- 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.
Add @deepseek-ai/dsh-host-directory-picker-auto, a node-half-only chooser
that samples the host situation once at boot (bind host via a new
httpServer.host getter, SSH markers, platform, DISPLAY/WAYLAND_DISPLAY)
and mounts the matching dual-face backend (-native or -browse) as a real
Loader entry in the in-memory root tree; the effect disposer removes it.
Entry-level mounting keeps the seam's one-row-swaps-both-faces invariant:
the client module table discovers the mounted backend's browser half
exactly as a config row's. apps/cli now composes -auto as its
directory-picker row; composing a backend row directly remains the pin.
The `ctx.tui` source line moved when e029ffb88 retired two imports from
`packages/ui/tui/src/index.ts`. Regenerated; `verify-cordis-catalog` is
green again, which is the gate CI caught.
Review follow-ups on the append-origin transcript projection.
The live/replay equivalence claim was stated unconditionally but does not
cover `tool/call`: only replay re-derives call pairing, because a call
event carries no `surfaceOp` of its own and inherits transcript
membership from the `assistant/message` that advertised it — which the
live listener has necessarily just rendered. Narrow the claim in the TUI
README and Agent Note, and record at `rebuildTranscript` why the filter
is replay-only rather than a missing live branch.
Add `surface-replayed-compaction`: the three existing fixtures all come
from the live path, leaving the resume case the bug report leads with
pinned only by a unit test. The new checkpoint mounts with the
replacement already stored and records byte-identical to
`surface-after-compaction-wide`, so the two fixtures now pin the
equivalence they assert. The shared fixture appends move into
`appendPreCompactionLog` / `appendCompactionCheckpoint`.
`MESSAGE_TYPES` is not "human message event types" — it includes
`assistant/message`. Say what the code distinguishes (append-origin
conversation messages vs. model-only replacement copies) at the const,
the `paginate` and `session.history` JSDoc, the apiproxy README, and the
Agent Note.
Also: spell the replace shape as `Extract<SurfaceOp, { op: 'replace' }>`
for symmetry with the module's two other uses; document why
`isCompactCheckpoint` keeps a replacement check that is redundant at both
call sites; say that Ctrl+R toggles reasoning, which rebuilds the
transcript; and qualify "the sole source of derived history" as derived
*model* history now that the transcript is the other projection.
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.
The terminal and history pagination both treated the model-visible surface as
the human transcript. A landed compaction replacement therefore erased the
conversation it summarized — messages the reader had already seen — and a
model-only replacement copy consumed a page's `maxMessages` quota, which could
also split a compaction's provenance from the replacement citing it.
`dsh-session` now exports the marker split `isAppendSurfaceEvent` /
`isReplacementSurfaceEvent`. The terminal replays append-origin surface events,
keeps a shadowed step's tool cards paired through its append-origin assistant
message, and renders one dim marker where a compaction landed; the checkpoint is
recognized through the compaction seam's `isCompactCheckpointSource` contract,
not the shape of the replacement. `session.history` counts only append-origin
human messages. Everything model-facing keeps reading `session.surface`.
An empty or omitted providers dict is now the valid dormant posture: the
adapter mounts with zero routes and no catalog entries, registers routes
the moment the llm-pi-ai settings section supplies profiles, and drops
them when it empties. The TUI demo mounts the adapter bare, so adding an
openai/anthropic provider is purely a settings.yaml (or, next PR, web
form) operation with per-request apiKeyEnv credential resolution.
Whether concurrent active tasks are legitimate depends on runtime
concurrency the tool cannot observe, but whether a deployment's agents
ever fan out is knowable at composition time. `allowParallelInProgress`
(default true) therefore replaces the hardcoded policy: the flag moves
the model-facing instruction and the accepted input together, so a
deployment running strictly sequential agents can restore the
single-active discipline from cordis.yml.
The durable-log invariant does not follow the flag. A log written while
parallel work was allowed must still replay after a deployment tightens
the policy, so the invariant stays silent on the active count.
New credentials data-structure page (type-equiv manifested), group README,
rewritten llm-deepseek/llm-pi-ai READMEs (dynamic configuration, dict
profiles, credential chain), capability-seams/service-role registration,
Agent Note (bilingual), demo compositions mounting settings-local +
credentials-local with no inline key plumbing, installSettingsSection
consumer helper on the settings seam (deduplicating both adapters' wiring),
jscpd symmetry markers for the provider twins, runtime-closure additions for
python/sdk-runtime, and doc-budget ceilings AGENTS.md 1750→1755 /
packages/README.md 850→865 for the structural one-line group rows.
Register a global `/feedback` command so a user can record a remark about the
session without spending a model turn. `/feedback <text>` acknowledges; empty
or whitespace-only input returns a usage error.
The plugin appends no session event of its own. `dsh-commands` already writes
a `command/run` / `command/done` pair for every dispatched command, carrying
the verbatim text and the settled outcome, and both records are log-only and
non-surface. The feedback is therefore durably in the session log and invisible
to the model without this package touching the log format.
Text is never parsed, so `/feedback /plan felt slow` records that literal
content. Nothing consumes the records; capture is deliberately inert.
New group `packages/feedback/` — no existing group owns feedback capture. Its
row raises the packages/README.md word ceiling by 10, which had no headroom;
one redundant sentence there was removed to offset most of the cost.
Review round on #758.
bash.run() only promises to resolve for nonzero exits, timeouts, and aborts,
and bash.resolve() can reject on policy grounds, so either could escape the
serial agent/step listener and abort the model turn — contradicting the
plugin's documented failed-query no-op contract. Contain both and log a
warning instead; the location is optional context.
The Agent Note claimed an unchanged location suppresses the query. It does
not: only the interval floor is checked before the query, while change
suppression compares state the query returned. Corrected in both languages
and re-recorded the i18n pairs.
Conflicts were the parallel-edit set: packages/README group table (kept
master's session-title wording, re-added the settings row condensed),
doc-budgets AGENTS.md ceiling (took master's 1750), and regenerated
graph/catalog docs plus re-recorded translation pairs. packages/README.md
ceiling 845 -> 850: the table legitimately gained one row for the new
settings group and the row is already minimal.