Files
oikos/cmd/nomos/assent.go
dtoro d52968876a
Some checks failed
ci / build-test (push) Has been cancelled
ci / docker-build (push) Has been cancelled
feat: general gated run primitive + chat-assent approval (Layer 0)
Implements the first slice of plans/2026-07-10-general-gated-execution.md:
Nomos gets one general execution tool instead of only a fixed action enum,
gated by an automatic risk classifier, and approval can be granted by the
operator just replying in chat instead of clicking a button.

- internal/policy/command.go: ClassifyCommand(cmd, declaredRisk) — rule-based
  read-only allowlist + destructive denylist, default-escalate to
  config_mutation for anything else. Classification can only ESCALATE the
  caller's declared risk, never de-escalate it (destructive always wins even
  if declared read_only). Compound commands (&&, ;, |, $()) never qualify for
  the read-only fast path. Full test corpus.
- internal/mcp/server.go: new `run` MCP tool — target (host:/lxc:), command,
  purpose, optional declared_risk. Read-only commands execute immediately;
  everything else queues an approval exactly like pct_create today, executed
  via httpapi's existing executeApprovedAction. Also fixes a real latent bug:
  pct_exec resolved an LXC's host attribute without the "host:" prefix, so it
  could never find the Proxmox host — new resolveExecTarget/resolveRunTarget
  helpers (mcp + httpapi) fix this for both the new `run` action and existing
  actions that route through the same execution path.
- internal/httpapi/phase3.go: "run" case in executeApprovedAction; fixes two
  bugs found while wiring this up — (1) DecideApproval hardcoded risk_class to
  'config_mutation' on every approve, silently corrupting the audit ledger for
  every other risk class; (2) denying/revoking an approval never updated the
  linked execution's status, so it stayed 'pending_approval' forever instead
  of reflecting the decision.
- cmd/nomos/assent.go: deterministic (not LLM-judged) chat-assent detection.
  Scoped to the immediately-preceding assistant turn's pending approvals only
  — an old "yes" can't retroactively approve something new. Destructive-risk
  actions are excluded from loose assent. Approves via the same HTTP decision
  endpoint the UI button calls, so both paths share one audit trail.
- web/.../InlineApproval.svelte: self-healing poll — a pending approval card
  now picks up being decided via ANY path (chat assent, Ops page, Matrix), not
  just its own button. Previously the banner stayed stuck showing
  Approve/Deny even after the action had already run elsewhere.
- nomos/SOUL.md: `run` is now the general capability ("no fixed menu, only a
  risk gate"); documents chat-assent behavior and the destructive exception.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-10 09:28:00 +02:00

119 lines
4.4 KiB
Go

package main
import (
"bytes"
"context"
"encoding/json"
"fmt"
"net/http"
"regexp"
"strings"
)
// Chat-assent approval: the operator authorizes a proposed action by
// replying normally in chat ("go ahead", "yes", "do it") instead of clicking
// a separate Approve button. This is deterministic (not LLM-judged) so it
// can't be talked around by a model that misreads intent, and it only ever
// looks at the assistant turn immediately preceding the operator's reply —
// an old "yes" from three messages ago can never retroactively approve
// something new. Destructive-risk actions are excluded: they always need the
// explicit typed-confirmation flow, never loose assent.
// pendingApproval is one gated action proposed in the immediately-preceding
// assistant turn, extracted from its tool_result text.
type pendingApproval struct {
execID string
destructive bool
}
// executionQueuedRE matches the "execution <uuid> queued" phrasing shared by
// the run and request_execution/pct_create tool result messages.
var executionQueuedRE = 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})\s+queued`)
// extractPendingApprovals scans the tool results of one assistant turn for
// gated actions that are still awaiting a decision.
func extractPendingApprovals(calls []persistedCall) []pendingApproval {
var out []pendingApproval
for _, c := range calls {
text := c.resultText()
m := executionQueuedRE.FindStringSubmatch(text)
if m == nil {
continue
}
out = append(out, pendingApproval{
execID: m[1],
destructive: strings.Contains(strings.ToUpper(text), "DESTRUCTIVE"),
})
}
return out
}
// negationWords, checked first: any of these anywhere in the message means
// the reply is NOT assent, even if a positive word also appears (e.g. "no,
// don't restart it yet" contains neither "yes" nor "go ahead", but "wait"
// alone should also block a stray "yes" a sentence later — checking negation
// first and returning false errs toward re-confirming rather than assuming
// consent, per "when in doubt, escalate").
var negationWords = []string{
"no", "nope", "don't", "do not", "stop", "wait", "hold on", "hold off",
"not yet", "cancel", "nevermind", "never mind", "actually don't", "skip that",
}
// assentWords, checked only if no negation matched.
var assentWords = []string{
"go ahead", "goahead", "yes", "yep", "yeah", "yup", "do it", "proceed",
"approve", "approved", "confirm", "confirmed", "ship it", "sounds good",
"lgtm", "run it", "execute", "ok go", "okay go", "please do",
}
// isAssent reports whether msg is a plain-language authorization of a
// pending proposal. Deliberately simple and auditable: a fixed word list,
// not a model judgment call, so behavior is predictable and can't be
// prompt-injected via the pending action's own content.
func isAssent(msg string) bool {
m := " " + strings.ToLower(strings.TrimSpace(msg)) + " "
for _, w := range negationWords {
if strings.Contains(m, " "+w+" ") || strings.HasPrefix(strings.TrimSpace(m), w+" ") || strings.HasPrefix(strings.TrimSpace(m), w+",") {
return false
}
}
for _, w := range assentWords {
if strings.Contains(m, w) {
return true
}
}
return false
}
// approveExecution grants (or denies) a pending execution via the same HTTP
// endpoint the chat UI's Approve button calls, so both paths share one code
// path server-side (executeApprovedAction) and one audit trail. Returns the
// decided status, or an error if the request failed outright (a 4xx for an
// already-decided/expired approval is reported via ok=false, not a hard err,
// since that's an expected race, not a bug).
func (a *agent) approveExecution(ctx context.Context, execID string) (ok bool, status string, err error) {
if a.apiBase == "" {
return false, "", fmt.Errorf("no API base configured")
}
body, _ := json.Marshal(map[string]string{"decision": "approve"})
req, err := http.NewRequestWithContext(ctx, http.MethodPost,
a.apiBase+"/api/v1/approvals/"+execID+"/decision", bytes.NewReader(body))
if err != nil {
return false, "", err
}
req.Header.Set("Content-Type", "application/json")
resp, err := a.httpClient.Do(req)
if err != nil {
return false, "", err
}
defer resp.Body.Close()
if resp.StatusCode != http.StatusOK {
return false, "", nil // already decided / expired / not found — not a hard failure
}
var out struct {
Status string `json:"status"`
}
json.NewDecoder(resp.Body).Decode(&out)
return true, out.Status, nil
}