@tangle-network/traces 0.11.4 → 0.11.6

package/dist/index.js

@@ -14,6 +14,7 @@ import {
   CodexTaskScopeError,
   CopilotAdapter,
   DEFAULT_HARNESS,
+  DEFAULT_SANDBOX_BASE_URL,
   EMPTY_SPAN_RUN_SNAPSHOT,
   EXPORT_RULE,
   EmptySessionError,
@@ -24,6 +25,7 @@ import {
   JsonSourceError,
   JsonlParseError,
   OpencodeAdapter,
+  POLICY_EVIDENCE_NOTE,
   PartialUploadError,
   PiAdapter,
   QwenAdapter,
@@ -46,19 +48,26 @@ import {
   analyzeSpans,
   analyzeSteeringChain,
   applyRedactor,
+  assembleSessionBundle,
+  assertSandboxReachable,
   buildFixPrompt,
   buildPolicyEvidenceRecord,
+  buildRetryFixPrompt,
   buildSessionIndex,
   buildSessionIndexFromRows,
   buildSpanRunTree,
   buildTraceFindingPacket,
   capText,
   chatTrajectoryToSpans,
+  classifyEvidenceOnlyPath,
   classifyLiveActions,
   classifyReaction,
+  classifyVerdict,
   claudeActor,
+  clipText,
   cmdReplayVerify,
   cmdReplayVerifyBatch,
+  cmdVerifyFindings,
   codexActor,
   collectPolicyEvidence,
   collectSessionIndex,
@@ -68,6 +77,7 @@ import {
   commandRedactor,
   conformanceOfSpans,
   connectors,
+  countScriptCommands,
   countSkillRunsJsonl,
   cwdMatchesSelection,
   defaultTraceLiveAnalysts,
@@ -76,6 +86,8 @@ import {
   describeSessionRelationship,
   dockerImagePreparer,
   enumerateReplayableCases,
+  environmentAttributes,
+  environmentFromSpanAttributes,
   equivalentGitCwds,
   executeUpload,
   exportTraceEvidenceFile,
@@ -83,12 +95,18 @@ import {
   exportTraceEvidenceText,
   extractFixCommand,
   fileRunContextSupervisorRunReader,
+  findContextRoot,
+  findSessionTextLeaks,
   findSpanSourceFiles,
+  findingReplayStep,
+  findingTrajectoryId,
   forEachTreeNode,
   generateFixCommand,
   goldIncorrectSteps,
   haloAnalyzer,
+  hashSessionBundleFiles,
   hodoscopeAnalyzer,
+  httpJsonTransport,
   importCodeTraceBench,
   incompleteInputsNote,
   indexSessionIdsByTrace,
@@ -103,6 +121,7 @@ import {
   isSubmitAction,
   knownHarnesses,
   listAdapters,
+  listSessionBundleFiles,
   loadState,
   loadTracesConfig,
   locateSessions,
@@ -114,6 +133,7 @@ import {
   otlpRowToSpan,
   outboundHash,
   parseCorpusFlag,
+  parseIncorrectStepsSubject,
   parseIsoToEpochMs,
   parseObservationOutput,
   parseRecordedReturncode,
@@ -121,10 +141,14 @@ import {
   parseReplayVerifyArgs,
   parseSession,
   parseSince,
+  parseVerifyFindingsArgs,
   planTraceAgenticRoute,
   planUpload,
+  primeAnalyzer,
+  projectSessionBundle,
   readCoordinationLog,
   readFileRunContext,
+  readFindingsFile,
   readJsonFile,
   readJsonl,
   readLabelEntries,
@@ -132,6 +156,7 @@ import {
   readRunContextJournal,
   readRunContextSnapshot,
   readRunWatchSnapshot,
+  readSessionBundleManifest,
   readSessionIndexFile,
   readSpanRunSnapshot,
   recordSessionCorruption,
@@ -143,11 +168,15 @@ import {
   renderRunWatchSnapshot,
   renderSpanRunSnapshot,
   renderValidation,
+  renderVerifiedFindingsSection,
   replayBatchUsage,
   replayVerify,
+  replayVerifyFinding,
   replayVerifyUsage,
   resolveAdapter,
   resolveCaseResources,
+  resolveFindingInvocation,
+  resolveFindingReplayability,
   resolveOtlpInputFiles,
   resolveRepoAttrs,
   resolveRunContextDir,
@@ -155,6 +184,7 @@ import {
   resolveSessionRepoAttrs,
   runCommand,
   runExternalAnalyzers,
+  runFixLoop,
   runPipelines,
   runReplayBatch,
   runTotals,
@@ -175,9 +205,11 @@ import {
   sessionHash,
   sessionIdFromAttributes,
   sessionIntegrityAttributes,
+  sessionTextSignatures,
   skippedAnalyses,
   span,
   spanRunTotals,
+  stampEnvironmentAttrs,
   stampRepoAttrs,
   stampSessionIdentity,
   stampSessionIntegrity,
@@ -201,6 +233,8 @@ import {
   userPromptSpan,
   validateOtlpSpans,
   validationExitCode,
+  verifyFindings,
+  verifyFindingsUsage,
   watchRunTarget,
   watchSessions,
   wrap,
@@ -209,12 +243,13 @@ import {
   writeInspectionReportFile,
   writeOtlpFile,
   writePolicyEvidenceFile,
+  writeSessionBundleManifest,
   writeSessionIndexFile,
   writeTraceEvidenceExportFile,
   writeTraceImprovementArtifacts,
   zaiChatCaller,
   zeroPricedCost
-} from "./chunk-AGPOIWPQ.js";
+} from "./chunk-MEAEXKUV.js";
 
 // src/collect.ts
 async function collectSessions(opts = {}) {
@@ -230,79 +265,6 @@ async function collectSessions(opts = {}) {
   return out;
 }
 
-// src/replay-wire.ts
-var SUBJECT_PATTERN = /^incorrect-steps-(\d+)-(\d+)-(escaped|unescaped)-consequence-(\d+)$/;
-function parseIncorrectStepsSubject(subject) {
-  const match = SUBJECT_PATTERN.exec(subject);
-  if (!match) return null;
-  return {
-    firstStep: Number(match[1]),
-    lastStep: Number(match[2]),
-    escapeStatus: match[3],
-    consequenceStep: Number(match[4])
-  };
-}
-function resolveFindingInvocation(finding, corpora) {
-  const subject = parseIncorrectStepsSubject(finding.subject);
-  if (!subject) {
-    throw new Error(
-      `replay-wire: subject '${finding.subject}' is not incorrect-steps-<first>-<last>-<escaped|unescaped>-consequence-<step>`
-    );
-  }
-  const failures = [];
-  for (const corpus of corpora) {
-    const resolution = resolveCaseResources(corpus, finding.trajId);
-    if (resolution.resolved) {
-      const at = subject.firstStep;
-      const step = resolution.resources.steps.find((s) => s.step_id === at);
-      if (!step) {
-        throw new Error(
-          `replay-wire: finding step ${at} is outside ${finding.trajId} (${resolution.resources.steps.length} steps)`
-        );
-      }
-      return { resources: resolution.resources, subject, at };
-    }
-    failures.push(`${corpus.name}: ${resolution.reason}`);
-  }
-  throw new Error(
-    `replay-wire: trajectory ${finding.trajId} is not replayable in any corpus \u2014 ${failures.join("; ")}`
-  );
-}
-async function replayVerifyFinding(finding, options) {
-  if (options.fixCaller && options.fixCommand !== void 0) {
-    throw new Error("replay-wire: pass fixCaller or fixCommand, not both");
-  }
-  const invocation = resolveFindingInvocation(finding, options.corpora);
-  let fixCommand = options.fixCommand ?? null;
-  if (options.fixCaller) {
-    const generated = await generateFixCommand(options.fixCaller, {
-      taskStatement: invocation.resources.taskStatement,
-      steps: invocation.resources.steps,
-      k: invocation.at
-    });
-    if (!generated.succeeded) {
-      throw new Error(`replay-wire: fix generation failed \u2014 ${generated.error}`);
-    }
-    fixCommand = generated.value.command;
-  }
-  const verdict = await replayVerify({
-    stepsPath: invocation.resources.stepsPath,
-    image: invocation.resources.image,
-    at: invocation.at,
-    fixCommand: fixCommand ?? void 0,
-    cwd: invocation.resources.cwd,
-    out: options.out,
-    caseId: invocation.resources.trajId,
-    stepTimeoutMs: options.stepTimeoutMs ?? invocation.resources.recordedStepTimeoutMs ?? void 0,
-    prefixLimit: options.prefixLimit,
-    backend: options.backend,
-    apiKey: options.apiKey,
-    baseUrl: options.baseUrl,
-    onProgress: options.onProgress
-  });
-  return { invocation, fixCommand, verdict };
-}
-
 // src/index.ts
 import {
   AnalystRegistry,
@@ -329,6 +291,7 @@ export {
   CopilotAdapter,
   DEFAULT_HARNESS,
   DEFAULT_REDACTION_RULES,
+  DEFAULT_SANDBOX_BASE_URL,
   EMPTY_SPAN_RUN_SNAPSHOT,
   EXPORT_RULE,
   EmptySessionError,
@@ -339,6 +302,7 @@ export {
   JsonSourceError,
   JsonlParseError,
   OpencodeAdapter,
+  POLICY_EVIDENCE_NOTE,
   PartialUploadError,
   PiAdapter,
   QwenAdapter,
@@ -361,20 +325,27 @@ export {
   analyzeSpans,
   analyzeSteeringChain,
   applyRedactor,
+  assembleSessionBundle,
+  assertSandboxReachable,
   buildDefaultAnalystRegistry,
   buildFixPrompt,
   buildPolicyEvidenceRecord,
+  buildRetryFixPrompt,
   buildSessionIndex,
   buildSessionIndexFromRows,
   buildSpanRunTree,
   buildTraceFindingPacket,
   capText,
   chatTrajectoryToSpans,
+  classifyEvidenceOnlyPath,
   classifyLiveActions,
   classifyReaction,
+  classifyVerdict,
   claudeActor,
+  clipText,
   cmdReplayVerify,
   cmdReplayVerifyBatch,
+  cmdVerifyFindings,
   codexActor,
   collectPolicyEvidence,
   collectSessionIndex,
@@ -385,6 +356,7 @@ export {
   commandRedactor,
   conformanceOfSpans,
   connectors,
+  countScriptCommands,
   countSkillRunsJsonl,
   createHostedClient,
   cwdMatchesSelection,
@@ -394,6 +366,8 @@ export {
   describeSessionRelationship,
   dockerImagePreparer,
   enumerateReplayableCases,
+  environmentAttributes,
+  environmentFromSpanAttributes,
   equivalentGitCwds,
   executeUpload,
   exportTraceEvidenceFile,
@@ -401,13 +375,19 @@ export {
   exportTraceEvidenceText,
   extractFixCommand,
   fileRunContextSupervisorRunReader,
+  findContextRoot,
+  findSessionTextLeaks,
   findSpanSourceFiles,
+  findingReplayStep,
+  findingTrajectoryId,
   forEachTreeNode,
   generateFixCommand,
   goldIncorrectSteps,
   haloAnalyzer,
+  hashSessionBundleFiles,
   hodoscopeAnalyzer,
   hostedClientFromEnv,
+  httpJsonTransport,
   importCodeTraceBench,
   incompleteInputsNote,
   indexSessionIdsByTrace,
@@ -422,6 +402,7 @@ export {
   isSubmitAction,
   knownHarnesses,
   listAdapters,
+  listSessionBundleFiles,
   loadState,
   loadTracesConfig,
   locateSessions,
@@ -442,10 +423,14 @@ export {
   parseReplayVerifyArgs,
   parseSession,
   parseSince,
+  parseVerifyFindingsArgs,
   planTraceAgenticRoute,
   planUpload,
+  primeAnalyzer,
+  projectSessionBundle,
   readCoordinationLog,
   readFileRunContext,
+  readFindingsFile,
   readJsonFile,
   readJsonl,
   readLabelEntries,
@@ -453,6 +438,7 @@ export {
   readRunContextJournal,
   readRunContextSnapshot,
   readRunWatchSnapshot,
+  readSessionBundleManifest,
   readSessionIndexFile,
   readSpanRunSnapshot,
   recordSessionCorruption,
@@ -466,6 +452,7 @@ export {
   renderRunWatchSnapshot,
   renderSpanRunSnapshot,
   renderValidation,
+  renderVerifiedFindingsSection,
   replayBatchUsage,
   replayVerify,
   replayVerifyFinding,
@@ -473,6 +460,7 @@ export {
   resolveAdapter,
   resolveCaseResources,
   resolveFindingInvocation,
+  resolveFindingReplayability,
   resolveOtlpInputFiles,
   resolveRepoAttrs,
   resolveRunContextDir,
@@ -480,6 +468,7 @@ export {
   resolveSessionRepoAttrs,
   runCommand,
   runExternalAnalyzers,
+  runFixLoop,
   runPipelines,
   runReplayBatch,
   runTotals,
@@ -500,9 +489,11 @@ export {
   sessionHash,
   sessionIdFromAttributes,
   sessionIntegrityAttributes,
+  sessionTextSignatures,
   skippedAnalyses,
   span,
   spanRunTotals,
+  stampEnvironmentAttrs,
   stampRepoAttrs,
   stampSessionIdentity,
   stampSessionIntegrity,
@@ -526,6 +517,8 @@ export {
   userPromptSpan,
   validateOtlpSpans,
   validationExitCode,
+  verifyFindings,
+  verifyFindingsUsage,
   watchRunTarget,
   watchSessions,
   wrap,
@@ -534,6 +527,7 @@ export {
   writeInspectionReportFile,
   writeOtlpFile,
   writePolicyEvidenceFile,
+  writeSessionBundleManifest,
   writeSessionIndexFile,
   writeTraceEvidenceExportFile,
   writeTraceImprovementArtifacts,

package/docs/replay-verify.md

@@ -77,6 +77,17 @@ A corpus is a gold label file plus a prepared directory (`normalized/<traj_id>/s
 
 Fix generation (`--fix generate`) runs ONE chat completion per arm-A-reproduced case against an OpenAI-compatible endpoint (default: glm-5.2 on z.ai). The prompt carries the gold step's action and observation, ±3 surrounding steps, and the task statement; the reply must be a single corrected shell command. Cases beyond `--max-fix-cases` are excluded by a seeded random sample and marked `sampled-out` in the report. Arm B replays the prefix in its own fresh sandbox and executes the corrected command; `failureVanished` = exit 0 with the failure signature absent.
 
+### Iterative fix loop (`--fix loop`)
+
+`--fix loop` (in `src/replay-fix-loop.ts`) replaces the single shot with an execution-feedback loop of up to `--fix-attempts` attempts (default 3) per case:
+
+- Attempt 1 is byte-identical to the one-shot prompt, so the report's `fixFlipAttempt1` stays directly comparable to `--fix generate`.
+- When an arm fails (nonzero exit or the failure signature persists) or the model call itself fails, the next prompt carries every prior command with its REAL executed stdout/stderr (clipped tails), never a paraphrase.
+- Retries may answer with a short script — at most 5 commands in one fenced block — executed as ONE `/bin/sh` unit; a longer script is rejected as a failed attempt and the rejection feeds the next prompt.
+- Every attempt runs in its own fresh sandbox with the same replayed prefix. A used sandbox is never mutated mid-arm, so a flip always proves the corrected step against the recorded prefix state, not against debris from an earlier attempt. Wall-time cost is one extra prefix replay per attempt.
+
+The report adds `fixFlipAttempt1` (flips at attempt 1 over cases whose attempt 1 executed), `flipsByAttempt`, the full per-attempt trail on each case row (`fix.attempts`), and one `armB-attempt<N>-result.json` per executed attempt (`armB-result.json` keeps the last executed attempt — the flipped one when the loop flips).
+
 Honest-reporting notes baked into the report:
 
 - A gold step whose recorded observation carries **no returncode** can never satisfy "arm A reproduces the recorded returncode": it counts against the replayability rate and the per-case table says why.
@@ -99,6 +110,12 @@ const { invocation, fixCommand, verdict } = await replayVerifyFinding(
 
 The subject grammar is the analyst benchmark's `incorrect-steps-<first>-<last>-<escaped|unescaped>-consequence-<step>`; `--at` is the finding's **first** incorrect step (the finding's claim, which may differ from the gold label). The wire resolves the trajectory across the given corpora, generates the arm-B fix through the same one-call generator (or accepts a pre-supplied `fixCommand`), and returns the full `ReplayVerdict`. It throws with a precise reason when the finding cannot be replayed (malformed subject, unknown trajectory, non-SWE case, step out of range) — the product surfaces that reason instead of a proof.
 
+## Product surface — verified findings
+
+`traces verify-findings` (and `traces analyze --verify-findings`) runs this proof per recorded analyst finding and annotates each with `reproduced | fix-flipped | divergent | not-replayable` plus a receipt directory.
+Unlike the wire above it accepts the shapes analysts actually emit (`incorrect-step-<n>` subjects, `metadata.block_first_step`, `trace://` evidence refs) and never throws on a finding-shaped dead end — the dead end becomes the finding's honest `not-replayable` receipt.
+See [Verified findings](./trace-analysts.md#verified-findings-executed-replay).
+
 ## Orchestrator prerequisites
 
 replay-verify talks to a sandbox API (`--base-url`); in local development that is the sandbox SDK adapter in front of an orchestrator running the docker driver.

package/docs/trace-analysts.md

@@ -68,9 +68,12 @@ traces analyze --all --last 20 --analyzer hodoscope
 
 # Run any installed command that accepts an OpenInference file path.
 traces analyze --last 1 --analyzer my-trace-tool
+
+# One-shot prime-RLM analysis through a local OpenAI-compatible bridge.
+traces analyze --last 1 --analyzer prime --analyzer-prompt "find unsupported completion claims"
 ```
 
-`--model` is forwarded to the built-in model-assisted analysts, HALO, and Hodoscope.
+`--model` is forwarded to the built-in model-assisted analysts, HALO, Hodoscope, and prime.
 HALO and Hodoscope use their own provider clients and credentials.
 The Hodoscope adapter pins version `0.2.4` and uses Python 3.11 through `uvx`.
 
@@ -87,6 +90,41 @@ Raw JSON is still a `report`.
 Only an adapter that validates the full finding shape may return `findings`.
 Hodoscope always returns `discovery`, and each candidate has `status: 'needs_review'` plus its source trace and span.
 
+### Prime engine
+
+`--analyzer prime` runs a one-shot analyst over the emitted OTLP artifact through an OpenAI-compatible bridge, such as cli-bridge's prime backend.
+Unlike the `--llm` analysts, which drill into the trace with paged tools, prime has no REPL and no trace tools: the full span projection is inlined into a single prompt as JSON.
+Oversized projections are re-rendered with a per-attribute character cap; if the projection still exceeds the inline budget the engine fails loud instead of silently dropping spans.
+
+Prerequisites:
+
+- A running bridge that accepts `POST /v1/chat/completions` and routes the configured model to the prime backend.
+- No key handling here: the bridge owns provider credentials.
+
+Configuration (flags first, then environment, then defaults):
+
+| Setting | Source | Default |
+| --- | --- | --- |
+| Bridge root URL | `TRACES_PRIME_BRIDGE_URL` | `http://localhost:4181` |
+| Model | `--model`, then `TRACES_PRIME_MODEL` | `prime/zai/glm-5.2` |
+| Per-call deadline | `TRACES_PRIME_TIMEOUT_MS` | `1200000` (20 min) |
+| Question | `--analyzer-prompt` | a general diagnosis question |
+
+```bash
+traces analyze --last 1 --analyzer prime
+TRACES_PRIME_BRIDGE_URL=http://localhost:4181 traces analyze --last 1 --analyzer prime --analyzer-prompt "find unsupported completion claims"
+```
+
+Output expectations:
+
+- The reply contract is one fenced JSON block of short strings citing span ids verbatim; a structurally malformed reply gets exactly one bounded repair turn that carries the malformed reply and the contract, never the trajectory.
+- A still-malformed reply after repair is a failed result (`ok: false`) with the raw reply preserved for inspection; one failed engine never discards the other results.
+- Valid rows become full `findings` with `trace://` span evidence, validated against the artifact like every other findings-kind engine; rows citing unknown or ambiguous span ids are rejected with a recorded reason.
+- Zero findings from a well-formed reply is an honest null, not a failure.
+- Bridge-reported token usage and call counts are recorded in the result output; cost stays uncaptured because this adapter has no pricing table.
+
+The scored prime-vs-dspy comparison — same trajectories, same scoring — lives in `@tangle-network/agent-eval`'s analyst benchmark (`runAnalystBenchmark`); this engine is the capture-side entry point, not the scoreboard.
+
 ## Write one analyst
 
 An analyst receives a paged trace store and returns typed findings.
@@ -210,6 +248,34 @@ console.log(renderAnalystBenchmarkMarkdown(result))
 Public labels test the measurement code, not the quality of every built-in analyst automatically.
 A real quality claim requires running the analyst over the corresponding trajectories, retaining all rows, and comparing it with named alternatives at equal model and request limits.
 
+## Verified findings (executed replay)
+
+An analyst finding is a cited claim until something executes it.
+`traces analyze --verify-findings` (and the standalone `traces verify-findings`) replays each finding's trajectory prefix in a real sandbox, re-runs the accused step, and annotates every finding with an executed verdict:
+
+| Verdict | Meaning |
+|---|---|
+| `reproduced` | the recorded failure signature (returncode + stable output substring) reproduced when the accused step re-ran |
+| `fix-flipped` | reproduced, and a supplied corrected command made the failure vanish in a fresh replay |
+| `divergent` | the step executed but the recorded failure did not reproduce — evidence against the finding, or against replay fidelity (the receipt carries prefix divergences so you can tell which) |
+| `not-replayable` | the finding could not be executed; the receipt names the precise reason (no step subject, unknown trajectory, no docker image, submit step, …) |
+
+```bash
+# Verify the findings an analyze run produced (marks each finding in the report):
+traces analyze --last 1 --llm --verify-findings \
+  --replay-corpus holdout=labels.json::prepared/ --verify-out ./receipts
+
+# Verify findings recorded earlier (e.g. extracted from an eval result.json):
+traces verify-findings --findings findings.json --out ./receipts \
+  --steps normalized/<traj>/steps.json --image <replay-ready-image> --cwd /app
+```
+
+Findings are matched by the shape analysts emit: subject `incorrect-step-<n>` (or the wire form `incorrect-steps-<f>-<l>-…`), `metadata.block_first_step`, and `trace://<trajectory>/…` evidence refs.
+Findings accusing the same step share one executed proof; each finding still gets its own receipt directory (`receipt.json` plus, when executed, `replay-verdict.json` and `report.md` with real stdout/stderr).
+Verification is execution, not generation: no LLM is involved unless you pass `--fix-command`.
+A missing sandbox is an error when any finding is replayable — verification never silently skips.
+Sandbox setup, execution semantics, and honest limits are in [Replay verification](./replay-verify.md).
+
 ## Turn findings into improvement
 
 Do not train or rewrite policy from an analyst's own prose alone.

package/examples/external-engines.ts

@@ -10,6 +10,7 @@ import {
   commandRedactor,
   haloAnalyzer,
   hodoscopeAnalyzer,
+  primeAnalyzer,
   writeOtlpFile,
 } from '@tangle-network/traces'
 
@@ -41,7 +42,19 @@ console.log(
     : `hodoscope unavailable: ${discovery.error}`,
 )
 
-// 3) External REDACTOR: scrub prose with your own PII model before upload. The
+// 3) PRIME ENGINE: one-shot RLM through an OpenAI-compatible bridge (run
+//    cli-bridge's prime backend locally, or set TRACES_PRIME_BRIDGE_URL).
+//    The full span projection travels inline; findings come back with
+//    validated trace:// span evidence.
+const prime = primeAnalyzer({ defaultPrompt: 'find unsupported completion claims' })
+const primeResult = await prime.analyze(otlp)
+console.log(
+  primeResult.ok
+    ? `${primeResult.findings?.length ?? 0} prime finding(s)\n${primeResult.output}`
+    : `prime bridge unavailable: ${primeResult.error}`,
+)
+
+// 4) External REDACTOR: scrub prose with your own PII model before upload. The
 //    command reads a JSON array of strings on stdin and writes the scrubbed array
 //    on stdout (a 3-line wrapper adapts openai/privacy-filter's `opf`).
 const redactor = commandRedactor({ command: 'my-pii-scrubber' })

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/traces",
-  "version": "0.11.4",
+  "version": "0.11.6",
   "description": "Point it at your coding-agent session traces (Claude Code, Codex, OpenCode, Gemini, Pi, …) and get failure-mode + efficiency findings. CLI + SDK over the @tangle-network/agent-eval analyst suite — observe live sessions, run your own analysts, redact, and upload to the Tangle Intelligence Platform.",
   "type": "module",
   "license": "MIT",
@@ -44,13 +44,14 @@
     "README.md"
   ],
   "engines": {
-    "node": ">=22"
+    "node": ">=22.13.0"
   },
   "dependencies": {
-    "@tangle-network/agent-eval": "0.142.2",
-    "@tangle-network/agent-runtime": "0.123.0",
+    "@tangle-network/agent-eval": "0.145.22",
+    "@tangle-network/agent-interface": "1.0.0",
+    "@tangle-network/agent-runtime": "0.137.0",
     "@tangle-network/agent-trace-contract": "^1.0.2",
-    "@tangle-network/sandbox": "0.17.2"
+    "@tangle-network/sandbox": "0.27.1"
   },
   "devDependencies": {
     "@types/node": "^22.0.0",

package/skills/build-trace-analyst/SKILL.md

@@ -48,6 +48,19 @@ Use `runAnalystBenchmark` with `registryBenchmarkRunner` from `@tangle-network/a
 Use AgentRx or CodeTraceBench labels when they match the finding.
 Compare the candidate with a no-findings baseline and relevant maintained or external analysts on identical cases and limits.
 
+Two shipped engines are available as comparison arms.
+`--analyzer prime` inlines the whole span projection into one prompt and needs no REPL.
+`--analyzer halo` and `--analyzer hodoscope` drill into the trace with paged tools.
+Compare an inline arm with a paged arm to separate the prompt from the navigation.
+
+```bash
+traces analyze --last 1 --analyzer prime
+```
+
+Assert the served model, not the requested id.
+A gateway can answer one id with another model, so a panel can report families it never ran.
+Use `assertServedModel` and `assertCrossFamilyServed` from `@tangle-network/agent-eval` before you trust a per-model or cross-family number.
+
 Use at least 20 independent cases before comparative statistics.
 Retain every case row and report exact-step recall, precision, F1, root-step accuracy when labeled, citation coverage and validity, clean-case false positives, repeat agreement, failures, latency, calls, all token fields, and known or missing cost.
 

package/skills/inspect-agent-traces/SKILL.md

@@ -22,10 +22,9 @@ traces analyze --otlp results/sessions --out .traces/all.md   # a directory of e
 Use `--harness` only for coding agents whose on-disk format we do not control.
 Those adapters are the legacy edge, not the way to integrate a system you own.
 
-A directory is a run directory: it holds the span export beside raw event, stream and
-SDK logs that are also `*.jsonl`. Only the OTLP files are read, and the rest are listed
-with what they actually hold. An `otlp/` subdirectory, when the producer made one, is
-read on its own.
+A run directory holds the span export beside raw event, stream and SDK logs that are
+also `*.jsonl`. Only the OTLP files are read; the rest are listed with what they hold.
+An `otlp/` subdirectory, when present, is read on its own.
 
 Any section headed `inputs incomplete`, or carrying an `Inputs incomplete` line above
 its table, is computed from a field the trace does not carry everywhere. Report those
@@ -73,6 +72,33 @@ traces improve --harness codex --current --latest-turn --workflow \
 `improve` writes findings, evidence, a report, and spans.
 It does not edit an agent, repository, memory store, or knowledge base.
 
+Write one session's durable evidence directory for a later reader:
+
+```bash
+traces bundle --harness claude-code --session <id-or-path> --out .traces/bundle
+```
+
+`bundle` copies the transcript, the derived report and spans, the `.evolve` ledger rows inside the session window, and a `manifest.json` with a SHA-256 per file.
+It spends no model call.
+A missing transcript stops the assembly.
+An absent optional input is recorded in `manifest.absent` with the probed path.
+
+## Pick the view for the reader
+
+That bundle is the FULL view (`manifest.view: "full"`), and it holds the whole session transcript.
+Never give it to a writer that must not see an earlier conclusion: the transcript holds the text of every file the session wrote, so a check for the earlier report FILE passes while its CONTENT is still readable.
+
+Project the writer's copy instead:
+
+```bash
+traces bundle-view .traces/bundle --view evidence-only --out .traces/writer
+```
+
+That view carries `derived/session-index.json`, `derived/evidence.jsonl`, and the structured `ledger/` records.
+It excludes the transcripts, the report, the spans, and every prose ledger file by name, with rules and hashes in `manifest.excluded`.
+Before writing, it compares each carried file against each excluded file for shared 8-word runs of prose, and drops any file that repeats one.
+`manifest.view` names which copy you hold.
+
 ## Report
 
 - State the source (`--otlp <path>` or the harness), selected task boundary, session and span counts, and integrity warnings.