Files
coder/coderd/agentapi/context.go
T
Kyle Carberry bca0ce04ca feat: integrate agent context snapshots into chats (#26389)
Makes the chat context foundation from #26385 live. That PR added the
storage columns, writer queries, and a dormant
`agentapi.ContextDirtyMarker` trigger with no production callers; this
PR wires them together end to end.

When a workspace agent pushes a context snapshot, bound chats now
hydrate to that snapshot's hash, and a later push with a different hash
flips already-pinned chats to dirty (emitting a `context_dirty` watch
event after the transaction commits). Chat creation pins the agent's
latest snapshot when one already exists. The experimental chat API
exposes this as `Chat.Context` (`*ChatContext` with `dirty`,
`dirty_since`, `error`), and a new `PUT
/api/experimental/chats/{chat}/context` endpoint re-pins the agent's
latest snapshot and clears the dirty marker.

`context_dirty_resources` stays NULL (the resource-level diff is
deferred to the UI phase) and the live per-turn context pull is
unchanged.

The end-to-end test provisions a workspace agent via the echo
provisioner, connects it over the Agent API v2.10, and exercises the
full path: an initial push hydrates a bound chat (clean), a second push
with a different hash marks it dirty, the API reports the dirty state,
and the refresh endpoint clears it.

<details>
<summary>Decision log</summary>

- **API shape — sub-struct.** Dirty state is surfaced as
`codersdk.Chat.Context *ChatContext { Dirty bool; DirtySince *time.Time;
Error string }` rather than flat fields, matching the RFC's named
`ChatContext` type and leaving room for future fields (resource diff,
sources). `db2sdk.Chat` populates it when the chat is context-tracked
(`len(ContextAggregateHash) > 0`), dirty, or carries a snapshot error,
and leaves it nil (`omitempty`) otherwise. `Dirty` mirrors
`context_dirty_since` being set.
- **Marker wiring.** The chat daemon is injected directly as the
`agentapi.ContextDirtyMarker`. It is unconditionally constructed (only
its background worker is gated), so the marker is always non-nil and the
wiring matches every other `api.chatDaemon` call site. `agentapi` still
treats a nil marker as "chatd absent", so `PushContextState` stays a
pure write path for any future caller that does not wire chatd in.
- **Refresh is atomic.** `RefreshChatContext` reads the agent's latest
snapshot and re-pins the chat in one repeatable-read transaction, so a
concurrent push cannot land between the read and the write and leave the
chat pinned to a stale hash with the dirty marker cleared.
- **Hydrate + dirty run inside the push transaction.** The fan-out
shares the push's transaction so a concurrent refresh cannot interleave
with the version gate; `context_dirty` watch events publish only after
commit. The pinned hash on dirtied chats is intentionally left unchanged
— the refresh endpoint re-pins it.
- **Dirtied chats keep their pinned hash.** Drift is advisory: a dirty
chat stays usable, and refreshing is the only path that advances the
pinned hash.
- **Test binds `chats.agent_id` directly.** In production the binding is
set lazily during a chat turn (`chatd.persistBuildAgentBinding`); the
test sets it via `dbgen` so it exercises the context flow rather than
turn resolution.

Plan: `coderd/x/chatd` context integration + E2E (sub-struct API,
create-time + push-time hydration, refresh endpoint;
`context_dirty_resources` and the per-turn pull untouched).

</details>

🤖 Generated by Coder Agents on behalf of @kylecarbs
2026-06-16 17:46:47 +00:00

430 lines
17 KiB
Go

package agentapi
import (
"context"
"database/sql"
"errors"
"math"
"sort"
"time"
"github.com/google/uuid"
"golang.org/x/xerrors"
"google.golang.org/protobuf/encoding/protojson"
"google.golang.org/protobuf/proto"
"cdr.dev/slog/v3"
agentproto "github.com/coder/coder/v2/agent/proto"
"github.com/coder/coder/v2/coderd/database"
"github.com/coder/coder/v2/coderd/database/dbtime"
"github.com/coder/quartz"
)
// Server-side caps on a single PushContextState request. The agent
// enforces its own caps (64KiB per resource payload, 2MiB aggregate,
// 500 resources; see agent/agentcontext/resolve.go), but coderd
// cannot trust a workspace process, so pushes are re-validated here
// with headroom above the agent caps:
//
// - maxContextResourcesPerPush allows excluded stub entries past
// the agent's 500-resource cap.
// - maxContextResourceBodyBytes covers protojson and base64
// expansion of a 64KiB payload.
// - maxContextAggregateBodyBytes matches the 4MiB DRPC message
// cap so the invariant survives transport changes.
// - The string and hash caps bound the remaining row columns;
// source doubles as a btree primary key column, which PostgreSQL
// limits to roughly 2704 bytes per index entry.
const (
maxContextResourcesPerPush = 1000
maxContextResourceBodyBytes = 256 * 1024
maxContextAggregateBodyBytes = 4 * 1024 * 1024
maxContextSourceBytes = 1024
maxContextErrorBytes = 4096
maxContextHashBytes = 64
)
// ContextAPI implements the v2.10 PushContextState RPC. It persists
// the latest pushed snapshot per workspace agent across two tables
// (workspace_agent_context_snapshots and
// workspace_agent_context_resources) so later phases can hydrate
// chats and surface drift to the dashboard.
//
// The handler is a pure write path: nothing else in coderd reads
// these rows yet. If a bug here returns errors the agent's RunPush
// loop backs off and the workspace keeps behaving exactly like it
// did before v2.10.
type ContextAPI struct {
AgentID uuid.UUID
// Workspace caches workspace fields for the duration of the agent
// connection so dbauthz can authorize against the workspace RBAC
// object without re-fetching the workspace on every push.
Workspace *CachedWorkspaceFields
Log slog.Logger
Clock quartz.Clock
Database database.Store
// DirtyMarker hydrates chats from, and marks chats dirty against, the
// snapshot persisted by a push. It is nil when chatd is not running,
// in which case PushContextState stays a pure write path.
DirtyMarker ContextDirtyMarker
}
// ContextDirtyMarker hydrates chats from, and marks chats dirty against, a
// freshly persisted agent context snapshot. It is implemented by chatd and
// injected at coderd construction so this package neither imports the chat
// domain nor performs chat-authorized writes directly.
type ContextDirtyMarker interface {
// HydrateAndMarkChatsDirty runs inside the PushContextState
// transaction using the supplied store. It hydrates chats for the
// agent that have no pinned hash yet (no dirty event) and flips
// already-pinned chats whose hash differs from aggregateHash. It
// returns a callback that publishes the resulting dirty watch events;
// the caller invokes it only after the transaction commits. The
// callback is a no-op when nothing transitioned to dirty.
HydrateAndMarkChatsDirty(ctx context.Context, tx database.Store, agentID uuid.UUID, aggregateHash []byte, snapshotError string, now time.Time) (publishDirty func(), err error)
}
// PushContextState persists a snapshot pushed by the workspace
// agent. The transaction upserts the snapshot row, upserts each
// resource, then deletes any resources whose source is not in the
// incoming set so the stored snapshot and resource table always
// agree. It runs at repeatable read isolation (with retries) so two
// concurrent pushes cannot interleave their writes; the loser of the
// conflict re-runs the version gate against the winner's committed
// state.
//
// Returns accepted = false (without writing) when the push is a
// replay or out-of-order resend: the agent's per-process version
// counter is monotonic, and only an initial = true push from a
// freshly-booted agent resets that baseline. Replays and stale
// retransmits leave the stored state untouched.
//
// Authorization happens in dbauthz: every query in the transaction
// authorizes the actor (the agent's token subject) against the
// workspace that owns the agent.
func (a *ContextAPI) PushContextState(ctx context.Context, req *agentproto.PushContextStateRequest) (*agentproto.PushContextStateResponse, error) {
if req == nil {
return nil, xerrors.New("agentapi: PushContextState request is nil")
}
if err := validateContextPushRequest(req); err != nil {
return nil, err
}
rows, err := validateAndConvertContextResources(req.Resources)
if err != nil {
return nil, err
}
// Attach the cached workspace RBAC object so dbauthz can take its
// fast path. On failure (or when unset, e.g. prebuilds) dbauthz
// falls back to fetching the workspace by agent ID.
if a.Workspace != nil {
injected, err := a.Workspace.ContextInject(ctx)
if err != nil {
a.Log.Debug(ctx, "failed to inject cached workspace RBAC object", slog.Error(err))
} else {
ctx = injected
}
}
clock := a.Clock
if clock == nil {
clock = quartz.NewReal()
}
now := dbtime.Time(clock.Now())
activeSources := make([]string, 0, len(rows))
for _, r := range rows {
activeSources = append(activeSources, r.Source)
}
sort.Strings(activeSources)
var accepted bool
// publishDirty is captured from the final (committed) attempt and
// invoked after the transaction commits; ReadModifyUpdate may re-run
// the closure on serialization conflicts.
var publishDirty func()
err = database.ReadModifyUpdate(a.Database, func(tx database.Store) error {
// The closure re-runs on serialization conflicts; reset any
// state carried over from a rolled-back attempt.
accepted = false
publishDirty = nil
existing, err := tx.GetLatestWorkspaceAgentContextSnapshot(ctx, a.AgentID)
switch {
case errors.Is(err, sql.ErrNoRows):
// No previous snapshot; first push always wins.
case err != nil:
return xerrors.Errorf("get latest snapshot: %w", err)
default:
// Accept either a fresh agent process (initial) or
// a strictly newer version. Out-of-order or replayed
// pushes leave the stored state untouched.
//
//nolint:gosec // existing.Version is a uint64 round-tripped via BIGINT; non-negative by construction.
if !req.Initial && req.Version <= uint64(existing.Version) {
return nil
}
}
_, err = tx.UpsertWorkspaceAgentContextSnapshot(ctx, database.UpsertWorkspaceAgentContextSnapshotParams{
WorkspaceAgentID: a.AgentID,
//nolint:gosec // Bounded by validateContextPushRequest.
Version: int64(req.Version),
AggregateHash: append([]byte(nil), req.AggregateHash...),
SnapshotError: req.SnapshotError,
ReceivedAt: now,
})
if err != nil {
return xerrors.Errorf("upsert snapshot: %w", err)
}
for _, r := range rows {
r.WorkspaceAgentID = a.AgentID
r.Now = now
_, err = tx.UpsertWorkspaceAgentContextResource(ctx, r)
if err != nil {
return xerrors.Errorf("upsert resource %q: %w", r.Source, err)
}
}
err = tx.DeleteStaleWorkspaceAgentContextResources(ctx, database.DeleteStaleWorkspaceAgentContextResourcesParams{
WorkspaceAgentID: a.AgentID,
ActiveSources: activeSources,
})
if err != nil {
return xerrors.Errorf("delete stale resources: %w", err)
}
// Hydrate and dirty chats against the snapshot just written, in the
// same transaction so a concurrent refresh cannot interleave with
// the version gate. Events are published only after commit.
if a.DirtyMarker != nil {
publishDirty, err = a.DirtyMarker.HydrateAndMarkChatsDirty(ctx, tx, a.AgentID, req.AggregateHash, req.SnapshotError, now)
if err != nil {
return xerrors.Errorf("hydrate and mark chats dirty: %w", err)
}
}
accepted = true
return nil
})
if err != nil {
return nil, err
}
if !accepted {
a.Log.Debug(ctx, "PushContextState dropped: replay or out-of-order",
slog.F("agent_id", a.AgentID),
slog.F("version", req.Version),
slog.F("initial", req.Initial),
)
return &agentproto.PushContextStateResponse{Accepted: false}, nil
}
// The snapshot committed; fan out dirty watch events to chats whose
// pinned context drifted from this push.
if publishDirty != nil {
publishDirty()
}
a.Log.Debug(ctx, "PushContextState accepted",
slog.F("agent_id", a.AgentID),
slog.F("version", req.Version),
slog.F("initial", req.Initial),
slog.F("resources", len(rows)),
)
return &agentproto.PushContextStateResponse{Accepted: true}, nil
}
// validateContextPushRequest enforces the request-level caps: counts
// and sizes a compromised workspace could otherwise inflate to DoS
// coderd or bloat the database.
func validateContextPushRequest(req *agentproto.PushContextStateRequest) error {
if req.Version > math.MaxInt64 {
return xerrors.Errorf("agentapi: PushContextState version %d exceeds int64 range", req.Version)
}
if len(req.AggregateHash) > maxContextHashBytes {
return xerrors.Errorf("agentapi: PushContextState aggregate hash is %d bytes, exceeds %d byte cap", len(req.AggregateHash), maxContextHashBytes)
}
if len(req.SnapshotError) > maxContextErrorBytes {
return xerrors.Errorf("agentapi: PushContextState snapshot error is %d bytes, exceeds %d byte cap", len(req.SnapshotError), maxContextErrorBytes)
}
if len(req.Resources) > maxContextResourcesPerPush {
return xerrors.Errorf("agentapi: PushContextState has %d resources, exceeds %d resource cap", len(req.Resources), maxContextResourcesPerPush)
}
return nil
}
// validateAndConvertContextResources translates wire resources into
// upsert parameters while rejecting structurally invalid input:
//
// - empty, oversized, or duplicate sources (the PK depends on
// uniqueness and indexes the source column),
// - unknown body variants (kept extensible by emitting the proto's
// reserved kinds via dedicated body messages),
// - unknown status enum values,
// - per-resource and aggregate body sizes past the server caps.
//
// Validation is deliberately strict here so a misbehaving agent
// cannot poison the snapshot table. Phase 2 readers can then trust
// that every row maps to a known proto variant.
//
// WorkspaceAgentID and Now are left unset; the caller fills them at
// upsert time.
func validateAndConvertContextResources(resources []*agentproto.ContextResource) ([]database.UpsertWorkspaceAgentContextResourceParams, error) {
rows := make([]database.UpsertWorkspaceAgentContextResourceParams, 0, len(resources))
seen := make(map[string]struct{}, len(resources))
aggregateBodyBytes := 0
for i, r := range resources {
if r == nil {
return nil, xerrors.Errorf("agentapi: PushContextState resource at index %d is nil", i)
}
if r.Source == "" {
return nil, xerrors.Errorf("agentapi: PushContextState resource at index %d has empty source", i)
}
if len(r.Source) > maxContextSourceBytes {
return nil, xerrors.Errorf("agentapi: PushContextState resource at index %d has %d byte source, exceeds %d byte cap", i, len(r.Source), maxContextSourceBytes)
}
if _, ok := seen[r.Source]; ok {
return nil, xerrors.Errorf("agentapi: PushContextState duplicate source %q", r.Source)
}
seen[r.Source] = struct{}{}
if len(r.GetSourcePath()) > maxContextSourceBytes {
return nil, xerrors.Errorf("resource %q: source path is %d bytes, exceeds %d byte cap", r.Source, len(r.GetSourcePath()), maxContextSourceBytes)
}
if len(r.Error) > maxContextErrorBytes {
return nil, xerrors.Errorf("resource %q: error is %d bytes, exceeds %d byte cap", r.Source, len(r.Error), maxContextErrorBytes)
}
if len(r.ContentHash) > maxContextHashBytes {
return nil, xerrors.Errorf("resource %q: content hash is %d bytes, exceeds %d byte cap", r.Source, len(r.ContentHash), maxContextHashBytes)
}
if r.SizeBytes > math.MaxInt64 {
return nil, xerrors.Errorf("resource %q: size %d exceeds int64 range", r.Source, r.SizeBytes)
}
kind, body, err := marshalContextResourceBody(r)
if err != nil {
return nil, xerrors.Errorf("resource %q: %w", r.Source, err)
}
if len(body) > maxContextResourceBodyBytes {
return nil, xerrors.Errorf("resource %q: body is %d bytes, exceeds %d byte cap", r.Source, len(body), maxContextResourceBodyBytes)
}
aggregateBodyBytes += len(body)
if aggregateBodyBytes > maxContextAggregateBodyBytes {
return nil, xerrors.Errorf("agentapi: PushContextState aggregate body size exceeds %d byte cap", maxContextAggregateBodyBytes)
}
status, err := contextResourceStatus(r.Status)
if err != nil {
return nil, xerrors.Errorf("resource %q: %w", r.Source, err)
}
//nolint:exhaustruct // WorkspaceAgentID and Now are filled by the caller at upsert time.
rows = append(rows, database.UpsertWorkspaceAgentContextResourceParams{
Source: r.Source,
SourcePath: r.GetSourcePath(),
BodyKind: kind,
Body: body,
ContentHash: append([]byte(nil), r.ContentHash...),
//nolint:gosec // Bounded above.
SizeBytes: int64(r.SizeBytes),
Status: status,
Error: r.Error,
})
}
return rows, nil
}
// marshalContextResourceBody picks the body variant set on the wire
// resource and returns the (body_kind, body_jsonb) pair stored in
// the resource row. The body is protojson encoded so the schema can
// be evolved by adding fields to the proto without coderd changes,
// and a future reader can round-trip back to the proto type by
// switching on body_kind.
//
// Body is always populated, even on non-OK statuses: the wire
// guarantees the oneof variant is set so coderd can still attribute
// the failure to a known kind. For variants with no content fields
// (mcp_config), an empty JSON object is stored.
func marshalContextResourceBody(r *agentproto.ContextResource) (kind database.WorkspaceAgentContextBodyKind, body []byte, err error) {
switch b := r.Body.(type) {
case *agentproto.ContextResource_InstructionFile:
payload := b.InstructionFile
if payload == nil {
payload = &agentproto.InstructionFileBody{}
}
body, err = marshalBody(payload)
return database.WorkspaceAgentContextBodyKindInstructionFile, body, err
case *agentproto.ContextResource_Skill:
payload := b.Skill
if payload == nil {
payload = &agentproto.SkillMetaBody{}
}
body, err = marshalBody(payload)
return database.WorkspaceAgentContextBodyKindSkill, body, err
case *agentproto.ContextResource_McpConfig:
payload := b.McpConfig
if payload == nil {
payload = &agentproto.MCPConfigBody{}
}
body, err = marshalBody(payload)
return database.WorkspaceAgentContextBodyKindMcpConfig, body, err
case *agentproto.ContextResource_McpServer:
payload := b.McpServer
if payload == nil {
payload = &agentproto.MCPServerBody{}
}
body, err = marshalBody(payload)
return database.WorkspaceAgentContextBodyKindMcpServer, body, err
case nil:
return "", nil, xerrors.Errorf("missing body variant; status %s requires a typed body", r.Status)
default:
return "", nil, xerrors.Errorf("unsupported body variant %T", r.Body)
}
}
// contextBodyMarshalOptions produces deterministic-ish JSON for the
// body so the stored value compares equal across pushes that yield
// equivalent protos. Strict canonicalization (RFC 8785) is not
// required here; the enum column plus the protojson round trip give
// us a stable enough store.
var contextBodyMarshalOptions = protojson.MarshalOptions{
UseProtoNames: true,
EmitUnpopulated: false,
}
// marshalBody is a small wrapper around protojson.Marshal that
// keeps the body encoding in one place; future phases that read
// these rows mirror the call with protojson.Unmarshal into the
// matching proto.Message.
func marshalBody(msg proto.Message) ([]byte, error) {
out, err := contextBodyMarshalOptions.Marshal(msg)
if err != nil {
return nil, xerrors.Errorf("marshal body: %w", err)
}
return out, nil
}
// contextResourceStatus translates the wire status enum to the
// database enum. STATUS_UNSPECIFIED is rejected: every well-formed
// snapshot row needs an explicit status so cache invalidation, dirty
// fan-out, and the Sources drawer can reason about partial pushes
// deterministically.
func contextResourceStatus(s agentproto.ContextResource_Status) (database.WorkspaceAgentContextResourceStatus, error) {
switch s {
case agentproto.ContextResource_OK:
return database.WorkspaceAgentContextResourceStatusOk, nil
case agentproto.ContextResource_OVERSIZE:
return database.WorkspaceAgentContextResourceStatusOversize, nil
case agentproto.ContextResource_UNREADABLE:
return database.WorkspaceAgentContextResourceStatusUnreadable, nil
case agentproto.ContextResource_INVALID:
return database.WorkspaceAgentContextResourceStatusInvalid, nil
case agentproto.ContextResource_EXCLUDED:
return database.WorkspaceAgentContextResourceStatusExcluded, nil
default:
return "", xerrors.Errorf("unknown status %d", s)
}
}