The mount-local baseline guard was seeded from "a baseline already exists in
the log", which a resumed session and a hot plugin remount both satisfy. That
made a resume skip its baseline, so offline AGENTS.md edits or removals never
reached the first resumed request — violating the documented resume contract.
Distinguish the two by agent/session-start: a startup or resume emits it before
the first step, while a remount attaches to an already-live session and never
witnesses it. Only a remount (no witnessed start, baseline already logged)
keeps the single logged baseline and skips; a resume falls through and
re-composes from current files. Adds a regression that resumes a session with
an offline baseline edit and asserts the fresh baseline reflects it.
goal-session rides retry turns and survives admission failures. A
recovery policy closes a goal round's failed turn and reopens its
history under a retry trigger; the attempt now adopts that turn and
drops the failed turn's provisional reason, so the round settles from
the retry's own outcome instead of blocking an armed goal with
turn-error after a successful response. A downstream admission hook
that throws (rather than blocks) used to strand the queued reservation
forever; the listener now clears a still-turnless matching attempt on
the rejection path and reschedules the round.
agent-loop contains a persistently rejecting step close in the catch
path the same way the finally contains the turn close, so the
post-finally tail always publishes the terminal status — previously a
double veto escaped run(), leaving status at running while whenIdle()
resolved. The whenIdle catch arm is annotated as the backstop it now
is: every driver rejection path is contained today.
workspace-context folds an already-appended baseline from the session
log when the plugin is hot-remounted over a live session, instead of
injecting a duplicate from its fresh mount-local guard.
The TUI's reference-admission discard listener installs before
followup(): admission runs synchronously inside it on the common path,
so a listener installed afterwards missed its own cleanup and leaked
one callback per referenced prompt.
Downstream packages lost the tests that exercised their agent-loop-facing
edges when the loop was rewritten. Restore 100% per-file coverage with
behavior tests through public seams: llm-retry config validation and
cancellation races, goal replay drift/staleness/teardown edges, plan-mode
disposed-flush, workspace-context empty-change commits, api-proxy
synchronous send failures, acp-snapshot spill-path extraction and refresh
write-back, ACP injection-triggered turns, cli-demo and tui inbox
lifecycle edges, and agent-loop retry/settlement/lifecycle branches.
The only source changes are narrowly-justified v8 ignore annotations on
invariant guards and one redundant-guard removal (workspace-context).
Replace send/steer/inject with one Agent.send primitive over the
(target × wakeup) matrix; followup/steer/inject become fixed-preset
alias methods on the now-abstract Agent class. Coalesce context/message
into user/message (injected context is a non-user source). Replace
agent/queued with agent/inbox/enqueue/dequeue/discard, add cancel
keepInbox, and add a FIFO-conservation invariant.
- workspace-context: a transiently unavailable but still-effective candidate
keeps its cached trimmed digest in the directory's dedup slot, so an
identical later sibling is not emitted as a duplicate set until the next
successful reconciliation
- app-boot: --resume rejects a following token that is itself resume syntax
instead of accepting it as a session id
- tui: the queued-steering badge tracks per-entry sources and a drain removes
one matching entry, so loop-authored steering (no agent/queued) cannot
consume a pending user message's slot
Load a per-directory local overlay in addition to the base instruction
file, matching the Claude Code AGENTS.local.md / CLAUDE.local.md
convention for git-ignored personal guidance.
- New config `localInstructionFileCandidates`, default
`['AGENTS.local.md', 'CLAUDE.local.md']`; empty disables the overlay.
The default lives in the plugin Config schema, so every front door
(TUI/ACP/headless) reads .local. files consistently.
- Per project directory the plugin loads the first-existing base
candidate, then additively the first-existing local candidate,
rendered after the base so it takes precedence within the byte budget.
- Base and local tiers get distinct scope keys via a NUL sentinel
(scopeKey/decodeScopeKey) so they never collide in the baseline map,
pending window, or version cache.
- The fixed user-global $DSH_HOME/AGENTS.md stays base-only.
Docs: README (config, lifecycle, Known Limitations), regenerated
config-catalog, and a new bilingual Agent Note cross-linked to the
owning workspace-context note. 100% per-file coverage retained.
Squash of the linearized fix/tui-color-scheme-v2 merge and the follow-up
catalog regeneration. Renders error cause chains at every diagnostic seam
(origin/fetch-failed-diagnostics) and adds color scheme detection with a
light-terminal palette.
Delete packages/ui/stdio and examples/repl-agent; rename stdio-demo to
@deepseek-ai/dsh-tui-demo (TUI-only, refuses pipes before Loader boot).
tui-agent owns the coding composition inline; echo-agent and the CI demo
smoke move to the one-shot cli-demo bin, which gains -p/--prompt. The
UI-independent with-key e2es move verbatim to tui-agent. SDK wizard's
'stdio' interface becomes 'tui'. PTY testing stays confined to TUI
surfaces; all other subprocess tests ride pipes.
See .agents/notes/implemented/simplification/2026-07-20-retire-readline-front-door.md
Delete @deepseek-ai/dsh-home and make dsh-paths the sole owner of the
single-root harness home ($DSH_HOME || ~/.dsh). Migrate tool-bash,
skill-local, and agent-spine-demo off dsh-home, and fold telemetry's
divergent globalConfigDir onto the shared resolver, dropping its second
XDG/APPDATA policy and the deepseek-harness namespace so the anonymous
id lives under the harness home. Add dshHomeDisplay() for symbolic
user-facing paths, replacing workspace-context's bespoke check.
context/message previously defaulted to a <context source="…">…</context>
wrapper. No model is trained on a <context> tag either, and message
framing does not belong on the session surface: the surface projects the
durable log, and a caller that wants a frame formats its own content —
which the one heavy producer (workspace-context) already does with its own
<system-reminder> frame, opting out via 'raw'. The tag only added
machinery — ContextEnvelope plus an envelope field threaded through
InjectOptions, HookContext, the context/message event, and the agent-loop
inject/additionalContexts plumbing.
context/message now projects its content verbatim as a user-role message,
sharing one deriveEventMessage case with user/message and steering/message.
ContextEnvelope and every envelope field are removed; context/message.meta
still carries durable, model-hidden JSON state. Regenerated catalogs and
website API; refreshed the three affected keyless snapshots (envelope field
only; timestamps unchanged).
Broadens and renames the steering Agent Note to cover both envelope
removals as one decision.
Agent Note: .agents/notes/implemented/simplification/2026-07-20-unwrap-injected-content-envelopes.md
_Kerman
ZiyaZhang
Turtle
Tianyi Cui