llm-deepseek gains a Loader+Include composition spec proving external
settings.yaml/.env edits reach the very next request, and a real-API e2e
where only a credentials-local document holds the key. The headless example
pins the first-run missing-credential UX as a keyless stream-json snapshot
(new credentials.cordis.snapshot.yml scenario); runLoaderSmoke learns
expectedExitCode so a designed failure surface can be pinned instead of
masked.
providers becomes a dict keyed by provider route, so the composition base
and the llm-pi-ai settings section merge per provider and the route set is
structural; the pre-release array shape and per-profile provider field fail
loud with migration directions. The adapter reads a profiles thunk once per
operation and resolves the credential per stream call (literal apiKey, then
apiKeyEnv through ctx.credentials with an ambient env fallback, then pi-ai's
provider-native discovery), so key, endpoint, and knob changes reach the
next request without restarts. Route-set or captured-retry-policy changes
re-register the same adapter instance in one synchronous section; an invalid
settings snapshot keeps the last good profiles.
The adapter takes an options thunk and a per-stream credential resolver
instead of frozen construction facts: base URL, catalog, defaults, idle
budget, and the API key re-resolve at each operation, so a settings or
credential change reaches the very next request while in-flight streams
keep the facts they started with. resolveAdapterOptions is the one explicit
resolve step (entry config fails loud at load; a live snapshot failing a
beyond-schema bound keeps the last good options). The plugin layers its
entry config under the optional llm-deepseek settings section and resolves
keys literal-first through ctx.credentials with an ambient env fallback; a
missing key now registers the route, warns, and fails each request with
actionable MISSING_CREDENTIAL instead of failing plugin load. The
registration-captured retry policy re-registers the route in place when it
changes.
Live environment wins read-only (shadowed writes reject instead of
appearing to succeed); the file is the writable source with byte-preserving
line edits, a quoting ladder dotenv reads back verbatim, atomic 0600
writes, wholesale snapshot replacement on reload, and write-drain
teardown.
Both packages import @deepseek-ai/dsh-llm at runtime (createUserMessage)
but declared it only in devDependencies; any resolver that honors
declared runtime dependencies resolves the import to a stale or missing
artifact. Found by the (since removed) source-launch declared-dependency
check.
Four review findings. Two are defects the previous two rounds introduced,
which the existing tests did not catch:
A backspace erased raw bytes, so one landing after an SGR reset ate part of
the escape: `\x1b[31mabc\x1b[0m\b\bXY` left `\x1b[` and repainted the rest of
the line with whatever the remainder parsed as. Backspaces now resolve over
VISIBLE characters — a CSI sequence is one indivisible unit a backspace steps
over on its way to the last printed character, so the surviving text keeps
the color its run authored.
The cwd normalizer popped a UNC share root: `\\server\share` with a `..`
became `/server`, losing the separators too. A UNC path's server and share
are its root, and Windows cannot climb above a share, so they are split off
and the remainder collapses against that root.
The other two are gaps the earlier fixes left:
The render-site fallback row still passed the args-derived summary, so any
terminal-declaring tool without its own keyed row (`terminal_send`) lost the
contract's above-card description. It now prefers the description exactly as
BashRow does.
A settled call read `call?.cwd`, which cannot tell "the call omitted a cwd"
from "the paging window dropped the call head". The second case has no cwd
anywhere and the original call may have used an explicit workdir, so it now
draws a bare `$` instead of naming the session workspace.
Clearing waitingApprovals in handleConnected raced the reconnect replay:
mux frames flow from stream open while onConnected waits for the
readiness handshake, so a replayed approval/requested could land first
and be wiped — amber dot and answerable card lost until the next
generation. The sweep moves to generation death (onStateChange
'reconnecting'), before any next-generation frame can exist, and now
also drops buffered answerable frames (approval/question pairs) whose
dead-generation rpcIds could never be answered — a session instantiated
later no longer replays zombie takeover cards. session/queued buffering
already re-baselines per generation; this closes the same window for
the interaction frames.
Two races from the #572 review, still live in the ported registry:
An ask whose signal aborted between the service's own check and the
microtask-deferred waterfall dispatch would register its abort listener
AFTER the signal fired — never invoked, entry pending forever, zombie
frame on every mux replay. The answerer now settles 'cancelled'
synchronously before publishing anything.
The audit back-scan let a callId-less ask claim the newest unclaimed
asked record even when that record carried another call's id. Pairing is
now shape-symmetric: callId-bearing asks take exactly their call's
record, callId-less asks take only callId-less records — neither can
steal under parallel asks.
writeFileAtomic: exclusive-create random-suffix temp + rename carrying the
caller-stated mode; settings-local persistSection now consumes it. The
credentials-local store shares it next.
Disposability parity with the question provider: a gateway disposed while
approvals are pending settles every registry entry as 'cancelled' (the
service's fail-closed vocabulary), so no ctx.approval ask dangles past the
proxy's lifetime and mux subscribers see the withdrawal. Spec mounts the
proxy on its own fiber and drives dispose with a live ask.
Addresses the ds-review-bot suggestion on PR #851.
A multi-line command repeated the cwd label on every prompt row, which
states something the view does not know: it carries ONE working directory —
where the call started — and a `cd` in the command moves later lines
elsewhere. `cd ~` then `ls` rendered both rows labelled with the session
workspace while `ls` actually listed the home directory.
The label now appears on the first row only, and later rows keep a bare `$`
so they still read as prompts. Same reasoning as the run-state dot: neither
a per-line directory nor a per-line exit status exists to report.
The built-bundle snapshot records the effect on fixture turn 60's two-line
command (`fixture echo done` becomes `$ echo done`).
A popup on a host command is not a second command — it is what that
command's BARE invocation does on this client. CommandContribution loses
hostBacked (contributions are pure client commands again; a host-name
collision fails loud, unchanged for /model), and the contract gains
CommandDecoration + command.decorate(): key = the HOST command name, no
catalog row, no claim participation. Dispatch consults decorations only on
the bare paths (menu pick / bare enter) after the host row resolves; space
and argued enter never see them — the two edges hostBacked had to guard
explicitly hold by construction in the decoration model. A decorated name
with no host row in the session's directory never fires (a decoration
cannot manufacture a command).
ui-permission switches register→decorate with zero behavior change
(options still read the permissions projection; a pick still submits
'/permission <preset>'). Specs rewrite to the decoration semantics: no
catalog row, bare-enter popup vs argued-enter host claim, space host
claim, no-host-row miss, unavailable fall-through, duplicate fail-loud.
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`.
Selecting a preset from the hero pushed the session into the conversation
view: the /permission switch logs its command/run + command/done pair, the
pair folds into flow nodes, and the composerPhase predicate counted ANY
node as conversation — so the hero (composerPhase === 'blank') collapsed.
The host-side blank bit was already correct (sessionBlank = no turn/start;
knob events open no turn), but the client derives its phase from window
content, and command rows are log-only records, not conversation.
derivePhase's hasContent now excludes command nodes — the client mirror of
the host predicate. The knob events themselves never fold (not
surface-eligible), so the pair was the only leak. Covers /plan on the hero
identically (same lifecycle pair, same predicate).
Specs: the host blank spec pins the three knob events as standalone
events; a session spec drives the /permission pair through the live path
and asserts phase stays 'blank' while the command node renders.
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.
The headless-agent test fixtures and the e2e test that depended on them
are out of scope for this PR. Unit tests in tmux-context.spec.ts cover
the plugin behavior.
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.
The headless-agent test fixtures were removed; move the driver, cordis.yml,
and mocks into packages/context/tmux-context/tests/fixtures/ so the e2e test
is self-contained.
Confirmed and fixed, each with a regression test that failed first:
- Concurrent writes to different namespaces lost whole sections on disk
(each persist rendered the full document from a stale text): the local
provider serializes render->write->rename->text-commit on one internal
persist chain shared by every namespace queue.
- One throwing settings/updated listener starved the rest (cordis emit
stops at the first throw): commit fans out per listener via
events.dispatch, contains individual failures, and rethrows the first
INVARIANT-coded error only after every listener ran.
- Write queues ignored fiber/service lifecycle: the base init now
registers a teardown that refuses new writes and drains queued chains;
queued tasks re-verify service liveness and namespace ownership before
running and again before committing, so a registrant disposed
mid-flight is never notified and a disposed service never commits.
- Async watcher invocations could interleave (a slow stale call applied
last): each watcher carries a serialized invocation chain — one call
at a time, in commit order; JSDoc/doc pages state the async timing.
- update/replace borrowed the caller's object until the queued task ran:
inputs are structured-clone snapshotted at call time; non-cloneable
plain objects reject with a typed error.
- Composition guard now proves the documented fallback: the consumer
uses the optional scoped-inject shape and boots both with the settings
entry (hot publish) and without it (entry-config resolution, no scope).
- core-data-structures index: settings.md row added to the sub-page
table in core.md/core.zh.md.
Both packages hold per-file 100% coverage across repeated runs.
Confirmed and fixed, each with a regression test that failed first:
- Concurrent update() lost patches (merge over one stale snapshot):
per-namespace serialized write queues; a failed write cannot poison
the queue for later writers.
- Fixed-name .tmp write followed planted symlinks and kept stale modes:
random-suffix sibling, exclusive-create (wx), 0600, cleanup on
failure, then rename.
- A throwing settings/updated listener escaped commit and permanently
wedged the provider reload chain (rejected refreshTask): commit now
contains listener failures (INVARIANT-coded errors still propagate),
async watcher rejections are adopted and contained
(watch callbacks are officially void | Promise<void>), and the
provider chains refreshes on a settled tail with an error log.
- No way to remove a user override: scope/service replace(section)
sets the user section wholesale; replace({}) re-inherits base and
schema defaults.
- The three-primitive provider contract did not hold (base never
called load()): the base Service.init loads and publishes once;
settings-local delegates via yield* super[Service.init]().
- Dispose did not quiesce: teardown flags closed, closes the watcher,
then awaits queued/in-flight reloads; closed is re-checked across
await points.
- Invariant now checks the authoritative relation with the seam's own
deepEqualJson: emitted next must equal settings.get(ns), and
next/prev must differ structurally (cosmokit dependency dropped).
- New docs/core-data-structures/settings.{md,zh.md} with type-equiv
blocks + manifest entries; catalog types moved from exemptions to
LINK_MAP; website page registered.
Both packages stay at per-file 100% coverage.
Two-package capability family mirroring session-persistence/:
- dsh-settings: abstract Settings service — namespace registry with
caller-fiber effect registrations, layered resolution (schema defaults
< composition base < user document), schemastery validation,
per-namespace deep-equal commit detection, and the settings/updated
event. Boot/registration validation fails loud; provider publishes
keep last-good per namespace.
- dsh-settings-local: settings.yaml/.json provider — resolveSpec
defaulting to $DSH_HOME/settings.yaml, chokidar hot reload,
content-equality self-write suppression, atomic 0600 tmp+rename
writes, comment-preserving YAML namespace patching.
Consumers register inside ctx.inject(['settings'], …), so every
composition works unchanged without a mounted provider. Real Loader +
Include composition test proves cordis.yml boot and external-edit hot
propagation; HMR disposal test proves registry cleanup. Both packages
hold per-file 100% coverage.
Doc budgets rise 1705→1710 (AGENTS.md) and 835→845 (packages/README.md):
one structural line per file for the new package group.
Agent Note: .agents/notes/implemented/architecture/2026-07-28-user-settings-seam.md
