The composer ring, percentage, and `~used / capacity` header read
`contextPressure.pressureTokens`, which moves only when a request reports
usage. Compaction reports none — compact-basic summarizes through a direct
`ctx.llm.stream()` call and appends only its own `compact/*` records plus the
replacement `user/message` — so the meter was frozen across the one action
taken to change it. Driving a real `compactNow` through the agent loop:
BEFORE compact: ring=4% header=~4227/100000 rows=[18, 0, 4365]
AFTER compact: ring=4% header=~4227/100000 rows=[18, 0, 286]
The composition rows fell 93%; the ring did not move, and would not until an
entire further turn completed. The panel then contradicted itself by more than
an order of magnitude at exactly the moment a reader opens it.
`contextPressure` now also publishes `projectedTokens`: the provider sample
plus the heuristic repricing of everything the surface gained or lost since
that sample, clamped at zero, folded through the shared `surface-fold.ts`. The
sample is stamped before the same event joins the surface, so an
`assistant/message` anchors against the surface its own request carried. Only
the delta is estimated, so the figure stays provider-anchored — the estimator's
CJK and JSON-schema underpricing stays out of the occupancy number — while
reacting the moment content lands or a span is shadowed. Same run after:
BEFORE compact: ring=4% header=~4323/100000 (pressure=4227, projected=4323)
AFTER compact: ring=0% header=~ 244/100000 (pressure=4227, projected= 244)
`contextOccupancy` prefers the projected figure and falls back to the bare
sample, so a projection restored from a pre-field checkpoint degrades to the
old behavior rather than disappearing. `stateVersion` moves to 3.
Assistant footers and the stats line gain TTFT/tok-per-second readings
folded from step timings; context occupancy moves off the stats line onto
a composer ring whose panel shows a heuristic system/tools/messages
breakdown from the new token-meter contextBreakdown session projection.
Injected context reached the transcript as one anonymous shape whose
expanded body was the whole message serialized as inline JSON, so the
escaping collapsed the only readable part — the model-facing prose —
into a single line.
`MessageSource` gains an optional producer-declared `form`, a small
semantic vocabulary of information shapes independent of `kind`: kind
says who produced the context, form says what shape it is, so several
producers may share one presentation. Two values ship.
`instructions` (workspace-context) lists the reconciled files above
their text and keeps the `<system-reminder>` framing verbatim, because
the framing is part of what the model read. `catalog` moves dsh-tool-skill
off the shared plugin kind onto a `skill-catalog` source carrying the
published name/description entries, and the body lists those instead of
re-parsing `<available_skills>` out of the prose. Catalog identity moves
with it: the republish digest now covers the durable entries, deleting
the text-slicing that recovered them from a logged message.
Everything else renders the opaque body — the model-facing text with its
real line breaks, then the remaining provenance as fields. That is the
documented default, not a leftover: a resumed, forked, or foreign log
must render whether or not its producer is mounted here, which is why
the classification lives in the durable source rather than a client-side
table keyed by producer.
Review follow-ups on the context provenance row:
- the producer name now follows ToolRow's aria-hidden separator dot instead
of a literal middot inside the text, so the two disclosure rows expose one
accessible-name shape;
- the English steering caption becomes the noun `Interjection`, parallel to
the Chinese 插话 and to the label this bubble carried before;
- the history fold gets its own provenance assertion, so the mapping it
declares separately from TranscriptAdapter is pinned on both sides;
- the superseded-in-part bullet on the no-interjection-chrome note now
enumerates the third Decision bullet too, whose no-badge and no-IconActions
clauses had both lapsed.
Every logged non-user user/message collapsed into one identical
「上下文注入」 row, and mid-turn steering rendered in exactly the bubble a
turn-opening prompt uses, so the transcript could not say what had been
added or which message interrupted a running turn.
TranscriptAdapter and the history fold now attach a provenance view to
each context node, computed by contextProvenance() from the durable
source alone: a role (inject, or recall for a cross-session snapshot)
and a producer name read out of the log — instruction paths for
workspace-instructions, session titles for session-reference, the plugin
id for a plugin source, and the bare kind for any other. No client-side
table of producer names, so a renamed or newly mounted producer stays
identifiable without a client release and a foreign log projects like a
live one. ContextInjectionRow titles itself from the role and shows the
name beside it; MessageItem captions durable and pending steering
bubbles.
The caption reverses one clause of the no-interjection-chrome decision,
which removed it because the composer could not steer; composer steering
shipped afterwards without amending that note, so this change supplies
the product decision its reintroduction clause required and corrects the
stale facts left in it.
Fixes#1291
The composer seat is one node laid out against two different edges: in Chat it
is a sticky child of the column's scroller and rides its content box, which a
space-consuming scrollbar shortens, while a view declaring a composer overlay
gets an absolutely positioned seat against the padding box, which no bar
reduces. With the transcript scrolling the two tabs disagreed by the bar's
width, so the centred input card moved 4px sideways on every switch — and the
same 4px inside Chat when a growing transcript began to scroll.
The column now reserves its scrollbar gutter unconditionally and states the
overlay branch a scroll container on the same axes, so both states measure
against the same width.
The READMEs of ui-primitives and ui-conversation still documented the
removed maxSources prop, DEFAULT_WEB_MAX_SOURCES, CHAT_WEB_MAX_SOURCES,
and the head/tail collapse; the 2026-07-30 frontend note described the
same as current behavior. Restate all of them as the fixed-height scroll
container and cross-link the two notes both ways.
Bound the model-visible/card-visible correspondence: spill-policy's
tools/post-execute replaces an oversized result's content while leaving
presentationMeta whole, so the card's contract is the view it receives,
not the model's context.
Narrow the unit test's name and header comment to what jsdom establishes
(one <ol>, no expand control); jsdom does not resolve CSS Modules layout,
so the scroll geometry belongs to an assembled browser case.
next-step items now split by origin: user-origin messages keep the
'steering' placement, while agent.inject context (approval notices, task
completion notices, attached snapshots) carries a new 'context' placement
that no surface renders until it is claimed as a durable user/message
context card. Widen the placement unions on the wire and in the client
runtime, add a projection test case, and document the split in the
apiproxy/ui-conversation READMEs and the web-steer chrome agent note.
Steer, inject, and followup now land as durable user/message events on the
session surface; the steering/message event type and its ConversationNode
kind are removed from the client projection. Update tests, docs, generated
catalogs, and agent notes to match, and align the steering e2e fixture and
prompt inventory assertions with the durable user/message landing.
Scope decision: previews for a browser that is not on the Host machine are
not supported. With that settled, host.openPath answers the supported case
completely — a file:// document in a real browser has full page capabilities
and no reach into /api — and the HTTP serving this branch had built answered
only the unsupported one.
Removed: the /f route and its listener, the workspace-file URL shape,
ApiProxy.workspaceRootOf, ConnectionHandle.fileUrl, and the port published
into the index page.
Kept, and finished:
- the produced-files row a turn ends with, derived from mutation locations;
- the path link now reads as a link at rest, not only on hover — the reported
"I can't open what it made" was this, sitting on a working capability;
- the Host opener prefers the default BROWSER for .html/.htm/.xhtml/.svg, so
a developer who binds .html to an editor still gets a rendered page
(macOS via the LaunchServices https handler, Linux via $BROWSER, every
failure falling back to the default application).
The retired designs and their measurements stay in the Agent Note, including
why same-origin serving was unsafe and why the sandbox that fixed it broke
the pages invisibly.
A sandbox header bought isolation by taking the document's origin away, and
measuring that cost decided against it: the reported artifact throws
SecurityError on load, and because an uncaught exception aborts the rest of
its <script>, every listener declared after that line — theme toggle, mobile
menu, model tabs — never binds. Two of the four artifacts in the reporting
user's workspace were dead pages under it, and they still looked right.
A second listener on the API's host, answering /f and nothing else, is the
same boundary without the amputation: cross-origin to /api (refused by the
Origin fence and by CORS), same-origin with itself (localStorage, cookies and
fetch all work). Its port is published into the index page; the browser half
reads it to address previews, and its absence — the keyless fixture lane — is
what makes a file row fall back to the Host opener instead of a dead tab.
fileUrl moves from IWorkspaces to ConnectionHandle: the transport owns both
the listener that serves the bytes and the port that addresses it.
Isolation is restored on the premise the review corrected: a workspace file
need not be agent-authored — a read row makes every file in a cloned
repository openable — and a same-origin active document was measured driving
/api/settings.describe to a 200 with full data. Script-capable documents go
back into an opaque origin; the preview's lost localStorage is the known cost,
and a separate serving origin is the way to retire it.
- confine(): a workspace rooted at a filesystem root has a realpath already
ending in the separator, and the doubled prefix 403'd every child.
- turnDeliverables(): reset on the turn boundary, not only at a closing
assistant, so an interrupted turn cannot spill into the next turn's row;
and recognize a mutation by render intent (diff card, or generic with
kind 'edit') so str_replace_editor's insert counts.
- 405 answers name the methods it allows.
- The e2e now cold-seeds a recorded WRITE turn, so the assembled application
covers the Produced row, its chip's served URL, and the isolation header.
- Agent Note matched to what shipped (the row is in this PR, not deferred);
ui-conversation README documents the new destination and the row; the
fixture lane's dead-tab quirk and the cold-path listing cost are recorded.
