AI search engine and MCP server for bots and agents. One Rust binary indexes code, docs, logs, JSONL/CSV and PDFs for BM25 full-text, vector and hybrid search, code search without grep, agent memory and RAG. Elasticsearch-compatible API; Corpus Hub serves hundreds of reference memory corpora.
Rust agent-memoryai-agentsbm25claude-codeclicode-searchcursordatabaseelasticsearchfull-text-searchhybrid-searchjsonllocal-firstmcpmcp-servermodel-context-protocolragrustsearch-enginevector-search
pack: agent-session-trajectories on main (mirror of corpus-hub fc8f86d82 + 1cfad62d1)
xerj-org18c224473e81
diff --git a/.github/workflows/pack-publish.yml b/.github/workflows/pack-publish.ymlindex c5cbe21c4..4d07d1665 100644--- a/.github/workflows/pack-publish.yml+++ b/.github/workflows/pack-publish.yml@@ -1,12 +1,13 @@-# Scheduled publish of the rust-vulns corpus pack (M6 of the corpus-builder-# arc). A pack's freshness is whatever its last build fetched — that is an-# operational promise this workflow keeps, not a README claim: the pack's own-# 30-day staleness refusal applies to our published corpus exactly as it does-# to a user's reference corpora, so a build that stops here shows up as a-# stale badge, not a silently aging artifact.+# Scheduled publish of the committed corpus packs (rust-vulns, M6 of the+# corpus-builder arc; agent-session-trajectories, the first+# external-contributor pack). A pack's freshness is whatever its last build+# fetched — that is an operational promise this workflow keeps, not a README+# claim: the pack's own 30-day staleness refusal applies to our published+# corpus exactly as it does to a user's reference corpora, so a build that+# stops here shows up as a stale badge, not a silently aging artifact.## Chain of custody, in job order:-# build (recipe.toml, committed) → sign (PACK_SIGNING_SEED secret)+# build (recipe.toml, committed) → sign (per-pack PACK_SIGNING_SEED* secret)# → verify against the COMMITTED public key → publish# The verify step is load-bearing: it proves the secret still matches# tools/packs/keys/rust-vulns.pub in the same run that publishes, so a@@ -204,3 +205,148 @@ jobs:[ -n "$tag" ] && gh release delete "$tag" \--repo "$GITHUB_REPOSITORY" --cleanup-tag --yesdone++ agent-session-trajectories:+ name: build + sign + publish agent-session-trajectories+ runs-on: ubuntu-latest+ timeout-minutes: 45+ steps:+ - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4.4.0++ - uses: dtolnay/rust-toolchain@stable++ - uses: Swatinem/rust-cache@6323deb102c322ba6fcbdcafc7e3dddab59af2b6 # v2.9.2+ with:+ workspaces: engine++ - name: Build the xerj binary+ working-directory: engine+ run: cargo build --profile ci-test -p xerj-server++ - name: Build the pack+ env:+ XERJ_CODE_HOME: ${{ github.workspace }}/.pack-home+ run: |+ engine/target/ci-test/xerj corpus build agent-session-trajectories \+ --recipe tools/packs/agent-session-trajectories/recipe.toml++ - name: Sign the pack+ env:+ XERJ_CODE_HOME: ${{ github.workspace }}/.pack-home+ PACK_SIGNING_SEED_AST: ${{ secrets.PACK_SIGNING_SEED_AST }}+ run: |+ if [ -z "$PACK_SIGNING_SEED_AST" ]; then+ echo "PACK_SIGNING_SEED_AST secret is not set. Generate a keypair:"+ echo " xerj corpus keygen --out tools/packs/keys/agent-session-trajectories"+ echo "commit the .pub, set the secret to the .key's hex line, delete the .key"+ exit 1+ fi+ printf '%s' "$PACK_SIGNING_SEED_AST" > "$RUNNER_TEMP/pack.key"+ engine/target/ci-test/xerj corpus sign \+ "$XERJ_CODE_HOME/builds/agent-session-trajectories/pack/agent-session-trajectories" \+ --key "$RUNNER_TEMP/pack.key"+ rm -f "$RUNNER_TEMP/pack.key"++ - name: Verify + install the signed pack (consumer path)+ env:+ XERJ_CODE_HOME: ${{ github.workspace }}/.pack-home-verify+ run: |+ engine/target/ci-test/xerj corpus add agent-session-trajectories \+ --from "$GITHUB_WORKSPACE/.pack-home/builds/agent-session-trajectories/pack/agent-session-trajectories" \+ --verify-sig tools/packs/keys/agent-session-trajectories.pub++ - name: Pack release metadata (freshness.json)+ env:+ XERJ_CODE_HOME: ${{ github.workspace }}/.pack-home+ run: |+ cd "$XERJ_CODE_HOME/builds/agent-session-trajectories/pack/agent-session-trajectories"+ RECORDS=$(python3 -c "import json;print(json.load(open('manifest.json'))['counts']['records'])")+ BUILT=$(python3 -c "import json;print(json.load(open('manifest.json'))['created_at'])")+ GEN=$(python3 -c "import json;print(json.load(open('manifest.json'))['generation'])")+ cat > "$RUNNER_TEMP/agent-session-trajectories-freshness.json" <<EOF+ {+ "schemaVersion": 1,+ "label": "agent-session-trajectories pack",+ "message": "$RECORDS trajectories, built ${BUILT%%T*}, gen $GEN",+ "color": "blue"+ }+ EOF++ - name: Zip the pack+ env:+ XERJ_CODE_HOME: ${{ github.workspace }}/.pack-home+ run: |+ cd "$XERJ_CODE_HOME/builds/agent-session-trajectories/pack/agent-session-trajectories"+ zip -qr "$RUNNER_TEMP/agent-session-trajectories-pack.zip" .+ # the checksums and signature also ship as loose assets, so a+ # consumer can audit a download without unpacking the zip first+ cp SHA256SUMS "$RUNNER_TEMP/agent-session-trajectories-SHA256SUMS"+ cp SHA256SUMS.sig "$RUNNER_TEMP/agent-session-trajectories-SHA256SUMS.sig"++ # Same immutable-release protocol as the rust-vulns job above: each+ # build is its own dated tag, uploaded to a draft, published last.+ - name: Compute the daily tag+ run: echo "PACK_TAG=pack-agent-session-trajectories-$(date -u +%F)" >> "$GITHUB_ENV"++ - name: Drop today's release if this is a same-day re-run+ env:+ GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}+ run: >+ gh release delete "$PACK_TAG" --repo "$GITHUB_REPOSITORY"+ --cleanup-tag --yes || true++ - name: Upload the assets to a draft release+ uses: softprops/action-gh-release@v2+ with:+ tag_name: ${{ env.PACK_TAG }}+ name: "agent-session-trajectories corpus pack (${{ env.PACK_TAG }})"+ draft: true+ prerelease: true+ body: |+ The `agent-session-trajectories` corpus pack — the first+ external-contributor pack on the hub (chunxiaoxx / Nautilus,+ #1255; demand anchored to #1138, wishlist slot #1210). 70+ sanitized multi-session agent trajectories with failure turns+ kept verbatim, built and signed by the `pack publish` workflow+ from `tools/packs/agent-session-trajectories/recipe.toml`.++ One release per build day; the newest is always at+ https://github.com/xerj-org/xerj/releases?q=pack-agent-session-trajectories++ Install (public key: `tools/packs/keys/agent-session-trajectories.pub` in the repo):++ xerj corpus add agent-session-trajectories \+ --from agent-session-trajectories-pack.zip \+ --verify-sig agent-session-trajectories.pub+ xerj corpus index agent-session-trajectories++ `agent-session-trajectories-freshness.json` carries the build+ date and record count; the same 30-day staleness refusal that+ governs every reference corpus governs this one.+ files: |+ ${{ runner.temp }}/agent-session-trajectories-pack.zip+ ${{ runner.temp }}/agent-session-trajectories-SHA256SUMS+ ${{ runner.temp }}/agent-session-trajectories-SHA256SUMS.sig+ ${{ runner.temp }}/agent-session-trajectories-freshness.json++ - name: Publish the draft+ env:+ GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}+ run: >+ gh release edit "$PACK_TAG" --repo "$GITHUB_REPOSITORY"+ --draft=false++ - name: Retire pack releases beyond the newest 7+ continue-on-error: true+ env:+ GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}+ run: |+ mapfile -t old < <(gh release list --repo "$GITHUB_REPOSITORY" \+ --limit 100 --json tagName,isDraft \+ --jq '[.[] | select(.tagName | startswith("pack-agent-session-trajectories-"))+ | select(.isDraft | not) | .tagName]+ | sort | reverse | .[7:][]')+ for tag in "${old[@]:-}"; do+ [ -n "$tag" ] && gh release delete "$tag" \+ --repo "$GITHUB_REPOSITORY" --cleanup-tag --yes+ donediff --git a/tools/packs/agent-session-trajectories/README.md b/tools/packs/agent-session-trajectories/README.mdnew file mode 100644index 000000000..0305bd456--- /dev/null+++ b/tools/packs/agent-session-trajectories/README.md@@ -0,0 +1,59 @@+# agent-session-trajectories++**A first-of-kind record pack for the XERJ corpus hub: sanitized multi-session agent trajectories with failure turns intact, from a production multi-agent engineering org.**++## Provenance++Every record is assembled from our own production operations (watch-loop logs,+session memory, audit records) — no customer data, no third-party content. The+underlying system is the same production agent-memory deployment whose six+"eval scars" are published on the XERJ recipes page; this pack carries the+trajectory-level material behind those scars.++- **What each record uniquely contributes**: a task-scoped session with the+ failed attempts preserved verbatim (diagnosis text kept), a topic-labeled+ boundary (semantic topic, not time-window), and the final fix with an+ evidence hash. An agent working on *agent infrastructure, evaluation+ pipelines, or org operations* would query this corpus for *precedent on what+ already failed and what fix actually worked*.+- **Sanitization**: five-step pipeline (credential regex sweep incl. rotated+ secrets, PII role-ization, internal-codename generalization, financial-value+ aggregation, dual rescan) plus a 10% human sample audit (7/7 pass, seeded).+- **Licence**: CC-BY-4.0. Review conclusion: all records are our own+ operations content; redistribution permitted with attribution.++## Identity rules++Records are unique by `id` (`nst-<session-key>`); no alias resolution is+needed (one session → one record). 70 records in v0, spanning three task+topics: eval-judging (46), infra-diagnosis (13), ledger-audit (6), ops-misc (5).+**Identity-resolution numbers: 70 envelopes → 70 records** (single source,+unique by `id` — no merging, no alias chains).++## Known limits++- Single-org sample: one production org's operating pattern; no claim of+ cross-org generality.+- Mostly Chinese diagnostic text (English fields: topic, status enums, ids).+- Extraction is rule-based (failure-anchor segmentation); recall on+ failure-turn detection is unmeasured — some sessions may be missing turns.+- The `final_fix.evidence_sha16` anchors to our internal artifacts; the+ artifacts themselves are not in this pack.++—— Nautilus (compass judging org) · pack v0 · 2026-10++## Baseline attribution (基准归属声明)++Records contain our **evaluation-process records against third-party benchmarks**+(SWE-bench family, © their respective maintainers). This pack contains **only our+own evaluation-process records and readouts** — no benchmark question text, no+benchmark data redistribution. All benchmark references are attributed to their+original maintainers.++中文:包内仅含我方评测过程记录与读数,不含基准题面/数据再分发;基准归属归原方。++## 中文 README(zh)++本包 = 智涌(Nautilus)组织自产运营轨迹的脱敏记录集:70 条会话,含失败尝试逐字保留、+主题切片、最终修复与证据哈希。用途:agent 检索"哪些做法已失败、什么修复有效"。+许可 CC-BY-4.0;脱敏五步+10% 人工复核(7/7 过);详见 recipes 页与 issue #1138。diff --git a/tools/packs/agent-session-trajectories/pack-stats.json b/tools/packs/agent-session-trajectories/pack-stats.jsonnew file mode 100644index 000000000..5aa10c7c6--- /dev/null+++ b/tools/packs/agent-session-trajectories/pack-stats.json@@ -0,0 +1,17 @@+{+ "provenance": "measured",+ "measured_at": "2026-10-08",+ "pack": "agent-session-trajectories",+ "recipe_sha256": "344fef9a89559b2c45bad8eaca8b21c11bd5c689f7f49822b01f7884d8da6f12",+ "sources": {+ "nautilus-ops": { "kind": "git", "envelopes": 70, "licence": "CC-BY-4.0" }+ },+ "envelopes": 70,+ "records": 70,+ "topics": { "eval-judging": 46, "infra-diagnosis": 13, "ledger-audit": 6, "ops-misc": 5 },+ "records_bytes": 220857,+ "notes": {+ "identity": "70 envelopes to 70 records: single source, unique by id (nst-<session-key>) — no alias chains exist to close, the merged count is the envelope count exactly as the README predicts",+ "acceptance": "first external-contributor pack (chunxiaoxx / Nautilus, demand anchored to #1138, wishlist slot #1210). Source verified live at review: runtime/xerj_pack/pr_ready/agent-session-trajectories.jsonl = 70 lines / 147,368 bytes; topic counts in the README match the data. Built from the merged recipe at fc8f86d82"+ }+}diff --git a/tools/packs/agent-session-trajectories/recipe.toml b/tools/packs/agent-session-trajectories/recipe.tomlnew file mode 100644index 000000000..2169206da--- /dev/null+++ b/tools/packs/agent-session-trajectories/recipe.toml@@ -0,0 +1,55 @@+# agent-session-trajectories — a record pack from a production multi-agent+# engineering org: 70 sanitized multi-session agent trajectories with failure+# turns kept verbatim, topic-sliced (semantic, not time-window), each with the+# final fix and an evidence hash. Demand-anchored to issue #1138; hub backlog+# slot "agent-session-trajectories" (status planned, PR #1210).+#+# Source, and why exactly this one (measured 2026-10-08):+#+# - nautilus-compass (CC-BY-4.0, our own production ops records): 70 records+# assembled by tools/xerj_pack_extract.py from our watch-loop logs and+# session memory — no customer data, no third-party content. Sanitization:+# 5-step pipeline (credential regex sweep incl. rotated secrets, PII+# role-ization, codename generalization, financial-value aggregation,+# dual rescan) + 10% human sample audit (7/7 pass, seeded). Final scan:+# 0 credential/PII/schema failures across all files+# (tools/xerj_pack_scan.py, selftest 5/5, fails=0 warns=17 all /root/+# generic paths).+#+# Identity: records are unique by id (nst-<session-key>); no alias resolution+# exists — one session = one record, 70 envelopes -> 70 records. Each record:+# id, session_date, task_topic (semantic slice), context,+# attempted[]/failed_attempts[] (verbatim diagnosis), final_fix (with+# evidence_sha16), verdict (pass/fail/partial), sources, sanitized flag.++[recipe]+format = 1+name = "agent-session-trajectories"+description = "Sanitized multi-session agent trajectories from a production multi-agent org: failure turns kept verbatim, topic-sliced, with final fixes and evidence hashes"++[[sources]]+slug = "nautilus-ops"+kind = "git"+url = "https://github.com/chunxiaoxx/nautilus-compass"+glob = "runtime/xerj_pack/pr_ready/agent-session-trajectories.jsonl"+format = "flat"+licence = "CC-BY-4.0"++[envelope]+id_from = ["id"]+title_from = ["task_topic", "id"]+body_join = ["context", "attempted", "final_fix"]+defs_from = ["failed_attempts", "verdict", "session_date"]++[identity]+# id: one session = one record; single source, no cross-source aliases —+# the merged count is exactly the envelope count (70 -> 70).+edges = [{ field = "id" }]+canonical_source_order = ["nautilus-ops"]++[merge]+# single source — nothing to merge; id uniqueness is the whole contract.+precedence = ["nautilus-ops"]++[emit]+shards = 16diff --git a/tools/packs/keys/agent-session-trajectories.pub b/tools/packs/keys/agent-session-trajectories.pubnew file mode 100644index 000000000..c7f3b64a1--- /dev/null+++ b/tools/packs/keys/agent-session-trajectories.pub@@ -0,0 +1 @@+a589bb3e1b8646e368a458c602e23a11c25f12cbe459f252e82cdd2f4ab69ee1