Skip to content

v0.3.9

Latest

Choose a tag to compare

@github-actions github-actions released this 16 Nov 22:20
· 115 commits to main since this release
09e4592

A command prints only what it makes now, and a check exits 1 when it finds what it checks for, so read getting-started's Updating before you update a script that reads one. Five parts of the app are redesigned. Needs you is one list of what waits on your answer: each card leads with what it asks and answers it there, and its panel shows the proposal itself. Library is one page of everything you keep, with a preview beside it. The Projects page says what needs you and what is running in each project, and a project's page opens on a box to start a chat there. A plan step is approved only as you read it, and a change to it after you read it is put to you again. The agent orb shows each state in a form of its own, from a still ring when it is ready to the claw mark when a turn is done.

⚠️ Breaking changes — read before upgrading

  • A check exits 1 when it finds what it checks for, its report still on stdout: status when no gateway is running, eval when a scenario fails, eval-gate on a regression, and security audit and memory audit when the scan finds something or could not open memory to scan. backup export and footprint --reclaim exit 1 when another run holds the lock, said on stderr (footprint --reclaim --json still prints its document), and project import when it left something unchanged, as restore does. learn add no longer takes --category, which no lesson stored, and learn list shows no category.
  • For apps, a lesson has no category: POST /api/lessons refuses one with a 400 that says so, memory_remember no longer takes one, and GET /api/lessons and memory_list show none, where every lesson read knowledge, and the Memory studio shows none. The Python client's add_lesson takes none and names its scope and workspace, and list_lessons and list_subagents return what the gateway lists, which they read as none. MemoryService.write_lesson takes no category, and every argument after the rule is named (negative=, source=), so a call that still passes a category raises TypeError (slack-channel builds one with over_vector_store and calls only get_lessons). A MemoryWrite hook, which fires for the lessons the after-turn review and consolidation write, is given key=lesson, where it was preference or the category consolidation's model chose (tool, preference or knowledge), so a Context matcher on key= no longer matches; one on scope= (after_turn_review, facet_veto, consolidation) tells them apart.
  • personalclaw token prints one line, this computer's sign-in link: the dashboard.url link is --remote, and what it is and how long it lasts is --explain, on stderr. pair, auth enroll, auth totp, inbound token create, inbound webhook create and spawn run --async print only the code, link, token or id they make, so | pbcopy and $(…) get it alone. config get KEY writes true, false and null as JSON does and exits 1 for a credential key without --reveal. auth totp refuses to replace an enrolled 2FA secret without --rotate. The warnings an action prints are on stderr.

Fixed

  • A loop that stalled or is blocked, a folder waiting for your trust, an action held for your OK, a run waiting after you declined a step and an escalated run are listed in Needs you and counted by its badge, and answering a stalled loop from its card starts it again.
  • run reads the gateway it starts to the end of its output, so a gateway that logs never stalls the run; spawn run's wait ends, saying why, when the agent or the gateway is gone, asks again when a poll times out, and signs in again when its hour runs out; status reads a gateway too busy to give its stats in 3 s as running (stats_unavailable timeout), where it read as gone; setup names the port the gateway serves; app new --list-types counts the providers the installed apps register; memory import and security deny-list refuse a file that is not JSON with a sentence, exit 1.
  • retrieval-eval --card > card.json is one document --label reads back, a folder memory with no records exports an empty document, and a refusing update, restore, study --run or project import prints nothing on stdout. update --to no longer pins a release it cannot install, footprint --reclaim --json always prints its document, and snapshot --list with no snapshots prints nothing. The dashboard finds marketplace skills again. auth disable no longer names a --clear that does not exist, and cron pause says why a job that is there was not paused. A sign-in sentence the dashboard shows another device, and doctor's line for a phone at the tailnet address, name personalclaw token --remote, whose link that device can open.
  • A chat starts helpers again once the helpers it had are stopped or have ended: after Stop fan-out or a stopped turn found them all waiting to start, after its last helper reached Max tokens / run, after five of its helpers failed in a row, or after a helper's start was denied or went unanswered. Until the gateway restarted, it refused every later helper with the old reason. A helper stopped while it runs, or ended at its time limit, now lets the next helper waiting to start begin at once; before, that helper waited for another one to end, and one asked for later could start first.
  • A project's work board lists a loop once, no longer again as each of its steps.
  • A workflow run's Max tokens / run and its failure breaker count every agent step of the run, one after another as well as side by side, until the run ends. A step that started after the one before it had ended began again at 0 tokens and 0 failures, so a run of steps in a row could spend Max tokens / run on each and fail five times per step. Once a run's steps are stopped (Stop fan-out, the ceiling or the breaker), its later steps are refused until the run ends.
  • A gateway whose start-up fails starts no helper that was waiting for its slot while it closes, after its app processes have stopped.
  • The gateway keeps answering while a code loop starts a stage's tasks: git checks the tree and cuts each task's worktree off the gateway's event loop, so a big repository no longer holds every request for seconds. A Stop, a Pause, a delete, or a task taken off the queue, cancelled or done meanwhile starts no worker for that task, and a stopped or deleted loop keeps no fresh checkout.
  • What an agent makes in a project's chat now belongs to that project and shows on its page: the artifacts and documents an agent CLI saves, and the images and videos any agent makes. A save there is no longer refused because another project holds an artifact of that name.
  • A folder in your artifacts with a name no artifact can have, or one that links out of the store, no longer makes every page that lists your artifacts fail: it is left out, and the rest are listed.
  • The gateway keeps answering while a code loop lands a finished task and on every poll: git merges the task's work back, reads it for its owner's review and answers the poll's checks off the gateway's event loop. A loop stopped or paused meanwhile is asked nothing, a merge already under way is recorded, and a task taken off the queue meanwhile is not merged.
  • Saving a fact again with the value it already has no longer hides the episodes that agree with it. Memory retires the episodes that mention a fact's old value only when the value changes, so a consolidation or an Add fact that restates a fact leaves them in recall.
  • /compact no longer calls a conversation "already short enough" when its context is more than a quarter full. When compaction can fold nothing there, the chat says how full the context is and that it holds the session's opening and its latest messages, which compaction keeps word for word.
  • The Tools page checks PersonalClaw's own tool server, personalclaw-core, from the gateway's own install. It started the first personalclaw in ~/.local/bin or on PATH, so on a machine with another install, a gateway from a development clone, a wheel under test or the desktop app ran that install's server against its own home and listed that install's tools.
  • The sidebar names a chat with no title by its first words with no Markdown mark left showing, in its row, its name and its tip, also when the cut falls inside bold text.
  • Running a turn again asks first about a step that was still running when the turn ended, so a command a restart cut off is not run twice unasked.
  • Auto-nudge stops by itself after 10 cycles unless you set Max cycles to 0, which the modal now says means no limit. Max cycles started at 0 with nothing saying so, so a nudge armed as the modal opened, or through POST /api/autonudge without max_cycles, ran an agent turn about every minute until you pressed Stop. The modal shows the limit beside the cycles fired, and says Stopped once a nudge has stopped.
  • A Code loop whose workspace has no branch checked out (a detached HEAD) merges nothing and starts no task until you check one out, and says so. Its merge review, merge records and conflict question named main whatever the repository's branches were called, and an Unattended loop merged its tasks onto the detached HEAD and deleted their branches, which left the work on no branch once a branch was checked out. They now name a detached HEAD as one, with its commit.
  • The gateway keeps answering while a code loop's stage gate reads what the stage changed: git reads it off the gateway's event loop, and a loop stopped, paused or deleted meanwhile is not judged.
  • A code loop paused while its stage gate runs the stage's check, asks its judge or scores the stage's quality is no longer advanced over the Pause: the gate reads the loop again after each, so the stage stays where it was, and the stage worker's nudge and the loop's owner are left as the Pause left them.
  • Accepting a proposed template saves it to Workflows, as the agent's save tool says.
  • A workflow step that asks you to pick, to type an answer or to fill in a form is answered in Needs you as it asks, where it showed Approve and Deny and refused the answer: its choices on its card, a typed line in its reply field, its form in its panel, and a parked step woken. A step that asks for a pick and lists no choices takes your own words, on its run's page too, and an empty answer is refused.
  • The Inbox's Proposals lens says beside each learning proposal's kind what its Accept does (saves the template, records that you agree), and offers no Approve on one that can't be accepted here. A refused Approve shows the gateway's sentence once, with no error's class name, and a learned template refused for a step holding something shaped like a credential says so, what to do, and that it still waits.
  • A new chat scrolls on a phone: the last suggestions and the composer are in reach under the first-run small-model card, where the page cut them off at the fold.
  • On a phone the Library's Power user view shows its tabs as a strip above the panel, which has the page's whole width: beside it, a 184 px column left each panel about 205 px and cut its headings and titles.
  • Automations › Week names the automations it cannot plot in full by the names the list shows, no longer by their internal ids.
  • Settings › Devices lists a sign-in link a browser opened as "Link opened in" that browser, with when, and that it still works as a token until it ends. It no longer reads as an unrecognised client among links no browser has opened.
  • A missed-run card settles once its automation has run on schedule since the slots it missed and that run did its work: the history records the missed slots let go, as a Dismiss does. A card about a run a restart cut off still waits for you.
  • Library settings' Being read row counts the items still being read or that couldn't be read, the ones its Show them lists, no longer the names the graph could not take without a model; those have a row of their own, with Set up a model. The Library settings tab and the Items count line count the same.
  • A chat's cost after each answer, the Usage page and the first models list after a start no longer freeze the gateway. The usage totals, rollup and fold read every row the usage ledger keeps (up to 50,000) on the gateway's event loop, so every other request waited up to 0.9 s a call on a full ledger, and a chat asks for its total on every load and after every answer; the first models list imported the OpenAI SDK there to ask an image provider if it can make images, about a second. They run in a worker thread.

Security

  • A folder outside the workspace, the folders you allowed and a workflow run's own folder is refused for a subagent before anything reads it, with the one refusal every such folder gets.
  • The credential mask covers a whole private key block, header to END line: RSA, DSA, EC, OpenSSH, PKCS#8, encrypted and OpenPGP keys, with their lines broken, escaped as JSON writes them, run onto one line, indented, marked as a diff, a quote or a # comment marks them, or written as string literals in code. A connection string's Pwd= or PWD= is masked as a password.
  • A credential's value written in quotes is masked by its name, as one without quotes is: "password": "…" in JSON, a quoted value in YAML, DB_PASSWORD="…" in an env file, and a keyword argument or a JavaScript object's field, a line break after its : or = included, for api_key, secret_key, access_token, refresh_token, client_secret, password, passwd and private_key, and for the SecretAccessKey and SessionToken the AWS CLI writes. The mask replaces what the quotes hold, so the name and the quotes stay and a JSON document still parses, and a tool call's card masks such a value in the call's input as its preview does.
  • A redaction warning names what kind of text it masked, and at most its length or a scheme or host the masked text keeps, never any of the value: a credential pattern's warning and a long URL query's warning held the first characters of what they matched.
  • A path rule written with ~ or a $VAR names the paths its home or variable names: a trailing /, a doubled /, a /./ or /../ inside it, or text glued to it, and a path or a rule that starts with two or more slashes is read from /.
  • The credential guard, the refusal of a change to what runs as you, and the question before a delete of your home folder, the filesystem root or the working folder compare a path in one Unicode form, so each is caught whether a letter in its path is written composed or decomposed.
  • The credential mask reads an AWS secret key's or session token's name in any case, the session token's older AWS_SECURITY_TOKEN included, and an access key id's before a value without quotes, so AWS_SECRET_ACCESS_KEY=… as an env file, printenv or the AWS CLI's env export writes it is masked as aws_secret_access_key = … is, with or without quotes. After an AWS name, a value that starts with $ and names where a secret is kept, such as a CI workflow's ${{ secrets.X }}, a shell's $(…), $VAR or an empty default, stays as it is, while $'…' and a literal default are masked, and an empty value never takes the next line's key as its value.
  • A credential set with :=, as Go and Make write it, or with Make's ::=, ?= or +=, is masked by its name, and so is a quoted value after a map's =>, as PHP, Perl and Ruby write it, under every name.
  • A plan step is approved only as it was shown: a change to a step after you read it, or after you approved it, is put to you again before anything builds or runs it, and so is a plan a step you approved is gone from. A deleted chat's plan goes with it.
  • An app you installed reads only the learning and skill proposals from conversations it started, and reads a proposal, a waiting item and a run's open asks masked as their own views show them, never with the token that answers a gate. A notification carries its row's ids, not the row's card or token; deleting a chat withdraws the proposals still waiting that were learned from it; and the second-opinion check reads a row's words masked and fenced.
  • A Library item's file is sent as a download unless it is a type that is shown inline (an image, audio, video or a PDF), and never runs script, even opened in a tab of its own.
  • A loop's own records are PersonalClaw's, not an agent's. An agent's change to a loop's brief, run state, steers and their history, ledger, merges or plan is refused whatever the chat or the loop allows, and your steer reaches its worker in the cycle's own message, sent again until that worker's next finding, rather than through a file the worker deletes.
  • A program an agent starts can no longer read the key that signs the security log, or change the log or its archive, at any sandbox level, in the home in use and in the default one, so the log's check no longer passes a row such a program wrote. The tool server an agent CLI runs hands its rows to the gateway to sign, marked as reported by it, and leaves them for the gateway's next start when no gateway answers.

Added

  • personalclaw token --bearer and --json, and --json on status, cron list, cron add, spawn list, skills list, skills search, incident status, memory partitions, auth status, snapshot, auth totp, inbound token create and inbound webhook create: one document, every key always there, null for what the command could not learn.
  • For apps, personalclaw.sdk.channel.SubagentManager has end_run_fanout(key): a workflow run's fan-out key (workflow:<run_id>) ends with the run, and what its agent steps shared (the breaker's count, a stop and the run's token total) ends once nothing of the run runs or waits.
  • For apps, personalclaw.sdk.channel.SubagentManager has stop_starting(): the manager starts no helper from then on, and one still waiting for a slot stays recorded for the next start. cancel_all begins with it.
  • The gateway lists your whole library in one read, GET /api/library: notes, files, links, artifacts, report write-ups and what each intent found, one row per thing, with a true total and the name of any store it couldn't read. A note or an artifact can move into a project, which takes it out of the one it was in, or leave its project. A note or a decision saved in a project's chat is filed under that project. An intent can write a report on a schedule, and its check of what you already saved can run in the background, with its progress and a way to stop it. The run that answers when it is done still works. A feed can be removed or checked at once. You can record which of two items that disagree to go by, and both are kept. Renaming or merging a tag keeps the reports that read it on the same items, and a report whose tags were all deleted stops running on its schedule and says why.
  • A link can open a new chat, or one you were in, with items from your library attached, and nothing is sent until you send it. ⌘K's Library group finds your notes, files, pages and documents, and what your intents found, in one search. A report's Inbox note and its notification open its write-up.
  • Library is one page of everything you keep: what you saved and what the agent made in one list, with a preview beside it, a Filter, a List or Grid, and Use in chat, Open, Pin and More on every item. What your intents find and the reports they write sit with everything else, and setting up feeds, intents and reports is part of the page, in the pane beside the list. Old Knowledge, Artifacts, Sources and Reports links open the same things in the library.
  • Library's Power user view, a switch in its header, adds the graph, your tags, every conflict, the decision journal and the library's own settings, and opens on the graph. It is remembered on every device, and turning it off hides no control.

Changed

  • The Nix flake installs 0.3.8.
  • Needs you's cards and a project's Running now say which project each piece of work is in, and a running card says what kind of work it is. A code session that made its own project is listed under it and reads its brief.
  • The agent orb shows each state in a form of its own: ready holds still, thinking swells, working turns two rings, waiting on you locks into a clasp, and done closes into the claw mark, with one bloom when the turn completed; a finished turn's line now shows it before it goes. A row's state in the sidebar, a project's chats and Home's agents is the orb's still mark.
  • Pages that list or search your artifacts answer faster when you have many: the Library's Artifacts tab and its search, Home's Pinned artifacts, the chat's artifact picker, and a project's or a loop's page.
  • The Projects page is one list. Each project says what needs you, what is running and when it last moved, with those rows under it, and opens in one click.
  • Approving a plan step names the step as you read it, in If-Match (its review_revision in the plan's read), as editing one names its draft. An approval without one is refused 428 revision_required, and one of a step that changed since its read 409 stale_write. Needs you's plan rows carry it in their answer's headers, so an answer sent as given approves the step as read.
  • Needs you is one list of what waits on your answer, across the whole width beside the sidebar and in columns when nothing is open, with proposals and notes after the asks that stop work. What runs, what finished and your chats moved to Home and the sidebar, where they already showed, and Home shows its pinned tiles again, in the section for what you keep on Home. The YOLO chip in every page's corner says what YOLO covers, and the badge counts what waits on your answer.
  • A project's page starts with a box to start a chat in it, then what needs you, what is running, what comes next, its tasks and the rest, with its brief, its instructions, its folder and what a new chat there is told beside it.
  • The desktop buddy draws the same claw ring, its poses in strokes: done is its open ring with a check, and the closed mark stays the chat's.
  • A project's page shows what its work made, as the library's cards under its tasks, and what its agents know, in About, with Add to write a note, add a link, record audio or upload files into the project. Each opens its preview beside the work.
  • Needs you's cards lead with what they ask: what a template does and its steps, what went wrong in a run, the change to a skill, with the kind as a small tag and the time left to answer. Every card that holds a decision has its own answer, asking once more where its panel does, so you open the panel only for a question with several parts or choices, or a form.
  • Needs you's panel shows the proposal itself, as its producer wrote it: a template's steps, a run's failure and its error, a skill's change as the block it appends, why it was proposed and where it came from, with what Accept does at its foot, and Next ask once it is decided. A question with several parts, a run's form and a loop's work waiting to merge or in conflict are shown and answered there too. Learning opens the proposal a link names.