Problem: `setup_problem` only cleared the output buffer when
`old_problem_id ~= problem_id`. If two different contests both had a
problem with the same ID (e.g. `a`), the condition was false and stale
output from the previous contest remained visible.
Solution: clear the output buffer at the top of `proceed()` whenever
`is_new_contest` is true, before any problem setup runs.
Problem: prompts showed the raw platform key (e.g. "codeforces")
instead of the proper display name (e.g. "CodeForces").
Solution: use `PLATFORM_DISPLAY_NAMES` for the username/password
prompts and the logout confirmation message.
Problem: `credentials.lua` cached username/password immediately on
input without any server-side validation. `scraper.lua` only used the
browser/FHS path for the `submit` subcommand.
Solution: rewrite `M.login()` to call `scraper.login()` with ndjson
status callbacks and only cache credentials on success. Generalize the
four submit-only guards in `run_scraper` to a `needs_browser` flag
covering both `submit` and `login` subcommands.
## Problem
After the initial submit hardening, two issues remained: source code was
read in Lua and piped as stdin to the scraper (unnecessary roundtrip
since
the file exists on disk), and CF's `page.fill()` timed out on the hidden
`textarea[name="source"]` because CodeMirror owns the editor state.
## Solution
Pass the source file path as a CLI arg instead — AtCoder calls
`page.set_input_files(file_path)` directly, CF reads it with
`Path(file_path).read_text()`. Fix CF source injection via
`page.evaluate()`
into the CodeMirror instance. Extract `BROWSER_SUBMIT_NAV_TIMEOUT` as a
per-platform `defaultdict` (CF defaults to 2× nav timeout). Save the
buffer
with `vim.cmd.update()` before submitting.
## Problem
AtCoder file upload always wrote a `.cpp` temp file regardless of
language. CF submit used `solve_cloudflare=True` on the submit page,
causing a spurious "No Cloudflare challenge found" error;
`_wait_for_gate_reload` in `login_action` was dead code. Stale cookies
caused silent auth failures with no recovery path. The `uv.spawn` ndjson
path for submit had no overall timeout.
## Solution
Replace AtCoder's temp file with `page.set_input_files` using an
in-memory buffer and correct extension via `_LANGUAGE_ID_EXTENSION`.
Replace CF's temp-file/fallback dance with a direct
`textarea[name="source"]` fill and set `solve_cloudflare=False` on the
submit fetch. Add a login fast-path that skips the homepage check when
cookies exist, with automatic stale-cookie recovery via `_retried` flag
on redirect-to-login detection. Remove `_wait_for_gate_reload`. Fix
`_ensure_browser` to propagate install errors. Add a 120s kill timer to
the ndjson `uv.spawn` submit path in `scraper.lua`.
## Problem
CSES submit was a stub returning "not yet implemented".
## Solution
Authenticate via web login + API token bridge (POST `/login` form, then
POST `/api/login` and confirm the auth page), submit source to
`/api/courses/problemset/submissions` with base64-encoded content, and
poll for verdict. Uses the same username/password credential model as
AtCoder — no browser dependencies needed. Tested end-to-end with a real
CSES account (verdict: `ACCEPTED`).
Also updates `scraper.lua` to pass the full ndjson event object to
`on_status` and handle `credentials` events for future platform use.
## Problem
`setup_problem` explicitly set `swapfile = true` on provisional buffers,
overriding the user's global `noswapfile` setting. The resulting `.swp`
files triggered E325 warnings on subsequent `:e` calls — especially
during the restore path, which redundantly re-opened the current buffer.
## Solution
Remove the `swapfile` override so the user's setting is respected, and
skip the `:e` call in `setup_problem` when the current buffer already
matches the target source file.
## Problem
`_submit_sync` was a 170-line nested closure with `_solve_turnstile` and
the browser-install block further nested inside it. Status events went
to
stderr, which `run_scraper()` silently discards, leaving the user with a
10–30s silent hang after credential entry. The NDJSON spawn path also
lacked stdin support, so submit had no streaming path at all.
## Solution
Extract `_TURNSTILE_JS`, `_solve_turnstile`, `_ensure_browser`, and
`_submit_headless` to module level in `atcoder.py`; status events
(`installing_browser`, `checking_login`, `logging_in`, `submitting`) now
print to stdout as NDJSON. Add stdin pipe support to the NDJSON spawn
path in `scraper.lua` and switch `M.submit` to streaming with an
`on_status` callback. Wire `on_status` in `submit.lua` to fire
`vim.notify` for each phase transition.
## Problem
Credentials lived in a top-level `_credentials` namespace, requiring
special
preservation logic in `clear_all()` and a separate key hierarchy from
the
platform data they belong to.
## Solution
Move credentials from `_credentials.<platform>` to
`<platform>._credentials`.
Migrate v1 caches on load, skip underscore-prefixed keys when
enumerating
contest IDs and summaries, and simplify `clear_all()` now that no
special
preservation is needed.
Stacked on #292.
## Problem
`:CP credentials login/logout/clear` is verbose and inconsistent with
other
actions that are all top-level (`:CP run`, `:CP submit`, etc.). The
clear-all
subcommand is also unnecessary since re-logging in overwrites existing
credentials.
## Solution
Replace `:CP credentials {login,logout,clear}` with `:CP login
[platform]`
and `:CP logout [platform]`. Remove the clear-all command and the
credentials
subcommand dispatch — login/logout are now regular actions routed
through the
standard action dispatcher.
## Problem
The `set` and `clear` subcommands don't clearly convey their intent —
`set`
reads like a generic setter rather than an auth action, and `clear`
overloads
single-platform and all-platform semantics in one subcommand.
## Solution
Rename `set` to `login`, split `clear` into `logout` (per-platform,
defaults
to active) and `clear` (all platforms).
New API:
- `:CP credentials login [platform]` — prompt and save credentials
- `:CP credentials logout [platform]` — remove credentials for one
platform
- `:CP credentials clear` — remove all stored credentials
Problem: closing the test editor left cp://test-N-* buffers alive,
causing E95 on reopen. The nofile buftype also rejected :w, which
was counterintuitive in an editable grid.
Solution: delete all test buffers in toggle_edit teardown. Switch
buftype to acwrite with a BufWriteCmd autocmd that persists test
cases and clears the modified flag. Hoist save_all_tests above
setup_keybindings so the autocmd closure can reference it.
Problem: <c-n>/<c-p> in the I/O view buffers required the cursor
to leave the source file to work, re-ran the solution on each
press, and gave no indication of which test was active. The
workflow is better served by :CP run <n> for a specific test or
:CP panel for full inspection.
Solution: remove navigate_test, next_test_key/prev_test_key config
options, and the associated current_test_index state field.
Problem: vim.json.decode maps JSON null to vim.NIL (userdata), but
cache.set_test_cases validates precision as number|nil, causing a
type error on every scrape where precision is absent.
Solution: guard the precision field when building the callback
table, converting vim.NIL to nil.
Problem: :CP login was a poor API — no way to clear credentials without
raw Lua, and the single command didn't scale to multiple operations.
Solution: replace login with a :CP credentials subcommand following the
same pattern as :CP cache. :CP credentials set [platform] prompts and
saves; :CP credentials clear [platform] removes one or all platforms.
Add cache.clear_credentials(), rename login.lua to credentials.lua,
update parse/dispatch/tab-complete, and rewrite vimdoc accordingly.
Problem: credentials were only set implicitly on first :CP submit.
There was no way to update wrong credentials, log out, or set
credentials ahead of time without editing the cache JSON manually.
Solution: add :CP login [platform] which always prompts for username
and password and overwrites any saved credentials for that platform.
Omitting the platform falls back to the active platform. Wire the
command through constants, parse_command, handle_command, and add
tab-completion (suggests platform names). Document in vimdoc under
the SUBMIT section and in the commands reference.
Problem: credentials were stored in a separate file,
cp-nvim-credentials.json, alongside the main cp-nvim.json cache.
Two files for one plugin's persistent state was unnecessary.
Solution: add get_credentials/set_credentials to cache.lua, storing
credentials under _credentials[platform] in the shared cache. Update
clear_all() to preserve _credentials across cache wipes. Remove the
separate file, load_credentials, and save_credentials from submit.lua.
Problem: luals flagged undefined-field on uv timer methods because
race_state.timer was untyped, and undefined-field on env_extra/stdin
because they were missing from the run_scraper opts annotation.
Solution: hoist race_state.timer into a typed local before the nil
check so luals can narrow through it; add env_extra and stdin to the
opts inline type in run_scraper.
Problem: lua typecheck flagged missing start_time field on ContestSummary;
ty flagged BeautifulSoup Tag/NavigableString union on csrf_input.get(),
a 3-tuple unpack where _extract_problem_info now returns 4 values in
cses.py, and an untyped list assignment in usaco.py.
Solution: add start_time? to ContestSummary LuaDoc, guard csrf_input
with hasattr check and type: ignore, unpack precision from
_extract_problem_info in cses.py callers, and use cast() in usaco.py.
Problem: toggle_interactive() had its condition inverted — it blocked
:CP interact on non-interactive problems while showing the message "This
problem is interactive", and passed through on interactive ones. The
panel guard in toggle_panel() was also missing a nil-check on
contest_data.index_map, which could crash if the index map was absent.
Solution: invert the toggle_interactive() guard to match the symmetrical
pattern in toggle_view(), fix the error message to say "not interactive",
and add the missing index_map guard. Also handle the stress panel type
in M.disable() so :CP stress can be toggled off.
Add command parsing and dispatch for :CP race, :CP race stop, :CP stress,
and :CP submit. Add tab-completion for race (platform/contest/--lang),
stress (cwd executables at arg 2 and 3), and race stop. Add
<Plug>(cp-stress), <Plug>(cp-submit), and <Plug>(cp-race-stop) keymaps.
Add submit.lua that reads credentials from a local JSON store (prompting
via vim.ui.input/inputsecret on first use), reads the source file, and
delegates to scraper.submit(). Add language_ids.py with platform-to-
language-ID mappings for atcoder, codeforces, and cses.
Add stress.lua that auto-detects or accepts generator and brute solution
files, compiles C++ if needed, and launches scripts/stress.py in a
terminal buffer with session save/restore and cleanup autocmds.
Add scripts/stress.py as a standalone loop that runs generator → brute →
candidate, comparing outputs and exiting on the first mismatch.
Add race.lua with a 1-second vim.uv timer that counts down to a contest
start time and auto-calls setup.setup_contest() at T=0. Exposes
M.start(), M.stop(), and M.status() for command dispatch and statusline
integration.
Add KattisScraper and USACOScraper with contest list, metadata, and
test case fetching. Register kattis and usaco in PLATFORMS,
PLATFORM_DISPLAY_NAMES, and default platform configs.
Problem: problem pages contain floating-point precision requirements and
contest start timestamps that were not being extracted or stored. The
submit workflow also needed a foundation in the scraper layer.
Solution: add extract_precision() to base.py and propagate through all
scrapers into cache. Add start_time to ContestSummary and extract it
from AtCoder and Codeforces. Add SubmitResult model, abstract submit()
method, submit CLI case with get_language_id() resolution, stdin/env_extra
support in run_scraper, and a full AtCoder submit implementation; stub
the remaining platforms.
Problem: the tolerance field for floating-point comparison was named
`epsilon`, which is an implementation detail, not the user-visible concept.
Solution: rename to `precision` in run.lua type annotations, internal
variables, and comparison logic.
Problem: uv downloads glibc-linked Python binaries that NixOS cannot
run, causing setup_python_env to fail with exit status 127.
Solution: detect NixOS via /etc/NIXOS and bypass the uv sync path,
falling through directly to nix-based Python discovery.
Problem: the hooks API conflated distinct lifecycle scopes under a flat
table with inconsistent naming (setup_code, before_run, setup_io_input),
making it hard to reason about when each hook fires.
Solution: introduce two namespaces — hooks.setup.{contest,code,io} for
one-time initialization and hooks.on.{enter,run,debug} for recurring
events. hooks.setup.contest fires once when a contest dir is newly
created; hooks.on.enter is registered as a buffer-scoped BufEnter
autocmd and fires immediately after setup.code. The provisional buffer
setup_code callsite is removed as it ran on an unresolved temp buffer.
Problem: after apply_template writes a file's content to the buffer,
cursor positioning was left entirely to the user's setup_code hook,
forcing everyone to reimplement the same placeholder-stripping logic.
Solution: add an optional templates.cursor_marker config key. When set,
apply_template scans the written lines for the marker, strips it, and
positions the cursor there via bufwinid so it works in both the
provisional and existing-file paths.
Problem: CpPlatformOverrides lacked a template field and merge_lang()
never copied ov.template into the effective language config, so
per-platform template overrides were silently dropped.
Solution: add template? to CpPlatformOverrides and forward it in
merge_lang(), matching how extension is handled.
Problem: output comparison used exact string equality after whitespace
normalisation, causing correct solutions to fail on problems where
floating-point answers are accepted within a tolerance (e.g. 1e-6).
Solution: add an optional ui.panel.epsilon config value. When set,
actual and expected output are compared token-by-token: numeric tokens
are compared with math.abs(a - b) <= epsilon, non-numeric tokens fall
back to exact string equality. Per-problem epsilon can also be stored
in the cache and takes precedence over the global default.
Problem: new solution files were always created empty, requiring users
to manually paste boilerplate or rely on editor snippets that fire
outside cp.nvim's control.
Solution: add an optional template field to the language config. When
set to a file path, its contents are written into every newly created
solution buffer before the setup_code hook runs. Existing files are
never overwritten.
Problem: test cases were executed sequentially, each waiting for the
previous process to finish before starting the next. On problems with
many test cases this meant wall-clock run time scaled linearly.
Solution: fan out all test case processes simultaneously. A remaining
counter fires on_done once all callbacks have returned. on_each is
called per completion as before; callers that pass on_each ignore its
arguments so the index semantics change is non-breaking.
Problem: M._cache = cache_data captured the initial empty table reference
at module load time. After M.load() reassigns cache_data to the decoded
JSON, M._cache is permanently stale and returns the wrong table.
Solution: remove the field assignment and expose get_raw_cache() which
closes over cache_data and always returns the current table.
Problem: vim.loop is deprecated since Neovim 0.10 in favour of vim.uv.
Five call sites across scraper.lua, setup.lua, utils.lua, and health.lua
still referenced the old alias.
Solution: replace every vim.loop reference with vim.uv directly.
Also fix contest-change detection so URL open logic triggers when either platform or contest changes. This makes :CP next/:CP prev and problem jumps open the correct page when open_url is enabled.
Co-authored-by: Codex <noreply@openai.com>
Problem: after an install or update, the on-disk cache may contain data
written by an older version of the plugin whose format no longer matches
what the current code expects.
Solution: embed a CACHE_VERSION in every saved cache file. On load, if
the stored version is missing or differs from the current one, wipe the
cache and rewrite it. Corrupt (non-decodable) cache files are handled
the same way instead of only logging an error.
Add LuaCATS annotations to the env conversion helper and drop the table.sort call since ordering is not required by uv.spawn.
Co-authored-by: Codex <noreply@openai.com>
Neovim/libuv spawn expects env as a list of KEY=VALUE strings. Passing the map from vim.fn.environ() can fail process startup with ENOENT, which breaks NDJSON test scraping and surfaces as 'Failed to start scraper process'.\n\nConvert env map to a deterministic list before uv.spawn in the NDJSON scraper path.
Co-authored-by: Codex <noreply@openai.com>
