recovery-probe 0.4.0-preview.3 → 0.4.0-preview.5

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -4,6 +4,18 @@ All notable changes to Recovery Probe are documented here. The project follows [
4
4
 
5
5
  Version 0.2.0 is the first npm release. Earlier 0.1.x revisions were repository-only prototypes.
6
6
 
7
+ ## 0.4.0-preview.5 (2026-10-16)
8
+
9
+ - Repair on Windows: the isolated checkout now holds the committed bytes instead of the CRLF line endings Git for Windows writes by default. A multi-line proposal with LF line breaks previously failed with "find text must match exactly once". `eol` rules in `.gitattributes` still apply.
10
+ - Repair on Windows: a committed symlink selected as a source is rejected again. Without symlink rights, Git for Windows checks it out as a plain file, which passed the old check; sources are now checked against the index.
11
+ - CI runs the repair pipeline on Windows.
12
+
13
+ ## 0.4.0-preview.4 (2026-10-15)
14
+
15
+ - `createFaultFetch` passes relative URLs, such as `fetch('/api/x')`, to the supplied fetch untouched. Previously they threw `Invalid URL`.
16
+ - The HTML reports are built from template files in `src/report-templates`. Text slots are escaped, a missing slot throws, and the rendered text is unchanged.
17
+ - The README leads with one idea, whether apps show the right data, and with the demo. It drops from 2,380 to about 1,070 words. The fault kinds and library API move to `docs/api.md`, and the JSON-configured runner to `docs/runner.md`.
18
+
7
19
  ## 0.4.0-preview.3 (2026-10-15)
8
20
 
9
21
  - Restore an existing instance-level `webContents.send` (for example an app's tracing wrapper) when an observation ends or the hook is disposed, instead of deleting it. Previously that wrapper was lost and later pushes bypassed it. A wrapper installed after the observation began is left in place.
package/CONTRIBUTING.md CHANGED
@@ -20,9 +20,14 @@ npm test
20
20
  npm run test:package
21
21
  npx playwright install chromium
22
22
  npm run test:browser
23
+ npm run test:repair # safety, authenticated portal, before/after pipeline
24
+ # After installing the optional Electron test runtime:
25
+ npm run test:repair:electron
23
26
  ```
24
27
 
25
- Code and JSON are formatted with Prettier (`npm run format`) and linted with ESLint (`npm run lint`); CI fails on either. Functions passed to `page.evaluate` run in the browser: keep them self-contained, and list new in-page modules in `eslint.config.mjs` so they get browser globals. TypeScript checks the JavaScript in `src`, `bin` and `examples` (`npm run typecheck`, `checkJs` without emitting); add JSDoc types where inference is not enough. Strict mode is off for now and can be tightened file by file. Markdown, HTML fixtures and workflow files are left as written. `.gitattributes` keeps LF line endings on every platform so local checks match CI. The repository-wide reformat is listed in `.git-blame-ignore-revs`; run `git config blame.ignoreRevsFile .git-blame-ignore-revs` to skip it in local blame.
28
+ Repair integration tests use a **controlled mock model response** but real apps, fault injection, source changes and reruns. They validate orchestration, not live AI diagnosis quality.
29
+
30
+ Code and JSON are formatted with Prettier (`npm run format`) and linted with ESLint (`npm run lint`); CI fails on either. Functions passed to `page.evaluate` run in the browser: keep them self-contained, and list new in-page modules in `eslint.config.mjs` so they get browser globals. TypeScript checks the JavaScript in `src`, `bin` and `examples` (`npm run typecheck`, `checkJs` without emitting); add JSDoc types where inference is not enough. Strict mode is off for now and can be tightened file by file. Markdown, HTML fixtures, report templates and workflow files are left as written. Report markup lives in `src/report-templates`: `{{slot}}` takes text and escapes it, `{{{slot}}}` takes markup built with `escapeHtml`. `.gitattributes` keeps LF line endings on every platform so local checks match CI. The repository-wide reformat is listed in `.git-blame-ignore-revs`; run `git config blame.ignoreRevsFile .git-blame-ignore-revs` to skip it in local blame.
26
31
 
27
32
  ## Install an unpublished source preview
28
33
 
package/README.md CHANGED
@@ -1,87 +1,55 @@
1
1
  # Recovery Probe
2
2
 
3
- **Test whether web and desktop apps show the right data: recovery after failed reads, cached content, and screens that lag behind their data.**
3
+ **Test whether web and desktop apps show the right data.**
4
4
 
5
- Recovery Probe works in two ways. It injects bounded faults into data reads, compares healthy and faulted screens, and records whether content survives or returns. It also observes reads and the screen without any fault, timing what the user sees after an action against the data the app has already received. It helps developers and coding agents reproduce problems such as disappearing lists, stuck loading states, missing retries, content served from a cache and row states that stay stale after fresh data arrives. Installing the package does not add recovery behavior to your application.
5
+ Apps can show the wrong data without any error: a list that disappears after one failed read, a cached screen that never picks up a change, or a row whose state lags behind data the app already received. Recovery Probe reproduces these cases on demand. It injects bounded faults into data reads and compares the screen with a healthy run, or it watches reads and the screen without any fault. Each run writes a local report that separates what it measured from what it suspects.
6
+
7
+ Installing the package does not add recovery behavior to your application.
6
8
 
7
9
  [![CI](https://github.com/BojanKovachki/recovery-probe/actions/workflows/validate.yml/badge.svg)](https://github.com/BojanKovachki/recovery-probe/actions/workflows/validate.yml)
8
10
  [![npm](https://img.shields.io/npm/v/recovery-probe.svg)](https://www.npmjs.com/package/recovery-probe)
9
11
  [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
10
12
 
11
- ## Supported applications
12
-
13
- | Application path | What can be tested | Setup |
14
- | --- | --- | --- |
15
- | Web apps in Chromium-based browsers | Renderer GET fetch/XHR failures, UI recovery and structural content changes | Launch a test browser or attach to an authorized development browser |
16
- | Electron desktop apps using renderer HTTP | The same request and UI checks | Attach through a loopback debugging endpoint |
17
- | Electron desktop apps using IPC reads | Renderer behavior after selected IPC rejections or null results | Add a development-only handler-registration hook and enable local inspection |
18
- | JavaScript code using fetch | Recovery logic around an explicitly supplied fetch wrapper | Import `recovery-probe/fetch`; Playwright is not required |
19
- | Web and Electron apps, without faults | Selected reads, cached content and tracked row states after user actions, including delayed UI updates | Attach as above; IPC observation uses the same development hook |
13
+ ## See it work
20
14
 
21
- Desktop support currently means Electron, not arbitrary native applications. IPC faults do not exercise native/Rust networking, HTTP status handling or token refresh. An unsupported data path is inconclusive, never a successful recovery check.
22
-
23
- ## Install
15
+ The package includes a deliberately broken app and a corrected one. Both pass the normal loading check; only the corrected one recovers from all three faults.
24
16
 
25
- **0.3.0** is the stable release target on npm's `latest` channel. The commands below pin the version for reproducible reports. Prereleases use `next` and are opt-in.
17
+ ![Recovery Probe catches a broken retry flow and verifies the corrected flow](docs/demo.svg)
26
18
 
27
19
  ```bash
28
20
  mkdir recovery-probe-tools && cd recovery-probe-tools
29
- npm view recovery-probe@0.3.0 version
30
21
  npm install --save-dev --save-exact recovery-probe@0.3.0 playwright@1.62.1
31
- npx recovery-probe --version
22
+ npx playwright install chromium
23
+ npx recovery-probe --demo
32
24
  ```
33
25
 
34
- If the version check reports that the version is unavailable, npm publication is still pending; use the [source installation](CONTRIBUTING.md#install-an-unpublished-source-preview). A GitHub merge alone does not publish the package. Use Node.js 22 or newer. A separate tools folder keeps the tester out of your application's production dependencies. Browser/desktop commands require Playwright; attached targets use their existing browser runtime. The guided web flow downloads Chromium when needed. See [CHANGELOG.md](CHANGELOG.md) for version history.
35
-
36
- For the 0.4 observation preview, install `recovery-probe@0.4.0-preview.3` (the `next` channel) with the same Playwright version. Preview.3 restores an app's own `webContents.send` wrapper after an observation and updates the project description; preview.2 added formatting, lint and implementation type checks; the action/row observation features were introduced in preview.1.
26
+ `--demo` exits successfully only when it observes that contrast: the happy path passes, the broken recovery is caught, and the corrected flow passes. Add `--json` to print the full report or `--out report.json` to save it.
37
27
 
38
- ## Choose a workflow
28
+ ## Test your own app
39
29
 
40
- **New in 0.3.0:** guided setup, HTTP/IPC discovery, refresh experiments and [fault-free observation](docs/observe.md) extend the original recovery helpers. Observation records selected reads and content while you navigate without reloading or clearing caches. Lightweight sampling, phase timings and separate cleanup evidence make capture cost and scope explicit. Use `"detail": "region"` for full DOM diagnostics.
41
-
42
- Content comparisons preserve the first observed result and summarize later samples: a temporary difference can subsequently match the baseline, or diverge again. These are observations, not ownership, freshness or stability verdicts. See the guide for updating the development IPC hook before capturing cached state; see the changelog for upgrading from 0.2.0.
43
-
44
- | Goal | Entry point | Guide |
30
+ | Goal | Start with | Guide |
45
31
  | --- | --- | --- |
46
- | Set up a first request-recovery check interactively | `recovery-probe start web URL` or `start desktop` | [Guided start](docs/start.md) |
47
- | Discover and test renderer reads in an existing browser or Electron window | `recovery-probe desktop --cdp URL` | [Attachment and HTTP discovery](docs/desktop.md) |
32
+ | Set up a first recovery check interactively | `recovery-probe start web URL` or `start desktop` | [Guided start](docs/start.md) |
33
+ | Discover and test the reads behind a screen in a browser or Electron window | `recovery-probe desktop --cdp URL` | [Attachment and HTTP discovery](docs/desktop.md) |
48
34
  | Compare an Electron screen under IPC faults | `recovery-probe ipc --discover --config FILE` | [IPC setup and discovery](docs/ipc.md) |
49
35
  | Test a refresh after content has already loaded | `recovery-probe refresh --config FILE --json` | [Refresh experiments](docs/refresh.md) |
50
- | Observe cached content without a reload | `recovery-probe observe --config FILE` | [Observation](docs/observe.md) |
51
- | Time row states after a user action, with optional scripted steps | `recovery-probe observe --config FILE [--steps FILE]` | [Actions, rows and delayed UI updates](docs/observe.md#actions-tracked-rows-and-delayed-ui-updates) |
52
- | Run a repeatable synthetic recovery check in CI | `npm run test:ci-example` (source checkout) | [CI example](docs/ci.md) |
53
- | Add an explicit recovery assertion to a test | `withFault` or `createFaultFetch` | Examples below |
36
+ | Watch cached content and row states after user actions, without faults | `recovery-probe observe --config FILE [--steps FILE]` | [Observation](docs/observe.md) |
37
+ | Run a repeatable recovery check in CI | `npm run test:ci-example` (source checkout) | [CI example](docs/ci.md) |
38
+ | Assert recovery inside your own tests | `withFault` or `createFaultFetch` | [API](docs/api.md) |
54
39
  | Propose and verify a source change | Opt-in repair workflow | [Repair setup and limits](docs/repair.md) |
55
40
 
56
- `desktop` is also the current command name for attaching to an existing Chromium-based web browser. Use only an authorized test session, with debugging endpoints on loopback. Keep the target window visible and uncovered.
57
-
58
- Guided setup asks you to choose a request and loaded content that proves success; later runs reuse the saved scenario. Structural discovery compares healthy and faulted screens without requiring a hand-written content selector, but still needs a selected read target and a valid baseline. Neither workflow infers your product's intended recovery policy.
59
-
60
- HTTP and IPC structural discovery calibrate three healthy reloads for the full `baselineTimeoutMs` window before injecting faults (45 seconds with the default 15-second window). All three stable end states must agree. This avoids treating a briefly stable loading shell as the healthy screen, accepts legitimate empty regions, and keeps an unsuccessful calibration inconclusive. Later controls may finish early once they match the calibrated baseline. Content arriving after the window remains outside the evidence.
61
-
62
- **Session-isolation boundary:** recovery/discovery/refresh workflows reload the renderer to establish controls. That clears in-memory caches and can hide state carried between sessions. Those modes cannot prove account isolation; content preservation within one session is not an ownership check. The new observation workflow collects evidence without resetting that state, but does not infer identities or data ownership.
63
-
64
- ## Refresh experiments
65
-
66
- Refresh mode tests an **already populated screen** using an explicitly configured synthetic window event, such as `online`. It first checks the event without a fault, then injects the next selected read failure and observes the screen without reloading. An optional second event tests whether another refresh repairs the screen.
67
-
68
- ```bash
69
- npx recovery-probe refresh --config ./refresh.json --out ./refresh-run1 --json
70
- ```
71
-
72
- This dispatches an event; it does not disconnect the network, change `navigator.onLine`, simulate sleep/resume or expire authentication. Spontaneous recovery and recovery after the second tool event are reported separately. See the [configuration, results and cleanup contract](docs/refresh.md).
41
+ `desktop` also attaches to an existing Chromium-based web browser.
73
42
 
74
- ## Reports and coding agents
43
+ **Where it works:**
75
44
 
76
- Reports separate measured facts—fault counts, successful reads, timings and UI changes—from suspected defects. A `CONTENT_LOSS` finding needs review. An unconsumed fault, unstable control or unsupported path is inconclusive. All non-recovery observations are limited to the configured time window.
45
+ - **Web apps** in Chromium-based browsers: launch a test browser or attach to an authorized development one.
46
+ - **Electron apps that read over HTTP** in the renderer: attach through a loopback debugging endpoint.
47
+ - **Electron apps that read over IPC**: add a development-only hook to the main process.
48
+ - **JavaScript code that takes a fetch function**: use `recovery-probe/fetch`, without Playwright.
77
49
 
78
- Agents can run saved scenarios, inspect JSON evidence, propose changes and rerun the same experiment. Recovery Probe supplies repeatable experiments and evidence; it is not a general bug detector or production self-healing system. The optional AI repair workflow requires selected local source files, launch/test commands and your own model credentials. It produces a reviewable candidate, not a guaranteed fix.
50
+ Desktop support means Electron, not arbitrary native applications. IPC faults do not exercise native networking, HTTP status handling or token refresh. An unsupported data path is inconclusive, never a successful recovery check.
79
51
 
80
- Reports remain local. Scenarios, endpoints, channels, selected text and authentication files may contain private information; do not publish them blindly. Tests and discovery do not call an AI service.
81
-
82
- ![Recovery Probe catches a broken retry flow and verifies the corrected flow](docs/demo.svg)
83
-
84
- ## Use in a Playwright test
52
+ ## Add a recovery check to a Playwright test
85
53
 
86
54
  Use `withFault` inside an existing Playwright test. The callback contains the real user action and the application-specific recovery assertion.
87
55
 
@@ -105,154 +73,37 @@ test('profile recovers after a transient 503', async ({ page }) => {
105
73
  });
106
74
  ```
107
75
 
108
- If the endpoint does not match, Recovery Probe throws `FAULT_NOT_TRIGGERED` instead of allowing the test to pass silently. If the fault is applied but the recovery assertion fails, it throws `RECOVERY_ASSERTION_FAILED` and retains the original assertion error as `cause`.
109
-
110
- ## Run the demonstration
111
-
112
- The repository includes deliberately broken and corrected applications. Both pass their normal loading check; only the corrected application recovers from all three faults.
113
-
114
- ```bash
115
- npm install
116
- npx playwright install chromium
117
- npx recovery-probe --demo
118
- ```
119
-
120
- The command prints a concise comparison:
121
-
122
- ```text
123
- broken synthetic fixture
124
- PASS Normal connection (OK)
125
- CAUGHT HTTP 503 once (RECOVERY_ASSERTION_FAILED, injected 1)
126
- CAUGHT Interrupted request once (RECOVERY_ASSERTION_FAILED, injected 1)
127
- CAUGHT Malformed JSON once (RECOVERY_ASSERTION_FAILED, injected 1)
128
-
129
- fixed synthetic fixture
130
- PASS Normal connection (OK)
131
- PASS HTTP 503 once (OK, injected 1)
132
- PASS Interrupted request once (OK, injected 1)
133
- PASS Malformed JSON once (OK, injected 1)
134
- ```
76
+ If no request matches, `withFault` throws `FAULT_NOT_TRIGGERED` instead of letting the test pass silently. Code that takes a fetch function can use `createFaultFetch` instead. See the [API reference](docs/api.md) and the [JSON-configured runner](docs/runner.md).
135
77
 
136
- `--demo` exits successfully only when it observes the expected contrast: the happy path passes, the broken recovery is caught, and the corrected flow passes. Add `--json` to print the full report or `--out report.json` to save it.
137
-
138
- ## Library HTTP faults
139
-
140
- | Kind | Injected behavior |
141
- | --- | --- |
142
- | `http-error` | Fulfils the matched request with an HTTP error status; defaults to `503`. |
143
- | `connection-failure` | Aborts the matched Playwright request with `connectionfailed`, or throws from the fetch adapter. |
144
- | `invalid-json` | Returns status `200` with a deliberately malformed JSON body. |
145
-
146
- Faults are bounded by `count` (default `1`, maximum `20`). The count is reserved synchronously, so concurrent matching requests cannot over-inject.
147
-
148
- ## Fetch-only tests
149
-
150
- The lightweight `recovery-probe/fetch` entry point has no runtime dependencies and does not require Playwright. It matches one exact absolute URL, including its query string, and never patches global `fetch`.
151
-
152
- ```js
153
- import { createFaultFetch } from 'recovery-probe/fetch';
154
-
155
- const probe = createFaultFetch(fetch, {
156
- url: 'http://127.0.0.1:3000/api/profile',
157
- kind: 'connection-failure',
158
- });
159
-
160
- await exerciseApplication(probe.fetch);
161
- probe.assertApplied();
162
- ```
163
-
164
- `exerciseApplication` represents your own test adapter or application controller. Calling `assertApplied()` is essential: it proves that every requested fault was consumed.
165
-
166
- ## Original JSON-configured runner
167
-
168
- The CLI supports pages that automatically load GET data, expose a visible Retry button after failure, and show a ready element after recovery.
169
-
170
- ```json
171
- {
172
- "name": "Profile recovery",
173
- "url": "http://127.0.0.1:3000/profile",
174
- "requestPattern": "**/api/profile",
175
- "readySelector": "[data-testid=profile-ready]",
176
- "retrySelector": "[data-testid=profile-retry]",
177
- "timeoutMs": 3000
178
- }
179
- ```
180
-
181
- ```bash
182
- npx recovery-probe --config recovery-probe.json --out artifacts/profile-recovery.json
183
- ```
184
-
185
- The runner first checks the normal baseline in a fresh browser context. If that baseline fails, fault comparisons are skipped and reported as inconclusive. Each fault then receives its own fresh context with service workers blocked so Playwright can intercept the request.
186
-
187
- Exit codes:
188
-
189
- - `0`: every configured check passed, or the controlled demo observed its expected broken-versus-fixed result.
190
- - `1`: a recovery check failed, was inconclusive, or was skipped.
191
- - `2`: invalid arguments, configuration, or environment setup.
192
-
193
- ## Original runner result semantics
194
-
195
- | Code | Meaning |
196
- | --- | --- |
197
- | `OK` | The configured assertion completed and the expected fault count was observed. |
198
- | `RECOVERY_ASSERTION_FAILED` | The fault occurred, but the recovery assertion did not pass. |
199
- | `FAULT_NOT_TRIGGERED` | The requested fault count was not observed. Check the route, method, cache, and service worker. |
200
- | `BASELINE_FAILED` | The ordinary flow failed, so the runner could not make a useful fault comparison. |
201
- | `INJECTION_ERROR` | Playwright could not apply or clean up the route reliably. |
202
-
203
- A passing result proves the configured recovery behavior for that test. It is not a certification of overall application resilience.
204
-
205
- ## API
206
-
207
- ### `withFault(page, options, exercise)`
78
+ ## Install
208
79
 
209
- Installs one temporary Playwright route, runs `exercise`, verifies the requested fault count, and removes only that route handler.
80
+ **0.3.0** is the stable release on npm's `latest` channel, and the commands above pin it for reproducible reports. The 0.4 observation preview is `recovery-probe@0.4.0-preview.5` on the `next` channel; install it with the same Playwright version.
210
81
 
211
- ```ts
212
- interface FaultOptions {
213
- match: string | RegExp;
214
- kind: 'http-error' | 'connection-failure' | 'invalid-json';
215
- status?: number;
216
- count?: number;
217
- method?: string;
218
- }
219
- ```
82
+ Use Node.js 22 or newer. A separate tools folder keeps the tester out of your application's production dependencies. Browser and desktop commands need Playwright; attached targets use their existing browser runtime. If a version is not on npm yet, use the [source installation](CONTRIBUTING.md#install-an-unpublished-source-preview): a GitHub merge alone does not publish the package. See [CHANGELOG.md](CHANGELOG.md) for version history.
220
83
 
221
- Register the fault after your own route mocks. The helper uses `route.fallback()` for requests it does not inject, allowing earlier handlers to continue processing them.
84
+ ## Reports, agents and privacy
222
85
 
223
- ### `createFaultFetch(baseFetch, options)`
86
+ Reports separate measured facts, such as fault counts, successful reads, timings and UI changes, from suspected defects. A `CONTENT_LOSS` finding needs review. An unconsumed fault, an unstable control or an unsupported path is inconclusive. Observation findings such as `DELAYED_UI_UPDATE` are leads to check against the application logic, not verdicts.
224
87
 
225
- Creates an isolated fetch wrapper for one exact URL and method. It returns `{ fetch, summary, assertApplied }` and leaves the supplied fetch function unchanged.
88
+ Coding agents can run saved scenarios, read the JSON evidence, propose a change and rerun the same experiment. Recovery Probe supplies repeatable experiments and evidence; it is not a general bug detector or a production self-healing system.
226
89
 
227
- ### `probeRecovery(browser, config)`
90
+ Reports stay local, and Recovery Probe sends no telemetry. Scenarios, endpoints, channels, selected text, screenshots, login state and command logs may be private, so do not publish them blindly. Request and response bodies and headers are not collected, and observation never records read contents.
228
91
 
229
- Runs the baseline and the three default recovery checks used by the JSON CLI. For custom interactions or stronger assertions, prefer `withFault` inside your own test.
92
+ Tests and discovery do not call an AI service. Only the opt-in repair workflow, with `--allow-source-upload`, sends the selected source files and bounded observations to your configured OpenAI model; login state and command logs are not part of that upload. It sets `store: false`, which is not a promise of zero provider retention, so review your employer's policy and the provider's data terms first. A repair is a reviewable candidate, not a guaranteed fix.
230
93
 
231
94
  ## Scope
232
95
 
233
- Recovery Probe focuses on deterministic request-level recovery checks and on fault-free observation of whether the screen reflects the reads it received. It does not simulate whole-device offline states, WebSocket interruptions, token expiry, browser sleep, slow-network profiles, or duplicate-submission safety. Observation findings such as `DELAYED_UI_UPDATE` are leads that need review against the application logic, not verdicts: read contents are never recorded.
96
+ Recovery Probe checks request-level recovery deterministically and observes whether the screen reflects the reads it received. It does not simulate whole-device offline states, WebSocket interruptions, token expiry, browser sleep, slow networks or duplicate submissions.
234
97
 
235
- Playwright already provides the routing primitives used here. Recovery Probe adds bounded recipes, fault-occurrence verification, cleanup, consistent result codes, a fetch adapter, and a runnable broken-versus-fixed example. If a few direct `page.route()` calls are clearer for your test, use them; this package is most useful when teams want the recovery pattern to be repeatable.
98
+ Structural discovery compares healthy and faulted screens without a hand-written content selector, but it needs a selected read and a stable healthy baseline. No workflow infers your product's intended recovery policy.
236
99
 
237
- The original assertion-based runner requires explicit selectors and endpoint matching; structural discovery and refresh modes use healthy-screen comparisons instead. It does not discover an application's recovery policy automatically. Request/response bodies and headers are not collected by the checker. Preview reports include endpoint origin/path, selectors, expected text and configuration; optional screenshots, login state and command logs may be sensitive. Do not commit or share these blindly.
100
+ The recovery, discovery and refresh workflows reload the renderer to establish controls. That clears in-memory caches, so those modes cannot prove account isolation. Observation keeps that state, but it does not infer identities or data ownership.
238
101
 
239
- Recovery Probe sends no telemetry. Tests/discovery do not call an AI service. Only explicit repair generation with `--allow-source-upload` sends the selected source contents and bounded observations to the configured OpenAI model. Login state and command logs are not part of that upload bundle. `store: false` is set on the API request; this is not a promise of zero provider retention. Review employer policy and provider data terms first.
102
+ Attach only to authorized test sessions, keep debugging endpoints on loopback, and keep the target window visible and uncovered.
240
103
 
241
104
  ## Development
242
105
 
243
- ```bash
244
- npm ci
245
- npm test # fetch/core tests
246
- npm run test:package # isolated tarball install + TypeScript + CLI checks
247
- npx playwright install chromium
248
- npm run test:browser # real-browser integration tests
249
- npm run test:repair # safety, authenticated portal, before/after pipeline
250
- # After installing the optional Electron test runtime:
251
- npm run test:repair:electron
252
- npm run demo:browser
253
- ```
254
-
255
- CI runs package checks on Node.js 22 and 24, Chromium integration, and real Electron checks. Repair integration tests use a **controlled mock model response** but real apps, fault injection, source changes and reruns. They validate orchestration, not live AI diagnosis quality. See [CONTRIBUTING.md](CONTRIBUTING.md) and [SECURITY.md](SECURITY.md).
106
+ See [CONTRIBUTING.md](CONTRIBUTING.md) for setup, checks and test suites, and [SECURITY.md](SECURITY.md) for reporting security issues. CI runs package checks on Node.js 22 and 24, Chromium integration, and real Electron checks.
256
107
 
257
108
  ## License
258
109
 
package/bin/ipc.mjs CHANGED
@@ -3,14 +3,8 @@ import { readFile, mkdir, writeFile, mkdtemp } from 'node:fs/promises'
3
3
  import { resolve, join } from 'node:path'
4
4
  import { discoverIpc, discoveryConfig } from '../src/ipc-discover.mjs'
5
5
  import { checkIpc } from '../src/ipc-check.mjs'
6
- const escape = (value) =>
7
- String(value).replace(
8
- /[&<>"']/g,
9
- (c) =>
10
- ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' })[
11
- c
12
- ],
13
- )
6
+ import { renderTemplate } from '../src/report-template.mjs'
7
+
14
8
  export async function ipcMain(args) {
15
9
  if (!args.length || args.includes('--help')) {
16
10
  console.log(`Recovery Probe IPC preview
@@ -71,7 +65,14 @@ No HTTP/Rust transport faults or AI calls. See docs/ipc.md for installation and
71
65
  JSON.stringify(report, null, 2),
72
66
  { mode: 0o600 },
73
67
  )
74
- const html = `<!doctype html><meta charset="utf-8"><title>Recovery Probe IPC report</title><style>body{max-width:1000px;margin:40px auto;font:16px system-ui;padding:20px}pre{white-space:pre-wrap;overflow-wrap:anywhere;background:#f4f5f7;padding:20px}</style><h1>${report.ok ? 'Selected IPC experiments completed without findings requiring review' : 'IPC recovery needs investigation'}</h1><p>${escape(report.limitation)}</p><p>Final reload: ${escape(report.finalReset)}. Cleanup cannot turn an earlier failure into a pass.</p><pre>${escape(JSON.stringify(report, null, 2))}</pre>`
68
+ const html = renderTemplate('ipc', {
69
+ heading: report.ok
70
+ ? 'Selected IPC experiments completed without findings requiring review'
71
+ : 'IPC recovery needs investigation',
72
+ limitation: report.limitation,
73
+ finalReset: report.finalReset,
74
+ evidence: JSON.stringify(report, null, 2),
75
+ })
75
76
  await writeFile(join(directory, 'report.html'), html, { mode: 0o600 })
76
77
  if (options.discover) {
77
78
  await writeDiscoveryReport(directory, report, config)
package/bin/refresh.mjs CHANGED
@@ -13,14 +13,8 @@ import {
13
13
  httpRefreshAdapter,
14
14
  ipcRefreshAdapter,
15
15
  } from '../src/refresh-adapters.mjs'
16
- const escape = (value) =>
17
- String(value).replace(
18
- /[&<>"']/g,
19
- (c) =>
20
- ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' })[
21
- c
22
- ],
23
- )
16
+ import { escapeHtml, renderTemplate } from '../src/report-template.mjs'
17
+
24
18
  export async function refreshMain(args) {
25
19
  if (args.includes('--help')) {
26
20
  console.log(
@@ -136,7 +130,13 @@ export async function refreshMain(args) {
136
130
  )
137
131
  await writeFile(
138
132
  resolve(directory, 'report.html'),
139
- `<!doctype html><meta charset="utf-8"><title>Recovery Probe refresh</title><style>body{font:16px system-ui;max-width:1000px;margin:40px auto;padding:20px}pre{white-space:pre-wrap;overflow-wrap:anywhere}</style><h1>Refresh experiment</h1><p>${escape(report.limitation ?? report.error)}</p>${report.results.map((r) => `<p>${escape(r.summary ?? r.code ?? r.stage)}</p>`).join('')}<details><summary>Evidence</summary><pre>${escape(JSON.stringify(report, null, 2))}</pre></details>`,
133
+ renderTemplate('refresh', {
134
+ limitation: report.limitation ?? report.error,
135
+ results: report.results
136
+ .map((r) => `<p>${escapeHtml(r.summary ?? r.code ?? r.stage)}</p>`)
137
+ .join(''),
138
+ evidence: JSON.stringify(report, null, 2),
139
+ }),
140
140
  { mode: 0o600 },
141
141
  )
142
142
  const summary = {
package/docs/api.md ADDED
@@ -0,0 +1,57 @@
1
+ # Library API
2
+
3
+ The library adds bounded faults to tests you already have. `recovery-probe` needs Playwright; `recovery-probe/fetch` has no runtime dependencies.
4
+
5
+ ## Fault kinds
6
+
7
+ | Kind | Injected behavior |
8
+ | --- | --- |
9
+ | `http-error` | Fulfils the matched request with an HTTP error status; defaults to `503`. |
10
+ | `connection-failure` | Aborts the matched Playwright request with `connectionfailed`, or throws from the fetch adapter. |
11
+ | `invalid-json` | Returns status `200` with a deliberately malformed JSON body. |
12
+
13
+ Faults are bounded by `count` (default `1`, maximum `20`). The count is reserved synchronously, so concurrent matching requests cannot over-inject.
14
+
15
+ ## `withFault(page, options, exercise)`
16
+
17
+ Installs one temporary Playwright route, runs `exercise`, verifies the requested fault count, and removes only that route handler. See the [README example](../README.md#add-a-recovery-check-to-a-playwright-test).
18
+
19
+ ```ts
20
+ interface FaultOptions {
21
+ match: string | RegExp;
22
+ kind: 'http-error' | 'connection-failure' | 'invalid-json';
23
+ status?: number;
24
+ count?: number;
25
+ method?: string;
26
+ }
27
+ ```
28
+
29
+ If the endpoint does not match, it throws `FAULT_NOT_TRIGGERED` instead of allowing the test to pass silently. If the fault is applied but the recovery assertion fails, it throws `RECOVERY_ASSERTION_FAILED` and retains the original assertion error as `cause`.
30
+
31
+ Register the fault after your own route mocks. The helper uses `route.fallback()` for requests it does not inject, allowing earlier handlers to continue processing them.
32
+
33
+ ## `createFaultFetch(baseFetch, options)`
34
+
35
+ Creates an isolated fetch wrapper for one exact URL and method. It returns `{ fetch, summary, assertApplied }` and never patches global `fetch`. The URL must be absolute and matches exactly, including its query string. Every other request, including a relative one such as `fetch('/api/x')`, reaches the supplied fetch untouched.
36
+
37
+ ```js
38
+ import { createFaultFetch } from 'recovery-probe/fetch';
39
+
40
+ const probe = createFaultFetch(fetch, {
41
+ url: 'http://127.0.0.1:3000/api/profile',
42
+ kind: 'connection-failure',
43
+ });
44
+
45
+ await exerciseApplication(probe.fetch);
46
+ probe.assertApplied();
47
+ ```
48
+
49
+ `exerciseApplication` represents your own test adapter or application controller. Calling `assertApplied()` is essential: it proves that every requested fault was consumed.
50
+
51
+ ## `probeRecovery(browser, config)`
52
+
53
+ Runs the baseline and the three default recovery checks used by the [JSON-configured runner](runner.md). For custom interactions or stronger assertions, prefer `withFault` inside your own test.
54
+
55
+ ## When plain Playwright routes are enough
56
+
57
+ Playwright already provides the routing primitives used here. Recovery Probe adds bounded recipes, fault-occurrence verification, cleanup, consistent result codes, a fetch adapter, and a runnable broken-versus-fixed example. If a few direct `page.route()` calls are clearer for your test, use them; this package is most useful when teams want the recovery pattern to be repeatable.
package/docs/observe.md CHANGED
@@ -108,7 +108,7 @@ Tracking ends at the next operator interval or capture termination, including in
108
108
 
109
109
  ## Actions, tracked rows and delayed UI updates
110
110
 
111
- Available from **0.4.0-preview.1**. The current preview target is **0.4.0-preview.3** on npm `next`; install it explicitly with `npm install --save-dev --save-exact recovery-probe@0.4.0-preview.3 playwright@1.62.1`. The development IPC hook must come from the same version.
111
+ Available from **0.4.0-preview.1**. The current preview target is **0.4.0-preview.5** on npm `next`; install it explicitly with `npm install --save-dev --save-exact recovery-probe@0.4.0-preview.5 playwright@1.62.1`. The development IPC hook must come from the same version.
112
112
 
113
113
  These options are opt-in and extend the same fault-free capture. They answer "how long after the user's action did the screen show the right state, and did data arrive before it?"
114
114
 
package/docs/repair.md CHANGED
@@ -75,7 +75,7 @@ These `web`/`desktop` checks never call an AI provider. Read their local reports
75
75
 
76
76
  ## Configure a repair run
77
77
 
78
- The repair runner must launch a **separate build from its own checkout**, not attach to the existing app. Stop whatever is using the configured test port. Commit or stash all changes, including untracked files, in the source repository first. Ignored `.env` files, dependencies and local credentials are not copied into the checkout. Use a disposable dev environment with nonproduction settings. Your configured setup commands must prepare whatever that build needs; secret environment variables are not inherited automatically.
78
+ The repair runner must launch a **separate build from its own checkout**, not attach to the existing app. Stop whatever is using the configured test port. Commit or stash all changes, including untracked files, in the source repository first. Ignored `.env` files, dependencies and local credentials are not copied into the checkout. The checkout holds the committed bytes: Git line-ending conversion is off there, so on Windows its files keep LF line endings where your working tree has CRLF. `eol` rules in `.gitattributes` still apply. Use a disposable dev environment with nonproduction settings. Your configured setup commands must prepare whatever that build needs; secret environment variables are not inherited automatically.
79
79
 
80
80
  Create `portal-repair.json` in the tools folder, outside the app repository:
81
81
 
package/docs/runner.md ADDED
@@ -0,0 +1,40 @@
1
+ # JSON-configured runner
2
+
3
+ The original CLI runner suits pages that automatically load GET data, expose a visible Retry button after failure, and show a ready element after recovery. For other screens, use [guided setup](start.md), [discovery](desktop.md) or `withFault` in your own test ([API](api.md)).
4
+
5
+ ```json
6
+ {
7
+ "name": "Profile recovery",
8
+ "url": "http://127.0.0.1:3000/profile",
9
+ "requestPattern": "**/api/profile",
10
+ "readySelector": "[data-testid=profile-ready]",
11
+ "retrySelector": "[data-testid=profile-retry]",
12
+ "timeoutMs": 3000
13
+ }
14
+ ```
15
+
16
+ ```bash
17
+ npx recovery-probe --config recovery-probe.json --out artifacts/profile-recovery.json
18
+ ```
19
+
20
+ The runner first checks the normal baseline in a fresh browser context. If that baseline fails, fault comparisons are skipped and reported as inconclusive. Each fault then receives its own fresh context with service workers blocked so Playwright can intercept the request.
21
+
22
+ The runner requires explicit selectors and endpoint matching. It does not discover an application's recovery policy.
23
+
24
+ ## Exit codes
25
+
26
+ - `0`: every configured check passed, or the controlled demo observed its expected broken-versus-fixed result.
27
+ - `1`: a recovery check failed, was inconclusive, or was skipped.
28
+ - `2`: invalid arguments, configuration, or environment setup.
29
+
30
+ ## Result codes
31
+
32
+ | Code | Meaning |
33
+ | --- | --- |
34
+ | `OK` | The configured assertion completed and the expected fault count was observed. |
35
+ | `RECOVERY_ASSERTION_FAILED` | The fault occurred, but the recovery assertion did not pass. |
36
+ | `FAULT_NOT_TRIGGERED` | The requested fault count was not observed. Check the route, method, cache, and service worker. |
37
+ | `BASELINE_FAILED` | The ordinary flow failed, so the runner could not make a useful fault comparison. |
38
+ | `INJECTION_ERROR` | Playwright could not apply or clean up the route reliably. |
39
+
40
+ A passing result proves the configured recovery behavior for that test. It is not a certification of overall application resilience.
package/package.json CHANGED
@@ -1,8 +1,8 @@
1
1
  {
2
2
  "name": "recovery-probe",
3
- "version": "0.4.0-preview.3",
3
+ "version": "0.4.0-preview.5",
4
4
  "type": "module",
5
- "description": "Test whether web and Electron desktop apps show the right data: recovery after failed reads, cached content, and screens that lag behind their data.",
5
+ "description": "Test whether web and Electron desktop apps show the right data.",
6
6
  "license": "MIT",
7
7
  "author": "Bojan Kovachki",
8
8
  "homepage": "https://github.com/BojanKovachki/recovery-probe#readme",
@@ -58,7 +58,9 @@
58
58
  "docs/ipc.md",
59
59
  "docs/refresh.md",
60
60
  "docs/observe.md",
61
- "docs/ci.md"
61
+ "docs/ci.md",
62
+ "docs/api.md",
63
+ "docs/runner.md"
62
64
  ],
63
65
  "sideEffects": false,
64
66
  "engines": {
@@ -127,7 +129,7 @@
127
129
  "test:repair": "node --test test/repair.test.mjs test/repair.integration.test.mjs test/web.integration.test.mjs",
128
130
  "test:repair:electron": "node --test test/repair.electron.test.mjs",
129
131
  "test:guided": "node --test test/guided-start.test.mjs test/guided-start.integration.test.mjs",
130
- "test:ipc": "node --test test/ipc-probe.test.mjs test/ipc-control.test.mjs test/ipc-observation.test.mjs test/visibility.test.mjs test/refresh.test.mjs test/observe.test.mjs test/observation-report.test.mjs test/observe-features.test.mjs",
132
+ "test:ipc": "node --test test/ipc-probe.test.mjs test/ipc-control.test.mjs test/ipc-observation.test.mjs test/visibility.test.mjs test/refresh.test.mjs test/observe.test.mjs test/observation-report.test.mjs test/report-template.test.mjs test/observe-features.test.mjs",
131
133
  "test:ipc:electron": "node --test --test-concurrency=1 test/ipc.electron.test.mjs test/observe.electron.test.mjs test/observe-queue.electron.test.mjs",
132
134
  "benchmark:observe": "node scripts/benchmark-observe.mjs",
133
135
  "test:ci-example": "node examples/ci/verify.mjs"
@@ -1,29 +1,40 @@
1
1
  import { mkdir, writeFile } from 'node:fs/promises'
2
2
  import { resolve, join } from 'node:path'
3
+ import { escapeHtml, renderTemplate } from './report-template.mjs'
3
4
 
4
- const escape = (value) =>
5
- String(value ?? '').replace(
6
- /[&<>"']/g,
7
- (c) =>
8
- ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' })[
9
- c
10
- ],
11
- )
12
- export function desktopReportHtml(report) {
13
- const rows = report.results
5
+ export function desktopReportHtml(report, target = 'desktop') {
6
+ const results = report.results
14
7
  .map(
15
8
  (row) =>
16
- `<tr><td>${escape(row.kind)}</td><td class="${escape(row.outcome)}">${escape(row.outcome.toUpperCase())}</td><td>${escape(row.code)}</td><td>${row.applied}</td><td>${row.successfulResponsesAfterFault}</td><td>${row.retryClicked ? 'Yes' : 'No'}</td></tr>`,
9
+ `<tr><td>${escapeHtml(row.kind)}</td><td class="${escapeHtml(row.outcome)}">${escapeHtml(row.outcome.toUpperCase())}</td><td>${escapeHtml(row.code)}</td><td>${row.applied}</td><td>${row.successfulResponsesAfterFault}</td><td>${row.retryClicked ? 'Yes' : 'No'}</td></tr>`,
17
10
  )
18
11
  .join('')
19
- const images = report.results
12
+ const screenshots = report.results
20
13
  .filter((row) => row.screenshot && /^[a-z0-9-]+\.png$/.test(row.screenshot))
21
14
  .map(
22
15
  (row) =>
23
- `<figure><figcaption>${escape(row.kind)} — ${escape(row.code)}</figcaption><img alt="Renderer at the end of the recovery check" src="${escape(row.screenshot)}"></figure>`,
16
+ `<figure><figcaption>${escapeHtml(row.kind)} — ${escapeHtml(row.code)}</figcaption><img alt="Renderer at the end of the recovery check" src="${escapeHtml(row.screenshot)}"></figure>`,
24
17
  )
25
18
  .join('')
26
- return `<!doctype html><html lang="en"><meta charset="utf-8"><meta name="viewport" content="width=device-width, initial-scale=1"><title>Recovery Probe — desktop report</title><style>body{font:16px/1.6 system-ui,sans-serif;background:#f6f7f9;color:#17212f;margin:0;padding:40px}main{max-width:1100px;margin:auto}h1{font-size:32px;letter-spacing:-1px}code{overflow-wrap:anywhere}table{width:100%;border-collapse:collapse;background:white;font-size:14px}th,td{text-align:left;border-bottom:1px solid #ddd;padding:12px}section{background:white;padding:20px;margin:24px 0;border-radius:10px}.pass{color:#087443}.fail{color:#b3261e}.inconclusive,.skipped{color:#755300}img{max-width:100%;border:1px solid #ddd}figcaption{font-weight:600}small{color:#536174}.scroll{overflow:auto}</style><main><p>RECOVERY PROBE · DESKTOP PREVIEW</p><h1>${report.ok ? 'Selected recovery checks passed' : 'Recovery needs investigation'}</h1><p><code>GET ${escape(report.endpoint)}</code></p><p>Recovery: <strong>${escape(report.recovery)}</strong> · Deadline: ${report.timeoutMs} ms · ${escape(report.generatedAt)}</p><div class="scroll"><table><thead><tr><th>Injected fault</th><th>Outcome</th><th>Result</th><th>Injected</th><th>Later successful responses</th><th>Retry clicked</th></tr></thead><tbody>${rows}</tbody></table></div><section><h2>How to read this</h2><p><strong>PASS:</strong> the fault occurred, the configured action completed, a later matching request succeeded, and the selected healthy UI marker appeared. <strong>FAIL:</strong> after the fault and action, recovery was not observed by the deadline. Investigate the app and the chosen assertion. <strong>INCONCLUSIVE:</strong> setup, baseline, interception or the recovery action could not be verified; this is not a confirmed app bug.</p><p>Final reload reset: <strong>${escape(report.finalReset)}</strong>. A reload that rescues a stuck page does not turn the original failed recovery check into a pass.</p><p>${escape(report.limitation)}</p></section><section><h2>Reproduce and investigate</h2><p>From this output directory, run <code>node reproduce.mjs</code> while your development app is on the same screen. See <code>scenario.json</code> and <code>fix-brief.md</code>.</p><p>No source patch was applied. No data was uploaded. Review local artifacts before sharing them.</p></section>${images}<small>HTTP status success and a visible marker do not certify response semantics. Choose a marker that only appears when this endpoint's data has actually loaded; configure readyText/busySelector for a stronger assertion.</small></main></html>`
19
+ return renderTemplate('desktop', {
20
+ target,
21
+ targetLabel: target.toUpperCase(),
22
+ heading: report.ok
23
+ ? 'Selected recovery checks passed'
24
+ : 'Recovery needs investigation',
25
+ endpoint: report.endpoint,
26
+ recovery: report.recovery,
27
+ timeoutMs: report.timeoutMs,
28
+ generatedAt: report.generatedAt,
29
+ results,
30
+ finalReset: report.finalReset,
31
+ limitation: report.limitation,
32
+ reproduceWhere:
33
+ target === 'web'
34
+ ? 'against the configured development URL'
35
+ : 'while your development app is on the same screen',
36
+ screenshots,
37
+ })
27
38
  }
28
39
  export async function writeDesktopReport(directory, report, config) {
29
40
  const dir = resolve(directory)
@@ -43,23 +54,14 @@ export async function writeDesktopReport(directory, report, config) {
43
54
  for (const [name, value] of Object.entries({
44
55
  'report.json': JSON.stringify(report, null, 2) + '\n',
45
56
  'scenario.json': JSON.stringify(config, null, 2) + '\n',
46
- 'report.html': desktopReportHtml(report),
57
+ 'report.html': desktopReportHtml(
58
+ report,
59
+ config.target === 'web' ? 'web' : 'desktop',
60
+ ),
47
61
  'fix-brief.md': brief,
48
62
  'reproduce.mjs': reproduction,
49
63
  })) {
50
- await writeFile(
51
- join(dir, name),
52
- config.target === 'web' && name === 'report.html'
53
- ? value
54
- .replaceAll('desktop report', 'web report')
55
- .replaceAll('DESKTOP PREVIEW', 'WEB PREVIEW')
56
- .replace(
57
- 'while your development app is on the same screen',
58
- 'against the configured development URL',
59
- )
60
- : value,
61
- { mode: 0o600 },
62
- )
64
+ await writeFile(join(dir, name), value, { mode: 0o600 })
63
65
  }
64
66
  return dir
65
67
  }
@@ -1,13 +1,7 @@
1
1
  import { writeFile } from 'node:fs/promises'
2
2
  import { join } from 'node:path'
3
- const escape = (value) =>
4
- String(value).replace(
5
- /[&<>"']/g,
6
- (c) =>
7
- ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' })[
8
- c
9
- ],
10
- )
3
+ import { escapeHtml, renderTemplate } from './report-template.mjs'
4
+
11
5
  export async function writeDiscoveryReport(directory, report, config) {
12
6
  await writeFile(join(directory, '.gitignore'), '*\n', { mode: 0o600 })
13
7
  await writeFile(
@@ -20,10 +14,10 @@ export async function writeDiscoveryReport(directory, report, config) {
20
14
  JSON.stringify(report, null, 2),
21
15
  { mode: 0o600 },
22
16
  )
23
- const rows = report.findings
17
+ const experiments = report.findings
24
18
  .map(
25
19
  (f) =>
26
- `<tr><td>${escape(f.id)}</td><td>${escape(f.classification)}</td><td>${escape(f.kind)}</td><td>${escape(f.confidence ?? 'not-rated')}</td><td>${f.partial ? `${f.completedRuns}/${f.plannedRuns} (partial)` : f.confirmations}</td><td>${escape(f.observedInterpretation?.classification ?? '')}</td></tr>`,
20
+ `<tr><td>${escapeHtml(f.id)}</td><td>${escapeHtml(f.classification)}</td><td>${escapeHtml(f.kind)}</td><td>${escapeHtml(f.confidence ?? 'not-rated')}</td><td>${f.partial ? `${f.completedRuns}/${f.plannedRuns} (partial)` : f.confirmations}</td><td>${escapeHtml(f.observedInterpretation?.classification ?? '')}</td></tr>`,
27
21
  )
28
22
  .join('')
29
23
  const summary = {
@@ -42,7 +36,12 @@ export async function writeDiscoveryReport(directory, report, config) {
42
36
  }
43
37
  await writeFile(
44
38
  join(directory, 'report.html'),
45
- `<!doctype html><meta charset="utf-8"><title>Recovery Probe discovery</title><style>body{font:16px system-ui;max-width:1100px;margin:40px auto;padding:20px}td,th{padding:10px;border-bottom:1px solid #ddd;text-align:left}pre{white-space:pre-wrap;overflow-wrap:anywhere}</style><h1>Recovery Probe discovery</h1><p>${escape(report.limitation)}</p><table><tr><th>Experiment</th><th>Interpretation</th><th>Kind</th><th>Confidence</th><th>Runs</th><th>Partial observation</th></tr>${rows}</table><h2>Run summary</h2><pre>${escape(JSON.stringify(summary, null, 2))}</pre><details><summary>Observations and bounded timelines</summary><pre>${escape(JSON.stringify(report, null, 2))}</pre></details>`,
39
+ renderTemplate('discovery', {
40
+ limitation: report.limitation,
41
+ experiments,
42
+ summary: JSON.stringify(summary, null, 2),
43
+ evidence: JSON.stringify(report, null, 2),
44
+ }),
46
45
  { mode: 0o600 },
47
46
  )
48
47
  }
@@ -1,5 +1,14 @@
1
1
  const kinds = new Set(['http-error', 'connection-failure', 'invalid-json'])
2
2
 
3
+ // A relative or malformed URL cannot be the absolute target, so it is never matched.
4
+ const absoluteHref = (value) => {
5
+ try {
6
+ return new URL(value).href
7
+ } catch {
8
+ return null
9
+ }
10
+ }
11
+
3
12
  /** Fault-injected fetch for an exact, absolute URL. Each instance is isolated. */
4
13
  export function createFaultFetch(baseFetch, options) {
5
14
  if (typeof baseFetch !== 'function')
@@ -54,14 +63,13 @@ export function createFaultFetch(baseFetch, options) {
54
63
  },
55
64
  async fetch(input, init) {
56
65
  const inputUrl = input instanceof Request ? input.url : input
57
- const requestUrl = new URL(inputUrl)
58
66
  const requestMethod = (
59
67
  init?.method ?? (input instanceof Request ? input.method : 'GET')
60
68
  ).toUpperCase()
61
69
  const signal =
62
70
  init?.signal ?? (input instanceof Request ? input.signal : undefined)
63
71
  if (
64
- requestUrl.href !== target.href ||
72
+ absoluteHref(inputUrl) !== target.href ||
65
73
  requestMethod !== method ||
66
74
  signal?.aborted
67
75
  ) {
@@ -1,4 +1,5 @@
1
1
  import { createRequire } from 'node:module'
2
+ import { escapeHtml, renderTemplate } from './report-template.mjs'
2
3
  const { version } = createRequire(import.meta.url)('../package.json')
3
4
 
4
5
  export function observationMetadata() {
@@ -306,16 +307,18 @@ export function comparisonDigest(report) {
306
307
  )
307
308
  }
308
309
 
309
- const escape = (value) =>
310
- String(value).replace(
311
- /[&<>"']/g,
312
- (c) =>
313
- ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' })[
314
- c
315
- ],
316
- )
310
+ const trackedRowHtml = (r) => {
311
+ const changes = r.transitions
312
+ .map(
313
+ (t) =>
314
+ `${escapeHtml(t.from)} → ${escapeHtml(t.to)} at ${seconds(t.afterMs)}–${seconds(t.byMs)} s`,
315
+ )
316
+ .join('<br>')
317
+ return `<tr><td>${escapeHtml(rowName(r.row))}</td><td>${seconds(r.firstSeen.fromMs)}–${seconds(r.firstSeen.toMs)} s as ${escapeHtml(r.firstSeen.state)}</td><td>${changes || 'none'}</td><td>${escapeHtml(r.lastState)}</td></tr>`
318
+ }
319
+
317
320
  export function renderObservationReport(report) {
318
- const rows = report.intervals
321
+ const intervals = report.intervals
319
322
  .map(
320
323
  (i) =>
321
324
  `<tr><td>${i.index}</td><td>${i.reads?.calls ?? 'unknown'}</td><td>${i.reads?.successes ?? 'unknown'}</td><td>${i.reads?.errors ?? 'unknown'}</td><td>${i.firstContent ? `${i.firstContent.fromMs - i.startMs}–${i.firstContent.toMs - i.startMs} ms` : 'not sampled'}</td></tr>`,
@@ -325,27 +328,58 @@ export function renderObservationReport(report) {
325
328
  const comparisons = (report.comparisons ?? [])
326
329
  .map(
327
330
  (c) =>
328
- `<p>Interval ${escape(c.interval)}: ${escape(c.summary)} ${c.completions ? `(${c.completions.successes} successful, ${c.completions.errors} failed read completions)` : `(${c.reads.calls} selected reads started in this interval)`} ${escape(comparisonFollowUpSummary(c))}</p>`,
331
+ `<p>Interval ${escapeHtml(c.interval)}: ${escapeHtml(c.summary)} ${c.completions ? `(${c.completions.successes} successful, ${c.completions.errors} failed read completions)` : `(${c.reads.calls} selected reads started in this interval)`} ${escapeHtml(comparisonFollowUpSummary(c))}</p>`,
329
332
  )
330
333
  .join('')
331
- const phases = Object.entries(sampling?.phases ?? {})
334
+ const phaseRows = Object.entries(sampling?.phases ?? {})
332
335
  .map(
333
336
  ([name, s]) =>
334
- `<tr><td>${escape(name)}</td><td>${s.meanMs}</td><td>${s.maxMs}</td></tr>`,
337
+ `<tr><td>${escapeHtml(name)}</td><td>${s.meanMs}</td><td>${s.maxMs}</td></tr>`,
335
338
  )
336
339
  .join('')
337
340
  const effects = report.actionEffects ?? []
338
341
  const actionsHtml = effects.length
339
- ? `<h2>Actions</h2>${effects.map((e) => `<p>${escape(actionSummary(e))}</p>`).join('')}`
342
+ ? `<h2>Actions</h2>${effects.map((e) => `<p>${escapeHtml(actionSummary(e))}</p>`).join('')}`
340
343
  : ''
341
- const rowsHtml = report.rows?.length
342
- ? `<h2>Tracked rows</h2><table><thead><tr><th>Row</th><th>First seen</th><th>State changes</th><th>Last state</th></tr></thead><tbody>${report.rows.map((r) => `<tr><td>${escape(rowName(r.row))}</td><td>${seconds(r.firstSeen.fromMs)}–${seconds(r.firstSeen.toMs)} s as ${escape(r.firstSeen.state)}</td><td>${r.transitions.map((t) => `${escape(t.from)} → ${escape(t.to)} at ${seconds(t.afterMs)}–${seconds(t.byMs)} s`).join('<br>') || 'none'}</td><td>${escape(r.lastState)}</td></tr>`).join('')}</tbody></table><p>Row keys are hashed unless an alias was configured. States outside the configured allowlist are recorded as “other”.</p>`
344
+ const trackedRows = report.rows?.length
345
+ ? `<h2>Tracked rows</h2><table><thead><tr><th>Row</th><th>First seen</th><th>State changes</th><th>Last state</th></tr></thead><tbody>${report.rows.map(trackedRowHtml).join('')}</tbody></table><p>Row keys are hashed unless an alias was configured. States outside the configured allowlist are recorded as “other”.</p>`
343
346
  : ''
344
347
  const notes = (report.notes ?? [])
345
- .map((n) => `<p>Note: ${escape(n.note)}</p>`)
348
+ .map((n) => `<p>Note: ${escapeHtml(n.note)}</p>`)
346
349
  .join('')
347
350
  const operatorHtml = report.operator
348
- ? `<p>Scripted steps: ${escape(report.operator.status)}${report.operator.error ? ` — ${escape(report.operator.error)}` : ''}.</p>`
351
+ ? `<p>Scripted steps: ${escapeHtml(report.operator.status)}${report.operator.error ? ` — ${escapeHtml(report.operator.error)}` : ''}.</p>`
352
+ : ''
353
+ const samplingHtml = sampling
354
+ ? `<p>Configured delay: ${sampling.requestedDelayMs} ms. Completed samples: ${sampling.completedSamples}. Mean sample duration: ${sampling.sampleDuration.meanMs ?? 'unavailable'} ms. Mean start-to-start cadence: ${sampling.startToStart.meanMs ?? 'unavailable'} ms (maximum ${sampling.startToStart.maxMs ?? 'unavailable'} ms).</p><p>${escapeHtml(sampling.note)}</p>`
349
355
  : ''
350
- return `<!doctype html><meta charset="utf-8"><title>Recovery Probe observation</title><style>body{font:16px system-ui;max-width:1000px;margin:40px auto;padding:20px}td,th{text-align:left;padding:8px}pre{white-space:pre-wrap;overflow-wrap:anywhere}</style><h1>Fault-free observation</h1><p>Capture: ${escape(report.status)}. This is not a recovery or account-isolation verdict.</p><p>Recovery Probe ${escape(report.metadata?.recoveryProbeVersion ?? 'unknown')} · Node ${escape(report.metadata?.runtime?.node ?? 'unknown')} · Browser ${escape(report.metadata?.browserVersion ?? 'unknown')} · IPC hook ${escape(report.metadata?.ipcHook?.version ?? 'unknown / not applicable')}</p>${sampling ? `<p>Configured delay: ${sampling.requestedDelayMs} ms. Completed samples: ${sampling.completedSamples}. Mean sample duration: ${sampling.sampleDuration.meanMs ?? 'unavailable'} ms. Mean start-to-start cadence: ${sampling.startToStart.meanMs ?? 'unavailable'} ms (maximum ${sampling.startToStart.maxMs ?? 'unavailable'} ms).</p><p>${escape(sampling.note)}</p>` : ''}<p>Sampling detail: ${escape(report.detail ?? 'region (legacy)')}. Marker counts describe the configured content selector; generic region item counts are heuristics.</p>${phases ? `<table><thead><tr><th>Measured phase</th><th>Mean ms</th><th>Maximum ms</th></tr></thead><tbody>${phases}</tbody></table><p>Renderer read time is included in the UI round trip, not additional to it.</p>` : ''}<p>${escape(report.limitation ?? report.error)}</p><table><thead><tr><th>Interval</th><th>Reads started</th><th>Successes</th><th>Errors</th><th>First content sample bounds</th></tr></thead><tbody>${rows}</tbody></table>${report.findings.map((f) => `<p>Interval ${f.interval}: ${escape(observationSummary(f))} <strong>${escape(f.code)}</strong> — ${escape(f.basis)}</p>`).join('')}${operatorHtml}${actionsHtml}${rowsHtml}${notes}${comparisons}<p>Comparisons concern sampled marker text and count, not ownership or freshness. A completion does not guarantee the UI has consumed its result.</p><p>Cleanup: ${escape(report.cleanup ?? 'unknown')}. The observation snapshot is from before cleanup; cleanupEvidence records the separate cleanup check when available.</p><details><summary>Evidence and coverage</summary><pre>${escape(JSON.stringify(report, null, 2))}</pre></details>`
356
+ const phases = phaseRows
357
+ ? `<table><thead><tr><th>Measured phase</th><th>Mean ms</th><th>Maximum ms</th></tr></thead><tbody>${phaseRows}</tbody></table><p>Renderer read time is included in the UI round trip, not additional to it.</p>`
358
+ : ''
359
+ const findings = report.findings
360
+ .map(
361
+ (f) =>
362
+ `<p>Interval ${f.interval}: ${escapeHtml(observationSummary(f))} <strong>${escapeHtml(f.code)}</strong> — ${escapeHtml(f.basis)}</p>`,
363
+ )
364
+ .join('')
365
+ return renderTemplate('observation', {
366
+ status: report.status,
367
+ version: report.metadata?.recoveryProbeVersion ?? 'unknown',
368
+ node: report.metadata?.runtime?.node ?? 'unknown',
369
+ browser: report.metadata?.browserVersion ?? 'unknown',
370
+ ipcHook: report.metadata?.ipcHook?.version ?? 'unknown / not applicable',
371
+ sampling: samplingHtml,
372
+ detail: report.detail ?? 'region (legacy)',
373
+ phases,
374
+ limitation: report.limitation ?? report.error,
375
+ intervals,
376
+ findings,
377
+ operator: operatorHtml,
378
+ actions: actionsHtml,
379
+ trackedRows,
380
+ notes,
381
+ comparisons,
382
+ cleanup: report.cleanup ?? 'unknown',
383
+ evidence: JSON.stringify(report, null, 2),
384
+ })
351
385
  }
package/src/repair.mjs CHANGED
@@ -336,11 +336,18 @@ export async function prepareRepair(
336
336
  await mkdir(dir, { mode: 0o700 }) // Never overwrite or reuse a run.
337
337
  const baseCommit = (await git(repo, ['rev-parse', 'HEAD'])).trim()
338
338
  const workspace = join(dir, 'checkout')
339
+ // Check out the committed bytes, whatever this machine's Git converts. Git for Windows
340
+ // writes CRLF by default, and a model's multi-line edit with LF breaks would not match.
341
+ // Explicit eol attributes in .gitattributes still apply.
339
342
  await git(dir, [
340
343
  'clone',
341
344
  '--local',
342
345
  '--no-hardlinks',
343
346
  '--no-checkout',
347
+ '--config',
348
+ 'core.autocrlf=false',
349
+ '--config',
350
+ 'core.eol=lf',
344
351
  '--',
345
352
  repo,
346
353
  workspace,
@@ -349,7 +356,18 @@ export async function prepareRepair(
349
356
  await git(workspace, ['remote', 'remove', 'origin']) // No accidental push target.
350
357
  const sources = []
351
358
  for (const path of config.sourceFiles) {
352
- await git(workspace, ['ls-files', '--error-unmatch', '--', path])
359
+ // The index says what a path is. Without symlink rights, Git for Windows checks out
360
+ // a committed symlink as a plain file holding the link text.
361
+ const [mode] = (
362
+ await git(workspace, [
363
+ 'ls-files',
364
+ '--stage',
365
+ '--error-unmatch',
366
+ '--',
367
+ path,
368
+ ])
369
+ ).split(' ')
370
+ if (mode === '120000') throw new Error('Source symlinks are not allowed')
353
371
  const content = await sourceFile(workspace, path)
354
372
  sources.push({ path, content, sha256: sha(content) })
355
373
  }
@@ -0,0 +1,33 @@
1
+ import { readFileSync } from 'node:fs'
2
+
3
+ const entities = {
4
+ '&': '&amp;',
5
+ '<': '&lt;',
6
+ '>': '&gt;',
7
+ '"': '&quot;',
8
+ "'": '&#39;',
9
+ }
10
+
11
+ /** Escapes text for HTML element content or a quoted attribute value. */
12
+ export const escapeHtml = (value) =>
13
+ String(value ?? '').replace(/[&<>"']/g, (c) => entities[c])
14
+
15
+ /**
16
+ * Fills `report-templates/<name>.html`. `{{slot}}` takes text and escapes it;
17
+ * `{{{slot}}}` takes markup the caller already built with escaped values.
18
+ */
19
+ export function renderTemplate(name, slots) {
20
+ const template = readFileSync(
21
+ new URL(`./report-templates/${name}.html`, import.meta.url),
22
+ 'utf8',
23
+ )
24
+ return template.replace(
25
+ /\{\{\{(\w+)\}\}\}|\{\{(\w+)\}\}/g,
26
+ (_, markupSlot, textSlot) => {
27
+ const slot = markupSlot ?? textSlot
28
+ if (!Object.hasOwn(slots, slot))
29
+ throw new Error(`Report template ${name} needs a value for ${slot}`)
30
+ return markupSlot ? String(slots[slot]) : escapeHtml(slots[slot])
31
+ },
32
+ )
33
+ }
@@ -0,0 +1,134 @@
1
+ <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8" />
5
+ <meta name="viewport" content="width=device-width, initial-scale=1" />
6
+ <title>Recovery Probe — {{target}} report</title>
7
+ <style>
8
+ body {
9
+ font: 16px/1.6 system-ui, sans-serif;
10
+ background: #f6f7f9;
11
+ color: #17212f;
12
+ margin: 0;
13
+ padding: 40px;
14
+ }
15
+ main {
16
+ max-width: 1100px;
17
+ margin: auto;
18
+ }
19
+ h1 {
20
+ font-size: 32px;
21
+ letter-spacing: -1px;
22
+ }
23
+ code {
24
+ overflow-wrap: anywhere;
25
+ }
26
+ table {
27
+ width: 100%;
28
+ border-collapse: collapse;
29
+ background: white;
30
+ font-size: 14px;
31
+ }
32
+ th,
33
+ td {
34
+ text-align: left;
35
+ border-bottom: 1px solid #ddd;
36
+ padding: 12px;
37
+ }
38
+ section {
39
+ background: white;
40
+ padding: 20px;
41
+ margin: 24px 0;
42
+ border-radius: 10px;
43
+ }
44
+ .pass {
45
+ color: #087443;
46
+ }
47
+ .fail {
48
+ color: #b3261e;
49
+ }
50
+ .inconclusive,
51
+ .skipped {
52
+ color: #755300;
53
+ }
54
+ img {
55
+ max-width: 100%;
56
+ border: 1px solid #ddd;
57
+ }
58
+ figcaption {
59
+ font-weight: 600;
60
+ }
61
+ small {
62
+ color: #536174;
63
+ }
64
+ .scroll {
65
+ overflow: auto;
66
+ }
67
+ </style>
68
+ </head>
69
+ <body>
70
+ <main>
71
+ <p>RECOVERY PROBE · {{targetLabel}} PREVIEW</p>
72
+ <h1>{{heading}}</h1>
73
+ <p><code>GET {{endpoint}}</code></p>
74
+ <p>
75
+ Recovery: <strong>{{recovery}}</strong> · Deadline: {{timeoutMs}} ms ·
76
+ {{generatedAt}}
77
+ </p>
78
+ <div class="scroll">
79
+ <table>
80
+ <thead>
81
+ <tr>
82
+ <th>Injected fault</th>
83
+ <th>Outcome</th>
84
+ <th>Result</th>
85
+ <th>Injected</th>
86
+ <th>Later successful responses</th>
87
+ <th>Retry clicked</th>
88
+ </tr>
89
+ </thead>
90
+ <tbody>
91
+ {{{results}}}
92
+ </tbody>
93
+ </table>
94
+ </div>
95
+ <section>
96
+ <h2>How to read this</h2>
97
+ <p>
98
+ <strong>PASS:</strong> the fault occurred, the configured action
99
+ completed, a later matching request succeeded, and the selected
100
+ healthy UI marker appeared. <strong>FAIL:</strong> after the fault
101
+ and action, recovery was not observed by the deadline. Investigate
102
+ the app and the chosen assertion. <strong>INCONCLUSIVE:</strong>
103
+ setup, baseline, interception or the recovery action could not be
104
+ verified; this is not a confirmed app bug.
105
+ </p>
106
+ <p>
107
+ Final reload reset: <strong>{{finalReset}}</strong>. A reload that
108
+ rescues a stuck page does not turn the original failed recovery check
109
+ into a pass.
110
+ </p>
111
+ <p>{{limitation}}</p>
112
+ </section>
113
+ <section>
114
+ <h2>Reproduce and investigate</h2>
115
+ <p>
116
+ From this output directory, run <code>node reproduce.mjs</code>
117
+ {{reproduceWhere}}. See <code>scenario.json</code> and
118
+ <code>fix-brief.md</code>.
119
+ </p>
120
+ <p>
121
+ No source patch was applied. No data was uploaded. Review local
122
+ artifacts before sharing them.
123
+ </p>
124
+ </section>
125
+ {{{screenshots}}}
126
+ <small>
127
+ HTTP status success and a visible marker do not certify response
128
+ semantics. Choose a marker that only appears when this endpoint's data
129
+ has actually loaded; configure readyText/busySelector for a stronger
130
+ assertion.
131
+ </small>
132
+ </main>
133
+ </body>
134
+ </html>
@@ -0,0 +1,46 @@
1
+ <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8" />
5
+ <title>Recovery Probe discovery</title>
6
+ <style>
7
+ body {
8
+ font: 16px system-ui;
9
+ max-width: 1100px;
10
+ margin: 40px auto;
11
+ padding: 20px;
12
+ }
13
+ td,
14
+ th {
15
+ padding: 10px;
16
+ border-bottom: 1px solid #ddd;
17
+ text-align: left;
18
+ }
19
+ pre {
20
+ white-space: pre-wrap;
21
+ overflow-wrap: anywhere;
22
+ }
23
+ </style>
24
+ </head>
25
+ <body>
26
+ <h1>Recovery Probe discovery</h1>
27
+ <p>{{limitation}}</p>
28
+ <table>
29
+ <tr>
30
+ <th>Experiment</th>
31
+ <th>Interpretation</th>
32
+ <th>Kind</th>
33
+ <th>Confidence</th>
34
+ <th>Runs</th>
35
+ <th>Partial observation</th>
36
+ </tr>
37
+ {{{experiments}}}
38
+ </table>
39
+ <h2>Run summary</h2>
40
+ <pre>{{summary}}</pre>
41
+ <details>
42
+ <summary>Observations and bounded timelines</summary>
43
+ <pre>{{evidence}}</pre>
44
+ </details>
45
+ </body>
46
+ </html>
@@ -0,0 +1,30 @@
1
+ <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8" />
5
+ <title>Recovery Probe IPC report</title>
6
+ <style>
7
+ body {
8
+ max-width: 1000px;
9
+ margin: 40px auto;
10
+ font: 16px system-ui;
11
+ padding: 20px;
12
+ }
13
+ pre {
14
+ white-space: pre-wrap;
15
+ overflow-wrap: anywhere;
16
+ background: #f4f5f7;
17
+ padding: 20px;
18
+ }
19
+ </style>
20
+ </head>
21
+ <body>
22
+ <h1>{{heading}}</h1>
23
+ <p>{{limitation}}</p>
24
+ <p>
25
+ Final reload: {{finalReset}}. Cleanup cannot turn an earlier failure into
26
+ a pass.
27
+ </p>
28
+ <pre>{{evidence}}</pre>
29
+ </body>
30
+ </html>
@@ -0,0 +1,74 @@
1
+ <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8" />
5
+ <title>Recovery Probe observation</title>
6
+ <style>
7
+ body {
8
+ font: 16px system-ui;
9
+ max-width: 1000px;
10
+ margin: 40px auto;
11
+ padding: 20px;
12
+ }
13
+ td,
14
+ th {
15
+ text-align: left;
16
+ padding: 8px;
17
+ }
18
+ pre {
19
+ white-space: pre-wrap;
20
+ overflow-wrap: anywhere;
21
+ }
22
+ </style>
23
+ </head>
24
+ <body>
25
+ <h1>Fault-free observation</h1>
26
+ <p>
27
+ Capture: {{status}}. This is not a recovery or account-isolation verdict.
28
+ </p>
29
+ <p>
30
+ Recovery Probe {{version}} · Node {{node}} · Browser {{browser}} · IPC
31
+ hook {{ipcHook}}
32
+ </p>
33
+ {{{sampling}}}
34
+ <p>
35
+ Sampling detail: {{detail}}. Marker counts describe the configured
36
+ content selector; generic region item counts are heuristics.
37
+ </p>
38
+ {{{phases}}}
39
+ <p>{{limitation}}</p>
40
+ <table>
41
+ <thead>
42
+ <tr>
43
+ <th>Interval</th>
44
+ <th>Reads started</th>
45
+ <th>Successes</th>
46
+ <th>Errors</th>
47
+ <th>First content sample bounds</th>
48
+ </tr>
49
+ </thead>
50
+ <tbody>
51
+ {{{intervals}}}
52
+ </tbody>
53
+ </table>
54
+ {{{findings}}}
55
+ {{{operator}}}
56
+ {{{actions}}}
57
+ {{{trackedRows}}}
58
+ {{{notes}}}
59
+ {{{comparisons}}}
60
+ <p>
61
+ Comparisons concern sampled marker text and count, not ownership or
62
+ freshness. A completion does not guarantee the UI has consumed its
63
+ result.
64
+ </p>
65
+ <p>
66
+ Cleanup: {{cleanup}}. The observation snapshot is from before cleanup;
67
+ cleanupEvidence records the separate cleanup check when available.
68
+ </p>
69
+ <details>
70
+ <summary>Evidence and coverage</summary>
71
+ <pre>{{evidence}}</pre>
72
+ </details>
73
+ </body>
74
+ </html>
@@ -0,0 +1,28 @@
1
+ <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8" />
5
+ <title>Recovery Probe refresh</title>
6
+ <style>
7
+ body {
8
+ font: 16px system-ui;
9
+ max-width: 1000px;
10
+ margin: 40px auto;
11
+ padding: 20px;
12
+ }
13
+ pre {
14
+ white-space: pre-wrap;
15
+ overflow-wrap: anywhere;
16
+ }
17
+ </style>
18
+ </head>
19
+ <body>
20
+ <h1>Refresh experiment</h1>
21
+ <p>{{limitation}}</p>
22
+ {{{results}}}
23
+ <details>
24
+ <summary>Evidence</summary>
25
+ <pre>{{evidence}}</pre>
26
+ </details>
27
+ </body>
28
+ </html>