| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "fmt" |
| 6 | "strings" |
| 7 | |
| 8 | "reasonix/internal/event" |
| 9 | "reasonix/internal/evidence" |
| 10 | "reasonix/internal/tool" |
| 11 | ) |
| 12 | |
| 13 | // hostReceiptsMaxItems bounds each rendered line so a long child run cannot |
| 14 | // crowd the parent's context out with paths and commands. |
| 15 | const hostReceiptsMaxItems = 8 |
| 16 | |
| 17 | const hostReceiptsHeader = "Host receipts (recorded by the host as the sub-agent ran, not claimed by it):" |
| 18 | |
| 19 | const hostReceiptsViolationLabel = "OUTSIDE DECLARED write_paths" |
| 20 | |
| 21 | // splitHostReceipts separates a child's own prose from the host attestation |
| 22 | // appended to it. Aggregates truncate prose to fit a budget; the attestation is |
| 23 | // bounded already and must never be the part that gets cut. |
| 24 | func splitHostReceipts(answer string) (prose, receipts string) { |
| 25 | idx := strings.LastIndex(answer, hostReceiptsHeader) |
| 26 | if idx < 0 { |
| 27 | return answer, "" |
| 28 | } |
| 29 | return strings.TrimRight(answer[:idx], "\n"), strings.TrimSpace(answer[idx:]) |
| 30 | } |
| 31 | |
| 32 | // boundedHostReceipts trims an attestation to fit limit bytes. The header and |
| 33 | // any violation line always survive: a parent may lose the detail of what |
| 34 | // changed, but never the fact that a write left the declared claim. |
| 35 | func boundedHostReceipts(receipts string, limit int) string { |
| 36 | if receipts == "" || len(receipts) <= limit { |
| 37 | return receipts |
| 38 | } |
| 39 | lines := strings.Split(receipts, "\n") |
| 40 | var violations []string |
| 41 | for _, line := range lines[1:] { |
| 42 | if strings.Contains(line, hostReceiptsViolationLabel) { |
| 43 | violations = append(violations, line) |
| 44 | } |
| 45 | } |
| 46 | if len(violations) == 0 { |
| 47 | return utf8Prefix(lines[0], limit) |
| 48 | } |
| 49 | // A claim escape outranks the header it would normally sit under. |
| 50 | if withHeader := strings.Join(append(lines[:1:1], violations...), "\n"); len(withHeader) <= limit { |
| 51 | return withHeader |
| 52 | } |
| 53 | return utf8Prefix(strings.Join(violations, "\n"), limit) |
| 54 | } |
| 55 | |
| 56 | // decorateExecutionReceipt records what the host itself observed about one tool |
| 57 | // call. The output length and the process outcome never come from model |
| 58 | // arguments, which is what makes the resulting attestation trustworthy. |
| 59 | func decorateExecutionReceipt(rec *evidence.Receipt, result string, ex *tool.ShellExecution) { |
| 60 | if rec == nil { |
| 61 | return |
| 62 | } |
| 63 | rec.ObserveOutput(result) |
| 64 | if ex == nil { |
| 65 | return |
| 66 | } |
| 67 | if ex.ExitCode != nil { |
| 68 | code := *ex.ExitCode |
| 69 | rec.ExitCode = &code |
| 70 | } |
| 71 | rec.Verification = ex.Verification |
| 72 | rec.Interrupted = ex.State == tool.ShellStateCancelled |
| 73 | } |
| 74 | |
| 75 | // composeSubagentAnswer assembles everything the parent is shown for one child |
| 76 | // run: the model's completion report when the child submitted one, the |
| 77 | // child's own prose, then the host's receipts. |
| 78 | func composeSubagentAnswer(ctx context.Context, answer string, sub *Agent, claims WritePathSet, delegationText string) string { |
| 79 | summary := sub.EvidenceSummary() |
| 80 | report, hasReport := sub.CompletionReport() |
| 81 | if hasReport { |
| 82 | answer = strings.TrimSpace(formatCompletionReport(report) + "\n\n" + answer) |
| 83 | } |
| 84 | recordDelegationAudit(ctx, summary, claims, hasReport, delegationText) |
| 85 | return appendHostReceipts(answer, summary, claims) |
| 86 | } |
| 87 | |
| 88 | // recordDelegationAudit emits one structured receipt per child run. It reports |
| 89 | // what the host observed, so an orchestration |
| 90 | // benchmark can separate real gains from extra tokens spent. delegationText is |
| 91 | // the parent-authored task before host framing, which is what makes the |
| 92 | // evidence-origin split a host record rather than a claim. |
| 93 | func recordDelegationAudit(ctx context.Context, summary evidence.ChildEvidenceSummary, claims WritePathSet, hasReport bool, delegationText string) { |
| 94 | audit := evidence.DelegationAudit{ |
| 95 | Depth: SubagentDepth(ctx), |
| 96 | ToolCalls: len(summary.Receipts), |
| 97 | MutationPaths: summary.MutationPaths(), |
| 98 | ClaimViolations: len(claimViolations(summary, claims)), |
| 99 | HasReport: hasReport, |
| 100 | } |
| 101 | audit.Mutations = len(audit.MutationPaths) |
| 102 | audit.ClassifyEvidenceOrigin(delegationText, summary.EvidencePaths()) |
| 103 | _, sink, _, _ := CallContext(ctx) |
| 104 | event.RecordDelegationAudit(sink, audit) |
| 105 | } |
| 106 | |
| 107 | // formatCompletionReport labels model claims separately from execution facts. |
| 108 | func formatCompletionReport(report evidence.CompletionReport) string { |
| 109 | var b strings.Builder |
| 110 | b.WriteString("Model-reported status: ") |
| 111 | b.WriteString(string(report.Status)) |
| 112 | b.WriteString("\nsummary: ") |
| 113 | b.WriteString(report.Summary) |
| 114 | for _, c := range report.Criteria { |
| 115 | b.WriteString("\n " + c.ID + " " + string(c.Status)) |
| 116 | if proof := criterionProof(c); proof != "" { |
| 117 | b.WriteString(" — " + proof) |
| 118 | } |
| 119 | } |
| 120 | for _, u := range report.Unresolved { |
| 121 | b.WriteString("\nunresolved: " + u) |
| 122 | } |
| 123 | return b.String() |
| 124 | } |
| 125 | |
| 126 | func criterionProof(c evidence.AcceptanceCriterion) string { |
| 127 | var parts []string |
| 128 | for _, e := range c.Evidence { |
| 129 | switch { |
| 130 | case strings.TrimSpace(e.Command) != "": |
| 131 | parts = append(parts, e.Command) |
| 132 | case len(e.Paths) > 0: |
| 133 | parts = append(parts, strings.Join(e.Paths, " ")) |
| 134 | case strings.TrimSpace(e.Summary) != "": |
| 135 | parts = append(parts, e.Kind+": "+e.Summary) |
| 136 | } |
| 137 | } |
| 138 | return joinBoundedReceipts(parts) |
| 139 | } |
| 140 | |
| 141 | // appendHostReceipts attaches the host's own attestation to a child's answer. |
| 142 | // A child cannot write, suppress, or contradict these lines. An empty block is |
| 143 | // omitted entirely, so read-only research children stay exactly as cheap as |
| 144 | // they were before. |
| 145 | func appendHostReceipts(answer string, summary evidence.ChildEvidenceSummary, claims WritePathSet) string { |
| 146 | block := formatHostReceipts(summary, claims) |
| 147 | if block == "" { |
| 148 | return answer |
| 149 | } |
| 150 | if strings.TrimSpace(answer) == "" { |
| 151 | return block |
| 152 | } |
| 153 | return answer + "\n\n" + block |
| 154 | } |
| 155 | |
| 156 | // claimViolations returns the mutations the host observed outside the write |
| 157 | // claim the child declared. Tool-level confinement already refuses these, so a |
| 158 | // non-empty result means a write reached the workspace through a surface the |
| 159 | // claim could not bind — the parent must not treat the run as scoped. |
| 160 | func claimViolations(summary evidence.ChildEvidenceSummary, claims WritePathSet) []string { |
| 161 | if claims.Empty() { |
| 162 | return nil |
| 163 | } |
| 164 | var out []string |
| 165 | for _, p := range summary.MutationPaths() { |
| 166 | if !claims.AllowsPath(p) { |
| 167 | out = append(out, p) |
| 168 | } |
| 169 | } |
| 170 | return out |
| 171 | } |
| 172 | |
| 173 | // formatHostReceipts renders only what the host is willing to attest to: files |
| 174 | // the child really changed, and commands whose outcome the host observed. |
| 175 | // Ordinary reads and greps are excluded on purpose — they are not claims a |
| 176 | // parent has to adjudicate, and every rendered line costs parent context. |
| 177 | func formatHostReceipts(summary evidence.ChildEvidenceSummary, claims WritePathSet) string { |
| 178 | changed := receiptDisplayPaths(summary.MutationPaths()) |
| 179 | commands := hostReceiptCommands(summary) |
| 180 | violations := receiptDisplayPaths(claimViolations(summary, claims)) |
| 181 | if len(changed) == 0 && len(commands) == 0 { |
| 182 | return "" |
| 183 | } |
| 184 | var b strings.Builder |
| 185 | b.WriteString(hostReceiptsHeader) |
| 186 | if len(changed) > 0 { |
| 187 | b.WriteString("\n changed: ") |
| 188 | b.WriteString(joinBoundedReceipts(changed)) |
| 189 | } |
| 190 | if len(commands) > 0 { |
| 191 | b.WriteString("\n commands: ") |
| 192 | b.WriteString(joinBoundedReceipts(commands)) |
| 193 | } |
| 194 | if len(violations) > 0 { |
| 195 | b.WriteString("\n " + hostReceiptsViolationLabel + ": ") |
| 196 | b.WriteString(joinBoundedReceipts(violations)) |
| 197 | } |
| 198 | return b.String() |
| 199 | } |
| 200 | |
| 201 | func receiptDisplayPaths(paths []string) []string { |
| 202 | out := make([]string, len(paths)) |
| 203 | for i, value := range paths { |
| 204 | out[i] = strings.ReplaceAll(value, `\`, "/") |
| 205 | } |
| 206 | return out |
| 207 | } |
| 208 | |
| 209 | // hostReceiptCommands keeps only shell receipts carrying an outcome worth |
| 210 | // attesting: a host-classified verification, or a command that did not succeed. |
| 211 | func hostReceiptCommands(summary evidence.ChildEvidenceSummary) []string { |
| 212 | var out []string |
| 213 | seen := map[string]bool{} |
| 214 | for _, r := range summary.Receipts { |
| 215 | cmd := strings.TrimSpace(r.Command) |
| 216 | outcome := hostReceiptOutcome(r) |
| 217 | if cmd == "" || outcome == "" || seen[cmd] { |
| 218 | continue |
| 219 | } |
| 220 | seen[cmd] = true |
| 221 | out = append(out, cmd+outcome) |
| 222 | } |
| 223 | return out |
| 224 | } |
| 225 | |
| 226 | func hostReceiptOutcome(r evidence.Receipt) string { |
| 227 | var parts []string |
| 228 | switch r.Verification { |
| 229 | case evidence.VerificationPassed: |
| 230 | parts = append(parts, "verification passed") |
| 231 | case evidence.VerificationFailed: |
| 232 | parts = append(parts, "verification failed") |
| 233 | } |
| 234 | switch { |
| 235 | case r.ExitCode != nil && (*r.ExitCode != 0 || len(parts) > 0): |
| 236 | parts = append(parts, fmt.Sprintf("exit %d", *r.ExitCode)) |
| 237 | case r.ExitCode == nil && !r.Success: |
| 238 | parts = append(parts, "did not complete") |
| 239 | } |
| 240 | if len(parts) == 0 { |
| 241 | return "" |
| 242 | } |
| 243 | return " (" + strings.Join(parts, ", ") + ")" |
| 244 | } |
| 245 | |
| 246 | func joinBoundedReceipts(items []string) string { |
| 247 | if len(items) <= hostReceiptsMaxItems { |
| 248 | return strings.Join(items, ", ") |
| 249 | } |
| 250 | return strings.Join(items[:hostReceiptsMaxItems], ", ") + |
| 251 | fmt.Sprintf(" (+%d more)", len(items)-hostReceiptsMaxItems) |
| 252 | } |
| 253 |