0.36.1

Web view: press o or run /agentmon open to see the session in your browser. The agent flow shows who started whom, the waterfall shows every call, and the top band reads context, quotas, cache hits, cost and turns against their limits or your usual sessions. Sections cover tool calls, verification, subagents, files and the session log. Every running session is one click away in the top bar.

It needs node on your PATH; everything else works without it. It runs on 127.0.0.1 with no login, so set web to off on a machine shared with other people.

README: rewritten around seeing what the agents do, with a moving picture, one section per web view item, and a new Cost and privacy section.

No release was published between 0.24.0 and this one.

0.24.0

Test review: a green run says nothing about a test edited to pass. When a task changes or removes an assertion, or adds a skip, and a later run over that test passes, the pane now shows a Focus item with the changed lines beside the result. It asks a person to look and sends nothing back, because some test updates are right. i drafts a question that asks Claude to justify each change. The list stays across turns until x puts it away.

Done gate: a run with fewer tests is now a review item when that turn edited a test file. With no test file edited, it still sends the turn back.

Limits: detection uses patterns, not a parser. Edits through a shell command and changes made outside the session are not seen. Every legitimate change to an existing assertion also raises the item.

0.23.0

Installing no longer opens the options dialog, so it works on a small screen. Everything starts on its defaults.

Language follows your terminal (LC_ALL, LC_MESSAGES, LANG) unless you set language. Saved options keep working.

To change options, ask Claude in chat, or edit pluginConfigs in ~/.claude/settings.json. Claude's edits apply at once, your own at the next prompt, with no /reload-plugins. debugLog, gitPollMinutes and gitFetch apply at the next session.

The options are no longer in /config or /plugin, and project, --settings and managed settings are not read.

0.22.0

Done gate: any passing test run no longer counts as evidence. Each edited source file belongs to its nearest package root, and that package needs a run after the last edit that reached it and executed at least one test. An optional .agentmon-tests file declares targets (glob, dir, full, min=N). Sent back as new gaps: an edited package no test reached, a run that executed no tests, and a command that ran fewer tests or skipped more than before the last edit, so a deleted test file or an added .skip no longer passes. Denied runs are not evidence; every test in a chained command counts. The pane, Fable's material and the bench's "verified" use the same check.

Also since 0.20.3 (0.21.0 had no release):

Pane: calm layout by default, classic with v or the header button (option layout). Calm keeps sections in a fixed order and hides insights, skills and usage unless usage needs attention. Past four running subagents, the rest fold behind a. Option refreshRate (live, 5, 15, 30, 60 s).

Done gate and privacy: a Fable answer leading with OK passes again; "OK, but ..." does not. Fable's suggested fix is labelled as Fable's when sent to Claude. Masking covers JSON-quoted keys, Authorization: token, more token shapes, URL signatures, whole PEM private keys and password flags.

Learning: /agentmon learn skips sessions that already have a bench row.

0.20.3

Focus: one error behind different calls shows as a loop (first error line, numbers and paths taken out), and an error met in three or more sessions suggests a CLAUDE.md note; a turn with far more tool calls than this project's turns usually take is called out; the cache hint judges this turn's requests from the second on; re-reads reset at a compaction; the approval hint names this turn's calls only.

Bench: cost per session, each request priced from its tokens at its model's list price, so sessions read from transcripts get a figure too; tool calls per prompt.

Done gate: Fable is asked only when it can judge, a stop it did not judge is no pass, and after a failed answer it is left alone for five minutes; a declined call is no tool error; /agentmon clear mid-turn resets the whole turn; a turn under way at a plugin reload keeps its clock and gate.

Learning: a turn is saved even when the engine gives no usage; the stored copy is trimmed; /agentmon learn parses each transcript once, backfills error signatures, and writes only keys that survive the prune; a half-done migration no longer overwrites a newer key; git fetch names refs/heads/.

0.19.0

Debug log: each session writes a metadata-only log under agentmon-live/debug//, pruned after seven days; /agentmon debug prints the path.

Learning: one store key per session and per bench row, so sessions in one project no longer overwrite each other; a full store gives up the oldest keys of other projects; old single entries migrate.

Internal: locales split into hooks/lang/.ts; turn clock, release checks and the done gate moved out of register.tsx; review fixes.

0.18.2

Release check: after an install or update the mod asks GitHub again at
the next session start, instead of trusting a release it remembered from
before. A newer release found is kept a day, anything else an hour.
Before, a remembered older release could hide the update button for up
to a day.

0.18.1

Memory check: the memory and instruction files are checked again at the
end of each turn, not only at start and after an Edit or Write. A file
trimmed with Bash, an editor or a git checkout now clears its Focus item
(such as "large memory file") at the next turn end. /agentmon memory
still checks right away.

0.18.0

Base branch check: Focus says when origin's default branch (else
origin/main) has commits this branch lacks. When those commits changed
files edited in the session, the item turns red and names them; i
drafts a prompt to show those changes, rebase and rerun the checks.
It checks every gitPollMinutes (default 5, 0 is off) and after each turn.
By default it compares against your last fetch; gitFetch fetches the base
branch first (never with CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC set,
never asking for credentials).

Also: secrets in camelCase or prefixed names (accessToken, PGPASSWORD)
are masked, quoted values with spaces included; a turn any Stop hook
sent back stays the same turn for the done gate.

0.17.0

Update check: when a newer release is out, the header shows update
next to the version. Click it (or press u) to draft the update commands
into the prompt; nothing is sent until you send it. changelog (l) folds
out the new release's notes.

The mod asks GitHub's releases API at most once a day and skips it while
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC is set; a failed check shows
nothing.