Bug guide: where to look and how to fix

01All bugs at a glance

02Bug 1: alerts with a link but no tickets

Four different paths in the code produce this, all in or around the alert loop of run_once in poller.py. Fix all four; any one of them alone keeps sending empty alerts.

Cause A: a seat read that finds nothing still alerts. The only "do not send" rule is at line 314: skip when groups were found before and none are new. When the seat reader returns an empty list, either because the show really has no group of GROUP_SIZE free seats or because the page never finished loading (the "Please wait" spinner), the count is zero, the rule does not match, and the generic "Seats opened up" alert goes out with the link and no seat lines. The except at line 308 does the same after any error ("falling back to MVP alert"). This is the most likely source of your blank alerts.

Cause B: "went up" alerts never check the Book button. Lines 247 to 252 queue an alert whenever the availability number rises. Only the "new show" path at line 244 checks r["bookable"], which comes from whether the API says the button opens the seat page or shows a sold-out message. So a show whose number flips while its button still says sold out gets alerted.

Cause C: the availability number is a rough hint. It is the colour BookMyShow paints the showtime (0 sold out, 1 few, 2 filling fast, 3 available). A jump from 0 to 1 can mean one single seat that is booked again by the time you tap. Without the V2 seat read the poller cannot tell.

Cause D: the link can be built with an empty movie code. _bms_url (line 166) takes row["eventCode"] from each theatre block of the answer. When a block has no code, the link becomes /seat-layout//AMBH/113045/20260801, which BookMyShow shows as an empty page.

How to find out which one hit you. Add one log line before notify() and let it run for a day:

log.info("alert sid=%s kind=%s prev=%s curr=%s bookable=%s v2=%s scrape_ok=%s groups=%d new=%d url=%s",
         r["sessionId"], kind, prev, curr, r["bookable"], v2_enabled, scrape_ok,
         pair_count_total, pair_count_new, url)

Then search the log for those lines:

grep "alert sid=" logs/poller.out.log

sudo journalctl -u bms-poller --no-pager | grep "alert sid="

Select-String -Path logs\poller.out.log -Pattern "alert sid="

A line with groups=0 points to cause A, bookable=False to cause B, and url=...seat-layout// to cause D.

The fix. Remember whether the seat read worked, require the Book button, and skip shows that were checked and found empty:

# poller.py, inside `for kind, r, prev, curr in alerts:`
scrape_ok = False
if v2_enabled:
    try:
        pairs = scrape_pairs_sync(...)
        scrape_ok = True
        ...  # the existing group bookkeeping stays
    except Exception as e:
        log.warning("scrape failed for %s: %s", r["sessionId"], e)

if not r["bookable"]:
    log.info("skip %s: not bookable yet", r["sessionId"]); continue
if v2_enabled and scrape_ok and pair_count_total == 0:
    log.info("skip %s: no free group of %d seats", r["sessionId"], group_size); continue
if v2_enabled and not scrape_ok and os.environ.get("ALERT_ON_SCRAPE_FAILURE", "0") != "1":
    log.info("skip %s: scrape failed, unverified", r["sessionId"]); continue
if v2_enabled and pair_count_total > 0 and pair_count_new == 0:
    continue  # the existing rule

And in _bms_url:

event = row["eventCode"] or os.environ["EVENT_CODE"]
if not (event and row["venueCode"] and row["sessionId"]):
    raise ValueError(f"incomplete row, cannot build link: {row}")

For cause C when the V2 reader is off: fetch the showtimes once more 10 seconds later and alert only if the rise is still there. With V2 on, the seat read is the proof, so make V2_SCRAPER_ENABLED=1 the default.

Test before shipping. Write test_poller_alerts.py using pytest. Replace bms.fetch_showtimes with a function that returns the saved answer in dump_showtimes.json (270 shows, 53 sold out), replace scrape_pairs_sync with one that returns [], replace notify with one that only counts calls, and check the count is zero. Then make the seat read return two groups and check one call happened with "Row" in the text. pytest's monkeypatch guide shows how to swap functions like this.

03Bug 2: the BookMyShow token is pasted by hand

The pasted token is not needed for polling at all, and for reading seats it can be replaced by a one-time login. This was tested on 11 October 2026:

• The showtimes API answered 200 with every login header left empty. Only the city headers and a normal browser "User-Agent" matter (a bare curl gets a Cloudflare 403).
• The seat page does need a login. With no cookies the "Please wait" spinner never clears. With the login saved on 25 July it still read 7 rows and 51 seats, eleven weeks later.
• The whole login is one cookie named ud. It holds JSON; its LSID field is both x-access-token and x-lsid, its MEMBERID is x-member-id, MOBILE and MEMBEREMAIL are the phone and email. x-bms-id is a separate cookie bmsId. x-advertiser-id exists nowhere and an empty value is accepted.

Where it lives. Session.from_env in bms.py (lines 47 to 59) crashes unless all seven BMS_* lines exist in .env. build_state.py turns a DevTools dump (../test.txt) into bms_state.json. main() in poller.py (lines 372 to 384) sends the "recapture from DevTools" push after a 401.

Step 1: stop requiring the token for polling. Make every BMS_* setting optional (os.environ.get(k, "")), send a header only when it is non-empty, and delete the seven lines from .env.example. After this the basic poller works on any machine with only a movie code.

Step 2: replace the DevTools dump with a one-time login window. Add login.py. It opens a visible Chrome with its own profile folder, you log in on BookMyShow with phone and OTP as usual, and the profile stays on disk. Anyone can run it with their own account. That is what "works on any account" means in practice.

# login.py  (run once per account, needs a screen)
from pathlib import Path
from playwright.sync_api import sync_playwright
PROFILE = Path.home() / ".bms-poller" / "profile"
with sync_playwright() as p:
    ctx = p.chromium.launch_persistent_context(str(PROFILE), headless=False,
            args=["--disable-blink-features=AutomationControlled"])
    page = ctx.new_page(); page.goto("https://in.bookmyshow.com/explore/home/hyderabad")
    input("Log in (phone + OTP) in the window, then press Enter here... ")
    ctx.storage_state(path="bms_state.json")   # keeps the old file format working too
    ctx.close()

Step 3: build the headers from the saved login instead of .env. A small session.py reads the cookies from bms_state.json and fills the Session object:

import json, urllib.parse
def session_from_storage(path, region_code, region_slug) -> Session:
    cookies = {c["name"]: c["value"] for c in json.load(open(path))["cookies"]}
    ud = json.loads(urllib.parse.unquote(cookies["ud"]))
    return Session(access_token=ud["LSID"], lsid=ud["LSID"], member_id=ud["MEMBERID"],
                   bms_id=cookies.get("bmsId", ""), advertiser_id="",
                   mobile=ud.get("MOBILE", ""), email=ud.get("MEMBEREMAIL", ""),
                   region_code=region_code, region_slug=region_slug)

The seat reader then opens the same profile folder without a window (launch_persistent_context(str(PROFILE), headless=True)) instead of loading a saved file. BookMyShow's own page code keeps the login fresh on every visit.

Step 4: when the login expires, ask for a re-login, not a re-paste. After a 401, or when the seat spinner fails twice in a row, send one push: "BookMyShow login expired, run python login.py". Retry every 5 minutes instead of sleeping an hour. For the Oracle server, which has no screen, run login.py on your own computer and copy the profile folder across. The profile is not tied to one machine. Replace YOUR_SERVER with the server's address:

scp -r ~/.bms-poller/profile ubuntu@YOUR_SERVER:~/.bms-poller/

scp -r ~/.bms-poller/profile ubuntu@YOUR_SERVER:~/.bms-poller/

scp -r "$HOME\.bms-poller\profile" ubuntu@YOUR_SERVER:~/.bms-poller/

Check it worked. python -c "import session; print(session.session_from_storage('bms_state.json','HYD','hyderabad').member_id)" prints your member id. python poller.py with an empty BMS_ACCESS_TOKEN still prints showtimes: ... sessions match filter. python test_v2_flow.py reads seat rows using the profile.

04Bugs 3 to 10

Each row has the smallest fix that closes it. Fix bug 3 first: without it, a poller with a past date keeps failing and nobody is told.

Two smaller things: after a 401 the poller sleeps a whole hour, so a fresh login is ignored for up to 60 minutes (make it 5), and the log file never rotates (1.4 MB of tracebacks in five weeks; add a RotatingFileHandler).

05How to debug, step by step

Reproduce from saved data first, change one thing at a time, and prove the fix with a test before touching the running service. Every bug above can be replayed offline with the files already in the folder.

1. Freeze the evidence. Copy logs/poller.out.log, state.json, notified_pairs.json and the alert email or ntfy message into a folder bug-YYYYMMDD/ before restarting anything.

2. Stop the background service so it does not overwrite the state files while you look:

   macOSLinuxWindows

   Copy

   launchctl unload ~/Library/LaunchAgents/com.sriram.bms-poller.plist

3. Find the alert in the log. The lines just before ALERTING show the showtimes: a/b sessions match filter (c bookable) counts for that cycle.

   macOSLinuxWindows

   Copy

   grep -n "ALERTING\|scrape failed\|skip \|spinner" logs/poller.out.log

   4. Replay the showtimes comparison offline. In a Python shell, load dump_showtimes.json, call _pluck_showtimes from poller.py, and compare with the frozen state.json. Same input always gives the same output, and it takes a second. 5. Replay the seat reading offline (with the virtual environment on, the same on every system). python canvas_parse.py diag/canvas_draw_log.json prints each row as dots and blocks. python adjacency.py diag/canvas_draw_log.json prints the groups. If the picture does not match diag/screenshot.png, the bug is in the seat reader, not the poller. 6. Read one show live, with a window. python seat_scraper.py --diagnose --headed --event ET00505091 --venue AMBH --session 113045 --date 20261012 opens a real browser window and saves a screenshot, the page and the drawing log into diag/. 7. Write the failing test first (pip install pytest): one test per cause, using the saved files as input, red before the fix and green after. 8. Run the real loop for one cycle. Set POLL_INTERVAL_SEC=20, run python poller.py in a terminal, watch two cycles, press Ctrl-C. 9. Ship. Commit on a branch, open a pull request, merge, then start the background service again (Getting started, section 8) and watch the log for ten minutes.

Two habits that save repeat work: keep the alert sid=... log line from bug 1 on every alert, and never test against the live site faster than every 20 seconds, because Cloudflare blocks the whole network after a burst.

06What to read for each kind of bug

One page per kind of bug. The BookMyShow-specific facts above come from this session's own tests; no public documentation of their API exists.