| Age | Commit message (Collapse) | Author |
|---|
|
The audit fixes were required to add no comments; two doc blocks slipped through
review. Removed, leaving only comments that predate this task.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
runOnElement checked only whether Runtime.callFunctionOn reached Chrome, not
whether the function it ran threw. A page-side exception left result.value
undefined, the result guard rejected it, and browser_fill, browser_focus and
browser_select_option all reported "No element matched: <ref>" for an element
that was there — sending the agent to re-snapshot a ref that was fine and
discarding the real error. The throw is now surfaced as cdp_error carrying the
exception text, which is distinguishable from both the no-match and stale-ref
paths; those still fire for a genuinely missing or stale element. The selector
path already reported throws correctly via evaluateJs.
Report: docs/superpowers/reports/2026-07-26-bug-audit.md#f2
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
browser_fill and browser_fill_form carried separate in-page implementations that
disagreed on the same element. The worst case was silent: browser_fill on a
checkbox fired the HTMLInputElement value setter, which writes the submit value
and leaves `checked` alone, then reported status ok — a no-op the agent had no
reason to doubt. Both tools now share one engine, with browser_fill's documented
<select> rejection kept as an explicit pre-check and per-tool focus behaviour kept
as a parameter.
Deliberate, user-approved behaviour changes:
- browser_fill on a DISABLED input now returns an error instead of success. The
old success was the bug; browser_fill_form already refused it.
- browser_fill on a checkbox/radio now toggles `checked` instead of the value.
- browser_fill_form on a contenteditable now uses execCommand("insertText"),
so it drives ProseMirror/Slack-style editors as browser_fill already did.
browser_fill keeps appending interactiveDiff and republishing refs; fill_form
still does not.
Report: docs/superpowers/reports/2026-07-26-bug-audit.md#f5
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
interactiveDiff re-assigned refs by tree position after every fill, click and
select, so any element appearing or disappearing shifted the numbering and a ref
the agent was still holding silently resolved to a different element — while the
tools kept promising that refs survive re-renders. Refs are now carried over by
backendNodeId and only genuinely new elements draw a new number.
Report: docs/superpowers/reports/2026-07-26-bug-audit.md#f6
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Types the event-queue iterator's TReturn as undefined so the terminal
result needs no cast; declares the isError field the pi runtime reads by
duck-typing; binds socket to a local const in the daemon transport;
replaces the @ts-ignore on the optional sharp import with a non-literal
specifier so src/ still typechecks where sharp was not installed.
Adds src/util/guards.ts as the single home for narrowing over unknown —
isRecord plus scalar/array readers — and converts the two private
isObject copies in cdp/types.ts and daemon/protocol.ts along with every
inline `as Record<string, unknown>` in domains/js.ts and domains/render.ts.
Gives Accessibility.getFullAXTree a real result schema, closing the last
unvalidated CDP boundary: only nodeId is required, every object carries
additionalProperties, and each AX value stays Unknown, so the tree
decodes as a whole or fails as a whole rather than being filtered.
Guards the five tool-result `details` readers, the daemon attach
response, the persisted session state, the browser-namespace flag, and
the fs errno narrowing, each feeding an already-present failure path.
The daemon-transport socket binding is not a bug fix: send() runs
synchronously inside sendWithTimeout's promise executor, so socket cannot
be reassigned between the check and the write. It removes the assertion
and makes the write immune to a later refactor that adds an await.
Two assertions survive by design and are documented in the task report:
the never->unknown re-widening in util/tool.ts, which the type-erased
tool registry makes unavoidable, and the AsyncFunction constructor lookup
in domains/js.ts, which JavaScript exposes no typed binding for.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Ten sites cast or implicitly typed JSON.parse output; JSON.parse("null")
would have satisfied all of them. parseJson returns a Result and validates
against a typebox schema. js.ts keeps unknown — user script output has no
fixed shape. daemon/bridge.ts assigned JSON.parse straight to CdpRawMessage
with no guard at all; it now uses isCdpRawMessage like the transport does.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Three review fixes on the restored fill_form:
- FILL_FN matches a <select> option by value, label or text but returns
el.value, so filling by visible label ("Canada" -> value "ca") tripped
the read-back mismatch check and reported a correctly-set field as
"Filled 0/1 fields; issues: ref e4 (value mismatch)". The mismatch test
now also accepts res.text, which FILL_FN sets only for selects, so
mismatch detection is preserved for every other field kind.
- fill_form always returned ok(), so total failure surfaced as a
successful result reading "Filled 0/4 fields". It now returns
invalid_state when okCount is 0, carrying the per-field issue list as
the error message. Partial success stays ok().
- Removed the two comment blocks this task had added to
test/manual/forms-test.ts.
Neither behaviour fix has automated coverage; the manual suite that would
exercise them still needs a live Chrome and was not run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
domains/forms.ts was a stale fork of form.ts imported only by a manual
test. It held two working tools the registry never registered: fill_form
and set_checked. Ported both onto the current pattern and deleted the
fork; form.ts versions of the shared tools win.
The fork's tools took ref as Type.Integer (a raw CDP backendNodeId).
browser_snapshot has since moved to string refs ('e7') resolved through
the session ref map, so as written the tools were not callable at all.
Ported them onto resolveRefToObjectId and Type.String refs; the only
user-facing strings that changed are the four that documented the old
[ref=N] syntax. All page-side JS is carried byte-for-byte.
form.ts exceeded 250 lines, so fillForm moved to form-batch.ts per the
single-field/batch split. form.ts is still 425 lines — the pre-existing
three tools account for that, so the split does not reach the target.
test/manual/forms-test.ts is repointed and adapted to publish an 'eN'
ref map, but NOT executed — it needs a live Chrome. Type-checked only.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
The ref-vs-x/y branch turned out to exist in one tool, not four; it is now
one helper with tests covering the half-specified-coordinate and stale-ref
cases. cdpCall folds the repeated cdp_error mapping at 24 active-session
call sites; target-scoped callOnTarget/callBrowser sites are left alone.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
serialized: true was applied in registry.ts, away from the tools it
describes. Now a required concurrency field on each definition, with a
test pinning the serialized set so the classes cannot drift.
The lock check defaults to taking the mutex for any class that is not
explicitly "parallel", so a definition that bypasses defineBrowserTool at
runtime fails safe rather than racing on shared CDP state.
The test script gains --conditions=import: pi-coding-agent publishes no
"require" condition, so importing any tool definition from a CJS-loaded
test could not resolve it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Adopting isFocusProbe inverted the safety default: the guard was ANDed into
the condition gating the check, so a probe result the guard rejected skipped
the check and fell through to Input.insertText — text typed into nothing,
silently, which is the exact failure the block exists to prevent. The guard
now feeds a value that defaults to { ok: false }, so a rejected shape
produces the error instead of bypassing it.
The other four guard adoptions (isFillResult, isFocusResult, isSelectResult,
isPageResult) were already fail-closed: each feeds a value whose rejection
lands on a pre-existing error branch.
Also assert x/y in the box centre test — the fields snapshot's off-viewport
filter reads — and wrap the zero-size message to the surrounding width.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
ref-resolve.ts and snapshot.ts had the same box-model unwrap, both using
data.model! to work around an untyped response. Extracted to domains/box.ts;
the typed schema makes the assertions unnecessary. Both derived the centre
identically (quad corner + half the model width/height), so centreOf keeps
that derivation; the shared zero-size rejection moved into boxOf, and
snapshot keeps its own off-viewport (x/y < 0) filter.
evalInIsolatedTab now goes through evaluateJson with the tab's own sessionId
and a caller-supplied guard; both callers evaluate expressions ending in
JSON.stringify, so they pass isJsonText rather than a check that checks
nothing. Runtime.evaluate/callFunctionOn result casts in form.ts, forms.ts
and keyboard.ts are replaced by named guards over the in-page result shapes.
Runtime.callFunctionOn's params schema gained arguments/returnByValue/
awaitPromise/userGesture: four domain call sites legitimately pass them and
typebox's additionalProperties:true does not widen the static type, so the
params type was rejecting valid CDP fields. Params schemas are type-only
(only results are compiled), so this is a compile-time widening with no
runtime effect.
tsc is green again for both tsconfigs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Target.attachToTarget was decoded at 5 sites, Browser.getWindowForTarget
at 5, and Runtime.evaluate unwrapped at another. Each is now one typed
helper. Deletes the dead cdp/window-binding.ts.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
* feat: pin the agent to a chosen browser profile (#15)
Harness tabs were created with a bare Target.createTarget, which places
them in Chrome's defaultBrowserContextId — and that value follows window
focus. Which profile the agent acted as therefore depended on which
browser window the user last clicked, so the same task could run as a
work account on one run and a personal account on the next.
/browser-profile lists the connected browser's profiles as "Name (email)"
and pins one to ~/.pi/agent/browser-harness.json, so the choice outlives
the session. The same picker appears on first browser_setup when nothing
is pinned; with no interactive UI, or on cancel, setup continues with a
note and the previous behavior.
Chrome offers no direct route to another profile: Target.createTarget
rejects a foreign browserContextId, and openerId does not inherit the
opener's context. The window is opened through the browser's own command
line (--profile-directory, handed to the running browser by Chromium's
ProcessSingleton) and identified by a unique file:// sentinel page, since
a bare about:blank#token argument is discarded. Further tabs come from
window.open evaluated with userGesture: true — the only CDP-reachable way
to place a tab in a non-default browser context — correlated by a unique
token so concurrent isolated tabs stay distinct. A failed handshake is
reported, never silently redirected to another profile.
All four tab-creation sites now funnel through cdp/target-factory, so
open_urls and the isolated tabs behind web_search / read_page cannot run
with a different profile's cookies than the visible tabs.
Also fixed along the way: snap Chromium is discovered, Windows uses
%LOCALAPPDATA%, browsers launched with an explicit --user-data-dir are
found, the running browser is chosen by ranking candidates against the
connected user-data-dir rather than taking the first process found, and
an installed-but-closed browser no longer reads as running.
CI runs the profile suites and a real-browser end-to-end test on ubuntu,
macOS, and Windows — profile discovery is the one part of the harness
whose behavior genuinely differs per OS.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: single-flight harness window creation
Isolated tabs run concurrently (browser_read_page is registered
unserialized), so several could each find "no harness window" and create
one — two blank windows before this feature, two LAUNCHED profile windows
after it. ensureHarnessWindow now shares one in-flight creation per
client.
Sharing introduces a second hazard: freshlyCreated is a claim on the seed
tab, and if every waiter received it, they would all drive the same tab.
Only the caller that started the creation keeps the claim; the rest open
their own tab beside it.
Covered by test/manual/profile-concurrency-test.ts (five concurrent
isolated tabs, pinned and unpinned: distinct targets, one window, correct
profile), wired into CI on Linux only — the coordination is pure JS and
identical across platforms.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: tilde expansion differs by purpose on Windows
CI on windows-latest caught this: join("~", "x") yields "~\x" there, and
expandTilde only handled "~/".
The two uses are not the same, so they no longer share a helper. The
agent dir stays byte-compatible with pi's own expandTildePath ("~/" only,
every platform) — being more lenient would put the pin file somewhere pi
never looks. $CHROME_USER_DATA_DIR is ours to interpret, so it accepts
"~\" as well, which is what a Windows user would type.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
- Add forms domain with a universal field setter that auto-detects element
type and fires framework-compatible input events (React native-setter trick)
- Add window-binding helper to capture and bind the real Chrome windowId as the
session's durable window identity on every new-window code path
- Wire supporting changes across CDP session/ownership, client, navigate,
isolated-tab, prompt, and state
- Add manual tests for forms and window binding
- Bump version 0.8.3 -> 0.9.0
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
Both tools dumped full output into the TUI. Add a shared renderExpandableText
helper (src/domains/render.ts) mirroring browser_execute_js: a compact preview
by default, full body on expand via keyHint("app.tools.expand"). Tools stash a
{summary, body, fullOutputPath} render payload in details; the full text still
goes to the LLM via ToolOk.text. Also surfaces engine + result count in the
web_search summary line (closes the S4 observability gap).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
URL (or owned targetId) → clean main-article text with nav/ads/footer
boilerplate stripped, via a dependency-free readability heuristic. DOM walk
lives in an in-page capture expression; scoring/selection is a pure, fixture-
tested function (link-density + boilerplate-ancestor filtering, body-text
fallback for structure-less pages). Reuses the isolated-attach lifecycle;
registered unserialized (safe concurrent read). 10 extractor unit scenarios.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
Thin query→ranked-links primitive. Isolated-attach tab lifecycle mirroring
browser_open_urls (never touches the user's current tab); pure Google SERP
parser with semantic (a h3) extraction and redirect unwrapping; CAPTCHA and
no-results surfaced as invalid_state + details.reason. Fixture-based parser
tests (9 scenarios). Establishes the isolated-tab lifecycle and tsx+fixture
test convention that browser_read_page and deep-research build on.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
IIFE wrap (#9)
- browser_dispatch_key now sets keyCode/which on the synthesized KeyboardEvent
(both ref and selector paths) so handlers checking e.keyCode === 13 fire.
- browser_execute_js IIFE auto-wrap triggers only when the source starts with a
return statement, not on any substring "return " (which silently produced
undefined for bare expressions).
- browser_fill guidelines point at dispatch_key for tag-input submission and the
snapshot+click recipe for custom dropdowns.
Bump to 0.8.1.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
|
|
Interaction tools could only target elements by raw viewport coordinates
(browser_click) or model-guessed CSS selectors (browser_fill/focus/
select_option/upload). Neither survives a re-render: on React/SPA forms,
every save or edit reflows the page, so captured @(x,y) coordinates go
stale and guessed selectors miss — leaving the agent to fall back to
hand-rolled execute_js DOM walks and unnoticed failed saves.
Add a stable ref system, mirroring Playwright MCP / browser-use:
- snapshot assigns each interactive element a ref (e1, e2, …), shown as
[eN] in the outline, and publishes a per-tab ref → backendNodeId map
- interaction tools accept an optional `ref` that re-resolves the element
at call time (via DOM.resolveNode / getBoxModel), so it stays valid
across re-renders; coordinates/selectors remain as fallbacks
- mutating tools append a compact diff of changed / newly-appeared (*) /
removed interactive elements, so silent failed saves surface immediately
- stale refs return an actionable "re-snapshot" error instead of a miss
Ref map is replaced per snapshot (bounded memory); fill keeps the native
_valueTracker write path so React controlled inputs still update.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
|
|
* feat: Unix socket daemon with auto-reconnect for persistent Chrome CDP
Eliminates the 'Allow Remote Debugging' Chrome prompt that appeared on
every pi session by multiplexing pi instances through a single persistent
daemon process.
Key additions:
- Unix socket IPC (net.createServer + readline JSON framing)
- CdpBridge with ID multiplexer + event router
- DaemonTransport implementing CdpTransport interface
- Auto-spawn with ensureDaemon() (lazy, checks socket)
Fixes since initial implementation:
- Bridge auto-reconnects to Chrome with exponential backoff (1s-60s).
Previously the bridge tried once at startup and permanently failed
if Chrome wasn't ready — causing 'Chrome not connected' on every
browser tool call.
- Daemon spawn uses project-local tsx binary instead of global --import
flag, fixing silent spawn failure (ERR_MODULE_NOT_FOUND).
- Stale ctx crash: persistState in onOwnershipChange wrapped in
try/catch to prevent CDP event consumer crashes.
49 test assertions across 4 test files. Zero TypeScript errors.
* feat: on-demand browser initialization
Browser daemon no longer auto-starts on session_start. Instead:
- session_start creates the client lazily but does NOT connect or spawn
the daemon. Zero Chrome prompts for non-browser pi sessions.
- Browser tool wrapper checks for daemon socket before each call.
Socket missing → clear error: 'Run /browser-setup first'.
Socket exists → proceeds silently (daemon already running).
- /browser-setup is the explicit, one-time init point. It spawns
the daemon and connects to Chrome only when the user opts in.
- SKILL.md updated with Connection Setup instructions telling agents
to direct users to /browser-setup when socket is missing.
* fix: checkChromeRunning on macOS — match by substring not exact
ps -A -o comm= returns full executable paths on modern macOS
(e.g. /Applications/Google Chrome.app/Contents/MacOS/Google Chrome),
not the basename. The old exact-match check (browserProcesses.includes(l))
never matched, causing false 'No browser instance running' errors.
Changed to substring includes() with helper/renderer/crashpad filtering.
* feat: browser_setup as agent-callable tool
The agent can now self-recover from 'Browser harness not initialized'
errors by calling browser_setup directly — no user mediation needed.
- Extracted shared performSetup() from the slash command into setup.ts
- Created src/domains/setup.ts with ensureAlive:false to bypass the
socket guard (this tool IS the setup, can't require itself)
- Registered browser_setup in the tool registry
- Updated SKILL.md: agent calls browser_setup directly, not ask user
- Removed stale test/manual/spawn-test.ts
* release: v0.6.0 — per-tab isolation, Unix socket daemon, on-demand init
|
|
|
|
prompts
Adds two new tools and pivots the agent's default workflow away from
screenshots toward structured page data.
browser_snapshot — CDP accessibility tree (roles, names, states, hierarchy)
with click coordinates @(x,y) attached to every interactive element via
DOM.getBoxModel (parallel, 1.5s aggregate budget). Pass directly to
browser_click — no screenshot round-trip. Optional includeScreenshot:true
for visual confirmation. format:"json" returns a slim structure with full
box per node.
browser_network_requests — replaces the deprecated browser_get_network_log
placeholder. Filters by urlPattern (substring or /regex/), method, status,
resource type, recency. Optional response-body capture under a 5s aggregate
budget with 50KB per-body cap. Buffer is page-scoped (cleared on tab switch),
500 records/tab, with a bufferOverflowed flag on drain.
CDP plumbing: in-process aggregator in src/cdp/network-buffer.ts wired into
the existing single-consumer event loop. Accessibility domain enabled
alongside Page/DOM/Runtime/Network on every attach.
Ctrl+O expand/collapse on three tools' renderResult — browser_snapshot,
browser_network_requests, browser_execute_js. Compact-by-default views with
keyHint("app.tools.expand") that respects user keybinding remaps.
Tool prompts pivoted snapshot-first: browser_screenshot is documented as
LAST RESORT for visual verification only; browser_snapshot is the default
for understanding pages; browser_execute_js is the default for surgical
reads. browser_click and browser_open_urls guidance follows.
SKILL.md collapsed from ~342 to ~53 lines — keeping only what isn't already
in tool prompts: the decision tree, connection rules, browser_run_script
daemon bindings.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
|
Track which page targets this harness session opened so the agent can never
attach to, switch to, or close a tab the user opened themselves. New tabs
are spawned in a dedicated Chrome window (newWindow:true on first open,
openerId chaining after) so harness activity stays visually separate.
- new OwnershipRegistry (src/cdp/ownership.ts) tracks owned targetIds and
the harness window seed; mutations notify a single listener that persists
to BrowserState (ownedTargetIds, harnessWindowTargetId)
- attachFirstPage prefers an already-owned live target; otherwise opens a
fresh harness-owned tab in its own window instead of grabbing whatever
page the user is on
- session subscribes Target.setDiscoverTargets and reaps targetDestroyed
events so the registry stays in sync with reality
- listTabs returns owned:boolean per tab and reconciles dead ids
- browser_list_tabs defaults to scope:"owned"; pass scope:"all" to inspect
the user's other tabs (read-only — they cannot be switched to or closed)
- browser_switch_tab refuses non-owned tabs, with a clear remediation hint
- browser_new_tab opens inside the harness window via openerId, falling
back to a fresh window when the seed has been closed
- new browser_close_tab tool — owned tabs only, supports unique hex prefix
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
|
The Image text-fallback render does not respect the rendered width,
so a long file path can blow past the terminal width and crash the
host TUI. Wrap the component to truncate each rendered line with
truncateToWidth so it fits the available width.
Co-authored-by: amankumarsingh77 <amankumarsingh7702@gmal.com>
|
|
- Add lightweight async mutex (src/util/mutex.ts) for FIFO serialization
- Add serialized?: boolean to BrowserToolDefinition
- Wire mutex acquisition into registerBrowserTool execute wrapper
- Expose mutationMutex() on BrowserClient
- Mark all mutation tools (click, type, scroll, navigate, switch_tab, etc.) as serialized
- Leave observation tools (screenshot, page_info, execute_js, http_get, etc.) unguarded
- Update prompt and skill docs to reflect parallel-safe behavior
- Fix scroll tool: add Page.bringToFront, explicit button:none/buttons:0/pointerType:mouse
- Update .gitignore to ignore scripts/ and data/ folders
|
|
parallelization guidance to SKILL.md
- browser_screenshot now has renderResult that creates an Image component
for inline display in Kitty/iTerm2/Ghostty/WezTerm terminals
- Falls back to dim text placeholder on terminals without image support
- Updated system prompt table description to mention TUI inline rendering
- Added Parallelization section to SKILL.md teaching the agent to batch
independent operations (browser_open_urls, browser_http_get, etc.)
|
|
matching, download auto-mkdir, session recovery, docs updates
Bug fixes:
- list_tabs: show full 32-char targetIds (was truncated to 8 chars)
- switch_tab: prefix matching fallback for hex prefixes >= 8 chars
- download: auto-create directory with mkdir -p
- ensureAlive: Runtime.evaluate probe to recover from crashed page targets
Docs:
- dispatch_key: clarify DOM-only event behavior
- navigate: warn about Google anti-bot limitation
- network_log: update deprecated promptSnippet
- SKILL.md: document actual daemon API (evaluateJs, pageInfo, session.call)
- SKILL.md: add troubleshooting for sharp/maxDim and Google navigation
- sharp: add to optionalDependencies for screenshot maxDim auto-resize
|
|
timeout + return validation
|
|
(deprecation note)
|
|
|
|
|
|
print_to_pdf
|
|
|
|
|
|
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
|
crosshair overlay
|
|
|
|
|
|
|
|
matched-count validation
|
|
Implements Task 9: adds src/domains/click.ts with the new architecture,
updates registry.ts to register clickTool, removes the legacy browser_click
block from tools.ts, and wires BrowserClient alongside BrowserDaemon in
index.ts so both systems coexist during migration.
Also fixes AnyBrowserToolDefinition variance: uses never-arg renderCall in
the erased type and a boundary cast in registerBrowserTool so concrete tool
definitions (with specific Static<S> args) are assignable to the registry
array without widening.
|
