assurance-reach 0.1.0

pip install assurance-reach==0.1.0 · PyPI

• assurance reach <path>: what a change to a file or folder reaches, by a code graph. It reads a
  graph such as Graphify's, never builds one, and follows dependency edges backwards from every symbol
  in the path: each symbol that depends on the change, with the relation and the line of the call or
  import that carries the hop. The relations followed are the ones Graphify's own blast radius
  follows, and a package manifest's depends_on; any other is named, never guessed into a dependency.
  Symbols reached only through edges the producer inferred or marked ambiguous are listed apart from
  those reached by extracted edges alone. Every report says first how far the graph is behind the
  code, by each file's content when the producer left a manifest, and ends with what it could not
  determine: hops past the depth, relations it does not follow, symbols in no file, records it could
  not read. --json prints it as assurance.reach/1. Pinned to the graph Graphify 0.9.77 writes,
  whether run on a whole project or on a folder below where it ran.
• A graph chooses nothing that is read. It is often committed, so it is read as input from whoever
  wrote the repository: only regular files inside the folder it covers are opened, a recorded root
  outside that folder is set aside and said, and text from it is printed escaped.

assurance-mcp 0.5.6

pip install assurance-mcp==0.5.6 · PyPI

• Requires assurance-cli 0.6.5. No change to the server or its tools.

assurance-cli 0.6.5

pip install assurance-cli==0.6.5 · PyPI

• assurance reach reaches assurance-reach: what a change to a file or folder reaches, by a code
  graph such as Graphify's, and how far behind the code that graph is. Listed in --help, on the start
  screen, and in --version.

assurance-budget 0.2.19

pip install assurance-budget==0.2.19 · PyPI

• The Stop hook says when Claude claims the tests pass right after they failed, edit seen or not.
  Every finding was measured from the last code edit the hook recognises, so a session with none, or
  whose edits it cannot see, as through a heredoc, could run a failing test three times and end with
  "All tests pass now." while the hook said nothing, though assurance audit called it a loop and
  --fail-on-loop failed on it. When Claude's last message says the tests pass and the last test run
  of the project, or else its last check, failed before it, the hook now says so (review
  suggested), once.

assurance-budget 0.2.18

pip install assurance-budget==0.2.18 · PyPI

• The Stop hook says a finding once. A finding whose evidence had not changed was said again
  whenever Claude said again that the tests pass, turn after turn. It is now said once while the
  edit, the command and the time it rests on are the same; a change to any of them is a new sentence,
  and is said. A count of commands it could not classify that grows from turn to turn is not a new
  finding. The advice to declare one of those commands under [audit] is given only for a command a
  declaration could name, not for a script fed to Python on standard input or with -c, nor for a
  command inside $( ).
• --json no longer names a Claude Code edit as made without reading. Claude Code refuses an edit
  to a file the session has not read or written, so every file edited_without_read named for it was
  one the session read or wrote in a way this reader did not record: on one real session, both of the
  files it named, one read in full by cat, one created by a heredoc and read by sed, beside 658
  shell commands it could not classify. The text report stopped saying so in 0.1.1; the JSON, which programs and other
  models read without the caveat a person weighs, kept the bare list. edited_without_read is now
  empty for Claude Code and unchanged elsewhere, and the finding moves to
  edits_with_no_recorded_read: the files, whether the harness refuses an unread edit, the count of
  unclassified commands, and a sentence saying what they mean. Where a harness does allow an unread
  edit, the text report names it as "No read recorded before editing", beside the unclassified
  commands that could hold the read.

assurance 0.1.25

pip install assurance==0.1.25 · PyPI

• Requires assurance-budget 0.2.19: the Stop hook says when Claude claims the tests pass right after
  they failed, whether or not it saw an edit. The README's hook command and the plugin are pinned to
  this version.

assurance 0.1.24

pip install assurance==0.1.24 · PyPI

• assurance reach (assurance-reach 0.1.0, new, and assurance-cli 0.6.5): what a change to a
  file or folder reaches, by a code graph such as Graphify's, with each hop's call site, what the graph
  read kept apart from what it inferred, how far behind the code the graph is, and what could not be
  determined. The README's hook command and the plugin are pinned to this version.

assurance 0.1.23

pip install assurance==0.1.23 · PyPI

• The Stop hook says a finding once (assurance-budget 0.2.18), and the plugin's hook stands down
  when a hook in your own settings already runs the audit after every turn: Claude Code runs both,
  and each said every finding, so each showed twice.
• Requires assurance-budget 0.2.18: assurance audit --json no longer lists a Claude Code edit
  under edited_without_read, since that harness refuses an edit to a file the session has not read
  or written;
  the reader's unseen reads go under edits_with_no_recorded_read, with what they rest on. The
  README's hook command and the plugin are pinned to this version.

assurance-mcp 0.5.5

pip install assurance-mcp==0.5.5 · PyPI

• Listed on the MCP registry again. 0.5.4 reached PyPI, and its listing did not: its server.json still
  said 0.5.3, and the publish job refuses a manifest that disagrees with the release. A test now fails
  the pull request when they disagree. No change to the server or its tools.

assurance-cli 0.6.4

pip install assurance-cli==0.6.4 · PyPI

• assurance serve reaches assurance-budget's local endpoint, where any agent sends its traces or
  run records and any workflow asks for a run's audit. Listed in --help and on the start screen.