prreviewbuddy help mcp` now register the full path too.
+
+## 0.89.0
+
+**Coding agents can use PR Review Buddy directly.** `prreviewbuddy mcp serve` is a local MCP server
+for Claude Code, Codex and other MCP clients. An agent can start a review, follow it, read its
+findings, re-check it after fixing them, and ask Buddy AI about the change, without opening a
+browser. PR Review Buddy still does the analysis with the agent you configured, in an isolated
+checkout, and every review it starts is in the workspace as usual. Nothing is added to an agent's
+settings unless you add it: `claude mcp add --scope user pr-review-buddy -- prreviewbuddy mcp serve`,
+or `codex mcp add pr-review-buddy -- prreviewbuddy mcp serve`.
+
+**Reviews as JSON.** `--json` on `review`, and on the new `show`, `findings`, `finding`, `status` and
+`ask` commands, writes one JSON document to stdout and nothing else. Findings come in pages of
+compact records, filtered by kind, open or closed, and grade, and each has a `finding_id` that stays
+the same across every update of its review. A finding is identified by its review and its
+`finding_id` together. `review --json --no-wait` starts the work in the
+background and answers at once. `ask --json --no-wait` does the same for a question and returns
+its `ask_id`; `status <review> --ask <ask_id>` returns the answer once there is one.
+`review --review <review> --resume` carries on an analysis that stopped. The contract and its versioning are in `docs/agent-interface.md`.
+
+**Exit codes say what went wrong.** Every command now uses one table: 2 for a usage error, 3 for a
+review that does not exist, 4 for an operation declined, 5 for an analysis that failed, 6 when the
+workspace server cannot be reached, 7 for a timeout and 8 for a command refused inside an agent run.
+Failures that exited 1 still exit non-zero, with the same message. A review that finds problems
+exits 0.
+
+**An agent run cannot start a review.** Every agent PR Review Buddy starts runs marked, and in that
+environment the CLI and the MCP server read reviews but refuse to start reviews, updates or
+questions, on top of the flags that already keep PR Review Buddy's MCP server out of a nested session.
+
+Not covered by `--json` yet: lens reviews, several pull requests at once, `--reanalyse` and
+`--delete`.
+
 ## 0.88.1
 
 **One change to a review at a time, from anywhere on the Reviews list.** While a change made from a

package/README.md

@@ -91,6 +91,21 @@ prreviewbuddy help               every command
 Run `prreviewbuddy help` for the full command reference, and `prreviewbuddy review --help` for
 every option a review takes. How reviews, updates and lenses work is under Help in the workspace.
 
+## Use it from your coding agent
+
+PR Review Buddy can be a tool your coding agent calls: it starts the review, reads the findings,
+fixes what is real and checks again, while you can still open every review in the browser.
+
+Connect it from **Settings → AI integrations** in the workspace, or by hand:
+
+```
+claude mcp add --scope user pr-review-buddy -- "$(command -v prreviewbuddy)" mcp serve
+codex mcp add pr-review-buddy -- "$(command -v prreviewbuddy)" mcp serve
+```
+
+Then ask: "Review my changes with PR Review Buddy and fix the real problems." Scripts can use the
+same reviews as JSON: `prreviewbuddy review --json`, `prreviewbuddy findings <review> --json`.
+
 ## What you need
 
 - **Node 20 or newer**, and **git**.