diff --git a/CHANGELOG.md b/CHANGELOG.md index b6cc7bc..ccf9d43 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -12,6 +12,24 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 - On macOS and Windows, reading bridge records started one process per record. Before a record is read, prodex removes a temporary hard link that an interrupted record write can leave beside it, and where the store has no descriptor directory to work through (macOS and Windows), that cleanup runs in a child process, which was started for every record read whether or not there was anything to clean. Found 2026-11-09 with the descriptor path switched off on Linux, as on macOS: listing 200 receipts took 1956 ms against 264 ms with it, and the follow-up budget check every follow-up runs (`--continue`, MCP `pro_consult` with `continue_task` or `continue_thread`) 1929 ms against 273 ms. The child process is now started only when the record has a second link, so reads start none and a record write still costs one. - On Windows, replacing a bridge record failed with `EPERM` while another prodex process was reading it. Records are replaced by renaming a new file over them, and Windows refuses that rename while another handle has the old file open: found on CI 2026-11-09, a session rewritten while another store listed sessions failed this way, and the test for that race was then skipped on Windows. The rename is now tried again for up to 2 seconds while Windows reports the file in use (`EPERM`, `EACCES` or `EBUSY`), with the checks before it repeated, and then fails with the rename's own error. Other platforms are unchanged. The race test runs on Windows again. - A record that could not be read was not named. With a task record at mode 000, `prodex tasks list` failed as `EACCES: permission denied, open '/proc/self/fd/20/task_...json'`, a descriptor path, and with a hard-linked record as "Target path is hard linked and cannot be used through safe file operations", naming no file. Found 2026-11-09. They now say `Task record is not readable: .bridge/tasks/.json (EACCES)` and `Task record is hard linked: .bridge/tasks/.json has another name on disk`, with what to do (give your user access to it, or remove the other name, or move it aside, then retry), the same for results, sessions and receipts. The path is the repo path, so MCP clients are not given the machine's absolute path. +- One untrusted result hid every other result from `prodex results show latest`, `prodex results artifact latest` and MCP `bridge_list_results`. Found 2026-11-09: with two finished tasks and the older one's `task_completed` receipt moved away, all three failed with "Result record is untrusted: .bridge/results/.json has no trusted task_completed receipt", while `prodex results show ` worked. `latest` now resolves to the newest trusted result, stepping past untrusted ones as `prodex pro latest` does, and fails with the untrusted record's reason only when no result is trusted. `bridge_list_results` lists the trusted results and adds `untrusted_skipped` (how many were left out) and `untrusted` (each one's task id and reason, with repo-relative paths). Showing or fetching an untrusted result by its task id still refuses it. +- `prodex pro blockers --cwd ` reported a repo with one corrupt `.bridge/results/*.json` as having no consults at all. Found 2026-11-09: a repo reported as "3 of 3 consults" printed "No consults recorded in this repo." with rc=0 once the corrupt file was added, and `--json` said `"roots": 0`. The record's error was swallowed. The command now fails with "Result record is corrupt: .bridge/results/.json. Move it aside or fix the JSON, then retry." Without `--cwd`, the other bridge roots are still reported and the one skipped is named on stderr with the same message. +- `prodex sessions cancel ` made the cancelled session the latest one. Cancelling stamped the session as just used: found 2026-11-09, after cancelling an older session by id, `prodex sessions show latest` showed that cancelled session, and `prodex sessions cancel latest` failed with "already blocked; it was cancelled before" while the newer session was still running. Cancelling now leaves the session's last-used time alone, so `latest` and the order of `sessions list` stay as they were; sending through a session still makes it the latest. +- `prodex pro blockers --since 2026-02-30` was read as 2026-03-02. A day the month does not have rolled over into the next month (found 2026-11-09). Such a date is now refused with "--since 2026-02-30 names a day that does not exist"; real dates, including 29 February in a leap year, dates with a time, and ages like `7d` or `24h` work as before. +- An option that takes one value, given twice, kept the first value without a word. Found 2026-11-09: `prodex tasks create --title First --title Second` saved the title "First", and `prodex ask --timeout-ms 300000 ... --timeout-ms 900000 --model Pro --model "GPT-5.6 Sol"` sent with a 300000 ms budget on Pro, so an agent that appended an override after its defaults got the default, and a long send could time out early. Every command now refuses the second one with "--timeout-ms was given twice; give it once, with the value you mean." The options meant to repeat (`--file`, `--attach`, `--tool`, `--command`, `--artifact`) still add one value each time. +- A value that starts with `--` was refused as missing: `prodex tasks create --title Fix --prompt "--dry-run should not write files"` said "--prompt requires a value", and `--prompt=x` was an unknown option with "Did you mean `--prompt`?". Found 2026-11-09. The first now says the value starts with `--`, which is read as an option, and to reword it so it does not (or to put `./` before a path); a value cannot start with `--` on any option. `--prompt=x` now says that options take their value as the next argument and shows `--prompt "x"`. A forgotten value followed by the next option (`--title --prompt ...`) still reads "--title requires a value". +- `prodex ui --help` failed with "Unknown option for ui: --help" and `prodex ui help` with "Unexpected argument for ui: help". Found 2026-11-09. Both, and `-h`, now print what `prodex ui` does. +- A help request carrying one of the command's own flags failed on that flag: `prodex tasks list --json --help` said "Unknown option for tasks list: --json". Found 2026-11-09. The help check listed the command's value options but not its switches, for `tasks list`, `receipts list`, `sessions list`, `pro blockers`, `pro list`, `pro latest`, `pro show` (`--json`) and `pro report-issue` (`--confirm`). Each now prints its help. +- `prodex version --bogus` printed the version and exited 0. Found 2026-11-09. `version` (and `--version`, `-v`) now refuses an argument it does not take, as every other command does, and `prodex version --help` prints its help. +- An argument that cannot be an option, such as `-` or `--`, was answered with advice for `prodex ask` whatever the command: `prodex tasks list -` said "If this is the prompt, put `--` before it: prodex ask -- "-"", and `prodex tasks create ... -- extra` the same. Found 2026-11-09. The hint is now given only by the commands that take a prompt, and names the one that was run (`prodex pro ask -- "-5"`, `prodex pro browser ask -- "-5"`, `prodex ask -- "-5"`); other commands say only that the option is unknown. +- `prodex setup --help` said the model rows "cannot be clicked", and `prodex pro browser help` that "no coordinate reaches them, so any other name warns"; `prodex pro help` said models such as GPT-5.6 Sol are rejected. Since 0.40.34 a model the picker lists is selected, and the 0.40.38 wording fix missed these lines. Found 2026-11-09. They now say that prodex picks the row with the keyboard, that a model the picker does not offer is a warning, and that a Chat step given as a model (`--model Pro`) moves the power slider. +- `prodex pro help` said "Creating a new project from the CLI is planned; for now create it in ChatGPT and pass --project", while the same help listed `--project-new "name"`, which has created projects since 0.40.3. Found 2026-11-09. The sentence is gone, and the selection options list `--project-new`. +- What `--effort max` means depends on its letter case: `Max` is the Work surface's Max rung, while `max` and `MAX` mean 매우 높음 (Extra High). The invalid-effort error and the `--effort` lines of `prodex pro help` and `prodex pro browser help` listed `max` among the aliases without saying so. Found 2026-11-09. They now say that max in any letter case but "Max" means 매우 높음, while "Max" is the Work rung; the mapping itself is unchanged. +- While `prodex ui` waited for an answer, its progress bar filled against a fixed 20 minutes, whatever the send's own budget was. Found 2026-11-09: the line read "0s / 20m prompt sent, waiting for answer (budget 5m)", so a send with no model selected, which waits 5 minutes, timed out with a quarter of the bar filled. The bar now fills against the budget the send settles on, the one it names in "(budget ...)"; 20 minutes stands in only until the send has said. +- In `prodex ui`, a list longer than the terminal was printed whole and scrolled the top of the screen away. Found 2026-11-09 with 25 projects at 80x24: the screen started at "4 Project number 4", so the question, the settings panel and the cursor on row 1 were off screen; with colour on, the logo and the panel alone took 15 of the 24 rows. Rows from 10 on were also a column to the right of rows 1 to 9. The list is now a window that follows the cursor, with a line above and below it saying how many entries are out of view; on a short terminal the logo gives way first, then the panel; and the row numbers are right-aligned. +- The "attach files?" line of `prodex ui` split some paths it should have kept whole, and the send then refused them as missing files. Found 2026-11-09: `Bob's notes.pdf` dropped onto GNOME Terminal arrives as `'/home/u/repo/Bob'\''s notes.pdf'` and came back as three paths, `"My File".pdf` came back as `My File` and `.pdf`, and `C:\repo\#1\a.pdf` lost the backslash before `#1`. The line is now read the way a shell reads words, so quoted and unquoted parts next to each other make one path, and a path that starts with a drive letter or `\\` keeps every backslash. A backslash-escaped space (`My\ File.pdf`, from a macOS terminal) and an apostrophe inside a word (`it's.txt`) work as before. +- `prodex ui` counted some emoji one column wide that terminals draw two wide, so a row holding one overran the terminal by a column for each. Found 2026-11-09 with U+2705, U+2B50 and U+2764 U+FE0F. These and the other symbols Unicode lists as wide now count two columns, as does a symbol followed by U+FE0F. On a terminal narrower than 40 columns, rows were still cut at 40 and wrapped: at 32 columns every hint broke onto a line of its own, and the title and the footer wrapped too. Rows, titles and the footer are now cut to the terminal's width, down to 24 columns; a hint gives way before its label, and the logo is left out where it is wider than the terminal. +- `repo_write_file_apply` failed for a file whose name is longer than about 192 bytes, after `repo_write_file_dry_run` had accepted it. The new content is written to a hidden temporary file beside the target, whose name adds about 63 bytes to the target's, and a name may be at most 255 bytes: found 2026-11-09, a 203-byte name failed with `ENAMETOOLONG: name too long, open '/proc/self/fd/21/.aaaa....md..