Why: Backend emitted tool deltas from a separate goroutine, racing with serialized run events and occasionally losing a UI delta.
Evidence: Backend race detector reproduced concurrent event writes; Agent-Native identity test RED→GREEN; full vet and tests pass.
Impact: runtime tool delta events now carry provider slot and call ID for single-owner downstream projection.
Refs: runtime/loop.go; runtime/event_vocabulary_test.go
Why: a model tool finish could execute partial JSON or an unterminated SSE stream, while post-Usage errors could lose their charge.
Evidence: missing-DONE, malformed-JSON, identity, non-object and post-Usage RED→GREEN tests; GOWORK=off go vet ./... and go test -count=1 ./... passed.
Impact: invalid tool terminals fail without execution or automatic retry and retain returned Usage; valid text and multi-part tool streams keep working.
Refs: runtime/llm/sse.go; runtime/loop.go
Why: Mutual exclusion alone does not preserve order. With parallel
execution the host cannot keep order-sensitive tools (card streaming
begin/append/commit, plan updates) correct by locking a mutex, because
whichever goroutine grabs the lock first is not necessarily the first
declared call - a real failure observed in the backend card-stream test.
Evidence: go test ./... green; TestLoopSerialToolsKeepOrder asserts the
serial calls execute in declaration order (s1->s2->s3) while parallel
calls still fan out.
Impact: New Config.SerialTools set. In parallel mode, matching calls run
sequentially on the caller goroutine in model-declared order; others
stay concurrent. Empty set keeps the previous all-parallel behavior;
sequential mode is untouched.
Refs: Dev task feature/agent-e1-parallel-fixes (F5).
Why: With parallel tool execution the host-side tool wrapper can no
longer infer which invocation it is serving from a single in-flight
slot; tools need the call identity (ID/name/arguments) to pair process
state and per-call bookkeeping. Context injection keeps the tools
registry signature unchanged.
Evidence: go test ./... green; TestLoopToolSeesOwnCallInContext runs
two parallel calls and asserts each reads its own ID from ctx.
Impact: Additive WithToolCall/ToolCallFromContext helpers; runOne now
wraps the ctx before Execute. Callers that never inject keep ok=false
and are unaffected.
Refs: Dev task feature/agent-e1-parallel-fixes (F5 parallel wiring).
Why: Parallel tool execution makes same-turn tool events interleave, so
consumers can no longer pair starts and ends by order; the kernel Event
needs the call ID. MiniMax interleaved-thinking models stream their
reasoning inside content <think> blocks unless the OpenAI-compatible
request opts in with reasoning_split=true, which forces reasoning into
the dedicated reasoning_content field the SSE layer already captures.
Evidence: go test ./... green; new tests TestLoopToolEventsCarryCallID
and TestOpenAIChatBodyReasoningSplit pin both behaviors (regression
anchors for the backend parallel-work change).
Impact: Additive Event.CallID (empty for non-tool events; existing
consumers unaffected) and an opt-in Client.ReasoningSplit flag that only
adds a request field when set. No default behavior change.
Refs: Agent-System-Dev proposal 2026-10-10-agent-delegation-routing-sandbox.md
(F2/F5), Dev task feature/agent-e1-parallel-fixes.
Why: hosts accumulating per message_end could not see cache hits; the
fields now ride the same event as the per-request totals.
Evidence: GOWORK=off go test green.
Impact: additive event fields.
Refs: Agent-Native#35
Why: cache tokens were parsed but never accumulated, the usage event
carried only cumulative totals and was skipped on terminal paths
(audit N20 residual), and stream retry was host-opt-in so subagents,
the verifier, and squad runs were never retried.
Evidence: GOWORK=off go vet + go test -count=1 ./... green;
TestUsageEventSplitsLastAndCache (event carries Last*/cache),
TestUsageAccumulatesCacheTokens, TestBudgetStopEmitsTerminalUsage,
TestDefaultRetryAttemptsTwo (0 value = 2 retries with
auto_retry_start events), TestNegativeRetryAttemptsDisables (negative
= off), TestRetryNeverDuplicatesForwardedDeltas (the only-before-first-delta
invariant holds under the library default).
Impact: usage events now carry Last* and cache fields and are emitted
on budget/error terminal paths too; Config.RetryAttempts defaults to
2 (negative disables) so every host gets the existing retry semantics
without wrapping; default (unset) host wrappers should be retired to
avoid double retries.
Refs: Agent-System-Documents#43 (N3 plan WP6), audit N20
Why: the completion doc assigns the Memory port to the library; today
memory recall is a product-side interface assertion with no library
surface, AgentSpec.MemoryScope has no consumer, and the model has no
write-back path.
Evidence: GOWORK=off go vet + go test -count=1 ./... green;
TestLoopInjectsMemoryAndRegistersWrite (recall block lands after the
system section, memory_write executes through the loop),
TestEmptyRecallInjectsNothing (empty recall changes nothing),
TestBlockRenderOrderingAndCaps (deterministic order, expiry/empty
pruning), TestWriteToolValidation (fail-closed scope/type/content).
Impact: new opt-in Config.Memory/MemoryScope; nil port keeps behavior
byte-identical; injection is request-scoped (never persisted, never
accumulates) and deterministic ([记忆] text shape preserved for the
product migration).
Refs: Agent-System-Documents#43 (N3 plan WP5)
Why: acceptance clause 3 requires span names and attributes to follow
the convention; today names were bare words with two ad-hoc attribute
keys, and every span got a fresh random trace with no parentage.
Evidence: GOWORK=off go vet + go test -count=1 ./... green;
TestSpanNamesFollowGenAIConvention (name suffixes, gen_ai.* key enum,
intValue usage, service.name resource, prompt/reserved-key exclusion),
TestLoopSpansFollowGenAIConvention (invoke_agent root with shared
trace, chat/execute_tool children with parent links, usage attrs on
chat spans); the pre-existing prompt-body denylist test passes
unchanged.
Impact: span names become invoke_agent <agent> / chat <model> /
execute_tool <tool>; attributes move to gen_ai.* with usage as
intValue; one Run shares one trace with invoke_agent as root. New
Config.AgentName/ProviderName (documented defaults keep old hosts
meaningful); telemetry StartFunc gains WithServiceName/WithEndpoint
options (defaults preserve current behavior).
Refs: Agent-System-Documents#43 (N3 plan WP4)
Why: the product conversation path needs the intent journal and
recovery pairing, but making the transcript the authoritative context
would freeze the assembled system prefix (memory/plans/injections go
stale). JournalOnly keeps the session as journal-only: recovery runs
against it, the context always comes from History, and the recovered
outcomes are injected as a system note. Separately, the one-time
History seed was memory-only and silently lost from the second run on
(authority mode); it is now persisted except in JournalOnly mode.
Evidence: GOWORK=off go vet + go test -count=1 ./... green;
TestJournalOnlyRecoversWithoutTranscriptAuthority (journal gets the
replayed pair exactly once, history never enters the journal, the
recovery note precedes the new user message, seeds stay out);
persistedProjection honors JournalOnly; all pre-existing durable and
harness tests pass unchanged (default semantics untouched).
Impact: new opt-in Config.JournalOnly; default (false) behavior is
byte-identical.
Refs: Agent-System-Documents#43 (N3 plan WP2)
Why: a kill between persist(assistant) and commitToolIntent left an
unpaired tail that no recovery path repaired (acceptance clause 1's
library gap), and Stop.Replacement mutated the in-memory assistant
while the transcript kept the original (audit N6 residual).
Evidence: GOWORK=off go vet + go test -count=1 ./... green;
TestResumePairsOrphanTailAfterCrashWindow (no-intent window pairs with
crashInterrupted, run continues), TestResumeMixedIntentAndOrphanTail
(intent path and tail repair never double-pair),
TestRepairOrphanTailNoopOnHealthyTranscript; the harness assertion now
pins replacement-as-user-note (forged=4, notes=3).
Impact: every next Run repairs the trailing assistant's resultless
calls before the first request; replacement text now appears only as a
user message in both memory and transcript (intentional, recorded in
the N3 card).
Refs: Agent-System-Documents#43 (N3 plan WP1)
Why: the completion doc's N2 requires old tool results to become file
paths (ZCode compact pattern) and cache-friendly prefix rules; today the
micro-compactor only left a transcript pointer and nothing pinned prefix
stability. alignToolTail had no direct edge tests.
Evidence: GOWORK=off go vet + go test -count=1 ./... green; new
TestClearOldToolResultsSpillsToDisk / TestSpillResultSanitizesCallID /
TestAlignToolTailEdges / TestSystemTextDeterministicAcrossIdenticalBuilds;
legacy placeholder bytes unchanged when ResultsDir is empty
(TestClearOldToolResultsEmptyDirKeepsLegacyPlaceholder).
Impact: opt-in via Config.ResultsDir; the native adapter wires it to
<workDir>/.agentx/results. No behavior change for other hosts.
Refs: 2026-10-07-agent-stack-n2-plan WP1
Why: N1 increment three (WP1/WP4) - agents need a canonical spec
format with plugin restrictions, hosts need typed input/output/tool
guardrails, delegation needs lifecycle events, and squad adapters
need to pass the permission ask port through.
Evidence: go vet clean; go test ./... 26 packages ok, including new
agentspec (12 tests), guardrails (5), loop_guardrails (5), and
subagent hook tests (3).
Impact: new runtime/agentspec and runtime/guardrails packages; loop
consumes UserPromptSubmit outcomes (deny skips the loop, modify
appends the rewrite); subagent gains RegisterWithHooks emitting
SubagentStart/Stop around delegation; GateOptions gains PermissionAsk
forwarded into the kernel config. All additions are opt-in: nil
ports preserve prior behavior exactly.
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md WP1/WP4
Why: The N1 harness phase needs the kernel to resolve ask decisions
through the host (paired HITL), emit the full hook lifecycle the plan
promises, and carry mode/rules/session-allow in one named object.
Evidence: go vet clean; go test ./... 24 packages ok, including new
loop_hooks_lifecycle_test.go (lifecycle order, prompt text, ask
allow/deny/legacy paths, stop continuation counts) and three new
policy tests (struct-form parity, session-allow store, nil store).
Impact: hooks adds PostToolBatch and SessionEnd events; loop emits
SessionStart/UserPromptSubmit/PostToolBatch/SessionEnd and PreCompact
lands in AgentSession.Compact; gateTool consults PermissionAsk on ask
and keeps ask=reject when the port is unset; policy gains Policy,
SessionAllows, and MemoryAllows. SubagentStart/Stop emissions are
deferred to the adapters assembly increment (factory site has no bus
access yet).
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md WP2/WP3
Why: a crash after a tool was about to run left the call unpaired, and workspace writes could not be rewound.
Evidence: GOWORK=off go test ./runtime/ ./runtime/llm/ ./runtime/checkpoint/ ./runtime/telemetry/ ./runtime/tools/ ./runtime/session/ ./adapters/native/
Impact: pending calls rerun only when both the stored and current replay declarations are safe; otherwise the result is an error. Empty fallback lists and a nil checkpoint log keep the previous behavior.
Refs: feature/durable-rewind
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Squad members never saw the direct-run stop, reminder, and before-tool gates, and there was no shared place to decide deny, ask, or allow.
Evidence: GOWORK=off go test ./runtime/hooks/ ./runtime/policy/ ./runtime/ ./adapters/native/ passed, including an A/B run where hook adapters match the old callbacks.
Impact: An empty bus leaves stop reason, reminder count, and deny count unchanged. Callers opt in by setting GateOptions.
Refs: feature/permission-hooks
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Old tool results were filling the window, compaction could split a tool group or commit a stale summary, and reminders arrived as extra user messages.
Evidence: Compaction retries a conflict and opens the circuit after three failures; clearing keeps the last five tool results; a stale file write is rejected; RPC reports started, queued, or handled.
Impact: Requests over the context window replace older tool results with a placeholder. Rewriting an existing file now requires a read in the same run. Subagent text over 32000 characters is capped.
Refs: feature/context-engineering
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Buffering a whole attempt hid live output, overflow detection missed common provider errors, and a cut-off tool call could still run.
Evidence: GOWORK=off go test -race -count=1 ./runtime/llm/ ./runtime/ passed.
Impact: WithRetry forwards events immediately and retries only when nothing has been forwarded. Length or an unclean end drops unfinished tool calls. Loop usage keeps a billing sum and the last request separately.
Refs: feature/llm-robustness
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: token budget let the current tool batch run, compaction could split a tool pair or race an append, writes were not crash-atomic, and several MCP and session edges only existed as comments.
Evidence: GOWORK=off go test -count=1 -race ./runtime/ ./runtime/llm/ ./runtime/session/ ./runtime/tools/ ./runtime/mcp/ ./adapters/native/ ./adapters/intentgate/ passed, including the TestR2* cases for budget, replacement, cancel, compaction, SSE, UTF-8, schema, MCP retry, session terminate, poison reload, and registry reload.
Impact: over-budget tool calls are skipped and paired; Stop.Replacement is appended to the session; the intent gate can follow Registry.Reload.
Refs: N5, N6, N7, N8, N9, N11, N12, N15, N16, N17, N18, N19, N21, I9
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: a tool-budget stop left tool calls without results, host callbacks raced, a canceled MCP read stole the next response, and fetch re-resolved after the SSRF check.
Evidence: GOWORK=off go vet ./... && go test -count=1 ./... && go test -race ./runtime/... ./adapters/native/... passed, including TestLoopBudgetToolCallCap, TestParallelHostCallbacksAreSerial, TestStdioSlowCancelThenEcho, TestPresetsAndClientForChatCompletions, TestFetchResolveChangesToLoopbackRejected, and TestFetchHostCheckRedirectStillRejectsPrivate.
Impact: over-budget calls get an error result and no tool event; one Run serializes host callbacks; MCP dispatches by id; anthropic and google presets use OpenAI-compatible chat endpoints; fetch dials the address it just checked. deps/agent-intent is 3130190; error-code-system and logging-system are a3ee67b and 49b5d8b.
Refs: N1, N2, N3, N4, N13, N14
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: A fixed continuation cap and a 4096-token Anthropic fallback stopped long replies before the agent finished.
Evidence: go test ./runtime/ ./runtime/llm/
Impact: MaxLengthContinuations of 0 keeps the old cap of 3; a negative value removes it. Unset Anthropic max_tokens defaults to 128000.
Refs: feature/stream-display-overhaul
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: the direct chat path still executes model-emitted calls that are not in the profile allow-list, while squad keeps the unknown-tool error when the fallback is unset.
Evidence: go test ./runtime/ passed, including TestLoopUnknownToolFallsThrough and the existing unknown-tool cases.
Impact: Config.UnknownTool is optional. Nil preserves today's registry error.
Refs: feature/agent-harness-p1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: the direct-path adapter accounts usage per turn, including runs that stop on a budget error before the final usage event.
Evidence: go test ./runtime/ ./runtime/llm/ after the field was added.
Impact: message_end and usage events carry TotalTokens. Callers that ignore the new field are unchanged.
Refs: feature/agent-harness-p1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: the direct chat path needs one shared loop for deliverable checks, tool pairing, reminders, and truncated replies, while squad callers keep today's behavior when the new fields stay unset.
Evidence: go test ./... in agent-native passed, including Stop cap, BeforeTool pairing, reminders absent from the session, repeat-guard cap, length continuation, and zero-value loop tests.
Impact: runtime.Config and llm.Event grow optional fields; AgentSession.Prompt copies them through. Squad sessions that leave them zero are unchanged.
Refs: feature/agent-harness-p1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Steering hooks existed on the loop but nothing owned the queues, abort, or a model change during a run.
Evidence: go test -count=1 ./runtime/
Impact: Run still emits the existing event names. message_start, tool_execution_update, and agent_settled are added beside them.
Refs: feature/pi-headless N3
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: product runners need register_artifact and card tools inside the kernel loop, and the UI needs token usage on the same event stream as text and tool calls.
Evidence: go test ./... in agent-native; TestHostToolAndUsageEvent and TestHostToolNameConflictFailsClosed.
Impact: Options.HostTools fail closed on name clashes; ProcessEvent and runtime events gain a usage type with token counts. Existing callers that ignore unknown event types stay compatible.
Refs: deliverable-quality-system B1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: the frozen exec port could not express per-request model, tool
allowlist, skills, budget, or identifiers, so squad roles and product
profiles had no way to steer one execution; contract v0.1.0 semantics
(budget/cancel) stayed unconsumed (parity-matrix #4/#10/#16).
Evidence: go vet ./... clean; go test ./... all ok, including new
loop budget caps (tool calls, prompt tokens, deadline), cancel
propagation into a blocking tool, registry Restrict fail-closed
policy tests, native adapter v2 tests (model override, allowlist,
deadline-bound context, execution-id log header, skills prompt line),
and settle classification marking budget_exceeded permanent.
Impact: exec.Request gains optional Model/ToolAllowlist/Skills/Budget/
SessionScope/ExecutionID/Env (zero value = legacy behavior); runtime
loop enforces a shared budget and emits budget_exceeded events;
adapters/native consumes all new fields. Consumers on the old port
keep compiling and behaving identically.
Refs: docs/parity-matrix.md rows #4/#10/#16 (Batch2 scope); deps/
development-guidelines standards/agent-execution.md (shared budget
across retry layers).
Contract codes now use the agent.* values, and runtime can emit structured logs without changing results when no logger is configured.
Co-authored-by: Cursor <cursoragent@cursor.com>
weichen
weichen