http_get called urllib.request.urlopen with no explicit SSL context, so
it fell back to whatever trust store the interpreter shipped with. On
python.org's Python 3.11 for macOS the CA keychain is empty until the
user runs Install Certificates.command, which nobody does — every HTTPS
call raised CERTIFICATE_VERIFY_FAILED. Agents were then writing
domain-skills that suggested subprocess curl or, worse, disabling TLS
verification entirely.
Load certifi's bundle explicitly via ssl.create_default_context. Same
approach requests and httpx use, works across macOS/Linux/Windows and
across Python installers.
* Reframe post-task ritual as the default "contribute back" procedure
Renames the section to "Always contribute back" and turns the guidance
into an imperative default. Adds concrete examples of what's worth a PR
(private APIs, framework quirks, stable selectors, URL patterns, waits,
traps), a schema for what a domain skill should capture (the durable
shape of the site, not the run narration), and an explicit do-not-write
list — most importantly banning raw pixel coordinates.
Narrows the scope to `domain-skills/` contributions only; no longer
nudges agents to update `interaction-skills/` or `helpers.py` as part
of this loop.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Replace "hesitate" with clearer cost framing
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix daemon attaching to invisible omnibox popup on fresh Chrome
When Chrome opens fresh, the only page targets are chrome://
internal pages and the omnibox popup (1px invisible viewport).
The daemon's attach_first_page() fell back to the popup, making
all subsequent work invisible to the user.
Fix: when no real pages exist, create an about:blank tab via
Target.createTarget instead of attaching to the omnibox popup.
Tested configurations:
- Fresh start with no real tabs → creates about:blank (1112x817)
- Navigate without AppleScript → works, tab visible
- Recovery from stale socket → auto-reconnects
- Chrome restart from scratch → creates about:blank
Also adds interaction-skills/connection.md documenting the
omnibox popup problem and startup sequence.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* add connection skill reference to main SKILL.md
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
CoinMarketCap (internal data-api/v3 fully open, no auth; 25 calls no rate limit),
Quora (full Chrome UA required; push() payloads double-encoded JSON; 3 SSR answers only),
Itch.io (http_get works; game cards via CSS selectors; RSS feeds exist),
Steam (appdetails single appid only; price in cents; ISteamApps/GetAppList dead in 2026),
HowLongToBeat (two-step token flow /api/find/init then POST; comp_* in seconds not hours).
Letterboxd (http_get works on film pages; JSON-LD CDATA gotcha; API needs OAuth),
Gutenberg (Gutendex REST API; text via /cache/epub/; .opf is 404, use .rdf),
Metacritic (internal backend API key in HTML; Nuxt __NUXT_DATA__ not __NEXT_DATA__),
RAWG (API needs key; window.CLIENT_PARAMS in HTML has full game data without key),
OpenLibrary (full free API; missing cover = 43-byte GIF not 404; description dual type).
Glassdoor (Cloudflare managed challenge; browser only; __NEXT_DATA__ + DOM fallbacks),
Medium (?format=json strips XSSI prefix; GraphQL /_/graphql no auth; RSS 10-item cap),
SoundCloud (oEmbed no-auth; __sc_hydration apiClient.id as client_id; API v2 with pagination),
Genius (OS token in UA bypasses 403; internal /api/songs no auth; strip first lyrics div header),
Dev.to (public REST API; burst limit 6 req then 429/1s; listings empty without auth).
Coursera (public API no auth, q=search is POST-only/405 on GET),
DuckDuckGo (Instant Answer API, skip_disambig=1 essential, widget answers unusable),
SEC EDGAR (company UA required for www.sec.gov; 10 req/s; XBRL frames for cross-company),
TrustPilot (http_get works; __NEXT_DATA__ has reviews; 10-page cap per filter).
Ran actual browser-harness sessions against each site and rewrote
the skill files from live test findings. Key corrections:
GitHub: wait(2) after wait_for_load() for React hydration; search API
separate 10 req/min limit; search/code needs auth (401 unauthed)
HackerNews: athing also matches comment rows (use 'athing submission');
job posts break naive score-zip; html.unescape() required for titles
Amazon: .zg-item-immersion gone from Best Sellers; #priceblock_ourprice
returns null (legacy); review count selector collides with cross-sell
widget — use [aria-label*='ratings'] instead
News: The Verge is Atom not RSS (namespace dict required); Reuters
hard-blocks http_get with 403 even with User-Agent; BBC shows no
consent banner from US IP; parallel fetch is 4.3x faster (0.16s vs 0.70s)
ProductHunt: goto() ERR_ABORTED — always use new_tab(); /posts/ URLs
don't exist (it's /products/); homepage has 30 fixed items no lazy load;
[data-test^='post-item-'] is the correct card selector
Captures three learnings from a real run that mis-reported $3,080 AUD for
a UACC-Rack-12U-Wall (real AU street ~$420-630): Tech Geeks is Shopify so
use /products/<handle>.js for canonical price and SKU; its .js `available`
flag is unreliable so cross-check the DOM for sold-out markers; and
sold-out pages there carry stale/junk prices that must never enter a
final table without a second-source sanity check.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A bare `bh` invocation blocks forever on sys.stdin.read(). Detect a TTY
stdin and exit with a one-line usage hint instead, so the failure mode
is obvious.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Saurav Panda