Review round findings:
- Clear the streaming attempt before the request-error waterfall: a cancel
landing during recovery (typically the llm/retry backoff, after clients
reset the streamed rendering) must not resurrect the failed stream's
prefix. Provider failures commit nothing, now including that window.
- Record interrupted: true on the cancellation-finalized assistant/message.
The chat projection keeps the settled prefix classified as interrupted
(Stopped chip, restored web goldens), and request inspection leaves the
request uncompleted so the step boundary classifies it as before.
- Pin the recovery-window and retry-discard semantics with content-bearing
failed streams in cancel.spec; update the ACP late-end expectation to the
finalized-prefix transcript.
- Mention interruptedBlocks() in the assembler stream-lifecycle JSDoc.
A turn cancelled mid-stream dropped everything the interrupted step had
streamed: chunks stayed in the log for replay, but no assistant/message
joined the surface, so the next request carried nothing the user had
watched stream. Follow-ups after cancel could not connect and forks
inherited the gap.
Keep the streaming attempt alive across the step's request loop and, when
an abort escapes with the attempt uncommitted, finalize its user-visible
prefix as the step's ordinary assistant/message citing the logged chunk
seqs. BlockAssembler.interruptedBlocks() owns the safe subset next to the
max-tokens rule: closed and open text/reasoning blocks with content, in
stream order; tool calls, empty blocks, and unknown open blocks drop.
Retry clears the attempt first, so an abort after llm/retry finalizes
nothing from the reset stream. Tool-phase cancellation and provider
failures keep their existing shapes.
The scaffold's ambient-root isolation proof still resolved a preset-realm
service; under the host registry it is a scope read on the composed
agent's view, and what it proves is unchanged.
The auto-merge kept recompose on the removed open re-link. It now moves
the binding this roster kept from the agent's mount; an agent that never
composed one has nothing to re-link, so the switch is its first bind —
exactly a mount — and once bound only the kept binding can move it.
setScopeParent could re-link any key from anywhere, leaving the
blank-session-only recompose rule entirely to caller discipline. The
relation now binds once — a second bind throws — and re-linking exists
only on the ScopeParentBinding returned to the original binder, the
private-capability shape the package conventions prescribe for a
single-caller operation. The preset roster keeps each composed agent's
binding in a WeakMap keyed by the agent, making it the sole authority
that can move an agent between standing compositions; the blank-session
contract itself stays with the gateway, which alone can see what a
session logged.
The persistence catalog recognizes SessionEventMap augmentations of
'@deepseek-ai/dsh-session/types' alone, so the selected event's old
root-specifier declaration compiled but silently vanished from the
generated catalog. The hermetic scaffold e2e still addressed the skill
registry as a preset-realm service; under the layered host registry the
composed agent's view is a scope read, and the ambient-root isolation it
proves is unchanged.
The persistence catalog recognizes augmentations of
'@deepseek-ai/dsh-session/types' alone, so the old root-specifier
declaration compiled but silently vanished from the generated catalog; the
Agent Note also drops its internal repository link for the public-links
gate, and the persistence catalog pair re-records with the restored
agent-preset section.
The shipped headless profile is base + headless (the direct-core front
door), and its installation-owned migration rewrites the old three-bundle
tuple down to that template, so a one-shot run composes no preset roster.
The runner's optional preset composition was compensating in the wrong
direction: it re-added roster reads the shipped profile never satisfies,
and its golden pinned a header field only a roster-bearing deployment
produces. The runner, the headless bundle manifest, and the dsh-run golden
return to the direct shape; skills in a direct run flow through the base
host rows' global layer, which is what the repository-plugin e2e proves.
The battery self-exclusion widens from references/ to the whole skill
directory (SKILL.md quotes the same leaked shapes), and the invocation
rules name the owning note's evidence self-hits; natural-language
battery lines gain -i so sentence-initial capitals hit; the actually/
wait false-positive family states that no committed line probes it.
The one test now says resolvability clears only this skill's bar —
current-state surfaces still reject resolvable change stories; the
never-deletion-alone rule is scoped to passages that carry factual
clauses. The first few-shot Fixed example links the real web
input-machine note path. The citations note's Verification exemption
(+zh, pair re-recorded) names the skill's files and the note's own
quoted evidence, making the claim mechanically honest.
The suite composes over the ambient DSH home, so a fixed session id
collides with the log a previous run persisted once the tool-execution
path checks persisted identity.
The skills subsystem page and dsh-skill README describe the host+per-scope
layers, SkillViewOptions joins the type-equiv corpus and the cordis-catalog
type pages, generated catalogs pick up the shifted source refs, and the
architecture Agent Note records the decision, the nearest-layer-wins
precedence choice, and the rejected alternatives.
SKILL.md carries the resolvability test, the 8-class taxonomy, the
sanctioned-keeps list, and the owner-first fix workflow; references/
carry 24 few-shot calibration blocks (both citation-boundary directions
plus the four overcorrection traps caught in the purge's review rounds)
and the EN+zh recall batteries with their false-positive families.
dsh-doc-standards' inline leakage probes collapse into a link to the
new skill; dsh-prose-standard gains the sibling pointer; the
committed-artifact-citations note (+zh, pair re-recorded) records the
skill as the operational workflow and exempts its calibration examples
from the battery expectation.
The shipped-Web e2e enables the bundled badge skill as a stand-in for a
repository plugin's deployment-level provider and asserts the layering both
ways: a standard-preset agent's view merges the global layer with its
preset's local discovery and its loader tool resolves the global skill,
while the host view stays global-only and a core-web agent can read the
layer but composes no loader tool.
This reverts commit f12c131ee9fdbc01b9294c8dc2c73cd36ab717cf.
The layered host skill registry closes the gap the bypass recorded: a
repository plugin's skill root registers into the global layer and reaches
every preset-composed session's catalog, so the fixture declares its skill
root again and the e2e proves the full skill, MCP, and TypeScript
contribution set. The accidental .claude/launch.json from that commit goes
with it.
The skills capability moves back to the host plane as one singleton, layered
with the ScopedLayers infrastructure the tools registry established: a
registration files into its calling context's layer (repository plugins and
host rows land global, a preset's skill-local lands in that preset's layer),
and a read merges the global layer with the viewing scope's chain, the
nearest layer winning duplicate names outright while rank keeps deciding
duplicates within one layer.
The web-app bundle re-enables the base skill registry row (skill-local and
tool-skill stay preset-owned), the standard preset drops its isolate skills
realm for bare rows over the host registry, tool-skill passes the calling
agent as the view scope, and the gateway's skills domain reads the host
registry in the presenter scope — a cold session now resolves its recorded
preset's standing key instead of failing.
creatixchu