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: 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: 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: Headless runs need the pi tool names, walked context files, prompt includes, and a trust gate before project settings are loaded.
Evidence: go test -count=1 ./runtime/tools/ ./runtime/resources/ ./runtime/skills/
Impact: Existing tool names stay registered. Aliases are added only when the underlying tool exists.
Refs: feature/pi-headless N4
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: external tool ecosystems speak MCP; the kernel had no way to
mount them, so every external capability needed a bespoke builtin.
Evidence: go vet ./... clean; go test ./... 11 packages ok. Tests run
against a REAL stdio server (test binary re-exec via
mcp.RunFakeServer) and httptest HTTP servers: initialize handshake,
tools/list with pagination cap, tools/call success / tool-level
isError / unknown-tool rpc error, JSON and SSE response modes,
Mcp-Session-Id capture, manager registration as mcp__<server>__<tool>,
namespace wildcard allowlist (mcp__fake__* keeps only that namespace,
unknown wildcard fails closed), unreachable server failing the step,
hung server bounded by per-request timeout, config validation
(duplicate names, dual/missing transport, bad name chars), and an
adapter end-to-end execution calling an MCP tool.
Impact: new runtime/mcp package; native.Options gains MCP *mcp.Manager
and Close() for process cleanup; tools.Registry.Restrict gains
prefix/* namespace entries; server-provided metadata is capped (64
tools/server, 1 KiB descriptions, 64 KiB text results) and treated as
untrusted input.
Refs: plan B5; agent-execution standard (tool providers get
authorized capability only).
Why: tavily_search/web_search existed as tools but were never
registered on the default execution path (parity-matrix #5 leftover),
and the kernel had no URL fetch capability at all.
Evidence: go vet ./... clean; go test ./... all ok. Fetch cases cover
private/loopback/link-local rejection (127/8, 10/8, 192.168/16,
169.254 metadata endpoint, ::1), file:// scheme rejection, DNS
resolution landing on private space, per-hop redirect validation
rejecting a redirect to loopback, allowed-host policy, HTML stripping
(script/style removed, entities decoded), content-type gating, stub
end-to-end fetch, and absence without FetchConfig (IP normalization
fix included: IPv4-mapped ::ffff:0:0/96 must not swallow all IPv4).
Adapter tests assert Options.Search registers both search tools and
passes queries to the provider, and that neither registers when
unconfigured.
Impact: native.Options gains Search and Fetch; both tool families are
opt-in by the composition root (fail-closed), closing parity-matrix
#5 wiring.
Refs: docs/parity-matrix.md #5; product settings tools.tavily.*
injection point.
Why: the kernel had no system-tool capability, so deliverable generation
via CLI/Python (the settled product direction) could not run inside
agent-native executions; squad's BashPermission gate had nothing to
guard.
Evidence: go vet ./... clean; go test ./... all ok. Cases cover echo
and non-zero exit reporting, env whitelist isolation (host secrets do
not leak into children), timeout and context-cancel killing the whole
process group promptly, blocklist rejection (sudo/destructive
patterns), 64 KiB head+tail output truncation, python script
persistence under .agent/scripts/, and fail-closed registration (no
interpreter configured = no tool). Adapter tests verify Options.Exec
wiring, per-request Env merge, allowlist stripping, and absence
without Options.Exec.
Impact: native.Options gains Exec *tools.ExecConfig; exec.Request.Env
now flows into child processes; workspace profile gains up to 2 tools
(only when configured). NOT an OS sandbox - isolation stays a host
responsibility, documented in tool descriptions.
Refs: parity-matrix #3/#4; settled product decision (deliverables via
CLI/Python + skills, see archived card feature-exec-interpreters).
Why: the five builtin tools could not express file discovery (no glob),
code search (literal-only grep), paged reads (200 KiB silent
truncation), or batch edits, capping the kernel below general coding-
agent parity.
Evidence: go vet ./... clean; go test ./... all ok. New cases cover
numbered line ranges and beyond-EOF rejection, regex/include/
ignore_case grep with invalid-regex errors, double-star glob
(**/*.go, docs/**, no-match), workspace confinement of glob, and
multi_edit atomicity (failed batch leaves the file untouched).
Impact: WorkspaceTools now registers 7 tools (adds glob, multi_edit);
read_file gains optional offset/limit and announces truncation; grep
gains regex/include/ignore_case while keeping literal default
behavior; existing callers see only additive changes.
Refs: docs/parity-matrix.md #3 (tool semantics negatives, Batch2).
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).
weichenandCursor