feat(discovery): discover OMP session transcripts in the pi adapter

.gitignore

@@ -4,3 +4,4 @@ node_modules/
 .DS_Store
 *.log
 package-lock.json
+graft/.cache/

README.md

@@ -217,15 +217,15 @@ It cannot be combined with
 
 backpass reads the local transcript stores of seven harnesses directly. No API, no upload.
 
-| Harness        | Store                                          | Repo tie                                            |
-| -------------- | ---------------------------------------------- | --------------------------------------------------- |
-| **claude**     | `~/.claude/projects/<munged-cwd>/<uuid>.jsonl` | per-line `cwd`                                      |
-| **codex**      | `~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl` | `cwd` + recorded `git.repository_url`               |
-| **pi**         | standalone and BB-managed Pi JSONL stores      | session-header `cwd`                                |
-| **opencode**   | `~/.local/share/opencode/opencode.db` (sqlite) | `session.directory` / `session_v2.directory`        |
-| **grok**       | `~/.grok/sessions/<encoded-cwd>/<uuid>/`       | `summary.json` `cwd` + `git_remotes`                |
-| **cursor CLI** | `~/.cursor/chats/<md5(cwd)>/<uuid>/`           | `meta.json` `cwd`                                   |
-| **hermes**     | `~/.hermes/state.db` (sqlite)                  | session cwd, with CLI prompt / ACP config fallbacks |
+| Harness        | Store                                           | Repo tie                                            |
+| -------------- | ----------------------------------------------- | --------------------------------------------------- |
+| **claude**     | `~/.claude/projects/<munged-cwd>/<uuid>.jsonl`  | per-line `cwd`                                      |
+| **codex**      | `~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl`  | `cwd` + recorded `git.repository_url`               |
+| **pi**         | standalone, OMP, and BB-managed Pi JSONL stores | session-header `cwd`                                |
+| **opencode**   | `~/.local/share/opencode/opencode.db` (sqlite)  | `session.directory` / `session_v2.directory`        |
+| **grok**       | `~/.grok/sessions/<encoded-cwd>/<uuid>/`        | `summary.json` `cwd` + `git_remotes`                |
+| **cursor CLI** | `~/.cursor/chats/<md5(cwd)>/<uuid>/`            | `meta.json` `cwd`                                   |
+| **hermes**     | `~/.hermes/state.db` (sqlite)                   | session cwd, with CLI prompt / ACP config fallbacks |
 
 Claude collection covers `$CLAUDE_CONFIG_DIR/projects` alongside the default store, so a
 relocated config dir does not hide its sessions. The variable is read from backpass's own
@@ -238,6 +238,25 @@ sessions under `~/.bb/pi-bridge-sessions/`. It also honors `PI_CODING_AGENT_DIR`
 set in backpass's environment. When roots overlap, backpass scans every applicable layout
 and reads each JSONL file once.
 
+OMP (Oh My Pi) collection is **off by default**. Set `discovery.includeOmp` to `true` in
+`.backpassrc.json` or your [personal config](#configuration) to also read
+`~/.omp/agent/sessions/` through the Pi adapter:
+
+```json
+{ "discovery": { "includeOmp": true } }
+```
+
+Then run `backpass scan --harness pi` (or a normal `backpass` run). The setting also
+applies to configured SSH hosts. For `--scope user`, set it in `user.discovery` and
+include `"pi"` in `user.discovery.harnesses`; user scope otherwise collects only Claude
+and Codex. Explicit Pi store environment overrides still work without this setting;
+it controls only the additional default OMP store, not the harness backpass invokes.
+
+OMP nests subagent JSONL files below each parent session, and a subagent's own subagents
+one level further down. Backpass analyzes each file separately, but uses the root session as
+their shared corroboration source; a session and its subagents cannot count as independent
+sessions.
+
 OpenCode collection reads both store layouts: OpenCode 1.x (`session`, `message`, `part`) and OpenCode 2.x (`session_v2`, `session_message`).
 For 2.x, session activity uses the later of the session's update time and its newest message's update time.
 An upgraded store keeps its 1.x tables beside the copies in `session_v2`, so a session found in both is read from `session_v2`.
@@ -289,12 +308,12 @@ OpenCode sessions with no recorded messages, such as unused agent probes, are no
 
 Every remaining session is labelled **interactive** or **non-interactive** (`src/interaction.js`).
 Codex `codex exec` / `originator: codex_exec`, Claude SDK, GitHub, action, and CI
-entrypoints, OpenCode child sessions (`parent_id`), and a cwd with a `.no-mistakes` path
-segment are non-interactive. Hermes gateway, cron, and WhatsApp sessions are classified the
-same way if they leak past collection's source filter. A no-mistakes pipeline run is just one
-kind of non-interactive session, not its own category. Missing harness metadata defaults to
-interactive. `backpass scan`, the proposal, and apply all print the mix so relevance is never
-silently computed against a robot-skewed pool.
+entrypoints, OpenCode child sessions (`parent_id`), OMP subagent transcripts, and a cwd
+with a `.no-mistakes` path segment are non-interactive. Hermes gateway, cron, and WhatsApp
+sessions are classified the same way if they leak past collection's source filter. A
+no-mistakes pipeline run is just one kind of non-interactive session, not its own category.
+Missing harness metadata defaults to interactive. `backpass scan`, the proposal, and apply
+all print the mix so relevance is never silently computed against a robot-skewed pool.
 
 ```sh
 backpass scan --since 7d --strict
@@ -812,6 +831,7 @@ CLI flags on top:
     "since": "30d",
     "worktreeGlobs": [],
     "cloneRoots": [],
+    "includeOmp": false,
     "minUserTurns": 2
   },
   "jobs": 4

src/acpx.js

@@ -906,8 +906,9 @@ function recoverPiUsage({ promptFile, cwd, startedAt }) {
     .filter((c) => c.mtimeMs >= since)
     .sort((a, b) => b.mtimeMs - a.mtimeMs);
 
+  const scanContext = piStore.createScanContext();
   for (const candidate of candidates) {
-    const descriptor = piStore.classify(candidate);
+    const descriptor = piStore.classify(candidate, { scanContext });
     if (!descriptor || !wanted.has(descriptor.cwd)) continue;
     const entries = readJsonl(candidate.path);
     const firstUser = entries.find((e) => e.type === "message" && e.message?.role === "user");

src/analyze.js

@@ -13,7 +13,7 @@ import { renderOpenGapIndex } from "./gap-ledger.js";
 import { evidenceKey, isEvidenceFresh, safeFileName } from "./state.js";
 import { emitProgress } from "./progress.js";
 import { UserError, color, info, warn } from "./logger.js";
-import { transcriptIdentity } from "./transcript.js";
+import { corroborationIdentityOf, transcriptIdentity } from "./transcript.js";
 
 /**
  * Stage 1 of the pipeline (design section 3): one cheap model call per transcript,
@@ -410,6 +410,10 @@ export async function analyzeTranscripts({
     harness: transcript.harness,
     id: transcript.id,
     identity: transcriptIdentity(transcript),
+    parentSessionId: transcript.parentSessionId || null,
+    corroborationIdentity: corroborationIdentityOf(transcript),
+    corroborationNativeId: transcript.corroborationNativeId || null,
+    corroborationStartedAt: transcript.corroborationStartedAt ?? null,
     path: transcript.path,
     mtimeMs: transcript.mtimeMs,
     bytes: transcript.bytes,

src/commands/propose.js

@@ -1,6 +1,11 @@
 import { consolidateGapLedger } from "../consolidate.js";
 import { foldEvidence } from "../fold.js";
-import { ledgerGapObservations, pruneGapLedger, recordGapObservations } from "../gap-ledger.js";
+import {
+  ledgerGapObservations,
+  normalizeGapLedgerSessions,
+  pruneGapLedger,
+  recordGapObservations,
+} from "../gap-ledger.js";
 import { synthesizeProposal } from "../synthesize.js";
 import { ProposalViolation } from "../proposal.js";
 import { formatCorpusMix, INTERACTIVE, NON_INTERACTIVE } from "../interaction.js";
@@ -13,7 +18,7 @@ import { printUsage } from "./usage.js";
 import { closeRemoteDiscovery, discoverForRun } from "./scan.js";
 import { capTranscripts } from "../sample.js";
 import { isEvidenceFresh } from "../state.js";
-import { transcriptIdentity } from "../transcript.js";
+import { corroborationIdentityOf, transcriptIdentity } from "../transcript.js";
 import { pruneHostCache } from "../discovery/cache.js";
 
 /**
@@ -29,11 +34,12 @@ import { pruneHostCache } from "../discovery/cache.js";
  * cap remain on disk. Folding those records would inflate `analyzedSessions` beyond the
  * sampled corpus or score positional instruction aliases against an index they never saw.
  * Legacy records stay excluded until ordinary discovery and analysis backfill them.
+ * Current discovery's corroboration fields are overlaid on admitted records, so a
+ * subagent analyzed before its parent appeared still folds under the parent's identity.
  */
 export async function foldForRun(ctx, memoryFile, memoryHash, skills = [], transcripts = [], { route = null } = {}) {
   const { state, minGapEvidence, gapLedgerMaxAge } = ctx.config;
   const selectedByIdentity = new Map(transcripts.map((transcript) => [transcriptIdentity(transcript), transcript]));
-  const selected = new Set(selectedByIdentity.keys());
   const evidence = state.listEvidence();
   const identitiesByLegacyId = new Map();
   for (const record of evidence) {
@@ -42,26 +48,41 @@ export async function foldForRun(ctx, memoryFile, memoryHash, skills = [], trans
     if (!identitiesByLegacyId.has(legacyId)) identitiesByLegacyId.set(legacyId, new Set());
     identitiesByLegacyId.get(legacyId).add(transcriptIdentity(record.transcript));
   }
-  const selectedGapSessions = new Set(selected);
+  const selectedGapSessions = new Set(selectedByIdentity.keys());
+  const legacyIds = new Set();
   for (const transcript of transcripts) {
+    selectedGapSessions.add(corroborationIdentityOf(transcript));
     const identities = identitiesByLegacyId.get(transcript.id);
     if (identities?.size === 1 && identities.has(transcriptIdentity(transcript))) {
       selectedGapSessions.add(transcript.id);
+      legacyIds.add(transcript.id);
     }
   }
-  const relevant = evidence.filter((e) => {
-    const currentTranscript = selectedByIdentity.get(transcriptIdentity(e.transcript));
-    return (
-      e.memoryPath === memoryFile.path &&
-      e.memoryHash === memoryHash &&
-      (e.transcript?.interaction === INTERACTIVE || e.transcript?.interaction === NON_INTERACTIVE) &&
-      currentTranscript &&
-      isEvidenceFresh(e, currentTranscript, memoryHash)
-    );
-  });
+  const relevant = [];
+  for (const record of evidence) {
+    const currentTranscript = selectedByIdentity.get(transcriptIdentity(record.transcript));
+    if (
+      record.memoryPath !== memoryFile.path ||
+      record.memoryHash !== memoryHash ||
+      (record.transcript?.interaction !== INTERACTIVE && record.transcript?.interaction !== NON_INTERACTIVE) ||
+      !currentTranscript ||
+      !isEvidenceFresh(record, currentTranscript, memoryHash)
+    ) {
+      continue;
+    }
+    const transcript = {
+      ...record.transcript,
+      parentSessionId: currentTranscript.parentSessionId || null,
+      corroborationIdentity: corroborationIdentityOf(currentTranscript),
+      corroborationNativeId: currentTranscript.corroborationNativeId || null,
+      corroborationStartedAt: currentTranscript.corroborationStartedAt ?? null,
+    };
+    relevant.push({ ...record, transcript });
+  }
 
   const ledger = state.readGapLedger();
-  recordGapObservations(ledger, relevant, { skills });
+  normalizeGapLedgerSessions(ledger, transcripts, { legacyIds });
+  recordGapObservations(ledger, relevant, { skills, legacyIds });
   // Consolidate after recording, so the pass sees this run's sightings too: two
   // sessions coining the same brand-new gap in one parallel fan-out can only line up
   // here. One bounded judged call; a failure degrades to lexical identity and the run

src/config.js

@@ -119,6 +119,8 @@ export const DEFAULT_CONFIG = {
      * is refused by name. Absolute paths; `~` is expanded.
      */
     opencodeStores: [],
+    /** Include Oh My Pi's default session store in Pi discovery; opt-in only. */
+    includeOmp: false,
     minUserTurns: 2,
     includeCursorIde: false,
   },
@@ -467,6 +469,9 @@ function validate(config, { kind = "project", repoRoot = null } = {}) {
   }
   config.discovery.harnesses = config.discovery.harnesses.filter((h) => known.has(h));
   parseSince(config.discovery.since);
+  if (typeof config.discovery.includeOmp !== "boolean") {
+    throw new UserError("config.discovery.includeOmp must be a boolean");
+  }
   if (!Array.isArray(config.discovery.cloneRoots) || config.discovery.cloneRoots.some((p) => typeof p !== "string")) {
     throw new UserError("config.discovery.cloneRoots must be an array of paths");
   }