@inceptionstack/pi-hard-no 1.0.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 InceptionStack
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,287 @@
+# pi-hard-no
+
+A [pi](https://github.com/badlogic/pi-mono) extension that automatically reviews code changes after each agent turn using a separate pi reviewer instance.
+
+## Install
+
+```bash
+pi install npm:@inceptionstack/pi-hard-no
+```
+
+Or manually:
+
+```bash
+cp index.ts ~/.pi/agent/extensions/pi-hard-no.ts
+```
+
+## How it works
+
+```
+Agent makes file changes (write, edit, bash)
+         │
+         ▼ agent_end fires
+         │
+         ▼ Extension detects file-modifying tool calls
+         │
+         ▼ Spawns a fresh pi instance (in-memory, isolated)
+         │
+         ▼ Sends per-file diffs + commit messages to reviewer
+         │  Reviewer reads each file itself via read(path) tool
+         │
+    ┌────┴────┐
+    │         │
+  LGTM    Issues found
+    │         │
+    │         ▼
+    │      Feeds back to main agent
+    │      Agent fixes → new review loop
+    │       (up to maxReviewLoops)
+    │
+    ▼ >1 file reviewed from git?
+    │
+    ├── No → done
+    │
+    └── Yes → Architect review
+              (cross-file consistency, architecture coherence)
+```
+
+The reviewer checks for:
+
+- Bugs, logic errors, off-by-one errors, race conditions
+- Security issues (injection, secret leaks, auth bypasses)
+- Missing error handling
+- DRY violations (Don't Repeat Yourself)
+- Single Responsibility Principle
+- Readability and maintainability
+
+## Configuration
+
+Config files are loaded from two locations. **Local takes precedence over global:**
+
+1. `cwd/.hardno/` — project-specific config
+2. `~/.pi/.hardno/` — global defaults
+
+All config files are optional. If missing, sensible defaults are used.
+
+Use `/scaffold-review-files` to generate config templates.
+
+### `.hardno/settings.json`
+
+```json
+{
+  "maxReviewLoops": 100,
+  "model": "amazon-bedrock/us.anthropic.claude-opus-4-6-v1",
+  "thinkingLevel": "off",
+  "architectEnabled": true,
+  "reviewTimeoutMs": 120000,
+  "toggleShortcut": "alt+r",
+  "cancelShortcut": "",
+  "judgeEnabled": false,
+  "judgeModel": "amazon-bedrock/us.anthropic.claude-haiku-4-5-20251001-v1:0",
+  "judgeTimeoutMs": 10000
+}
+```
+
+| Setting            | Type        | Default                                                        | Description                                                                                |
+| ------------------ | ----------- | -------------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
+| `maxReviewLoops`   | integer > 0 | `100`                                                          | Max review→fix→review cycles before stopping                                               |
+| `model`            | string      | `"amazon-bedrock/us.anthropic.claude-opus-4-6-v1"`             | Reviewer model (`"provider/model-id"`)                                                     |
+| `thinkingLevel`    | string      | `"off"`                                                        | `off\|minimal\|low\|medium\|high\|xhigh`                                                   |
+| `architectEnabled` | boolean     | `true`                                                         | Enable architect review (triggers when >1 file reviewed from git)                          |
+| `reviewTimeoutMs`  | integer > 0 | `120000`                                                       | Max wall-clock per review in ms                                                            |
+| `toggleShortcut`   | string      | `"alt+r"`                                                      | Key id for toggling review on/off                                                          |
+| `judgeEnabled`     | boolean     | `false`                                                        | Opt-in LLM gate that suppresses redundant reviews on read-only turns (see [Judge](#judge)) |
+| `judgeModel`       | string      | `"amazon-bedrock/us.anthropic.claude-haiku-4-5-20251001-v1:0"` | Model used by the judge (`"provider/model-id"`)                                            |
+| `judgeTimeoutMs`   | integer > 0 | `10000`                                                        | Max wall-clock per judge classification call in ms                                         |
+| `cancelShortcut`   | string      | `""` (none)                                                    | Key id for cancelling review (opt-in, see below)                                           |
+
+> **Note:** `roundupEnabled` is accepted as a legacy alias for `architectEnabled`.
+
+### `.hardno/review-rules.md`
+
+Custom review rules appended to the reviewer prompt. Only include review criteria — the surrounding prompt (tools, budget, workflow, response format) is handled automatically.
+
+```markdown
+## Architecture
+
+- All API endpoints must validate input with zod schemas
+- Database queries must use parameterized statements
+
+## Security
+
+- No console.log in production code (use logger)
+- No secrets in code — use environment variables
+```
+
+Use `/add-review-rule <text>` to quickly prepend rules, or `/hardno-rules` to open the file in pi's editor.
+
+### `.hardno/auto-review.md`
+
+Override the "what to review / what not to report" section of the review prompt. The surrounding prompt (tools, budget, workflow, response format) is always included automatically.
+
+### `.hardno/architect.md`
+
+Custom rules for the architect review (cross-file consistency check):
+
+```markdown
+## Architecture
+
+- Verify module dependency graph has no cycles
+- Check error handling is consistent across all modules
+- Flag any TODO/FIXME comments added during fix loops
+```
+
+> **Note:** `.hardno/roundup.md` is accepted as a legacy fallback.
+
+### `.hardno/ignore`
+
+Gitignore-style patterns to exclude files from review:
+
+```
+# Skip generated files
+*.generated.ts
+dist/
+node_modules/
+
+# Skip specific paths
+src/vendor/**
+```
+
+## UX
+
+### Status bar (bottom of pi)
+
+- `lgtm on (Alt+R toggle)` — idle, no pending files
+- `lgtm on 🔒 push blocked · will review 3 files (Alt+R toggle)` — edits accumulating, push blocked
+- `lgtm reviewing… 🔒 push blocked (/cancel-review)` — reviewer running
+- `lgtm on issues found 🔒 push blocked (Alt+R toggle)` — review found issues
+- `lgtm skipped — no files to review` — nothing to review after fix turn
+- `lgtm off (Alt+R toggle)` — disabled, push guard off
+
+### Review progress widget
+
+During reviews, an animated widget appears below the editor showing:
+
+- ASCII art senior dev with reading glasses
+- File list with active file highlighted and per-file tool usage counts
+- Elapsed time, model name, loop count
+
+### Commands
+
+| Command                   | Description                                                           |
+| ------------------------- | --------------------------------------------------------------------- |
+| `/review`                 | Toggle review on/off                                                  |
+| `/review N`               | Review the last N commits                                             |
+| `/review-all`             | Review all changes (pending diff → last commit → all files in cwd)    |
+| `/cancel-review`          | Cancel an in-progress review (works during architect review)          |
+| `/review-judge-toggle`    | Toggle the duplicate-review suppressor (judge) for this session       |
+| `/review-clean-logs`      | Wipe `~/.pi/.hardno/review.log` + `reviews/*.json` (config untouched) |
+| `/scaffold-review-files`  | Create `.hardno/` config templates in a git repo                      |
+| `/hardno-rules`           | Edit `.hardno/review-rules.md` in pi's built-in editor                |
+| `/add-review-rule <text>` | Prepend a custom rule to `.hardno/review-rules.md`                    |
+
+### Keyboard shortcuts
+
+| Key                | Default  | Configurable     | Action                                              |
+| ------------------ | -------- | ---------------- | --------------------------------------------------- |
+| Toggle shortcut    | `alt+r`  | `toggleShortcut` | Toggle review on/off                                |
+| Cancel shortcut    | _(none)_ | `cancelShortcut` | Cancel in-progress review                           |
+| `ctrl+alt+r`       | built-in | no               | Cancel review (fallback, terminals that support it) |
+| `ctrl+alt+shift+r` | built-in | no               | Full reset: cancel, reset loops, clear all state    |
+
+> **Note:** `/cancel-review` is the recommended cancel method. It works in all terminals. Keyboard shortcuts for cancel are opt-in via `cancelShortcut` in settings because many terminals (especially iTerm2 on macOS) don't reliably send modifier key combos.
+
+## Review loop behavior
+
+1. Agent makes changes → review triggers
+2. If issues found → agent fixes them → review triggers again
+3. If LGTM → loop counter resets
+4. If loop count reaches `maxReviewLoops` → stops with a warning
+5. Toggling off/on with `/review` resets the counter
+
+### Architect review
+
+After the review loop reaches LGTM, an **architect review** triggers automatically when more than one file was reviewed from git across the session. No heuristics or judge gating — it always runs for multi-file changes.
+
+The architect review:
+
+- Checks architecture coherence across all changes
+- Verifies cross-file consistency (naming, patterns, types)
+- Looks for accumulated tech debt from fix loops
+- Validates documentation is still accurate
+- Uses tools (`read`, `bash`, `grep`, `find`, `ls`) to explore the full codebase
+
+Disable with `"architectEnabled": false` in settings.
+
+## Judge
+
+The **judge** is an opt-in duplicate-review suppressor. When enabled, it runs a cheap classifier LLM (default: Claude Haiku 4.5) on each bash tool call the agent made this turn. If every bash call classifies as `inspection_vcs_noop` (reads state only — `git status`, `git log`, `echo`, inspection compounds, etc.) **and** no `write`/`edit` tool call ran, the full review is skipped with reason `judge_read_only`.
+
+**Why it exists:** the deterministic classifier in `changes.ts` uses a static allowlist. Commands using shell builtins outside the allowlist (e.g. `echo` in a compound) get flagged as "potentially modifying" and trigger an unnecessary review of already-reviewed content. The judge catches those false positives.
+
+**Fail-safe by design:**
+
+- Off by default.
+- Fail-open: any judge error (timeout, transport, parse) → review runs as normal.
+- `unsure` classification → review runs (same as "modifying").
+- Any `write`/`edit` tool call skips the judge entirely and goes straight to review.
+- A kill switch: set `"judgeEnabled": false` to disable instantly.
+
+**Enable in `.hardno/settings.json`:**
+
+```json
+{
+  "judgeEnabled": true
+}
+```
+
+See `eval/RESULTS.md` for the evaluation that picked Haiku 4.5.
+
+## What triggers a review
+
+Only fires when file-modifying tools were used during the agent turn:
+
+- `write` — new files
+- `edit` — file edits
+- `bash` — commands matching file operations (`cp`, `mv`, `rm`, `sed -i`, `cat >`, `tee`, `mkdir`, `echo >`)
+
+Pure read/search turns are skipped. Non-file-modifying bash commands (`git commit`, `curl`, `aws`, etc.) are also skipped.
+
+### Untracked (new) files
+
+Files created via `write` that haven't been `git add`ed are detected via `git ls-files --others --exclude-standard` and included in the review context, labeled as `(new file)`.
+
+## Cancellation
+
+You can cancel a review at any time:
+
+- **`/cancel-review`** — works in all terminals, recommended method
+- **Configured shortcut** — set `cancelShortcut` in settings if you want a hotkey
+- **`ctrl+alt+r`** — fallback, works in terminals that support the key combo
+
+Cancellation stops the current review immediately, including architect reviews. The agent continues normally.
+
+## Push guard
+
+The extension automatically blocks `git push` when:
+
+- **A review is in progress** — wait for the review to complete
+- **The last review found issues** — fix the issues and get LGTM first
+- **Files have been modified but not yet reviewed** — wait for the review to start and complete
+
+The status bar shows `🔒 push blocked` whenever push would be blocked.
+
+The block applies to any `bash` tool call matching `git push` (including `git -C <dir> push`, `git push origin main`, etc.). The agent sees a clear "Push blocked" message explaining why.
+
+The block clears automatically when:
+
+- The next review returns **LGTM**
+- The review **skips** with "no files to review" (issues resolved by deletion/revert)
+- You do a **full reset** (`Ctrl+Alt+Shift+R`)
+- You **disable** review (`Alt+R` toggle) — push guard is off when review is off
+
+No git hooks are needed — this is enforced at the extension level via pi's `tool_call` event interception.
+
+## License
+
+MIT

package/architect.ts

@@ -0,0 +1,128 @@
+/**
+ * architect.ts — Final "zoom out" architecture review after mini-review loops complete
+ *
+ * Triggered automatically when more than 1 file was actively reviewed by the
+ * review step. No heuristics or judge gating — if multiple files were
+ * touched, an architecture-level review always runs.
+ *
+ * Looks at the big picture: architecture coherence, cross-file consistency,
+ * accumulated tech debt, and documentation accuracy.
+ */
+
+import type { ReviewResult, ReviewRunner } from "./reviewer";
+import { readConfigFile } from "./settings";
+import { log } from "./logger";
+
+const DEFAULT_ARCHITECT_PROMPT = `You are a senior architect doing a final "zoom out" review. A series of code changes were just made and passed individual mini-reviews. Now step back and look at the big picture.
+
+You have tools available (read, bash, grep, find, ls) to explore the full codebase.
+
+## Architecture coherence
+- Do all the pieces fit together? Any orphaned code that nothing calls?
+- Is the module dependency graph clean? Any unexpected coupling?
+- Does the layering make sense (e.g. no circular dependencies)?
+
+## Cross-file consistency
+- Are naming conventions consistent across all changed files?
+- Are similar patterns handled the same way everywhere?
+- Are types/interfaces consistent and not duplicated?
+
+## Integration completeness
+- Is new code properly wired up? Exports used? Imports correct?
+- Are there any missing integration points?
+- Do tests cover the integration paths, not just unit-level?
+
+## Accumulated tech debt
+- Did the back-and-forth fix loops create any franken-code?
+- Any TODO/FIXME/HACK comments that were added?
+- Dead code or unused imports that accumulated?
+- Any functions that grew too large or do too many things?
+
+## Documentation
+- Is the README still accurate after all changes?
+- Are architecture docs (if any) still correct?
+- Do public APIs have adequate comments/types?
+- Are new files/modules properly documented?
+
+## Response format
+If everything looks good at the big-picture level, say "LGTM — architecture looks solid."
+If there are issues, list them as bullet points with severity (high/medium/low).
+Focus on systemic issues that individual mini-reviews would miss.
+Do NOT repeat issues that were already found and fixed in mini-reviews.`;
+
+/**
+ * Load architect review rules from .hardno/architect.md.
+ * Falls back to .hardno/roundup.md for backwards compatibility.
+ */
+export async function loadArchitectRules(cwd: string): Promise<string | null> {
+  // Try new name first, fall back to old name
+  const content = await readConfigFile(cwd, "architect.md");
+  if (content?.trim()) return content.trim();
+  const legacy = await readConfigFile(cwd, "roundup.md");
+  return legacy?.trim() || null;
+}
+
+export function buildArchitectPrompt(customRules: string | null): string {
+  let prompt = DEFAULT_ARCHITECT_PROMPT;
+  if (customRules) {
+    prompt += `\n\n## Additional project-specific architect review rules\n\n${customRules}`;
+  }
+  return prompt;
+}
+
+// ── Trigger logic ──────────────────────────────────
+
+/**
+ * Determine whether the architect review should run.
+ * Triggers when more than 1 file was actively reviewed AND the review
+ * content came from one or more git repositories.
+ */
+export function shouldRunArchitectReview(reviewedFiles: string[], isGitBased: boolean): boolean {
+  if (!isGitBased) {
+    log(`architect: skip — reviewed files are not from a git repo`);
+    return false;
+  }
+  const dominated = reviewedFiles.length > 1;
+  if (dominated) {
+    log(`architect: will run — ${reviewedFiles.length} files reviewed from git repo(s)`);
+  } else {
+    log(`architect: skip — only ${reviewedFiles.length} file(s) reviewed`);
+  }
+  return dominated;
+}
+
+// ── Full architect review ──────────────────────────
+
+export interface ArchitectReviewOptions {
+  signal: AbortSignal;
+  cwd: string;
+  model?: string;
+  customRules: string | null;
+  sessionChangeSummary: string;
+  /** Unique id for the architect review cycle. Separate from the senior review id. */
+  reviewId?: string;
+  /** Max wall-clock for the architect's LLM call, in ms. Passed through to the reviewer. */
+  timeoutMs?: number;
+  onActivity?: (description: string) => void;
+  onToolCall?: (toolName: string, targetPath: string | null) => void;
+}
+
+/**
+ * Run the final architect review.
+ */
+export async function runArchitectReview(
+  runner: ReviewRunner,
+  opts: ArchitectReviewOptions,
+): Promise<ReviewResult> {
+  const prompt = `${buildArchitectPrompt(opts.customRules)}\n\n---\n\nHere is a summary of all changes made in this session:\n\n${opts.sessionChangeSummary}\n\nPlease explore the codebase with your tools to verify everything fits together.`;
+
+  return await runner(prompt, {
+    signal: opts.signal,
+    cwd: opts.cwd,
+    model: opts.model,
+    reviewId: opts.reviewId,
+    timeoutMs: opts.timeoutMs,
+    onActivity: opts.onActivity,
+    onToolCall: opts.onToolCall,
+  });
+}