set() on an idle agent appends plan/mode at once — no request boundary
would arrive until the next prompt, so a queued intent used to hang as
pending forever (the composer showed a dead pending target). A running
agent keeps the boundary-flush path unchanged. set() now reports which
branch ran (committed/queued/cancelled/noop); the /plan handler's copy
follows the branch (idle: "Plan mode on/off", mid-turn: the next-step
wording), and both commit paths share the header-delta narration. The
invariant drops turn enclosure: plan/mode is a standalone whole-value
event (the synthetic log-only turns removal already established the
between-turns append shape). The fixture mirrors the idle commit.
CI's cold jsdom needs more than findByRole's 1s default between opening the
dialog and the fixture listing's first paint; the surrounding helpers already
wait 10s.
apps/cli/cordis.yml swaps bash-local/fs-local for the acp-agent composition
(sandbox-local + sandbox-policy + bash-sandbox + user-approval + permission
+ fs-sandbox; fs-policy composes on top unchanged). The deployment default
stays danger-full-access + never — byte-for-byte the old unconfined
behavior, so the replay e2e lane and demos are unaffected — while
DSH_PERMISSION_MODE opts a process into a confined default and the
/permission command switches per session. The permission preset table ships
the three product presets (read-only/ask, workspace-write/ask,
danger-full-access/never) explicitly in the deployment config.
PermissionSelect returns as the Access seat's wired occupant: options and
the current value read from the 'permissions' projection through the
standard-kit useProjection (no fetch, no mount timing — the resident
composer's mount-once fetch bug dies with the fetch), key absence renders
nothing (permission-less host, or a Draft with no session), and a pick
submits the '/permission <preset>' command line through the new
ComposerBarInjected.command callback (Session.command = command.execute
admission; the pushed projection frame lands the confirmed value). The
READONLY_OPTIONS placeholder and its local state leave InputBar.
The connection fixture mirrors the host: a permissions unit fold (three
knob events over the fixture preset table), the projections block +
baseline/push frames carry the key, and /permission joins the command
catalog with the same switch-through-knob-events handler shape.
The persisted row (sessionId, key, stateVersion, observedSeq, state)
becomes (sessionId, key, ver, seq, val) — the cache medium repeats these
three names for every unit of every session, so the long forms dominated
the JSON payload. ProjectionCheckpointRow and the checkpointRow zod spec
rename together; the domain spec bumps to v3 (cache semantics: the old
medium is discarded, not migrated). The unit-facing declaration keeps
stateVersion — only the persisted/checkpoint row shape changes.
A fresh injected face while the same request is open re-fires the effect;
the armed guard must not relaunch the chooser (the uncovered branch CI's
per-file gate flagged).
Drop the assertions eslint proved unnecessary (getByRole gains the element
type parameter instead, keeping tsc satisfied), declare the injected
setPlanMode as a function property (the seat face is this-free), and narrow
the fixture's command args without String()'s object stringification arm.
- While a picking flow is open (native chooser pending, browse dialog up) or
its pick is being adopted, every other menu action disables: a late outcome
must not race a concurrent selection or creation (ds-review-bot warning).
- ctx.directoryPicker joins the architecture Capability Services map (both
languages); neighboring rows condensed to keep the doc inside its ceiling.
- directory-picker-browse documents that its client half lands in the next
stacked PR: a -browse composition today hides the picking affordance (the
documented empty-hole default) rather than misbehaving (ds-review-bot
critical; the dialog itself ships in #821).
eslint --fix removed the no-unnecessary-type-assertion hits the review
knives introduced; the two await-thenable errors were vi.runAllTicks()
awaits in the cache spec, replaced with advanceTimersByTimeAsync(0) (the
fake-timer-safe microtask drain). Repo-wide eslint, dual-aggregate tsc,
and the focused 100% coverage set all green.
Static: the cache package.json files array matches the workspace
constraint shape, the unused dsh-storage-json devDependency is dropped
(tests run on the memory backend), and docs/module-graph.md is
regenerated for the new package edge.
Coverage: two unreachable branches deleted rather than tested —
coldSnapshot's floor-0 tail reuse (a baseSeq-0 restore never throws and
an unrelated record still carries a usable watermark) and flushSoft's
non-mandatory clean-skip (throttle triggers only fire dirty). New tests
close the real gaps: write() on a never-dirty session and the non-JSON
unit-state rejection, plugin disposal clearing armed interval timers,
cachedSnapshot's all-version-mismatched and cwd-identity arms, the
zero-units empty-log cut, the coordinator seek-hook ladder (suffix /
not-found / plain failure / abort-reason relay), and the superseded-
retirement race proving forget()'s exact-entry guard.
Review finding (PR #791): the column carried bare values (no seq), so the
client could not seed its value store without risking a stale list block
outranking newer push frames — and nothing consumed the column at all,
leaving cold titles absent after a restart. SessionSummary.projections is
now the same SessionProjectionsBlock as the history tail (values +
asOfSeq; attached rows cut the live registry, cold rows serve the cache's
identity-checked cachedSnapshot whose asOfSeq is the lowest served-row
watermark). SessionManager.refreshList seeds each row's block into the
per-session projection store via per-key apply — partial-baseline
semantics: an absent key never clears, and higher-seq-wins keeps stale
list blocks beneath push frames and tail baselines — so cold titles
surface in the sidebar without opening a session.
Review finding (PR #791): rows carried only version/watermark/state, so a
recreated session id, or a persistence store replaced under a surviving
cache, could pass every watermark check and seed state folded from an
unrelated log; a checkpoint racing ahead of an eager log flush could
likewise expose values no stored log contains. Records now store the
header identity (createdAt, cwd) they were folded from — reads validate it
against the live header (listing) or the tail's stored header (cold read)
and discard unrelated records whole (domain version 2 discards v1 media by
the pre-release stance). A live checkpoint additionally flushes the
session's buffered events durably before the cache row lands: the cache
can trail the log, never lead it. cachedValues is reshaped into
cachedSnapshot(meta): the identity witness plus the {asOfSeq, values} cut
the list carrier serves.
Review finding (PR #791): with no projection definitions registered,
restoreFloor() is undefined and the fast path returned a successful empty
snapshot without touching persistence — a nonexistent session 'succeeded',
violating the documented not-found contract in that supported topology.
The no-unit branch now probes readFrom(id, 0): an absent log rejects with
the seam's not-found, a present one dates the empty cut at its stored end.
The projection cache surfaced two substrate gaps, both properties of the
domain-KV stack it landed on. (1) The json backend's root is relative and
joined per-open against process.cwd(): sessions are patched global
(~/.dsh/sessions) but workspace.json/session_projcache.json land under
<launch dir>/.storages, splitting derived media from their source of
truth. Proposal: patch storage-json.root to ~/.dsh/storages beside the
session root (profile key storageRoot, mirroring persistenceRoot) and
resolve-once at backend construction, adopting the JSONL backend's
recorded rationale. (2) A damaged cache medium (truncated / version-bumped
/ schema-drifted) bricks fail-loud boot even though its content is fully
rebuildable from session logs. Proposal: DomainSpec grows
recovery: 'reject' | 'reset'; the facility, on exactly the damage-class
errors, destroys a declared-reset medium once and reopens empty —
workspace stays authoritative and loud. Bilingual pair recorded.
gen-cordis-catalog/api, config and persistence catalogs, and doc graphs
regenerated over the new sessionProjectionCache service and the registry's
checkpoint faces. Classifications: ProjectionCheckpoint joins the type-link
exemptions (owned by the projection package source), Partial joins the
foundation names, the cache service gets its capability-seam role row, and
the package takes the one-sentence Model Experience contract (host-side
read-model accelerator, no model surface).
The base's bilingual gate now covers this branch's new README: add the
Chinese counterpart of the session-projection-cache package README, the
cache row on the group README's Chinese side, and re-record the touched
pairs (group, cache, session-persistence — whose English side gained the
readFrom rows earlier on this branch). verify-translation-pairing: 541
pairs consistent.
The watermark cache is the registry's authoritative mutable state; a
checkpoint consumer holding the live reference could corrupt every
subsequent snapshot and frame through it. structuredClone at the read face
(total, by the unit plain-JSON contract) pins the boundary; a mutation test
proves the cache is unreachable through handed-out rows. restore and
viewCheckpoint only touch caller-owned rows — no other leak path.
SessionSummary grows an optional projections column (whole value per key,
same passthrough posture as the history-tail block): attached rows cut the
live registry watermark cache; cold rows view the persisted projection
cache's stored rows via the new registry viewCheckpoint face (version-
matching keys only, zero I/O) — the RFC's motivating scenario, every
session's title across a listing without loading one event log. The column
is fail-soft and absence-coded: no registry, no cache row, or a throwing
read serve the row without the column, never breaking the listing.
New package on the domain data form: one session_projcache record per
session (key → {stateVersion, observedSeq, state}), landing beside
workspace.json under the shipped json backend. Write policy: two mandatory
points (turn/end + session disposal) with count/interval throttling between
them (both Config fields required — flush cadence is a deployment choice);
every background write is fail-soft (log + stay stale, self-heal on the
next write or cold read). coldSnapshot(id) runs the read ladder — cached
rows + persistence readFrom from the registry's anchored restore floor +
registry restore + fail-soft write-back — detecting crash-repair-shrunk
logs via the one-below anchor and degrading to a single full re-read.
Mounted in apps/cli/cordis.yml (writeEveryEvents 200 / writeIntervalMs
5000).
The registry grows the state-level read ladder the persisted cache builds
on: checkpoint(session) snapshots every unit's {stateVersion, observedSeq,
state} row from the watermark cache; restoreFloor(checkpoint) anchors the
tail read one event below the lowest usable watermark (so a shrunk log is
provable); restore(checkpoint, events, baseSeq) refolds each unit from its
usable row (or from init over a full read), rejects rows a tail read cannot
fix (version mismatch / overreach with baseSeq > 0 => re-read from 0), and
returns both the snapshot and the refreshed rows for durable write-back.
ProjectionCheckpointRow/ProjectionCheckpoint are the persisted-row types
minus the record keys.
SessionPersistence grows readFrom(id, fromSeq, signal?): the non-mutating
read-from-seq primitive for checkpoint consumers (the persisted projection
cache folds only the tail past its watermark). Coordinator owns validation,
per-id serialization, and the sequential fallback (loadStored + forward
skip); SQLite implements the optional seek-capable loadStoredFrom hook
(WHERE seq >= ?), JSONL stays sequential by contract. Contract suite covers
suffix exactness, empty-tail, non-mutation, and cancellation; seam README
(both languages) documents the method and the hook.
The read side becomes the 'permissions' session projection: src/types.ts is
the key declaration's one home (PermissionSelect = whole select: table
options in declaration order plus a current-only 'custom'), served through
./types and the ./client re-export. The unit folds the three whole-value
knob events (permission/preset, sandbox/mode, approval/policy) into a plain
KnobState and views the select over the composition defaults the service
already owns; current() shares the same derive step, so the fold exists
once. The write side becomes the /permission command (the /plan
registration shape): bare invocation reports the current preset and the
table, a preset argument switches through set() immediately — no turn
anchoring (knob events need no enclosure), no dedicated RPC. Both children
activate only when their registry is composed.
directory-picker-browse becomes dual-face: its browser half fills
ui-workspace's two directory-flow holes with the Select Workspace Directory
dialog (figma Harness 813-23126 family — Miller two-column view, breadcrumb
with click-to-edit path zone, nested New-folder dialog), driving the node
half's host.listDirectory/host.createDirectory and owning its locale
namespace (directory-browser, zh default / en). The dialog moves here from
ui-workspace wholesale — the trigger surfaces keep only the flow-hole owner
conversation.
apps/cli flips its one directory-picker row -native -> -browse, swapping the
host backend and the client interaction together; picking now works for
remote deployments out of the box. The keyless workspace-flow snapshot boots
the browse bundle and drives menu -> dialog -> Documents -> project -> Open
against the fixture tree.
The simplified inheritance path uses Session creation seeds for policy events, and the public CreateSessionOptions JSDoc now names that supported role. The type-equivalent persistence catalog still described seeds as replay/fork-only, so doc-sync correctly rejected the mismatch.
Align the explanatory paragraph and exact type-equivalent block in both languages, then re-record the bilingual pair. This keeps the public catalog from understating the constructor seam that makes the simplification possible.
master's session-projection row landed the table at 874 words against the
870 ceiling this PR set; tighten the host/client rows this PR added instead
of raising the ceiling.
The implementation no longer stores inherited policy in SessionHeader or resolves a second baseline chain, but the feature, sandbox, and approval notes still described that machinery. Keeping those claims would make the smaller design look incomplete and invite reintroduction of the generic persistence surface.
Rewrite the owning feature note around the actual delegation snapshot: source-tagged policy events follow the optional fork prefix, ordinary last-event-wins folds establish precedence, and persistence captures the constructor seed with the first materialized batch. Condense the alternatives and consequences to the decisions and coverage that remain load-bearing.
Align the sandbox and approval notes plus the subagent-inprocess consumer README with that contract. Update the four Chinese counterparts minimally and re-record each pairing hash so both languages describe the same shipped mechanism.
