Skip to content

Release 1.5.1 - #475

Merged
LoopGlitch26 merged 140 commits into
mainfrom
agent/issue-151-release
Oct 30, 2026
Merged

LoopGlitch26 merged 140 commits into
mainfrom
agent/issue-151-release

Conversation

@LoopGlitch26

@LoopGlitch26 LoopGlitch26 commented Oct 29, 2026 •

Copy link
Copy Markdown
Member

Release 1.5.1. The full notes are in CHANGELOG.md, under "1.5.1 (2026-10-30)".

Why: a week of running Tracekit on its maintainer's laptop. The coding-agent hooks held 208 calls; each waited about 8 minutes for an answer nobody could see, then was blocked, and none was approved. The same week, tracekit view held a core at 100% on a 414k-record store, and the policy stopped calls for text they only wrote or searched for.

Held calls: the hook blocks a held call at once with the commands that answer it. Once a person approves, the agent's retry of the same call (same run, tool and arguments, same policy) runs once. A retry is told who said no, a yes is not reused once a later no answered the same call, and one approval completes one call (a second result is a signed gap). TRACEKIT_APPROVAL_WAIT keeps the old wait where a person answers in time.

New:

  • Admin-managed hooks for Claude Code (init --v2 --managed [--managed-only]), checked by doctor.
  • tracekit report (experimental): a week of a signer's store on one page.
  • Holds on moving money (Stripe, PayPal, Square: TK-A043 to TK-A045; server-comms TK-M005), whatever the API key; TK-A046 holds a download piped into an interpreter given code.
  • The viewer: HORIZON zooms and pans (folders show on hover or zoom), INFLUENCE answers in plain words first and draws only the links that matter, and docs/use-cases.md with videos.
  • A CI job for the Go verifier.

Policy: TK-D002 and TK-D007 read the parsed command, not raw text, so they no longer fire on text a command only writes or searches for. A sed script that writes, reads or runs, a heredoc run in the same call, a shell told to run stdin, a git alias's command (from -c, git config or the environment) and a path with . or .. parts are read as what they do. None of 9,924 held-out, E17 and session decisions is weaker than 1.5.0 on either engine; a 752-case adversarial sweep found only TK-A046's holds of inline code, which hold rather than deny by design. E21's known pool: 247 of 420 (default), 275 (strict), no new false alarm.

The viewer's cost: it reads only what a store appends, checks the bytes it hashed (a store edit timed around a check still turns the run FAILED), re-checks within a CPU budget, and keeps at most 64 MiB of growing runs' bytes. On the maintainer's 520,000-record store: 50-100% of a core for 14 minutes with the page in use, then 0-4% (it held 100% before).

Deep review before release: five independent areas (policy and parser, signer and approvals, the viewer's store reading, the page, verifiers/CLI/install), each fix re-reviewed by its finder, plus two Codex adversarial passes over the policy. Every finding is fixed with a test that fails without it.

Checks on 78ea370, on a laptop (the laptop run had all its code, the last two commits being the status message and a CHANGELOG line; since then the head, 5fd05d6, adds: docs/use-cases.md with a still frame per video (videos played from GitHub Pages) and the wiki script's image-in-link fix; seven tests made to pass on Windows; the TS adapter's MCP name trim (CodeQL ReDoS alert); two viewer test races fixed at their cause (a test-harness check counted as the viewer's, a taken-back row read as a row); a signed run's "being built" graph message in a viewer with why logs; and the veto delay's "wait 61 s"; each with a test that fails without it, and the last four reviewed independently):

  • make test-ts lint build test: 4,002 Python tests passed, 382 skipped, 0 failed; TypeScript 145 passed, 0 failed.
  • All 389 v2 bundles the tests exported verify.
  • The offline evals CI runs, re-run on this head: every gate passes; the changed results are committed and the README and docs follow.
  • Hands-on as a user: a fresh venv from the built wheel with an isolated HOME (init, doctor, status, a scripted agent's 12 calls with the expected verdicts, report, export), and the viewer on the maintainer's real store in a browser (HORIZON, INFLUENCE, DESK), at desktop width and in both themes.

After merge: tag v1.5.1 on the merge commit, which runs the Release workflow (container image, PyPI, then npm). Then publish the two advisories (PCM replay; witness count).

…e, a weekly report

- Held calls block at once with the commands that answer them; once a
  person approves, the agent's retry of the same call runs once.
- tracekit init --v2 --managed puts Claude Code's v2 hooks in its
  admin-managed settings; doctor checks them (D-HOOKS-MANAGED).
- tracekit report (experimental): a week of a signer's store on one page.
- TK-A014 holds only Bash, PowerShell and Monitor; coding pack
  2026.10-v2-40, server -10, strict -4.
- Trust configs: one witness is one key under one name with one class,
  in every verifier; PCM replay senders are keyed by public key.
- Viewer, browser verifier, CLI messages and the transcript reader:
  fixes from a hands-on QA pass. CI job for the Go verifier.

See CHANGELOG.md for the upgrade notes.
Stripe holds an agent's money calls for a person only for agent-tagged
keys; a coding agent with a developer's own key got no hold, and every
pack only flagged the network command.

- TK-A043 (coding pack): the Stripe CLI's money commands (refunds,
  payouts, transfers and topups create, charges create/capture,
  payment_intents create/confirm/capture, invoices pay, treasury
  outbound payments and transfers), stripe post to those endpoints,
  and a POST to them on api.stripe.com, to PayPal's payouts, captures
  and refunds, or to Square's payments and refunds, through curl,
  wget, HTTPie or xh. Reads, stripe listen/trigger/logs tail/login and
  GETs (curl -G as Stripe's docs list with it) aren't held.
- TK-A044: Stripe's MCP money tools (send_money, transfer_money,
  convert_currency, create_refund). TK-A045: stripe_api_write when it
  names a refund or one of those endpoints.
- Server pack: TK-M005 holds a payment tool such as issue_refund called
  with no amount (a full refund), which TK-M001 let through; TK-M002
  holds those names over MCP. With a numeric amount TK-M001 alone holds
  it, so agent-approver bounds on TK-M001 still apply.
- The hold says what to check: the amount, currency and payee in the
  command, and that a call with no amount moves what was set earlier.
- Packs: coding 2026.10-v2-41, server-comms -4, server -11, strict -5.
- E17: no verdict changes; two MCP create_refund cases now also name
  TK-A044. E21: recall and false positives unchanged; results
  re-recorded for the new policy hashes, with that run's latency
  (evaluation.md and the README's E21 p99 quote it).
- Docs: coding agents (calls that move money), policy-v2 rule lists,
  three limits rows.
…t_cents

- TK-A043: a POST option before the URL now reaches it across any words
  that can't start another POST option of that client, so an empty
  argument, an indented body or a long one-line body no longer breaks
  it; each stretch of the command is crossed once (linear, no bound).
  After the URL the option stays within sixteen space-separated pieces.
- POST options are each client's own: HTTPie data items only for
  http/https/xh/xhs (curl -w, -b, -H values and wget -e were held), and
  HTTPie's and xh's -F/-d and wget's -d/-nd are no longer body options.
- `stripe <resource> [<operation>] --help` (or -h) with nothing else
  isn't held; help anywhere else still is, since an option takes the
  next word as its value.
- transfer_reversals create, and any number of slashes after the host.
- TK-M005: an empty amount with an amount_cents was left to neither
  rule; the unless now reads amount_cents only with no amount key.
- The remedy says to check the command or the file it sends.
- Docs: the limits row names the window, the reach-stopping options,
  the Stripe CLI in a container, and the help form.
- E17: no decision changes. E21: recall and false positives unchanged;
  re-recorded for the new policy hashes.
A failed line's headline counts its problems instead of repeating the pass
text. A trust file that can't be used is UNVERIFIABLE (the trust file can't
be used), exit 2 as before, with its problem named and no exception names.
A damaged zip says the bundle file is damaged. A failed bundle's printed
assurance is not rated. The browser uses the CLIs' UNUSABLE BUNDLE word,
names a self-approval as the dev reason and says an approval expired. The
Go verifier names records by log seq, prints the reason before the verdict,
fails a log key pinned as a witness at the checkpoint (exit 1, as Python and
TypeScript do) and clamps witnesses_required before converting it.
The file pickers are labelled buttons with the chosen name beside them, so a
name no longer shows twice. The page gets the viewer's tokens (dark, and
light when the system asks), a verdict banner, coloured statuses and a phone
layout, all in the page's one hashed style block: no new dependency, still
offline.
TK-D008 denied any shell write in ~/.claude/projects/<project>/, so the
agent couldn't append to its own memory notes (memory/MEMORY.md and the
notes beside it); TK-F015 and TK-F009 flagged those notes as harness
settings. TK-D009 already let the file tools write them.

- TK-D008 and TK-F015: an `unless` for a .md file directly in a
  project's memory/ (the same notes TK-D009 leaves out), per write or
  delete target, so a command that also writes a transcript is still
  denied. TK-F009 the same for the file tools.
- effects: the link ln, cp -s or cp -l makes is also MAY_WRITE, which
  only TK-D008 reads, so a link made as a memory note (pointing at a
  transcript) is still denied, as it was before.
- Every transcript write the tests covered is still denied; new cases
  for an append, rm, truncate, mv of the project folder and of
  ~/.claude/projects, a subagent transcript, paths through memory/..,
  and links made in memory/.
- Coding pack 2026.10-v2-42. E17: no verdict changes (policy hash only).
  E21 rewrites: none.
The latency gate held each pack's p99 in microseconds to the recorded
run's plus 25%. Packs grew until dev's p99 sat at its cap (140 and 150
us), and the same packs measured 141 to 156 us on the recording laptop
under ordinary load, so make eval failed at random.

- Each call's fastest of five decisions is now followed, in each try,
  by a fixed reference workload (JSON, shlex and stdlib re over one
  command; no Tracekit code). p99_ref is the p99 of decision time over
  reference time, and the gate holds it within 25% of the recorded
  run's, on the same machine, re-measured once before it fails.
- On the recording laptop, every core busy raised p99 up to 1.7 times
  and p99_ref at most 10%; efficiency cores alone raised p99 up to 4.5
  times and p99_ref at most 14%. 60 literal-free rules added to dev.yaml
  doubled dev's p99_ref (5.53 and 5.18) and the gate failed; 15 raised
  it about 20% and it passed.
- The gates' max_p99_us becomes max_p99_ref, recorded from this run;
  CI's Linux job still skips the machine-bound gate and checks its own
  1 ms threshold. A failure says what to do.
- Results regenerated: the packs' new hashes, the new latency fields;
  recall and false positives unchanged. Docs and the limits row follow.
…it is

init --dev --v2 --no-hooks records no_hooks in the client config; doctor's
D-HOOKS-PRESENT then passes, says what is still recorded and how to add the
hooks later, instead of telling the person to install what they declined.
tracekit export printed the export's summary as raw JSON. It now prints the
bundle, what it holds and the commands that verify it (v2: signer trust and
verify --trust; v1: verify --key). --json keeps the JSON, which a test and
format-v2.md rely on.
tracekit pcm --help printed the experimental line above a help that says
(experimental) itself. The line is for running a command; help is not a run.
…'s commands use it

Started with --data-dir, the viewer's EXPORT and TRUST commands (and the
report's) named --config signer.yaml, a file the person had to write first;
the header's COPY gave them without the note that said so. export --v2,
signer trust and signer vkey now read a bare data dir as the viewer does, so
the copied commands run as they are. A data dir names no socket or
witnesses: export says it can't ask for a checkpoint, trust says it pins
none (docs/limits.md).
…ace says so

A retry inside an agent approver's veto delay read requested, and the hook
said to approve with tracekit approvals approve an approval already approved.
approval_wait's reason (in the RPC contract already) now says it is approved
and how many seconds are left; the hook looks once when it did not wait, and
says to wait that long, then retry.

When two retries race for one approval, the loser read "approval_id not bound
to this call"; the refusal now says the approval was already used by another
attempt of the same call.
A Bash move took the first word with a slash as its place, so
git checkout -b feature/x read "ran feature/x". In a Bash command HORIZON
now takes only a path no branch or ref looks like (~/, /, ./, ../, a dotfile,
dir/file.ext), skips a leading cd and /dev/ redirects, and otherwise names
the move by its program and verb (git checkout) on this machine. A Bash
move's rows show its command.
As root, the child that runs an edit as the agent's user ended with
os._exit, which flushes nothing: with stdout piped, its message ("Tracekit v2
hooks removed from ...") was lost. The parent flushes before the fork, so the
child repeats none of its output, and the child flushes before it exits.
With a long --data-dir path (a macOS dev data dir), the header's EXPORT and
TRUST commands took the width they wanted and left the run's name and times
a column of single words.
…that they are

Codex skips a hook that isn't managed until a person trusts it in /hooks
(learn.chatgpt.com/docs/hooks), and says nothing: after init --agent codex no
Codex call was recorded or decided. Init (v2 dev and system, v1 dev) now ends
with the step, and doctor's D-CODEX-HOOK-TRUST reads Codex's trust store
(~/.codex/config.toml, [hooks.state."<hooks.json>:<event>:<group>:<handler>"]
trusted_hash and enabled) for each of Tracekit's hooks, and says when it
can't read it.
Since 1.5.1 a coding agent's hook blocks a held call at once and the agent
goes on; the approval stays open so the agent's retry can run. NEEDS YOU,
the alerts, the toast and the badge still said "<agent> waits for a person"
and "WAITING FOR YOU". They now say the call is held and that approving it
lets the agent's retry run (badge: TO ANSWER n); APPROVALS, PERMISSIONS and
the run review say "not answered yet", the outcome view.held_outcome and
review() now count (view.OPEN), and tracekit approvals list no longer says
"nothing is waiting".
…heir last parts

With about 100 places, many long worktree paths, the desktop drew a callout
for every restricted place whether or not it had room, and the labels piled
into unreadable text. Callouts now go busiest first (a clicked place first),
each only where it covers nothing drawn, at any width; a long path shows its
last two parts, and hovering a place names it in full.

Also: a held call's status reads TO ANSWER (it read NOT ANSWERED YET, cut off
in the tape's STATUS column).
With the banner off for --help, 21 experimental commands said they were
experimental nowhere in their help. Each one's help now ends with the line,
drawn from the list the banner uses, so the two can't drift; pcm's
description no longer says it a second time.
client.json's hooks section now carries no_hooks, as dev mode's client config
does, and doctor's D-HOOKS-PRESENT passes with what is still recorded and how
to add the hooks, instead of "no hooks recorded in client.json" and a fix
that installs them.
…eckpoint covers it

A run waiting for its re-check after the store's bytes changed read "Pending: no checkpoint covers this run's last
record yet", as a run no checkpoint covers does. Its review sentence now says it is being checked again because the
store's bytes changed after the viewer read them.
…ame; a URL's host is read from the whole word; RECORDS waits for a re-check

- Replaying, the folders no move had reached yet were full circles with no name: empty circles that read as clutter.
  Until a move reaches it (or it's picked, or restricted), a folder is a hairline with no fill, dots or name.
- Re-review of the loopback fix: the words were split at = , ; & | ( > " ` ' before a URL's host was read, so
  http://localhost=@evil.example read as localhost (curl connects to evil.example) and stayed out of EGRESS. A URL's
  host is now read from the whole word: its authority to the first / ? or #, the host after its last @.
- A selected run being checked again kept its old review's RECORDS; it reads … until checked.
- test_view: the rows test waited for the watcher's review row, which can come from a check at an earlier checkpoint,
  so rows() checked the last records itself under load (1 in 2 full-file runs); it waits for the watcher's check of
  every record.
…he others keep theirs

Re-review of the cap: a run over BUNDLES_MAX alone was kept, then every other run's bytes went before its own, so
the signer's own run (174 MB, never final) emptied the cache at each of its checks and each other growing run was
read whole at its next one.
…reen, a PATH list is its first folder

From the owner's look at the 1.5.1 preview:
- The readouts read HOSTS OUTSIDE, RESTRICTED (N TRIED · N RAN), BLOCKED and HELD (EGRESS, FENCES HIT and HOLDS made
  people think, and DESK says BLOCKED and HELD); the side panel no longer repeats two of them, AGENTS and PLACES only.
- Zoomed in, the rings' arcs had no names (each was named once, upper left): a ring's name now goes on the part of it
  in view when its upper-left spot is off screen.
- `PATH=/x/.venv/bin:$PATH` drew a place named bin:$PATH; a PATH-style list is its first folder.
- Another tool's call is placed by its arguments in order again, its subject first: the previous commit read a URL in
  any argument before the subject (a Write to ~/.claude/settings.json was drawn at a host it mentioned).
…rs on hover, pick, restriction or zoom

The owner's look at the preview: every folder's circle at FIT made the map messy and hard to read, and nothing said
what the circles were. A folder's circle now shows when the pointer is on it or a place in it, when it is picked or
restricted (its shield), and for all folders from 1.6x; a folder with no circle is a dot. The key says what a folder
is and how to see one.
… how we know, move around)

The owner's look at the preview: the key read as one cluttered list. It now groups the marks (in two columns), the
map's shapes (restricted place, data leaving, folder, rings), the grades, and how to move around, each a few words.
The demos were organised by feature; people recognise their own problem first. docs/use-cases.md names eight
problems in the words of the teams that have them, what goes wrong without Tracekit, what Tracekit does (the rule or
feature), and how to see it: the cover-up, PocketOS, refund/evidence, health and OpenShell videos (the four new ones,
3-6 MB each, in docs/use-cases/, as docs/demo/ keeps the cover-up), the recorded runs in docs/demos/ and the commands.
The wiki lists it first on Home and opens its Use cases and demos section.
…ore were blocked or held

INFLUENCE's answer line says "Sub-agents' messages (untrusted) came before 48 blocked Bash calls in 3 runs". The
groups gave the stopped calls per tool but not the runs they were in, so one runaway run and a pattern across many
read the same. influence_graph's groups add stopped_runs: per tool, the distinct runs with a call of it that ended
blocked or held (flagged calls ran, so they are not counted).
…w over the table

A supervisor asks one thing of INFLUENCE: did something my agent read come before a risky call? The page was a table
of chips under "CITED BY" and "REACHES". It now opens with the answer: the untrusted content that came before the most
blocked or held calls ("Messages from another session (untrusted) came before 101 blocked Bash calls in 54 runs"), a
replay-confirmed cause first with its effect, each line opening its row; or "Nothing untrusted came before a blocked
or held call." Only a replay's confirmation says "caused"; everything else "came before" or "led to".

Under it, a flow: each kind of content on the left (its trust and sources), the tools called after it on the right.
Only the lines that matter are drawn, coloured as the chips (blocked red, held amber, flagged blue, a confirmed cause
bold red with its effect, a ruled-out link dashed), as thick as their runs on a log scale. Drawing every plain line
made a hairball at 12 kinds and 40 tools: they are hidden, counted in a note, and show for the content or tool pointed
at. The tools with a marked line come first, then four quiet ones, the rest one "+N more". Pointing brightens a node's
or line's lines and dims the rest by classes, with no redraw; clicking content or a line opens its row, scrolls to it
and marks the tool's chip; clicking a tool marks it in every row; empty space clears. Its data attributes are indices,
never a name an agent chose. Panels under 600 px leave it out.

The table is the detail, in plain words ("in front of the agent for 247,045 decisions", "led to"; a tool's output
kind reads "output of Read"). The page is put, not rewritten, so a reload of the same data draws nothing again.
…xt and other kinds have names in words

From the owner's preview: "Across all 1 run", "came before 1 blocked Bash calls", and "paste" where every other kind
of content had a name. Named too: skill instructions, files the prompt referenced, tool results and errors, saved
tool output, sandbox violation notices, Slack and channel messages, scheduled check-ins.
Right after init --dev --v2, tracekit status printed a raw [Errno 2] naming the dev signer's runtime socket, though the dev signer starts at an agent's first call. It now says so, and names tracekit up. A configured or system signer that can't be reached, and any other failure, keeps its own error.
…s; plain headings

GitHub serves a repo's .mp4 as a download, so the videos now live on gh-pages (videos/) and each use case links a
still frame to its copy there. The wiki script relinks an image inside a link, which it left as it was.
A Windows path in a regex, permission bits Windows doesn't have, a rename over a file a reader holds open (Windows
refuses it), a signer on a Unix socket, and a path compared as a string.
/^[_-]+|[_-]+$/g backtracked over each run of _ or - inside the name (CodeQL js/polynomial-redos, open since 12 Oct):
7.3 s for a 150,000-character name, 1.5 ms now, with the same result on 20,014 names. The test fails without it.
…p it from being imported

With --why (or a why/ folder in the data dir) the viewer imports no signed run, but OVERVIEW, GRAPH, WHY and
AGENTS of a verified run without a why log said "This run's graph is being built…" forever. The snapshot's
signer now carries `why` in that mode ("--why", or the data dir whose why/ is read), and the page says
"being built" only when the viewer imports signed runs; otherwise it says why the run has no graph and
names what shows it: `tracekit view` without --why, or `tracekit why import --store … --run … --out NEWDIR`
then `--why NEWDIR`. runCmds' signer flag is shared as signerArg. docs/viewer.md and the limits row say the same.
…ay not read, and never waits on a checkpoint

Review of the previous commit: with --why X on a data dir whose why/ is read too, the page named the data dir's why/;
it now says only that the viewer reads why logs. A run no checkpoint covers yet read "built once a checkpoint covers
its records" in why-log mode, where no checkpoint brings a graph. The commands say to keep the viewer's other options,
and the limits row says the views' command is a file store's.
… take them back when it catches the edit

A re-check after the store's bytes changed puts RETRACTED (an empty dict) in the run's place, so reading the tape after
the edit raised KeyError: 'session_id', or found the rows taken back and not yet shown again. The test now keeps each
row appended for the run and checks that the edited decision never reached the tape.
…ifier

The watcher checks runs on its own thread with the same patched bundle_records, so its entry set the shared flag while
rows()'s export ran the harness's check of the exported bundle (no memo: every record) on the test thread; those
schema checks read as rows() checking again (CI: seq 3 to 11). Forced, the old test counted seq 0 to 11.
…he rows check asserts rows() checked the run

From review: the tape test's report() read the tape through this helper, which hit RETRACTED ({}) as KeyError when the
viewer caught the edit first; and the rows check passed without rows() reaching its check at all.
…ot a second more

not_before is kept to the microsecond; the wait subtracted an unrounded clock, so a wait read at the answer's own instant (Windows' coarse clock) said 61 s for a 60 s delay. Both sides are microseconds now.
@LoopGlitch26
LoopGlitch26 merged commit 2fb5c06 into main Oct 30, 2026
57 of 58 checks passed
@LoopGlitch26
LoopGlitch26 deleted the agent/issue-151-release branch October 30, 2026 10:57
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant