Files
oikos/cmd/nomos/continue.go
dtoro 6f9998fa29
Some checks failed
ci / build-test (push) Has been cancelled
ci / docker-build (push) Has been cancelled
fix: auto-continuation silently dropped LLM errors + added outer retry
Verified live: the new auto-continuation worker (previous commit) worked end
to end for the happy path (provision -> auto-verify -> report success, zero
operator ticks). But testing a failure-recovery case (a destroy that failed
because the container was still running) surfaced a real bug: continueSession's
emit closure only captured "text" events, so when chatWith ended the turn on
an "error" event (LLM returned an empty/refusal response, internal retry also
empty), the worker persisted a completely blank, uninformative "auto" message
— no sign anything had gone wrong, undermining observability of the very
mechanism just built.

- Capture "error" events and, if the turn produced no text/tool_calls at all,
  persist an explanatory placeholder instead of blank.
- Add one outer retry of the whole chatWith call when the first attempt
  produces nothing — the principle behind this whole feature ("don't give up
  on the first error") should apply to the continuation mechanism itself, not
  just the homelab commands it's continuing.

Also verified live: recovery-from-failure works via the normal chat path once
prompted, and cleaned up the test container.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-10 14:30:23 +02:00

159 lines
6.4 KiB
Go

package main
import (
"context"
"encoding/json"
"fmt"
"log/slog"
"regexp"
"strings"
"time"
"github.com/google/uuid"
)
// execIDRe matches "execution <uuid>" in a tool result — the phrasing shared
// by request_execution / run when they queue or start a gated execution.
// Only these async executions need continuation; the synchronous auto-run
// path returns its output inline and is already observed in-turn.
var execIDRe = regexp.MustCompile(`(?i)execution\s+([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})`)
func extractExecutionIDs(toolResult string) []uuid.UUID {
matches := execIDRe.FindAllStringSubmatch(toolResult, -1)
seen := map[uuid.UUID]bool{}
var out []uuid.UUID
for _, m := range matches {
if id, err := uuid.Parse(m[1]); err == nil && !seen[id] {
seen[id] = true
out = append(out, id)
}
}
return out
}
// runContinuationWorker is the event loop that replaces the human typing
// "continue". It polls for gated executions that (a) were initiated by a chat
// session and (b) have just finished, and — while that agent has an open assent
// window (an approved plan is in flight) — feeds each result back into the
// agent so it proceeds to the next step or recovers from the failure, all
// without an operator tick. Blocks until ctx is cancelled.
func (a *agent) runContinuationWorker(ctx context.Context) {
if a.store == nil {
slog.Warn("nomos: continuation worker disabled (no store)")
return
}
slog.Info("nomos: continuation worker started")
ticker := time.NewTicker(4 * time.Second)
defer ticker.Stop()
for {
select {
case <-ctx.Done():
return
case <-ticker.C:
a.processContinuations(ctx)
}
}
}
func (a *agent) processContinuations(ctx context.Context) {
pending := a.store.pendingContinuations(ctx, 5)
windowOpen := a.store.assentWindowActive(ctx, a.agentID)
for _, p := range pending {
// Scope gate: only auto-continue while an approved plan is active.
// A finished one-off execution with no window is left as-is (marked
// continued so we don't re-check it forever) — the operator decides
// what happens next, as today.
if !windowOpen {
a.store.markContinued(ctx, p.ExecID)
continue
}
a.store.markContinued(ctx, p.ExecID) // stamp first: a failure here must not cause a re-continue loop
a.continueSession(ctx, p)
}
}
// continueSession re-invokes the agent for one finished execution, persisting
// the resulting assistant turn just like handleChat does. The operator sees it
// on their next load of the session (live push is a follow-up).
func (a *agent) continueSession(ctx context.Context, p pendingContinuation) {
note := buildContinuationNote(p)
slog.Info("nomos: auto-continuing session", "session", p.SessionID, "execution", p.ExecID, "status", p.Status)
var toolCalls []map[string]any
var finalText, errText string
// One retry if the LLM call itself produced nothing (transient flake /
// empty-response) — the whole point of this mechanism is "don't give up
// on the first error," which should apply to the continuation call
// itself, not just the homelab commands it's continuing. Found live: a
// destructive-recovery continuation hit an empty LLM response, its
// internal retry (chatWith's own maxLLMRetries=1) also came up empty, and
// without this outer retry the operator would see nothing at all.
cctx, cancel := context.WithTimeout(ctx, 10*time.Minute)
defer cancel()
for attempt := 0; attempt < 2; attempt++ {
toolCalls, finalText, errText = nil, "", ""
emit := func(ev agentEvent) {
if ev.Type == "tool_use" || ev.Type == "tool_result" {
if m, ok := ev.Data.(map[string]any); ok {
m["type"] = ev.Type
toolCalls = append(toolCalls, m)
}
}
if ev.Type == "text" {
finalText, _ = ev.Data.(string)
}
if ev.Type == "error" {
errText, _ = ev.Data.(string)
}
}
a.chatWith(cctx, p.SessionID, "", note, emit)
if finalText != "" || len(toolCalls) > 0 {
break
}
if attempt == 0 {
slog.Warn("nomos: auto-continuation produced nothing, retrying once", "session", p.SessionID, "execution", p.ExecID, "error", errText)
}
}
text := finalText
if text == "" && errText != "" {
text = fmt.Sprintf("(auto-continuation hit an internal error and did not respond: %s — the execution's own result is above; you may need to prompt the agent again)", errText)
slog.Error("nomos: auto-continuation produced no response after retry", "session", p.SessionID, "execution", p.ExecID, "error", errText)
}
assistantMsg, _ := json.Marshal(map[string]any{
"role": "assistant",
"text": text,
"tool_calls": toolCalls,
"auto": true, // marks this as an autonomous continuation, not an operator turn
})
a.store.saveMessage(ctx, p.SessionID, "assistant", assistantMsg)
}
// buildContinuationNote frames the finished execution for the model: what
// happened, and what to do about it. The persist-through-errors instruction
// lives here (and in SOUL) so the agent recovers instead of stopping.
func buildContinuationNote(p pendingContinuation) string {
action := p.Action
if i := strings.IndexByte(action, ':'); i > 0 && len(action) > 40 {
action = action[:i] // keep just the action verb for brevity; params are in the DB
}
result := p.Result
if len(result) > 3000 {
result = result[:3000] + "…[truncated]"
}
var b strings.Builder
fmt.Fprintf(&b, "[System: execution %s (%s) finished with status=%s.\nResult: %s\n\n",
p.ExecID, action, p.Status, result)
switch p.Status {
case "completed":
b.WriteString("It SUCCEEDED. Continue the approved plan: run the next step. If this was the final step, verify the end goal actually works (e.g. curl the service) and then report success to the operator. Do NOT stop and wait for the operator to say 'continue'.")
case "failed", "cancelled":
b.WriteString("It FAILED. Do NOT give up or hand back to the operator. Diagnose the cause from the result above (and by running read-only inspection commands if needed), form a hypothesis, fix it, and retry or take an alternative approach. You have an active assent window, so config_mutation steps run without re-approval. Only stop and ask the operator if you are genuinely blocked (need information only they have) or the fix would require a destructive action they haven't approved.")
default: // denied / revoked
b.WriteString("The operator denied or revoked this step. Stop executing this plan and briefly acknowledge.")
}
b.WriteString("]")
return b.String()
}