togglr

• create_client() returns the created client (a list), no longer the httr response (#54); toggl_update_entries() sends one PUT per entry and returns the updated entry, or a list of them (#56).
• A write refused by the API raises an error (#54); toggl_start() and toggl_create() return NA, with a warning, when the API accepted the entry but its answer could not be read.
• get_toggl_api_token() fails, outside an interactive session, instead of prompting, and never deletes the stored key (#51).
• set_toggl_api_token() returns the token invisibly (#84).
• simplify() is rewritten on the public API of jsonlite (#69).
• get_detailled_report(): cache = FALSE by default, max_page and memoise_cache_dir deprecated (#60, #67); a 7th positional argument now goes to cache.

Other changes

• get_detailled_report() reads Reports API v3 (/search/time_entries, enriched with the names of users, projects, clients and tags) instead of Reports v2, with the same columns and types. Pages of 1000 entries (options(togglr.page_size)) instead of 50: a month of a team of 20 is read in one or two requests. The period is checked against the total duration of /search/time_entries/totals, equal before and after the pagination and, up to the rounding of the API (1 s per entry), to the sum of the entries. New argument user_ids; user_agent is ignored (#76).
• The completeness checks of get_detailled_report() also treat as a movement any entry received that was edited after the read started: the pages are numbered by row, so a deletion followed by an insertion of the same duration shifts the rows without changing the totals. This replaces the total_grand tolerance for running timers, which could hide such a movement (#77). A running timer may only grow by the time between the two totals, measured by the clock of the server from its own start, not by the time spent waiting before the first total. One movement still escapes: an entry deleted and replaced by one of exactly the same duration that is not in the pages received.
• get_summary_report(), get_weekly_report() and get_dashboard() read Reports API v3 too, with the output of Reports v2. Reports v3 only returns ids: the names of users, projects and clients come from the Track API v9, archived projects included. get_weekly_report() returns an empty tibble for a week without any entry. An unreadable answer to a report request is an error, never an empty report. The amounts of the summary items (cur, sum, rate) are no longer given (NA). Entries without a visible project are grouped with an NA id and name (Reports v2 called them "Private Project") (#76).
• With a shared state file, an exhausted quota, or a spread of the requests, recorded by another process while this one was waiting is no longer erased: it is waited for in turn, or fails within max_wait (#78).
• The requests of Reports v3 (/reports/api/v3/workspace/<id>/) count against the quota of their workspace, like the requests of the API v9 (/workspaces/<id>/): an exhausted Reports quota no longer blocks /me.
• get_time_entries() is never truncated: /me/time_entries returns at most 1000 entries, so a full answer splits the period in two, recursively. A period starting more than 3 months ago (the limit of the API) fails before any request, pointing to get_detailled_report(user_ids = ) (#13). Its output gains a client_name column; project_name and client_name cover the archived projects too (#14). get_project_id_and_name() gains active (NULL, "false" or "both") and returns 0 rows, not an error, for a workspace without any matching project. since and until may be text dates again.
• get_workspace_id() returns the default workspace of the account (default_workspace_id of /me), and the first workspace only when no default is set; it used to be always the first one. New get_workspaces() lists every workspace, the default one flagged (#23).
• get_dashboard() and get_all_project_names() gain user_ids, to keep the projects some team members worked on (#12).
• New get_project_data(): the data of one project from its id (#28).
• toggl_start(tags = ) sends its tags, now covered by a test (#16).
• set_toggl_api_token() returns the token invisibly (it was printed when called at top level, e.g. in a script whose output is logged), and checks the new token before deleting the stored one. update_toggl_api_token() asks first: a cancelled input keeps the stored token (#84).
• Tests: coverage from 85 % to 93 %, a real multi-process test of togglr.state_file, no test depending on covr's instrumentation; new workflow measuring the coverage (#84, #32, #33).
• DESCRIPTION declares R (>= 4.1.0) and the minimal versions of its dependencies; the obsolete inst/dependencies.R is no longer shipped (#83).
• The API token is no longer printed: create_client() and toggl_create_project() used httr::verbose(), which shows the Authorization header (#50).
• Every request times out after options(togglr.timeout) seconds (120 by default), and network errors (timeout, DNS, connection reset) are retried like a 5xx (#45).
• get_detailled_report() returns fixed column types (e.g. task is always text, tid integer), whatever the content of the page: periods over a year with an empty year no longer fail to combine (#46).
• New option togglr.state_file: the pacing of the reports calls and the knowledge of an exhausted quota are stored in that file, under a lock, and shared by every R process using it (callr / future / targets workers, cron jobs). Without it, the state stays private to the session (#48).
• A 402 without quota headers (a feature of a paid plan) now stops at once with the message of the API, instead of blocking every request for an hour. An exhausted quota is remembered per token, as a hash: the token is never stored (#49).
• The api_token and workspace_id given to a function are now used by every call it makes: project, client and dashboard helpers, get_time_entries() (new workspace_id argument), toggl_start(), toggl_stop(), toggl_create(), toggl_create_project() no longer fall back to the keyring token nor to an extra /workspaces request. client_name_to_id() and client_id_to_name() gain a workspace_id argument (#51).
• get_toggl_api_token() no longer deletes the stored key when the keyring cannot be read (e.g. locked under cron); in a non interactive session it fails with a clear message instead of prompting (#51).
• toggl_start() and toggl_create() send the start time converted to UTC, instead of removing one hour: entries were one hour off in summer, and more outside Paris (#52, from #44 by @bradyrippon).
• get_current_duration() counts from the start of the running entry, not from its last edit (at); get_project_total() and get_current_project_task_detail() were affected (#53).
• Every write (POST, PUT, PATCH) goes through toggl_request(), with the policy of the reads: an error of the API is raised with its message (no more "$ operator is invalid for atomic vectors"), the quota and 429 are respected. A POST is never retried after a 5xx or a network error, as it may have been processed. toggl_start() / toggl_create() notify then raise the error; toggl_stop() only notifies STOP once the API accepted it. create_client() now returns the created client (a list) instead of the raw httr response (#54).
• toggl_create_project() works again: no crash without client, an existing client is reused instead of being created again, its client_id is sent, and the body follows the v9 format (private, color, active are no longer ignored) (#55).
• toggl_update_entries() sends one PUT per entry (the v9 API takes a single id per PUT), sends stop (it was accepted then dropped), sends the project as project_id, and no longer sends the read-only at. It returns the updated entry, or a list of them for several ids (#56).
• get_time_entries() sends its dates in UTC (...Z): the + of a time zone offset was sent unencoded in the url, and read as a space by the server (#57). Entries with several tags are covered by a test (#30).
• get_detailled_report() splits by calendar year as soon as a period spans 365 days or more across a new year: a 367 days period (bounds included) was sent as is, beyond the one year limit of the reports API v2 (#58).
• The pagination of get_all_projects() is bounded: it stops with an error if the API returns the same page twice, or after options(togglr.max_pages) pages (100 by default), instead of spending the quota in an endless loop (#59).
• get_detailled_report(max_page = ) now warns that the argument is deprecated and ignored (#60).
• The reports functions (get_detailled_report(), get_summary_report(), get_weekly_report(), get_dashboard()) accept since / until as a Date, a date-time (its day in its own time zone) or a text "YYYY-MM-DD", and fail with a clear message, before any request, on anything else or on since after until (#61).
• Default periods ("one year ago") are computed with %m-%: they were NA on February 29th (#62).
• options(togglr.min_interval) (1 s by default) now spaces every request, reads and writes, retries included, not only the pages of the detailled report (#64).
• A period that keeps moving during the pagination is read up to options(togglr.report_attempts) times (3 by default, 2 before) (#65).
• An entry seen twice during the pagination keeps its most recently updated version, not the first one read (#66).
• The reports cache (cache = TRUE) lives in its own folder, the reports sub-folder of rappdirs::user_cache_dir("togglr") or options(togglr.cache_dir), apart from the memoise files of 0.2.x. New argument cache_dir and function clean_cache(); memoise_cache_dir and clean_memoise_cache() are deprecated (#67).
• The quota announced by every answer (x-toggl-quota-remaining) is used: at 0, the next request waits for the reset (or fails within max_wait) instead of receiving a 402; with fewer than options(togglr.quota_spread) (10) requests left, they are spread until the reset (#68).
• simplify() no longer calls an internal function of jsonlite (getFromNamespace()): it uses the public toJSON() / fromJSON(), at full precision, with the same result (#69).
• globalVariables() no longer declares function names (case_when, full_join, tribble) nor unused names, which could hide a missing import (#70).
• Every request sends the user agent togglr/<version> in its header; get_dashboard() no longer sends a hardcoded api_test (#71). The user_agent argument of the report functions is kept for compatibility and ignored.
• get_workspace_id() fails with its message when the account has no workspace, instead of "subscript out of bounds". New tests cover the edge cases of the pagination, of the quota and of the cache (#73).
• Package metadata and documentation: new title and description, no more LazyData without data, targeted imports only (no more whole dplyr, httr, keyring, assertthat, getPass), get_detailled_report() has its own help page, the vignette no longer runs install.packages() during R CMD check. R CMD check is clean: 0 error, 0 warning, 0 note (#74).
• toggl_create_project() (and so toggl_start() on a new project) works again in a workspace without any client: get_all_client_info() returns the id and name columns, even with no rows (review of #75).
• The reports cache is kept per token (as a hash): a period fetched with one token is no longer returned to another token of the same workspace (review of #75).
• The spreading of the last requests of the quota waits at most max_wait, with a message, and only slows down the token that is short of quota (review of #75).
• The quota is remembered per token and per scope (the workspace or organization of the url, or the user for /me and /workspaces): an exhausted workspace no longer blocks /me or another workspace (review of #75).
• A request processed by the API is returned even if its quota cannot be recorded in the state file (warning): a time entry is no longer reported as not started, then created twice (review of #75).
• One pass on the state file per request instead of three, and a read no longer rewrites it (review of #75).
• A time entry accepted by the API but whose answer cannot be read is reported as started (with a warning, id NA), not as NOT STARTED (review of #75).
• A state file that cannot be replaced is an error instead of a silently lost update, and the temporary file is removed (review of #75).
• Network errors are no longer retried when the backoff exceeds max_wait (review of #75).