Why: the Q3 compiler maps core.WorkflowDef prompt steps onto squad graph
steps; the rendered prompt text needs a first-class carrier in the step
task brief (until now only leader delegations had a briefing channel).
Evidence: TestStepBriefRendersIntoTaskFile asserts the section presence,
body, and position (after targets, before the context table); full squad
suite green (SELF), CI authoritative (GATE).
Impact: additive StepSpec field brief + writeTaskFile section; no behavior
change when Brief is empty.
Refs: Q3 plan WP1 (2026-10-07-agent-stack-q2q3-plans.md); Dev task card
.task/feature-agent-stack-q3.md
Why: Q2 acceptance 1 requires a scripted-backend conformance for every one of
the 12 builtin topologies; task_board was the last unlit driver.
Evidence: wp4_conformance_test.go: diamond-dependency board respects the
dependency gate in claim order; orphan lease (claimed + expired) re-arms on
driver re-entry and re-claims with Attempts=2 via Resume; non-critical task
failure cascades dependents to skipped and finishes degraded; fail-closed
when tasks remain unclosable. TestTopologyConformance_allTwelve runs all 12
names (10 runnable to completion, handoff/mailbox rejected as primitives).
Full suite green with -race.
Impact: task_board lit in the topology registry (all 12 topologies now
covered); RunState gains Board (persisted in runstate.json); RunOptions gains
TaskBoardLeaseSeconds; leader topology now maps to mode before plan
confirmation so Topology=leader runs no longer hit the confirmation gate.
Refs: Q2 plan WP4 (2026-10-07-agent-stack-q2q3-plans.md); task card
.task/feature-agent-stack-q2.md
Why: Q2 acceptance requires leader runs to detect stuck delegation loops and
cap replans instead of spinning to the turn limit, and a mailbox primitive
whose messages can never resolve a pending permission approval (acceptance 3).
Evidence: squad conformance tests (leader_mailbox_conformance_test.go):
ledger stuck->replan, replan cap finalize, mailbox delivery after hold,
mailbox-cannot-approve-permission (structural + actor second defense),
post guards; full suite green with -race.
Impact: LeaderState gains Progress/StuckTurns/Replans/MaxReplans/ReplanNote
(persisted in leader.json); RunState gains Mailbox (persisted in
runstate.json); RespondDecision gains Actor - non-human actors are rejected
on choice gates; step briefs and leader task briefs include pending mail.
Refs: Q2 plan WP2 (2026-10-07-agent-stack-q2q3-plans.md); task card
.task/feature-agent-stack-q2.md
Why: the completion doc Q2 requires twelve builtin topologies each with
a scripted-backend conformance; today the engine has only the
standard/leader mode string with no topology abstraction, no loop
re-arm (cycles are fail-closed by design), no in-graph routing, and no
run deadline.
Evidence: GOWORK=off go vet + go test -count=1 green across squad and
all three adapter modules; ten conformance/validation tests:
sequential, parallel, loop-until-field (exact iteration count +
persisted LoopRuns), loop-budget-exhausted (loop_exhausted event, run
still completes), router-selects-branch (loser skipped, run NOT
degraded), router-no-match fail-closed, evaluator-optimizer (draft
re-armed via loop body until verdict=pass), long-running deadline
(pending steps skipped, degraded closeout), unknown/primitive/
unregistered-driver fail-closed, builtin registry completeness (12).
Impact: additive — Topology/Deadline on RunOptions, Topology on
Blueprint, Loop/LoopRuns/Branches/Handoff on StepSpec; empty topology
keeps byte-identical standard behavior; MaxIterations is the total
iteration cap (first run included); router-skipped steps are exempt
from degraded derivation; runParallel now preserves the first specific
RunFailed error instead of re-wrapping via failStep; deadline metadata
uses RFC3339Nano (second-precision RFC3339 truncated sub-second
deadlines into instant expiry).
Refs: Agent-System-Documents#44 (Q2 plan WP1)
Why: the exec bridge now threads Last*/cache (Agent-Native#39); squad
ProcessEvent carries them to hosts (the product's squad runner reads
them for usage accounting).
Evidence: GOWORK=off go vet green across squad and adapters/native;
nested pin at the merged native.
Impact: additive event fields.
Refs: Agent-Native#39, N3 plan WP6
Why: Q1 WP5 - platform's SQLite becomes an implementation of the
squad Store port so run state and orchestration state live in one
database, and platform's in-process execution goes through the
Executor port instead of bypassing it (X6).
Evidence: all five modules go vet clean and go test ok, including
engine-on-SQLite (state in DB, no run-level JSON files, second
engine recovers from the same DB), SQLiteStore round trip, and the
full squad compatibility net.
Impact: runState/hitlGate exported so the port is implementable;
Options.Store wires into engineDeps; leader persistence and
checkpoints route through the port; PrepareWorkspace no longer
writes run-level JSON (st.save owns it); platform RunInProcess
takes squad.Executor (callers updated); serve mounts the engine on
platform's RunStore - single orchestration.db for both.
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md Q1 WP5
Why: Q1 WP4/WP6/WP7 - run persistence goes through a Store port
(FileStore carries the current layout), superstep boundaries write
tiered snapshots with a graph fingerprint for recovery continuity,
and the four hardcoded stop conditions move behind one Termination
registry with And/Or composition.
Evidence: go vet clean; full suite ok including new checkpoint,
continuity-rejection, termination-primitive, grant-cap and the
Q23-updated leader closeout tests; cmd module green with
SweepExpired wired into the sweeper loop.
Impact: runState saves via Store (FileStore unchanged by default);
advance writes sync/async/on_exit checkpoints and Resume rejects
graph-shape changes from the crash window; leader turn cap now
degrades through finalize instead of a bare failure (Q23); manual
repair grants are bounded (Q26); wakeup cap and step timeout read
from the same registry; the dead SweepExpired is now called (Q6).
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md Q1 WP4/WP6/WP7
Why: Q1 WP3 - each step execution carries a stable attempt id
(runID/stepID/N) for audit and idempotency, and auto-mode
interrupted recovery must not re-execute a step whose products all
landed before the crash (side effects exactly once).
Evidence: go vet clean; full squad suite ok, including the new
kill-at-superstep-boundary test (completed steps never re-execute,
running-with-products judged complete by attempt) and
missing-product retry tests.
Impact: StepSpec gains AttemptID set when entering running and used
as the execution id; recoverInterrupted (HitlAuto) checks the
step's targets first - all present means completed with artifact
merge, otherwise the legacy pending retry. Transient retries share
the same AttemptID.
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md Q1 WP3
Why: Q1 WP1/WP2 - replace the linear step scan with per-round graph
compilation and layer advancement so Needs drives scheduling and
write-conflicting steps never run concurrently.
Evidence: go vet clean in all five modules; go test ok everywhere,
including the full compatibility net (engine, leader, recovery, r2
regression, intent, evaluator suites) and new graph tests.
Impact: new squad/graph.go (compile/layer/conflict); StepSpec gains
Writes; advance advances by superstep (parallel run of
conflict-free members, serial for write-conflicts, single-member
groups keep dynamic adjustment). Legacy semantics preserved:
adjacent-Parallel mapping, degraded-continue on non-critical
failure (failed counts as satisfied), empty status = pending,
cycle/unknown/dup needs fail-closed.
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md Q1 WP1/WP2
Why: the squad core has to compile without agent-native or agent-intent. Ports stay in the core; adapters and the CLI own the concrete clients.
Evidence: GOWORK=off go test -count=1 ./squad/ ./store/ ./textextract/ ./schema/ passed, and the intent, sqlite, native, and cmd modules passed their own go test.
Impact: core go.mod keeps error-code-system only. Nested gitlinks stay. Callers register the native backend through agent-squad/adapters/native.
Refs: D1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: parallel groups dropped a hold, HITL timeouts were unused, the planner failed closed on any model error, and event sequence numbers restarted at 1.
Evidence: GOWORK=off go test -count=1 -race ./squad/ ./squadintent/ ./store/ ./platform/ passed, including TestR2ExpiredDecisionUsesCancel, TestR2IntentStepSkipsRoleRegistry, TestR2PlannerFallsBackAndCapsSteps, TestR2TaskFileUsesLatestBlackboard, TestR2FanoutCancelDoesNotPanic, TestR2EventSeqContinuesAfterRestart, TestR2JournalModeWAL, and TestR2GateReloadSeesNewIntent.
Impact: a failed or held parallel sibling cancels the rest; expired requests synthesize cancel or fail; the intent gate follows Registry.Reload.
Refs: Q6, Q10, Q11, Q12, Q13, Q17, Q18, Q19, I9
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: A parallel group only checked the first step, free text cancelled confirmation and permission holds, and a failed reply deleted the pending request before the decision ran.
Evidence: GOWORK=off go vet ./...; go test -count=1 ./squad/ ./cli/; go test -race -count=1 ./squad/. New tests cover the second parallel member, free-text holds, restored pending, AtomicWrite, and listenGuard.
Impact: Any member that needs approval or bash permission keeps the whole group suspended. Free text on confirmation, permission, and quality review is stored as steering and the same request stays pending. A dispatch error restores pending.json.
Refs: feature/agent-stack-r0-r1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: A step confirmation cancelled the whole run, approvals and repair budget vanished on restart, and the HTTP server accepted anonymous and oversized requests.
Evidence: GOWORK=off go vet ./... && GOWORK=off go test -count=1 ./... && GOWORK=off go test -race ./squad/... passed. scripts/submodule-consistency.sh ok.
Impact: Approve and skip apply to one step; runstate.json and unix flock survive restart; wakeup leaves permission holds pending; serve defaults to loopback and requires a token off-loopback.
Refs: Q1 Q2 Q3 Q4 Q5 Q8 Q9 Q14 Q15 Q16
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Tests and product budgets need to drive each kernel step without a live model, and stop it when the tool-call cap is hit.
Evidence: go test ./squad/ -count=1 passed after fetching agent-native cfc840c.
Impact: AgentNativeConfig.Stream and MaxToolCalls reach native.Options. The nested kernel pin moves to cfc840c.
Refs: deliverable-quality-system B2
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Simple tasks should finish as one assistant step, while complex tasks stay in leader mode, and every kernel step needs the host's tools and events.
Evidence: go test ./squad/ -count=1 passed, including direct route, leader entry, skill override, and delegation score fields.
Impact: Hosts can inject HostTools and OnEvent; a Router can choose direct or plan without changing product roles.
Refs: deliverable-quality-system B2
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: RoleSpec.Tools/Skills were dead fields, BashPermission only gated
an HITL prompt with nothing to enforce, leader turns dropped process
events and tool profile, and the HTTP API could not start leader mode.
Evidence: go vet ./... clean; go test ./... 6 packages ok. New cases:
step requests carry ToolAllowlist/Skills/SessionScope/ExecutionID and
a step-wide shared deadline; buildAgentNativeOptions bash gating
(deny/unset get no exec tools even when configured, ask/allow do,
missing model fails closed); Derive inheriting tool_profile/tools/
skills/bash_permission with request overrides; role-level HitlPolicy
making a bash=ask step fail under auto instead of holding; leader
requests fully assembled with detail events bridged to the sink; HTTP
createRun accepting mode/leader_role/max_leader_turns (engine-driven
repro of the plan-validation precondition included).
Impact: RegisterAgentNative keeps its signature and delegates to a new
RegisterAgentNativeFull(AgentNativeConfig); nested agent-native
gitlink pinned to f9ab2ef (full-agent batch); llm planner derive
schema gains tool_profile/tools/skills/bash_permission.
Refs: Dev task feature/full-agent-r1 batch B9; agent-native PR #3.
HTTP failures now carry catalog codes and a trace id, and the parent replace directives are the only copies used at build time.
Co-authored-by: Cursor <cursoragent@cursor.com>
weichen
