Translate a mid-read AbortError from readFile into the seam's structured
FsError('FS_ABORTED') in readWholeText and readForEdit (the streaming/write
paths already did), and make the socket-type probe test reject on a listen
error instead of hanging where unix-domain sockets are unavailable.
Implements the split-the-filesystem-seam RFC. ctx.fs shrinks to a text-storage
provider seam (resolve/stat/readText/streamText/writeText/editText with branded
FsTargetKey/FsVersion and an explicit FsWriteExpectation); the new
dsh-file-context package owns the model-facing policy (read windowing,
observed-state, write/edit freshness) as the concrete ctx.fileContext service.
Authorization is now freshness-based rather than full/partial view: a windowed
read records the file version and authorizes a later edit when the file is
unchanged, removing the dead-end where reading lines 100-150 of a large file
could not edit line 120. editText stays a provider primitive so version guard +
literal match + atomic rewrite remain one critical section, and the stale check
runs before matching so a stale edit reports FS_STALE_VERSION. tool-fs injects
fileContext, never reaching around to ctx.fs (the no-bypass contract).
The compaction e2e never exercised compaction: its window/fixture combo
(contextWindow 8000, thresholdRatio 0.5 → threshold 4000; four small files)
peaked at ~1389 estimated tokens, so compactIfNeeded declined every pre-step
and compact/start never landed. Shrink the window (contextWindow 2400 →
threshold 1200; retainTokens 500 + summarizationMaxTokens 300 = 800 < 1200,
convergence holds) and grow the fixture to six files so a couple of bash steps
reliably cross the threshold. Verified compaction fires and the suite passes
across repeated real-API runs.
Sync docs left stale by the landed compaction work: list compaction.e2e.ts and
keyless-smoke.e2e.ts in the coding-agent README (and fix the wrong "Both
self-skip" count), add compaction to the examples with-key inventory, and
replace the hypothetical compaction/marker / "future plugin" naming in the
session README, session types JSDoc, and the core-data-structures catalog with
the real compact/start, compact/summary, compact/end events.
Manual review round, two non-blocking findings:
- CBR-005: _extractText's JSDoc claimed it "walks events in log order",
but it walks the seqs in surface order (the inline comment already said
so) — the exact distinction CBR-001 paid for, since after a replace a
high-seq checkpoint heads the surface before lower-seq retained nodes.
Corrected the JSDoc to match.
- CBR-006: the "HMR safety" suite only asserted registration; the actual
dispose-and-confirm-cleanup test lived under "llm inject", so a reader
searching by name could miss it. Added a disposal test to the HMR-safety
suite (mount via the real plugin fiber with LlmService present so inject
resolves, dispose, assert ctx.get('compact') is undefined) and reframed
the llm-inject test's trailing teardown to point at it.
Codex round 1 CBR-003: several docs still described compaction as an
`agent/request` waterfall concern, and the implemented compaction RFC
claimed "No changes to dsh-session or dsh-invariants" while the diff
changed both.
- Package READMEs / JSDoc (agent, agent-loop, system-prompt, compact,
compact-basic): compaction now lives on the serial `agent/pre-step`
seam (fired after turn/start, before step/start); the structural guard
is tool-pairing balance (`isToolPairingBalanced`), not step-alignment;
the convergence bound is strict (`>=` rejects).
- architecture.md / core-data-structures/compaction.md: same seam +
predicate + dispatch-mode updates; regenerated cordis catalog.
- Implemented compaction RFC, updated in place to describe shipped
reality: the seam is `agent/pre-step` (@mode serial) fired before
step/start; alignment is surface tool-pairing balance; the convergence
invariant rejects `>=`; and the "no dsh-session/dsh-invariants changes"
claim is corrected — dsh-session gains the tool-pairing predicate and
dsh-invariants drops its `start <= end` replace assertion (a positional
replace makes start > end normal).
Codex round 1 CBR-002: `resolveConfig` rejected only
`summarizationMaxTokens + retainTokens > threshold` (allowing equality),
but `compactIfNeeded` declines only when the estimate is `< threshold`.
At exact equality the post-compaction history sits at the threshold and
re-triggers on the very next check.
Make the bound strict (`>=` rejects), so post-compaction history is
guaranteed strictly below the threshold. Updated the boundary test (the
sum-equals-threshold case is now rejected, not accepted) and added an
"accepts just below the threshold" case; nudged one unrelated config that
incidentally sat at the equality boundary.
Codex round 1 CBR-001: a head-anchored compaction checkpoint was
mis-classified by the log-position step-alignment scan, so a second
auto-compaction over a checkpoint-headed surface silently failed.
Root cause: `isStepAlignedStart/End` scanned the LOG by seq, but a
`replace` op lands a checkpoint at a high log seq whose SURFACE position
is the head — its log neighbours (the open step's assistant/message) are
not its surface neighbours, so the forward scan wrongly reported mid-step.
Fix, per the agreed direction:
- Replace the two log-position predicates with one surface-anchored
helper `isToolPairingBalanced(nodes, events, beforeSeq)` in
`dsh-session` (renamed step-boundary.ts → tool-pairing.ts). A cut is
balanced when no unanswered tool-call precedes it on the surface; a
region is collapsible iff both edges are balanced cuts. The open-tail
and free-node cases fall out of the same counter. It also throws on a
corrupt surface (a tool/result with no matching call).
- Move compaction off the in-step seam to a new "pre-step" seam fired
after turn/start and before step/start, so a compaction's log-only
compact/* records and its replacement node land cleanly OUTSIDE any
step (the honest structure crash-safety relies on). Renamed the event
agent/pre-request → agent/pre-step and switched its dispatch from
parallel → serial (listeners mutate the surface as a side effect;
serial isolates them so concurrent appends can't interleave). Extended
the catalog generator to accept @mode serial.
Regression coverage: a real-loop test driving an auto-compaction asserts
the landed checkpoint is a balanced cut on both sides; unit tests pin the
checkpoint case, the mid-step injection case, multi-call steps, and the
corrupt-surface guard. Proven red on the old log-position logic.
Reform the compaction blueprint so a runaway turn survives and the design
stops drifting across review rounds:
- Drop in-flight-turn protection ("layer 2"). Retention is a uniform tail→head
whole-unit walk; the only structural guard is step-alignment. A single turn
that alone exceeds the window now compacts its own early closed steps instead
of being retained verbatim (the failure mode that motivated this).
- Move auto-compaction off the agent/request waterfall onto a new awaited
agent/pre-request loop seam, fired before history derivation. Compaction
mutates the surface; the loop derives once from the result — no double-derive,
and a listener structurally cannot act on not-yet-derived messages.
- Tighten compactIfNeeded to required (session, system, model, signal).
- Enforce a single-pass convergence invariant in resolveConfig: reject configs
where summarizationMaxTokens + retainTokens exceeds the threshold, so a
compaction can never immediately re-trigger.
- Document the crash vs recoverable failure taxonomy; core session repair stays
compaction-agnostic (a log-only orphaned compact/start is inert).
- Wire dsh-compact-basic into examples/coding-agent and add a with-key
compaction e2e (compaction's first real-world exercise + runaway net).
- Rewrite the RFC to encode the blueprint and move it to implemented/.
The runaway-turn snapshot is a named deferred follow-up: dsh-llm-replay cannot
yet serve the interleaved summarization model call.
Collapses the per-round review churn of the prior compact-basic branch into a
single clean baseline on top of compact-interface, so the upcoming retention
refactor lands as fresh, well-scoped commits rather than stacking on a history
of fixes that are being superseded.
P1: both merge parents shipped SCHEMA_VERSION=3 for different layouts (surface
columns vs seed_length), so an on-disk 3 was ambiguous and wrongly accepted.
Bump to 4 (merged layout) so the version check rejects both sibling v3s.
P2: a surface-eligible event with no surfaceOp lands in the log but vanishes
from deriveMessages() (surface is the sole derivation path). The typed append
overload enforces the marker only when the type arg is a literal; it collapses
to optional when widened to the union (a caller iterating raw events). Guard at
runtime in both append() and the seed constructor — no backward-compat for
surface-less logs. Shared seed fixtures carry surfaceOp explicitly and the
appendLog helper forwards it verbatim (no synthesized default). Exports
isSurfaceEligibleType. Regression tests for all three, each verified to fail
on the unfixed code.
Gates: typecheck, test (1115), snapshot (14), doc-sync, lint, build, hygiene green.
Reconciles the session-surface work (surfaceOp/sourceEventSeqs provenance as
the sole derivation path) with master's worktree-subagent series (fork-seed
boundary + out-of-process subagent backends).
Semantic reconciliations beyond the textual auto-merge:
- SQLite SCHEMA_VERSION: both sides bumped 2->3. Merged to a single v3 carrying
BOTH column families — master's seed_length on `sessions` and surface's
source_event_seqs/surface_op on `events`. writeRow + both INSERT sites bind
the full set; the schema doc lists all three added columns as the v2->v3 gap.
- agent-loop runStep request: master's `sessionId: session.id` and surface's
per-append surfaceOp/sourceEventSeqs coexist (different regions).
- Fork seed + surface: a fork seeds the child from the parent's LIVE events,
which now carry surfaceOp, so the child's surface rebuilds correctly. Verified
end-to-end — the subagent-fork replay recalls the inherited "SAFFRON" codeword
through the seeded prefix.
- Subagent snapshot fixtures (recorded pre-surface) re-enriched via KEYLESS
deterministic replay: only surfaceOp/sourceEventSeqs added onto existing
recorded lines (matched by seq), no recorded value changed. Not re-recorded
against the live API.
Gates: typecheck, test (1112), test:snapshot (14), doc-sync, lint, build,
hygiene all green.
A fork subagent seeds its child session with a prefix of the parent's log, and
that seed becomes the child's persisted log — so a fork child's .jsonl begins
with the PARENT's events, including the parent's assistant/chunk events. The
snapshot replay harness derived a child's script from its whole log, which would
replay the parent's recorded responses as the child's model calls. Spawn-only
scenarios never hit it, but a fork snapshot would mis-route silently.
Record the seed boundary and skip the inherited prefix at replay:
- SessionHeader gains an optional `seedLength` (how many leading events were
inherited via a seed), threaded through CreateSessionOptions/CreateAgentOptions
meta and stamped by the fork backend (= seeded-prefix length; absent for spawn).
It is EXPLICIT, never inferred from seed.length: a resume seeds the whole stored
log, so the resume path passes the persisted boundary back.
- Both persistence backends round-trip it: JSONL header line, SQLite seed_length
column. The SQLite table change bumps SCHEMA_VERSION 2->3; per the pre-release
stance the backend rejects an older user_version on open with NO migration.
- llm-replay's parseSessionHeader reads seedLength and loadSessionScripts derives
a child script from events AFTER the boundary. seedLength is 0 for spawn, so
spawn replay is byte-for-byte unchanged.
Closes the routing-correctness gap the per-session snapshot replay RFC under-
stated; a recorded fork scenario remains a future addition but now derives
correctly. RFC: docs/rfc/implemented/testing/2026-06-22-fork-child-replay-seed-boundary.md.
Regression coverage: a fork child fixture whose seeded prefix carries a parent
chunk (derived script must exclude it, proven red without the slice); a seedLength
persistence round-trip through the shared coordinator contract (both backends);
the fork backend stamping it; resume preserving it from the persisted header.
Two findings on the ACP backend:
Blocking: cancel() only sent session/cancel, so a child that ignores the notify
or wedges the prompt left result hung forever — the model-facing tool awaits
result before its finally disposes, so the parent cancellation hung and the
child stayed alive, violating the SubagentRun.cancel() contract (result settles
aborted). The result path now races the ACP drive against a cancelSettled
promise that requestCancel resolves, so result settles aborted the instant a
cancel is requested, regardless of the child. dispose() still kills+reaps the
process. New MOCK_IGNORE_CANCEL mock mode (receives cancel, never resolves the
prompt, never exits) drives a regression proven to hang without the race.
Nit: the drive-path catch was an empty broad catch that discarded the error
(AGENTS.md forbids). Because cancellation is now handled by the race arm, a
rejection reaching the catch is always a genuine child-level error — bind it,
flatten to error, and surface the original via a new AcpRunSpec.onError sink
that the provider wires to ctx.logger.warn, so a real fault is preserved.
addEventListener('abort') does not fire for a signal already aborted before the
listener is added, so a parent step cancelled before the subagent tool ran
would never reach the child — the tool leaned on each provider re-checking
request.signal itself, leaving the bridge's own claim incomplete for any
provider that relies on run.cancel(). Re-check exec.signal.aborted right after
registering and cancel explicitly. Regression test uses a spy provider that
only reacts to cancel() (never inspects the signal); proven to hang without the
fix (result never settles) and settle aborted with it.
Adds the @deepseek-ai/dsh-compact interface package: the abstract
CompactService (ctx.compact) with compactIfNeeded / compactRegion, the
compact/* session-event types via SessionEventMap declaration merging, and the
capability-seam RFC. Wires the package into the three root tsconfigs and the
cordis catalog. A backend implementation lands separately.
Master's #36 moved declaration output to lib/types (and types/exports/files
point there). The merge applied that to all pre-existing packages, but the
subagent backends introduced on this stack (subagent-inprocess, subagent-spawn,
subagent-fork) still used the old lib/ layout. Bring them onto the new
convention and add them to the single typecheck tsconfig.json references.
The shared run driver lived inside dsh-subagent-spawn, so the spawn package
carried fork-aware seeding logic and dsh-subagent-fork depended backward on
dsh-subagent-spawn — the two in-process backends were not independent.
Move the driver (startInProcessRun, depthOf, SubagentDepthError,
InProcessRunOptions) into a new pure-library package
@deepseek-ai/dsh-subagent-inprocess that registers nothing. spawn and fork now
both depend only on that driver and neither knows about the other; spawn no
longer re-exports it and fork no longer imports from spawn.
Also wire BOTH backends in examples/coding-agent/cordis.yml (config-only): load
dsh-subagent-spawn + dsh-subagent-fork + two dsh-tool-subagent instances with
distinct toolNames (subagent → spawn, subagent_fork → fork), demonstrating that
exposing multiple transports needs no code change.
Two round-3 findings:
(A) The EOF-quiesce window reused the 3000ms SIGTERM grace, the SAME value as
dsh-bash-local's own SIGTERM->SIGKILL grace. The child acp-agent's EOF teardown
disposes its loop, which stops child-owned bash -- and a SIGTERM-trapping bash
grandchild can hold that for up to ~3s before its own SIGKILL, then the child
still owes a final flush. With both graces equal, the parent's SIGTERM fired
exactly as the child reached its own SIGKILL+flush, cutting it off. Split the
EOF grace into its own knob (disposeEofGraceMs, default 6000ms) that exceeds a
single signal-grace of nested-teardown headroom. The child is an arbitrary ACP
agent, so the value is a standalone generous default, NOT derived from any
child's internals. Tier-1 test now uses a flush that outlasts the SIGTERM grace
but fits the EOF grace, so it lands only because the EOF tier honors its own
wider window (proven RED when tier 1 reuses the small SIGTERM grace).
(B) The middle-tier (SIGTERM) test only asserted dispose returned in time, so
an EOF->SIGKILL ladder with the rung removed would still pass. The mock's
MOCK_IGNORE_EOF mode now installs a SIGTERM handler that touches an observable
marker before exiting; SIGKILL is uncatchable, so removing the SIGTERM rung
leaves the marker absent (proven RED). The test asserts the marker exists.
dispose() ended stdin and sent SIGTERM in the same tick, so the child's
EOF-driven quiesce had no window to run. The real acp-agent has no SIGTERM
handler in a normal session — it flushes persistence and stops child-owned
work via the server bridge's connection-close path (conn.closed → per-agent
dispose → final session/flush), driven by stdin EOF, NOT by a signal. A prompt
response can resolve from a turn/end before that post-turn flush lands, so the
child still owes durable work when dispose runs; a same-tick default SIGTERM
terminated it mid-flush, orphaning child-owned bash and dropping the flush.
dispose now waits for the child's natural exit after stdin EOF first, then
escalates SIGTERM (grace), then SIGKILL — a three-tier ladder. Add an
`exitsWithin` helper for the bounded waits.
Regression coverage: a new mock mode (MOCK_FLUSH_ON_EOF) flushes a marker
asynchronously on EOF then self-exits; the tier-1 test asserts the marker
lands (proven RED on the same-tick-SIGTERM ordering — child killed mid-flush).
MOCK_IGNORE_EOF covers the middle tier (ignores EOF, dies on default SIGTERM);
the existing MOCK_TRAP_SIGTERM test covers the SIGKILL tier.
Two lifecycle findings from the review:
- A (blocker): dispose() could hang forever. It only sent SIGTERM and awaited
exit, with no escalation — a child that traps SIGTERM (or our acp-agent if it
doesn't quiesce on stdin EOF) would wedge dispose, stranding tool-subagent's
finally cleanup and orphaning child-owned work (e.g. bash subprocesses). dispose
now: ends stdin (graceful ACP close so the child can flush + exit), SIGTERM,
then escalates to SIGKILL if it doesn't exit within a grace period
(DEFAULT_DISPOSE_GRACE_MS, injectable via spec.disposeGraceMs), awaiting the
certain exit. Mirrors the bash executor's bounded teardown. Regression test
drives a SIGTERM-trapping mock subprocess and asserts dispose returns promptly
— proven to hang (red) without the escalation.
- B: an already-aborted request still spawned the configured binary. startAcpRun
now returns an inert already-aborted run BEFORE spawning, so a pre-cancelled
request launches nothing. Test points the command at `touch <sentinel>` and
asserts the sentinel never appears.
The dispose regression test exposed (via systematic-debugging) that the child
must signal trap-armed readiness before the test cancels — a bare timeout raced
the trap install and the default SIGTERM handler killed the child, making the
guard a no-op. The mock now touches its ready file once the trap is in place and
the test waits on that condition. The `cancelled` flag moved onto a holder object
so TS control-flow doesn't narrow the catch-time read to always-false.
The first OUT-OF-PROCESS subagent backend, proving the seam generalizes past the
in-process backends. @deepseek-ai/dsh-subagent-acp runs each child agent in a
spawned subprocess, driven over the Agent Client Protocol as the CLIENT — the
direction-inverted twin of the dsh-acp server bridge. Point the configured
command at the acp-agent example and the harness talks to its own process.
- Fresh process per run: start spawns, runs one ACP session (initialize →
newSession → prompt), dispose kills the subprocess and awaits its exit.
- Minimal client stub: advertises no fs/terminal; accumulates agent_message_chunk
text as the result output; auto-answers session/request_permission by a
configured policy (reject default / allow). No start-time capabilities (an
out-of-process child can't enforce the parent's depth/tool-filter); ignores
request.parent; injects only `subagents`.
- StopReason mapping (end_turn→completed, cancelled→aborted, …); result resolves
error/aborted on a child failure, never rejects (seam contract).
- Security: credential-shaped ambient env vars are scrubbed; the child's own key
is forwarded only via explicit config.env. A spawn-level error (ENOENT) is
captured and raced against the ACP drive so a bad command settles error rather
than crashing the parent.
Testing designed at every tier: keyless integration drives a scripted mock ACP
server subprocess (cancellation incl. the pre-newSession race and a
torn-pipe-after-cancel, permission auto-answer, non-message updates, spawn
failure, HMR, export shape) at 100% coverage; a with-key e2e drives the REAL
acp-agent example process (PONG + real file write, verified on disk) — the
harness driving itself. Snapshot coverage of an ACP child is deferred as
TODO(acp-subagent-replay) (each child is its own process with its own replay).
Stayed on @agentclientprotocol/sdk 0.25.1: the proposed 0.28.x bump only
deprecates the stable ClientSideConnection/AgentSideConnection API this layer
uses (33 sites incl. the server bridge), turning no-deprecated red across code
this PR shouldn't rewrite — that fluent-API migration is its own follow-up. The
backend needs nothing 0.28.x adds.
This completes the subagent seam stack (PR1 interface → PR2 in-process → PR2.5
snapshot infra → PR3 ACP); the seam RFC moves to implemented/, amended.
Reconcile the session-surface feature with master's package reorg and
simplifications:
- Adopt master's folded usage (assistant/message.usage; standalone `usage`
event dropped) and re-attach surface metadata (surfaceOp/sourceEventSeqs).
- Add surface opts to master's new max-tokens assistant/message append.
- Port surface columns onto the coordinator-refactored SQLite backend at its
new path; drop the dead v1->v2 migration (bump-and-reject, no migration per
pre-release policy).
- Move the session-surface RFC into implemented/architecture/ and refresh its
stale body (no migration, SESSION_FORMAT_VERSION=0, renamed package paths).
- Update the core-data-structures catalog SessionEvent blocks for the two new
surface fields; regenerate the cordis catalog.
- Re-harvest ACP snapshot fixtures (keyless replay) to carry surface metadata.
The createdAt+recordedId child sort comment over-claimed "tie-safe". Codex
flagged that a same-millisecond sibling tie would be broken by random session
id, which does not recover first-call order. In the current synchronous cut that
tie is unreachable — the subagent tool awaits one child's result and disposes it
before the parent starts the next, so siblings' createdAt values are strictly
ordered and match first-call order. Restate the comment to that real invariant
(at both the replay sort and the harvest sort), note that the id tiebreak only
makes a degenerate collision deterministic, and flag the concurrent-subagent cut
that would need a real first-call ordinal with XXX(concurrent-subagents). The RFC
records the same limitation. Comment/doc only — no behavior change.
Dudu-0223
Hypatia May
Yichen Jiang
Tianyi Cui
imccyu