The chat stats line took its token totals from the loaded conversation nodes,
so paging changed them and compaction erased the billing behind replaced
content. It also had no way to show context occupancy: the numerator and
capacity never reached the browser.
Both now come from token-meter session projections read through the standard
useProjection seat. Window nodes keep supplying turn and step counts plus LLM
and tool wall times, which are correctly window-scoped facts about what is on
screen; accounting no longer comes from there.
`tokenUsage` supplies billing and cache hit. `contextPressure` supplies
occupancy, pairing the newest provider-reported prompt size with the newest
capacity recorded by `request/context`. Deployments without token-meter drop
the token groups; a route whose adapter advertises no capacity drops the
occupancy group rather than rendering a placeholder.
Occupancy is deliberately approximate: the numerator and capacity are
independent last-wins fields, not one atomic request observation, so switching
models pairs a fresh capacity with the prior route's pressure until the next
request reports usage. It is a user-facing reference figure that nothing in the
harness makes decisions from, and it matches how the TUI status line has always
computed occupancy. The Agent Note and token-meter README state this as a
decision, including why the atomic alternative was implemented and rejected, so
it is not re-litigated as a defect.
Snapshot delta is one added `Context N% of 128K` segment across eight web
goldens; the preceding commit absorbed master's pre-existing golden drift.
- gate model selection on steering-placement image carriers from enqueue
until their steering/message event publishes; release the gate when an
admission ends idle without publication (both behaviorally asserted)
- reject session.updateQueue edits carrying non-text blocks at the RPC
boundary (queue edits cannot bypass image admission)
- extend the durable-directory walk past a first-created DSH_HOME to the
deepest pre-existing ancestor
- strip Windows-style separators from attachment display names on POSIX
- verify attachment reads with a header-only probe (digest already proves
the bytes decoded fully at admission); document the read path
- make SessionInputShell.addImages refusal observable and keep workspace
transfers/composer intake from leaking refused drafts
- own ONE recursive image walk (dsh-llm contentHasImage) across apiproxy,
pi-ai, compact-basic, and the DeepSeek text-only assertion
- drop the redundant canonical-base64 regex and the no-op role read
- move AttachmentId/AttachmentError out of types.ts (brand.ts/error.ts);
document why AttachmentError does not extend HarnessError
- document the hard attachments inject in both consumer READMEs
Conflict resolution notes:
- StatsLine: master redesigned the row into pipe-separated groups with LLM and
tool wall times. Kept that design and swapped only the token accounting
source, so counts and durations stay window-scoped while billing and context
occupancy read the durable projections.
- Generated artifacts (cordis catalog, module graph, event producer/consumer,
i18n pairing hashes) and web snapshots took master's side; they are
regenerated and re-recorded after this merge.
- Web e2e goldens and details-panel/timeline assertions took master's side:
that evolution is unrelated to this branch.
- ui-conversation package.json: kept master's devDependency ordering, re-adding
only the token-meter entry this branch needs.
Replace the transient `session/model-request` mux frame with ordinary durable
session state. Occupancy now rides two last-wins projection fields instead of a
non-replayable frame that needed removal tombstones and cross-stream fencing.
The frame was the only non-replayable class on the mux stream. Because host and
mux are independent SSE streams with no cross-stream order, a request emitted
before a removal could arrive after `host/session-removed`, and a legitimate
request for a new lifecycle reusing the same id could be fenced by a late
removal. Fixing that needed a lifecycle generation on every frame; the frame
itself was the problem.
Removed: the `session/model-request` frame and schema, the `agent/model-request`
core event, the ApiProxy measurement point, the client-side telemetry map and
removal tombstone, and the synthetic `cancelled` open error used to signal
reconnect through the error channel.
Added: `request/context`, a log-only session event recording the
registration-bound capacity of the route a request resolved to, appended beside
`request/header` from the lookup that already prepared the call and skipped when
the route is unchanged. Capacity stays out of `EpochHeader` because it is
adapter metadata about a route, not an input the request was built from, so it
must not join request reconstruction or header equality.
The `contextPressure` projection pairs the newest provider-reported prompt size
with the newest recorded capacity. The two are deliberately not one atomic
request observation: switching models can pair a fresh capacity with the prior
route's pressure until the next request reports usage. The figure is a
user-facing reference, and this matches how the TUI status line has always
computed occupancy.
Conflict resolutions:
- `session.list`: master's projection columns fold into the PR's cancellable,
batched `listVisibleSessionSummaries`, which `session.search` shares as its
visibility baseline; master's goal helpers stay beside it.
- Client sessions face: master narrowed `ctx.sessions` to `ISessions`, so the
search verb and its protocol-constant bound are declared there and the
test-runtime double implements them (recorded, empty page unless a scenario
stubs hits).
- `WorkspaceBrowser`: master's per-row Rename wiring rides the PR's search
results view; the tree keeps the PR's query-free derivations.
- `dsh web` bin: the PR's shutdown-handlers-before-readiness order with
master's boot-time LAN address snapshot.
- `session-query-sqlite`: master's `SCHEMA_VERSION` 7 stands; the PR's bump
carried no schema change.
- Specs: master wraps assistant/steering message payloads and requires an
`application/json` carrier request, so the search fixtures and tests follow.
- Web aria goldens keep master's recording plus the PR's search placeholder;
the navigation-panes inventory keeps master's terminal-card golden next to
the PR's search-results golden.
Eight compiler-locked methods: settings.describe/update/replace serve
redacted layered namespace views (secrets structurally absent from every
layer, write-only in the update direction) and fold seam refusals into
settings-rejected; credentials.describe/set/unset expose value-free views
with credential-rejected on shadowed writes; llm.providers merges the
configurable directory with live routes and llm.models claims the
host-scoped catalog reservation through the buildModelCatalog extraction
session.models now shares. Three HostFrame invalidations bridge the seam
events (host/settings-changed, host/credentials-changed,
host/models-changed), and the connection route generalizes the native-
dialog check into a privileged-method set covering all four writes. The
fixture and both fake clients grow the same face.
The fixture route shares the requireSession guard, reads the accepted seq off
the appended event, and gains the error-arm spec the workspace.rename case
set as precedent; WorkspaceBrowser drops the unchanged-title block — confirming
the current automatic title IS the pin gesture — and both touched client
READMEs document the new dialog and the unary-settle rule bilingually.
- llm-deepseek: the uncatalogued resolveModel fallback declares text-only
modalities — the wire route is text-only regardless of catalog
membership, so "unknown" must not let the host persist-then-fail images.
- session.selectModel also consults the pending-inbox mirror: a queued
image prompt enters the log only when claimed, after a switch would land.
- attachment store: ensureDurableDirectory syncs every ancestor entry up to
a caller-vouched boundary regardless of what mkdir reports — a raced
"already existed" is not "already durable".
- One image walker (imageBlockIn/imageInEvent) now serves both attachment
authorization and the selection gate; referencedImage therefore also
authorizes references inside wrapped message content.
- InputHub: the scope disposer resolves the conversation service optionally
(teardown/HMR must reach quiescence), and a send failing after its scope
died releases the in-flight drafts instead of restoring them onto a
disposed shell.
- http-bridge destroys declared-oversize requests with connection: close
instead of draining a body the client can trickle indefinitely.
- LlmService validates AND detaches modality arrays identically on the
advisory and exact routes; READMEs record the fourth INVALID_MODEL_INFO
rejection reason.
- CLI provider docs (JSDoc, README pair, Agent Note pair) describe the
reuse behavior; llm-route.spec now parses the SHIPPED cordis.yml through
the production extraction, pinning the row coupling.
- image-display lane pins gallery/rail shape in inline snapshots and the
object-URL scheme this environment must take; stale host.schema comment
dropped.
rename is a per-session verb on the outward session face (prompt/cancel
precedent), not a list-service verb: the Session calls session.rename and
applies the response {title, seq} to its projection store under
higher-seq-wins, so the list row updates before the push frame. The fixture
api and the test-runtime double follow the same face.
- resolveLlmRoute: reuse the yml pi-ai row for providers it already routes
(DUPLICATE_ADAPTER boot failure) and detect an unset model by origin, not
by comparison against one deployment default; covered by a new spec.
- LlmService.resolveModelInfoFor preserves (and validates) modality
metadata, arming the host image preflight for exact-route resolution.
- session.selectModel refuses a text-only target once the session log
carries an image on any replayed route; an accepted switch would strand
every later turn with no in-product recovery.
- The composer no longer gates image intake on the handshake activeModel
snapshot (wrong authority for a per-session decision); the host preflight
plus the error strip own capability, deployment limits stay client-side.
- InputHub shell teardown releases the scope's draft images (File objects
and object URLs leaked for the page lifetime).
- session.prompt image parts carry optional alt into the durable block;
ImageBlock documents assistant-side rendering as forward compatibility.
- Assembled built-client lane apps/web/tests/image-display.snapshot.ts pins
the history galleries over the authorized attachment route, the lightbox,
and the composer paste rail; the attachment rail is an accessible group.
- Docs: validateImage on the seam page, fixture byte metadata matches its
PNG, and the Agent Note claims now match the shipped coverage.
The multimodal sample turn carried the pre-send-unify event shape (bare
content/provenance instead of the message wrapper), crashing the built
client fold; it also sat after the todo_write turn, retiring the plan
projection the TodoPanel tests pin. Wrap both messages through the
fixture helpers and order the image turn before the todo turn; re-record
the code-mode trajectory ordinals the two added message nodes shift.
Nine findings, one critical. Terminal cases verified in a real terminal first.
CRITICAL: cells held the accumulated SGR history, so every state boundary
re-emitted the whole chain — output switching color without a full reset
emitted O(n^2) characters. Measured: 3200 such cells produced 25 MB, and the
reviewer's ~90 KB alternating-color case is well under bash's own output cap.
State is now a normalized record (foreground, background, attribute set) with
one canonical sequence per boundary, so the emitted text is linear in cells;
the 90 KB case parses in 36 ms. That also makes the attribute closers every
chalk-based tool writes actually close: `\x1b[1mbold\x1b[22mplain` leaves the
following write PLAIN, which a real terminal confirms.
Width follows emoji presentation, not the U+2600-U+27BF block: `A✓B` redrawn
with `XY` shows `XYB`, so the check every progress line writes is ONE column.
Taking the block as wide misaligned exactly the output this card exists for.
Writing over either half of a wide pair blanks the other, since a terminal
cannot leave one cell of a two-cell glyph standing.
`line\n\x1b[0m` does not end in a newline as a string yet its last parsed line
holds nothing visible, so the terminator check now reads the parsed lines — it
had added a blank row and inflated the collapse count.
A line with no cursor movement no longer builds a column buffer at all; only
its SGR is folded, so an `ls -R` or a 5k-line log allocates nothing per
character.
The `.terminalDescription` rule had been inserted into an existing grouped
selector, silently giving `.codeBody` description typography and changing its
bottom margin from 4px to 0 — a pre-existing surface this PR does not own.
Split out, `.codeBody`'s margin restored.
Three comments contradicted their code: the fixture's exit-marker line (still
claiming recovery from a marker deliberately removed), `bash-sample`'s header
(still routing a click to the details panel, and calling the consumer's cap the
block's own), and a DetailsPanel comment stacked above the wrong rule. The
ui-primitives README documented only the CR/BS half of the replay, so a reader
would expect `OK0%` where `100%\r\x1b[KOK` renders `OK`.
The native adapter's route was named deepseek, colliding with pi-ai's
catalog provider of the same name, so the two DeepSeek paths could never
be mounted side by side. The web settings page needs both configurable at
once. Compositions, fixtures, goldens, scaffolding defaults, and docs all
move together (pre-release, no shim); TUI/session-query-spill/
missing-credential goldens re-recorded through their keyless refresh
modes because provider-name length shifts box padding and spill
truncation points.
Eight findings, each terminal case verified in a real terminal first:
`\x1b[32mdone\rok\x1b[0m` then `plain` shows `okne` green and `plain` in the
DEFAULT color. The replay returned the last written cell's state, so a reset
landing after the final write vanished from both the text and the returned
state — and every build tool writes exactly that shape, so the color leaked
onto all later output. The replay now converges to the state the scan ended
in, which is also what it hands to the next line.
`abcd\b\x1b[1K|` shows ` |`: CSI 1K erases THROUGH the cursor column, and
the loop stopped before it. The erase mode also reads only the first
parameter now, since a terminal treats `1;2K` exactly as `1K`.
`éx\rYZ` shows `YZ`: a combining mark takes no column, so it attaches
to the cell already written instead of advancing the cursor and leaving the
`x` standing.
`中x\rA` shows `A x`: overwriting a wide character's lead cell leaves its
spacer as a blank rather than closing the gap, which would shift everything
after it one column left.
The banner lost its span when the gutter became padding — a plain block child
only reaches the content box, so the reserved column was painted in the body
color and the card's top-left radius drawn in it. Invisible in the light
theme, where banner and body share a token; visible in the dark one. The
header now pulls back across the gutter and re-insets by the same amount.
The replay trigger matches the same CSI shape the parser accepts, so a form
like `\x1b[1;2K` can no longer skip its own erase, and `replayLine`'s JSDoc
documents its new parameter and returned pair.
Docs: four places still described the dot as sitting left of the card surface,
which stopped being true when the gutter became the card's own padding, and
two fixture comments still referenced the exit marker that was deliberately
removed.
Six findings. Each terminal-semantics case was checked in a real terminal
(tmux, reading back the painted screen) before changing anything:
`100%\r\x1b[KOK` shows `OK`. Modelling the `\r` without its erase left the
previous frame's tail standing — a regression against the old truncate, since
`\r\x1b[K` is the single idiom every spinner and progress bar writes. Erase is
now part of the same replay, in all three parameter forms.
`a\tb\rXY` shows `XY b`. Counting a tab as one column produced `XYb` and
destroyed the alignment this card exists to hold, so the cursor now advances
by terminal columns: tabs reach the next 8-column stop and a wide character
takes two cells.
`\x1b[31mabc\rX\nnext` paints BOTH lines red. A newline does not reset the
graphic state, so state threads from one replayed line to the next instead of
closing at each line end.
Only `m` accumulates into a cell's style now. Folding cursor and erase
sequences in grew the state string per redraw and emitted boundaries anser
had to discard.
The empty check reads the parsed lines the card renders rather than the raw
text: output that is only escapes or control bytes survives `trim()` yet
parses to nothing, and drew blank rows plus a copy control for invisible
bytes instead of the placeholder.
The gutter is the card's own left padding rather than a margin. Every render
site rewrites `margin` wholesale for its own indent, which silently cancelled
the reservation and let a container clip the dot.
The fixture sample no longer carries an `[exit code: 1]` line: the real bash
presenter consumes that marker precisely because the card shows the exit as
its own pill, so the built-bundle snapshot had pinned a frame showing it
twice — one the product path cannot produce.
Three review findings, each verified against the presentation contract:
The call view's `description` was dropped, so a presenter that authors one
(`terminal_send` declares `Terminal <id>`) lost the contract's above-card text
and the row fell back to an unrelated args-derived summary. It now rides the
same derivation and outranks that summary.
A relative workdir was concatenated but never normalized, while the bash
executor resolves it before running: with session cwd `/w/app` and workdir
`..` the command runs in `/w`, yet the card displayed the label `..`. The
resolved path now collapses `.`/`..` segments, drops a `..` that would climb
past a root the way a filesystem does, and keeps a Windows path's separators
since the value is only ever displayed.
`run_code` sub-dispatches carry no presenter views on the shipped wire —
`session.ts` folds `tool/code-dispatch(-start)` with null views and the host's
`viewFor` presents only top-level call/result events — so a nested bash call
cannot reach a terminal card. The existing test only passed by injecting views
that path cannot produce; it now says so, and a second arm pins the no-view
shape the wire actually delivers.
Restoring master's fixture also fixed the todo snapshot lane, which my earlier
merge had broken by dropping the projection support the todo dock reads. The
terminal sample turn moved ahead of the todo turn, because the standing plan
retires at the next `turn/start` and a turn appended after it emptied the dock.
The card props are now nested under `card` so a render site spreads exactly the
primitive's own surface, and the fixture reads each sample's authored exit
status instead of re-implementing the bash tool's `parseExitStatus`.
