`run_script` and `run_streamed` both left their select loop on stdout EOF
and then ran unbounded post-loop drains, so `child.wait()` sat downstream
of a read that a surviving descendant can hold open forever. Pipe EOF is
not the child's to give: the write end is inherited by every descendant
spawned without its own redirection, and `hermes update` deliberately
runs its build steps with stdout inherited. One resident gateway stranded
the whole update, exit code included.
Both now go through one `pump_child`, which takes the exit status from
waiting on the process and bounds the drain from the moment it exits — a
slow child is not a stuck one, so nothing is metered while it runs. An
abandoned drain says so in the log rather than silently truncating.
Cancelling was not an escape hatch either: `start_kill` reaches the
child, not the grandchild with the handle, so the bounded drain is what
lets a cancel return at all.
Same bound `Invoke-HermesStep` grew in windows.ps1 (#90455), and the same
shape as Go's `exec.Cmd.WaitDelay`.
The loading placeholder had its own copy of the row grid — its own min-height,
its own two columns, `pl-2` and no trailing inset — so its right-hand block sat
several pixels off from where the real rows land. The list stepped sideways as
sessions resolved.
Compose the shared row chrome instead, which is where that inset lives.
"Load more" and a workspace's "show more" hang off the bottom of a list rather
than sitting in a row, so they never saw the shell's trailing inset and stayed
flush against the edge every row above them now stops short of.
The cron row had its own copy of the row grid — its own min-height, its own
`grid-cols-[minmax(0,1fr)_auto]`, its own `pl-2 pr-1`, and a comment explaining
that the numbers were chosen by hand to line up with the session rows above it.
They had already drifted apart on the right edge.
Compose SidebarRowShell / SidebarRowBody / SidebarRowLead / SidebarRowLabel
instead, so a cron job and a session share one definition of what a row is and
cannot drift again.
SidebarRowShell owned the row's height and, through the body, its leading
inset — but nothing owned the trailing one. The actions slot rendered with no
padding, so the age, chips and kebab sat on the row's border box. That is the
same pixel a working row paints its arc on (`.arc-row` sets `--arc-standoff:
0rem`), so the animation ran straight through the text.
Give the shell that inset. It is the only box containing both the one-line
row's actions column and the card variant's in-body cluster, so one class
covers every trailing thing a row can render. The card drops the body's
label-to-actions gap in exchange: it has no such column to clear, and keeping
the gap would pull its header in past every line below it.
Adopt upstream's one-hidden-forever-chat model and remove the per-bot hidden
session browser while preserving connection-bound remote Bot opening, the full
remote action menu, same-name isolation, and global Sessions navigation.
Disband removed the room log and each member's group membership, but
BotsPane's mergeServerMeta then overlaid the STALE cached roster snapshot
(fetched before the disband) whose ui_meta['hermes-bots'] still carried the
old groups array — spreading it back over local meta and re-listing the
group as an empty row. Any later meta write for the bot (pin, title,
canonical-chat pointer) re-uploaded the resurrected membership server-side,
making the ghost permanent. Same stale-overlay class could revert renames,
re-group left members, and undo pins/hides.
Fix (class-level, not per-field):
- saveBotMeta stamps a per-bot last-local-write time (re-stamped when the
profiles.configure write settles).
- useRoster stamps each snapshot with its fetch ISSUE time (fetchedAt).
- mergeServerMeta skips overlaying any bot whose local write post-dates the
snapshot; the next (fresh) fetch overlays normally, so server truth still
gets the last word. No-fetchedAt callers behave exactly as before.
- disbandGroupChat / renameGroupChat invalidate the roster query so all
surfaces converge on a fresh snapshot immediately.
Tests: new regression (stale snapshot cannot resurrect a disbanded
membership; fresh snapshot still wins) proven to fail without the fix via
sabotage run; fence-off compatibility test; hide-bots shape regex updated.
341/341 plugin tests pass.
`hermes update` runs in the pre-pull interpreter. The auto-restart phase
imports freshly-pulled gateway source, which resolves sibling imports
against the OLD sys.modules cache — so any update where an already-cached
module gained a new export ImportErrored the whole phase and left the
gateway serving pre-update code (2026-08-20 field failure: new gateway.py
needs cli_output.line_input, cached cli_output predates it).
Class fix replacing the per-symptom _UPDATE_RUNTIME_RELOAD_MODULES
approach: _purge_stale_hermes_modules() evicts every cached module under
the Hermes package prefixes (hermes_cli/gateway/tools/tui_gateway/agent)
right before the restart phase, so later lazy imports rebuild a
self-consistent module graph from the updated checkout. The updater's own
executing modules are exempt (purging them buys nothing; reload-in-place
is the unsafe op, and we never reload). Root-segment check spares
prefix-lookalike packages. Best-effort, never raises.
5 new tests incl. an end-to-end repro of the field failure shape
(stale module missing symbol -> ImportError -> purge -> import resolves).
tests/tools/test_website_policy.py and tests/cli/test_surrogate_sanitization.py
repeatedly failed under the parallel CI runner (process-teardown timeouts /
async-timeout flakes) across multiple unrelated PRs this cycle, while passing
locally. Removed per maintainer direction to stop the flake taxing every PR.
posix.sh now probes `update --help` before the real update call; the fake
counted the probe as call #1, shifting the exits.N mapping so the retry
gate never fired. Answer the probe out-of-band so counted calls remain
actual update attempts.
The desktop updater ran `hermes update --yes`, which auto-restored any
uncommitted source-tree edits onto the freshly updated checkout. On dirty
from-source installs this silently carried local modifications across every
update and could break the rebuilt app (field report: Windows update handoff
leaving the app 'crashed').
New `hermes update --keep-stash`: local changes are still autostashed so the
update can proceed, but are never re-applied — they stay parked in git stash
with printed recovery guidance. Both desktop handoff scripts (windows.ps1,
posix.sh) now pass it, probing `update --help` first so older installed
backends without the flag keep working. Failure paths are unchanged (stash
preserved, no restore); updates.non_interactive_local_changes: discard still
wins.
Tests: park/restore/failure-path coverage incl. a sabotage-verified
regression test; docs updated.
Builds on @fattchris resolve_turn_limit salvage (#67696): flips the default
from a numeric cap to unlimited across all construction paths (CLI, agent_init,
run_agent subagents), adds inf/infinity/null to the unlimited spellings, and
sets DEFAULT_CONFIG agent.max_turns to null. The turn cap caused more problems
than it solved (silent mid-task truncation).
Addresses teknium1 sweeper review on PR #67696:
1. Gateway bridge: Skip str(None) bridging when YAML value is Python None
(from or bare ). Previously str(None) → None → unlimited
instead of default 90. Now clears stale env var so resolver applies default.
2. TUI: Route _cfg_max_turns through resolve_turn_limit instead of bare
int(). Old code crashed on none/unlimited and swallowed 0 via
. HERMES_TUI_MAX_TURNS env var also routed through
resolver.
3. Docs: Document unlimited spellings (none/unlimited/infinite/0/-1) in
configuration.md.
4. Tests: Add TestGatewayBridgeNullHandling (4 tests) and TestTUIResolver
(8 tests) covering null handling, string spellings, env var override,
and legacy root-level config.
All 50 tests pass.
Previously agent.max_turns only accepted positive integers. Setting it to
'none', 'unlimited', or 0 — all natural ways to say 'no limit' — either
crashed int() or was silently skipped by `or` checks, falling back to 90.
This adds resolve_turn_limit() in hermes_cli/config.py as the single
normalization point. It accepts:
- int/float → int(raw) (floats truncated)
- numeric string ('120') → int(raw)
- 'none'/'unlimited'/'infinite'/'∞'/'-1'/'0' (case-insensitive,
whitespace-tolerant) → sys.maxsize sentinel
- YAML None/null → default (90)
- bool/list/dict/garbage → default (with debug log)
All config-reading sites (cli.py, gateway/run.py, cron/scheduler.py) now
call this instead of bare int(), so agent.max_turns: none in config.yaml
becomes a first-class supported spelling of 'unlimited'.
The sentinel (sys.maxsize) survives the str()→int() round-trip through
the HERMES_MAX_ITERATIONS env-var bridge in gateway/run.py and works in
every <, >=, remaining = max - used comparison without requiring call
sites to learn about a special value.
Includes 38 tests covering the full spelling table, the str→int env-var
round-trip, and sentinel properties.
The two deepEqual tests pin the exact env object, so the new spawn tag
field made them red. Assert the tag in both shapes and add direct
spawnTag() coverage (winms-derived seconds, dash fallback, non-winms
markers rejected).
Every long-lived Hermes process is now positively identifiable so reapers
never have to guess lineage from PPID archaeology or cmdline shape:
- hermes_cli/process_identity.py (new): HERMES_SPAWN tag build/parse,
spawn-ledger.json self-registration keyed on (pid, create_time) — PID
reuse cannot forge the pair — with #89298-style corrupt-file quarantine,
and a kill-on-close job-object self-attach (BREAKAWAY_OK preserved for
the existing CREATE_BREAKAWAY_FROM_JOB escape hatches).
- serve/dashboard (web_server.py) and the gateway entry point register
themselves at startup and attach to the job; Desktop legacy
HERMES_PARENT_PID/winms marker reused as spawner identity so lineage
works with every Desktop version.
- Desktop stamps HERMES_SPAWN on backend spawns (parent-process-identity.ts).
- hermes update gets a positive-identity rung ahead of the heuristic ones:
_ledger_reapable_backend_pids reaps holders the ledger PROVES are orphaned
backends (purpose reapable + recorded spawner provably dead) in ANY update
context. Ledger-unknown holders fall through to the existing rungs.
22 new tests, sabotage-verified.
The one-shot keyless extract rescue (d1eefe6ac) treated ANY whole-batch
failure as a backend outage. A website-policy refusal also arrives as a
failed batch, so blocked URLs were routed through the free-tier ring:
in CI the ring's live fetch attempt returned a result for the wrong URL
or a bare None error, turning test_website_policy reds on main (slices
8/12 and 12/12) — and in production it would fetch content the user
explicitly blocked.
_rescue_extract now partitions policy blocks (blocked_by_policy flag or
policy error text) out of the rescue set: they are preserved verbatim,
only genuine failures ride the ring, and order/merge parity is kept.
Two sabotage-verified regression tests pin the class.
Approvals (pending_approval) had the identical blind spot as clarify:
blocked server-side in the hidden member session, invisible until timeout.
- syncGroupClarify mirrors pending_approval alongside pending_clarify
(clarify outranks when both appear); entries carry kind, command, the
server's choice set (once/session/always/deny, fallback once/deny), and
the runtime session id approval.respond keys on.
- The room card renders approvals as command-in-code + choice buttons
(closed set, no free text; deny tinted destructive) and routes
approval.respond via the member's own source.
- 6 new tests incl. an end-to-end blocked-on-approval drive; sabotage
run (approval mirroring disabled) fails all 4 behavioral tests.
Group members run in hidden plumbing sessions, so a member's clarify tool
blocked server-side with no surface to answer it — the room showed
'@lead is thinking…' until the 300s clarify timeout (salihsungur's report).
- The turn poll and the stranded-harvest pass mirror each member's
`pending_clarify` resume field into $groupClarify and hold the turn
deadline open while a question waits (bounded by the existing hard cap).
- The room renders a question card per blocked member — choice buttons,
free-text, batch sub-questions — and answers route via clarify.respond
through the member's OWN source (requestForBot), so cross-connection
members work. The answered exchange echoes into the room log.
- needs-you badges the roster row while a question waits; disband/rename
clear mirrored cards; older backends without pending_clarify no-op.
- 7 new tests incl. an end-to-end blocked-turn drive, sabotage-verified
(disabling the poll gate fails the drive test).
The no-live-shim probe defaulted to True (proceed with the reap) when
_venv_scripts_dir() returned None or the concurrent-instance detection
raised. Flip the default and the except-arm to False so an unverifiable
shim state keeps the updater refusing, matching the fail-closed contract
stated in the PR. Also fix the docstring: the hand-off gate lives in the
caller, not a function parameter.
Field incident (2026-08-20): a Windows Desktop update hand-off
(update --yes --gateway --force) left a swarm of per-profile serve
backends (mr-tester, probe-inherit, turqoise, clippy, maroon, …) holding
cryptography/_rust.pyd. Some still had a live parent (the tearing-down
Electron process, or the venv launcher->worker two-hop chain mid-exit),
so the strict orphan-only reap (_orphaned_desktop_backend_pids, which
bails the instant ANY holder has a live parent) disqualified the whole
set and the venv-holder guard dead-ended. The user saw a ~12-minute hang,
force-closed, and the half-done state stranded bot sessions.
New rung: _handoff_reapable_backend_pids reaps surviving Hermes
serve/dashboard backends from this venv — live parent or not — but ONLY
in the hand-off context the caller gates on: args.gateway AND the
update-incomplete marker present AND no live hermes.exe shim. In that
window nothing legitimate supervises or respawns a serve backend (the
Desktop tree-kills its backends and parks any relaunch behind the marker,
#50238), so a surviving backend is a leak, not a race. A non-backend
holder (operator REPL, stray script) still disqualifies the whole set;
psutil-unavailable returns None (keep refusing). Wired as the final rung
before the existing dead-end, after the orphan-only reap.
#90732 removed right-click → Sessions (one forever-chat per bot is the
product contract); #90756 cleaned the last in-app copy. This removes the
remaining docs bullet describing the dead affordance.
parseBackendOwnership returned [] for unreadable JSON and reapOrphans
unconditionally rewrote survivors — one corrupt read replaced the roster
with [], permanently orphaning every backend it described. The sweep now
detects corruption, parks the file as .corrupt (evidence preserved), and
skips the rewrite; empty/missing files keep the legacy sweep behavior.
PR #90732 removed the per-bot Sessions browser (right-click -> Sessions),
but the Disband-group confirm dialog still told users they could open the
kept 'Group: X' sessions 'from each bot's session browser' - an affordance
that no longer exists. Drop the stale clause; also update the one test
comment that still described the removed workspace.
Review follow-up from #90732 (found by the 3-angle review pass).
One conflict, gateway/relay/adapter.py send_for_platform: main added the
turn-final draft-seal interception (_sfp_metadata with the _interim_send
marker stripped, seal-or-fall-through); this branch added format-hint
stamping on the same frame. COMPOSED: the plain-send frame now stamps
_with_format_hints_for_platform over _sfp_metadata (the stripped copy),
so both the seal fall-through contract and the cron-lane block hints
hold. Note: the seal frame itself (op:draft final) does not stamp hints
— cron sends are never open drafts, so the flagship path is unaffected;
noted as a connector-PR follow-up for streamed interactive finals.
_scale_to_zero_is_idle() consumed _running_agent_count(), but cron jobs
run through a standalone AIAgent on the scheduler's own thread pool and
API-server runs live on the adapter — both outside _running_agents (the
same blind spot the #60432 shutdown-drain fix addressed with
_active_work_count()). The idle predicate therefore read True DURING a
running cron job; a suspend at that moment freezes the job mid-flight.
Observed live on staging 2026-08-20: is_idle held True throughout the
10:45:04-22 cron run — only watcher-tick timing (next tick 9s after
completion) avoided a mid-job freeze.
Use _active_work_count() (agents + cron + API runs). New tests cover a
running cron job and an active API run each blocking idle, plus the
all-quiet True case; both blocking tests fail without the fix.
seed_mock.assert_not_called() alone could pass for the wrong reason — a
harness failure before delivery also leaves the seed uncalled. Assert
the real adapter recorded exactly one live send to the origin chat, so
the test pins the D6 thread-fallback decision, not an accidental
no-delivery.
flyd mounts the local Machines API (flaps) socket at /.fly/api owned
root:root 0755, but the gateway runs as the unprivileged hermes user.
The scale-to-zero self-suspend (gateway/scale_to_zero.py suspend_self)
therefore failed every attempt with EACCES and an opted-in machine could
never sleep — fail-awake held, but the feature was inert. Verified live
on staging 2026-08-20: repeated 'flaps suspend request failed: [Errno
13] Permission denied' until a manual chgrp/chmod on the socket, after
which the same watcher tick suspended the machine cleanly (flaps 200 ->
suspension/suspended).
stage2 runs as root before the supervision tree starts: chgrp hermes +
g+w on the socket when present. Minimal widening — the socket stays
root-owned; no-op off Fly.
The #90732 adoption scan used session.list's 200-row recency window. A busy
bot profile (group-chat traffic, routines, or accumulated fork spam) pushes
an older forever-chat past row 200, the scan misses it, and the mint path
re-enters the unique-title-conflict fork loop — same pathology, higher
trigger threshold.
Profile → Named Session is an exact registry (UNIQUE title index), so
consult it exactly:
- session.list gains a `title` param: indexed WHERE title = ? lookup,
window-free, hidden rows resolve, archived/deny-listed do not,
compression lineages resolve to the live tip (resolved_id), mirroring
profiles.list's preferred_session resolver.
- findExistingCanonicalChat sends title: 'Bot Chat'. Older gateways ignore
the unknown param and return the windowed listing — the local scan stays
as the compatibility rung.
- Adoption opens the lineage tip (resolved_id) while pinning the durable id,
same split as the preferred_session path.
Unspecced MagicMock/AsyncMock adapters fabricate
supports_inchannel_continuable_for_platform as a truthy callable, so the
scheduler's duck-typed D6 gate silently took the relay accessor branch in
every in-channel test — the native scalar fallback the fixtures describe
was never exercised, and setting supports_inchannel_continuable=False on
a mock could not force thread mode. Pin the accessor to None on both
mock fixtures (matching a real native adapter, which never defines the
method), and add a fallback-boundary test with a real minimal adapter
class: scalar False -> in_channel fails safe to thread, flat seed never
fires.
`hermes update` on Windows detached on every run, including the
`Already up to date!` no-op that never touches the venv. emozilla hit
the visible half: the shim exits, PowerShell takes the console back,
and a child prints the result under a fresh prompt — it reads as a
frozen update. The invisible half is worse: the hand-off sat ahead of
the fetch, so it also carried off the stash and branch-switch
questions, which #90205 then had to answer by closing stdin. Nobody
who mods Hermes got asked about their local changes again.
The shim lock is real and the child is still required — a launcher
holds venv\Scripts\hermes.exe open without FILE_SHARE_DELETE for the
whole command, so the quarantine rename is refused and uv fails with
os error 32. A parent that waits deadlocks against the handle it is
itself holding, and Windows has no exec to escape with.
But that lock only binds one step. Move the hand-off to the dependency
sync boundary, beside the native-module deferral that solves the same
"this process holds a file the sync must replace" problem — and for
the reason that placement already exists (#86735: a preflight ahead of
the fetch re-bricked the flow it was meant to protect). Everything
before the sync now runs foreground in the user's console: the
preflight, the stash question, the branch switch, git pull. An
up-to-date run never hands off at all.
Deferring to the next launch cannot substitute here the way it does
for a mapped .pyd: every future `hermes` launch is also the shim, so
the marker would defer forever. The child re-runs the update to keep
the node/web/lazy-refresh tail, and takes the sync it was spawned for
rather than the up-to-date early return.
The seed-key fix made the SESSION scoped, but the delivery leg still
dropped the scope: cron route_metadata carried only job_id (+thread),
DeliveryRouter stamps scope_id only for the configured HOME channel, and
the RelayAdapter's per-chat scope cache is cold after a gateway restart
(learned from inbound only). A scoped Slack origin that is not the home
chat therefore egressed with NO tenant discriminator, and the connector's
fail-closed guard could reject the brief before delivery — the
delivery-leg sibling of the seed-key scope gap.
Copy origin.scope_id into the live text and media routing metadata for
ORIGIN-MATCHING targets only (setdefault — never overrides router/home
stamping). Fan-out/broadcast targets are excluded by the origin gate: a
fan-out target's tenant is not the origin's, and a wrong scope is worse
than none.
Tests: restart-shaped positive (scoped non-home origin -> scope_id on
routed metadata, RED before this fix) and legacy negative (scope-less
origin stamps nothing).
Symptom: switching between bots forked a brand-new "Bot Chat" for the
returned-to bot on EVERY switch, burying the user's real forever-chat
(one report: a 930-message chat displaced by 7 forks in one morning).
Root cause is a self-perpetuating loop between three parties:
- state.db enforces UNIQUE(title): the first fork permanently squats
the "Bot Chat" title; every later mint's title request is silently
dropped by set_session_title (returns 0, no error to the caller).
- The post-turn LLM auto-titler then names the untitled fork from its
kickoff content ("Assistant introduction request #2", ...).
- openBotCanonicalChat's identity check is title-string matching, so
the renamed fork reads as "not plumbing" -> corrupted metadata ->
clear pin -> mint again. Grandfathered pre-convention chats (real
history, derived titles) hit the same branch and are forked away
from immediately.
Fix, two invariants:
1. Adopt-before-mint: createCanonicalChat first scans the profile via
session.list include_hidden:true for an existing "Bot Chat" row and
re-pins it instead of creating. The UNIQUE index makes this an exact
registry lookup (at most one match), not a heuristic. Older gateways
without include_hidden find nothing and fall through to mint.
2. A pin that resolves to a NON-plumbing session carrying real history
is the user's conversation - keep it and open it (title drift is
metadata damage, not ownership loss). Only a pin resolving to an
EMPTY stray draft is treated as corrupted and replaced (which now
goes through adoption first).
Also removes the right-click -> Sessions per-bot stored-session browser
(ProfileSessionsWorkspace and its atoms/query/rows). Bot Mode's product
contract is ONE forever-chat per bot; a browser listing every hidden
plumbing session contradicts that and confused users into opening dead
forks. The Sessions workspace test goes with it; the include_hidden
source-shape test now pins the adoption scan instead, and a new suite
(canonical-chat-adopt-before-mint.test.mjs) covers both invariants plus
the older-gateway fallback.
An unset checkbox's `.value` is "on" per the HTML spec, and the inventory read
the value before the state — so a ticked box and an empty one both came back as
`value: "on"`, and the agent had no way to tell them apart. Anything built on
reading a form back ("is the newsletter box already checked?") was answering
from a coin flip.
The state check now runs first, gated on the input's type rather than on
`checked` being defined — it is defined, as false, on every input including
text fields, which is what made the original ordering look reasonable. ARIA
checkboxes and switches built out of divs read `aria-checked`, which the old
branch would have missed anyway since a div has no `.checked`.
Found by putting the extracted helper under a direct test. It was unreachable
before: the only way to observe it was to run a whole action against a whole
document.
act-in-page.ts had grown to 1,054 lines: one function holding twenty-eight
closures, of which roughly four hundred lines were pure string and geometry
work that never touched the holder, the action, or any page state. None of it
could be tested. `slug`, `affinity`, `coin` and the rest were reachable only by
running a whole action against a whole document and inferring what they must
have done.
Four modules now, at the seams the dependency graph actually has:
types.ts the six shared interfaces, re-exported from act-in-page.ts so
no import site anywhere changes
naming.ts what an element is called — labels, slugs, stems, anchors
visibility.ts whether it is really there and whether it is on screen
identity.ts the re-bind ladder and handle minting
Each is a factory rather than a bag of exports, and that shape is load-bearing
rather than taste. These sources are stringified into the guest page, where
module scope does not exist; separate exports would be separate names for the
bundler to mangle independently of the call sites inside the stringified core.
A factory that stringifies whole has no cross-module reference to break. The
core takes the kits as a parameter for the same reason — it names nothing it
did not receive.
That failure mode deserves spelling out, because it is why this is not a plain
import. The renderer minifies. An imported binding referenced inside the core
would be renamed to something the injected bundle never declares, and it would
break in packaged builds ONLY — green in dev, green under vitest, broken for
users. `actEngineSource()` is now the single supported way to obtain the
source, so a partial injection is not something a caller can express.
The self-containment test moves with it, and gets stronger: it evaluates the
assembled bundle rather than one function, which is what the guest actually
receives. It caught the contract change on the first run.
Bodies moved unchanged; the only edits are closure captures becoming
parameters, and `coin` taking the counter it used to read off the holder. The
engine is 634 lines, all of it orchestration. The 53 existing engine tests pass
untouched, which is the evidence the move preserved behaviour, and 27 new tests
exercise the extracted helpers directly for the first time.
Every drive_preview action answered with the entire inventory — around 120
elements of ref, role, label, and an up-to-eight-rung `:nth-child` selector
chain. On a real app shell that was ~24.5k characters, re-sent after every
click, so a ten-step task paid for ten copies of a page that had barely moved.
Handles are now durable and legible. An element is named after what it is and
what it says — `btn-sign-in`, `inp-email`, `srch-search-projects` — minted once
per page and never reused, with duplicates disambiguated as `btn-edit`,
`btn-edit-1`. Each one remembers a stable attribute, its role, its accessible
name, and the nearest landmark it sits in, so when a framework destroys the
node and builds a new one the handle moves across and the agent is told
`rebound` rather than being handed a removal it has to react to and an addition
it has to re-read. The re-bind ladder is anchortree's (Apache-2.0), minus its
geometry rung, which can never clear the threshold on its own.
Because the handles hold, the first look at a page returns the inventory and
every look after it returns only what moved. `changed` carries the ref and
whichever of label/value/disabled actually shifted — role and selector are
absent by construction, since a change in either would mean the re-bind ladder
was looking at a different element. A delta gives way to a full re-read when
half the page is new, where there is nothing left to reuse.
The selector column is gone with it. It was 74% of the inventory on an
85-element page, nothing downstream ever read it, and a positional chain is
wrong the moment a sibling appears. An `#id` or `[data-testid]` survives when
the page offers one; everything else is addressed by handle.
Legibility is what makes the delta work rather than a nicety. `+ btn-sign-in`
on turn nine reads on its own, where `+ @e42` sends the model back to an
inventory twenty thousand tokens ago.
Measured on an 85-element app shell: 18,693 -> 4,930 characters for a baseline,
and a steady turn that moved two things costs ~200.
Driving someone's browser invisibly is unnerving, and this browser is the one
they are signed into. The pane now draws the field the agent can reach, a box
round what it is touching, a cursor that goes there, and a wipe over text it
just read — one cursor primitive and one mark primitive, in a closed shadow
root so the agent's own inventory cannot see them. Marks carry the same handle
the agent addresses them by, so the word on screen and the word in the
transcript are the same string.
The point is supervision rather than decoration: a person glancing at the pane
can tell what is about to happen to their live session, and stop it.
It also has to cover the waiting. The agent flashes through a click in under a
second and then sits idle for the twenty to a hundred seconds the model spends
deciding what to do next, which is most of the wall clock of any task — so the
surface used to look broken during the part where it was working hardest. A
think stage runs off the $busy edge, sparsely flashing elements from the field
the last action left behind, and rest stops it. It guards itself: started
before there is an overlay or a field, it idles until there is one, so it can
be raised on the turn boundary without knowing whether anything has been
inventoried yet.
read_preview had the same hole from the other side. Reading is the cheapest
thing the agent does — hundredths of a second between two model round trips —
so paging through a document left the pane dark for twenty seconds immediately
after the one moment that showed anything. It draws a top-to-bottom wipe over
the text it took, and that is the one stage allowed to be a wipe: reading is
the only thing the agent does to a page in an order a person could follow.
Both go through preview-nudge, which says a single stage to an overlay the page
already has rather than re-shipping the engine to narrate. On a page the agent
never acted on it is a no-op, which is the honest answer — chrome there would
be a lie about what it did.
Everything respects prefers-reduced-motion.
A dispatched MouseEvent is untrusted, so hover menus never opened and any
control that gates on isTrusted ignored it. The pane now sends input through
the webview itself: the pointer travels to its target and the page cannot tell
it from a hand.
The in-app browser was a one-way mirror. open_preview put a page in the pane
and read_preview read its text back, but nothing could touch it. A click meant
falling back to the browser_* tools, which drive a separate Chromium the user
cannot see — so "log into this and pull my invoices" happened in a different
browser from the one on screen, with none of the sessions the user is already
signed into.
Four pieces, and they only make sense together:
· an in-page engine that inventories what is interactable and performs the
verb, injected as source because it has to run inside the guest page;
· the preview.act.request bridge from the gateway into the pane;
· drive_preview, for acting: elements, click, type, scroll, press, and the
pane's own back/forward/reload;
· annotate_preview, for marking without acting.
Those last two started as one tool doing two unrelated jobs. Leaving a mark is
not an action — it outlives the turn that drew it — so it gets its own verb,
and the interaction verb gets a name that says what it does.
Gating is the existing surface rule: desktop_ui folds in on session
source: 'desktop', and the bridge refuses to act for a background session, so a
turn running behind the user's back cannot reach into the page they are working
in.
Two details worth a reviewer's attention. Typing assigns through the
prototype's value setter, because React shadows value with its own accessor and
ignores an input event whose value it believes it already wrote — a plain
el.value = … types into a field that snaps back on the next render. And
clicking replays the pointer/mouse pair before activation, because frameworks
bind to mousedown as often as to click.
Brooklyn Nicholson