Skip to content

Releases

v0.4.55 — Milestone M36 accepted: standard tracing export (OTel span)

Choose a tag to compare

@surlymochan surlymochan released this 19 Nov 03:01
b5a80c5

Milestone release. Milestone M36 "标准 tracing 导出(OTel span → Langfuse/LangSmith)" accepted — two merged PRs, verified end to end in CI and locally.

Milestone M36 — accepted (epic #279; acceptance detail #279 comment 6998343385; #279/#282 closed). openswarm trace export <run-id> maps a finished run (status.json + result.json + ledger row + per-attempt event files) to OpenTelemetry spans and exports OTLP/HTTP JSON to a local file (default, no network) and/or a configurable OTLP endpoint. Slice 1 (PR #281, merged 2026-11-19T02:13:24Z, squash d22c658) built the span model + export core; slice 2 (PR #283, merged 2026-11-19T03:00:06Z, squash b5a80c5) fixed three mapping gaps found on real runs (agy usage key names, ledger own-row priority, rework+escalation coverage).

Acceptance evidence, all measured and second-party verified by the main round, not worker claims. ① Three real runs exported (20261114-103312-1348a6, real LLM run, 5 spans; 20261112-020043-924da3, verify + 2 reworks + escalation pi→agy, 9 spans; 20261112-020045-310390, verify + 1 rework, 7 spans); 140-item independent check all passed; the main round re-exported two runs and got byte-identical payloads to the worker's evidence files, trace ids deterministic and unique. ② OTLP/HTTP verified against a local receiver: POST /v1/traces, 202, 7,436 bytes, payload deep-equal to the file export; failures never reported as sent. ③ Full suite independently re-run by the main round: 1410 passed, 6 skipped, 212 subtests; PR #281 CI 10/10, PR #283 CI run 44446933394 10/10. ④ Zero new dependencies; existing subcommand signatures and SKILL contract untouched.

Honest notes, stated plainly: the three real runs' verify records carry no seconds, so verify spans are zero-length points marking the completion instant — no duration invented. In run 924da3 the root span lasts 1.598s while result.seconds is 0.308s; the ~1.29s gap (verify + rework overhead) has no source field and was not force-allocated.

v0.4.54 — M35: bound Pi event-log growth

Choose a tag to compare

@surlymochan surlymochan released this 18 Nov 19:24
0359f53

M35 accepted — Pi executor event-log growth bounded

Slice 1 (#276) compacted plain-text deltas; slice 2 (#278) found the same growth in thinking/toolcall deltas on a real run and fixed it.

Measured evidence (real Pi runs, MiniMax-M3.1-Flash-Preview):

  • Defect run on merged main: 31,270,608 bytes in ~60 s (thinking_delta 88%, toolcall_delta 10.5%)
  • Offline replay of that stream through the fix: 31,270,608 → 832,651 bytes (37.6x); all 2,730 deltas compacted, 0 pass-through; held state peaks at 16,257 chars
  • Live re-runs: 49.75x and 51.46x compression; end-block reconstruction 24/24 (offline) and 23/23, 19/19, 12/12, 9/9 (live runs); pi_usage and watch sequences identical to rebuilt unfiltered streams; final text reconstruction exact
  • Interrupted runs (host restart kill, risk_blocked stops) replay cleanly; all lines parse

What changed: PiEventCompactor now compacts thinking/toolCall cumulative snapshots by contentIndex, tracks the consumed prefix so batched snapshots still compact, passes through anything unrecognised, and never drops *_end receipts. No subcommand, flag, or SKILL changes.

Acceptance verdict: #274. Evidence: #277.

v0.4.53 — M34: ledger/result encoding unified to UTF-8

Choose a tag to compare

@surlymochan surlymochan released this 18 Nov 10:00
14e9f15

M34 验收通过:ledger/result 编码统一 UTF-8

  • 切片 1 (#270):append_ledger 及全部 ledger/result 读写路径显式 UTF-8,三处 ledger 读路径加 errors="replace" 容错旧行。
  • 切片 3 (#273):fsencoding 为 ascii 时以 PYTHONUTF8=1 重执行一次,argv 原样透传、标记防循环;sh() 显式 UTF-8 解码(修 Windows 腿 cp1252 解码 format-patch 输出致 patch=None 崩溃)。
  • 验收证据:中文任务名 run(20261118-085947-dab0f1,commit 4a55b94381)与非 ASCII 片段 run(20261118-083537-5bee37)的 result.json 与 ledger 行逐字一致,两文件 strict UTF-8;全量测试与 PR CI、合并后主干 CI 均 10/10 全绿。
  • 未动:CLI 名、bin 路径、既有 OPENSWARM_* 变量语义、子命令签名、SKILL 契约、派单/重试/门禁语义、shipped 默认、历史账本。
  • 跟进项:hermes 默认模型待用户交互自选;fleet guard allowlist 无 hermes。

v0.4.52

Choose a tag to compare

@surlymochan surlymochan released this 18 Nov 05:44
29eac0c

M33: health-gate covers the auto-rework escalation pick.

  • The escalation pick (after retries are exhausted, route() picks a stronger backend) is now health-gated like the initial pick and the fallback re-route (stage: escalation). An unhealthy target blocks the rework dispatch with no exec attempt and no reroute to the exhausted origin backend or a third backend; the decision is recorded in ledger and result.json as auto_rework.escalation.health_gate (PR #264).
  • AUTH_ERROR_RE now covers the agy sign-in family, including agy 1.5.10's real error 'Please sign in to view available models.' (previously misclassified as unknown, so the gate could wave a logged-out backend through under the shipped on_unknown=allow default) (PR #267).
  • Evidence (all verified by second-party review of the original run records): healthy escalation to hermes dispatched with gate record (run 20261118-123023-1eb66b); unhealthy escalation to agy blocked with zero exec attempts under the shipped default (run 20261118-132529-3c00db, verdict unhealthy, action blocked, ledger and result.json identical).

v0.4.51

Choose a tag to compare

@surlymochan surlymochan released this 17 Nov 21:37
55d1a03

M32: health-gate shipped default flipped to enabled.

  • Shipped default routing.health_gate.enabled now true (M32 slice 1, PR #260, squash 55d1a03); the gate runs unless disabled by config or --no-health-gate.
  • Default-on full suite green: 1345 tests, 0 failures (PR CI 10/10 legs green; main CI green).
  • Real cross-host evidence (verified by second-party re-read): deepseek blocked (exit 2, no exec) with the reason matching an independently re-run probe; codex/pi/agy allowed with ledger rows; pi really executed on gocc (status done, verify rc 0, proof file read back).
  • Flip ruling: no rollback; the default stays enabled.
  • Honest notes: the pi execution ran via direct ssh plus the operator-authorized wrapper (fleet exec does not forward env and pi-isolated is Linux-only — pre-existing gap); codex real execution still awaits its quota reset on 2026-11-22.

v0.4.50: M31 fleet per-host health gate (no flip)

Choose a tag to compare

@surlymochan surlymochan released this 17 Nov 15:45
495a824

M31 accepted: fleet per-host health gate + default flip re-evaluation (no flip)

What shipped (PR #255 + #257):

  • Executor-side health probe (openswarm health-probe --backend <b> [--json]) and per-host gate decisions (fleet_probe_host_health(), health_gate_decide() as the single verdict path for run/batch/fleet); gate blocks unhealthy hosts and lets healthy hosts dispatch, recorded in the ledger as fleet_health_gate rows with probe_host / probe_side: executor.
  • Real cross-host evidence (gocc): deepseek blocked (RC=2, unhealthy — no DeepSeek API key) and codex dispatched and really executed (remote run 20261117-213237-5602ef, 28.8s, result done, usage 25838/122/71); full evidence in docs/m31-slice2-real-evidence.md.
  • Flip evaluation: NOT flipped. With the default flipped to enabled: true, the full suite ran 1345 tests with 55 failures + 9 errors — dispatch-path tests implicitly depend on the default being off (this VM's codex is not logged in, so the gate blocks dispatch). policy.json stays enabled: false. Re-flip requires migrating those 64 tests to green under default-on plus a user decision on the impact (every dispatch on a host with no logged-in default backend would exit 2).

Verification: PR CI 10/10 green; main CI run 44107892084 success (attempt 3 — macOS py3.13 leg was twice cancelled by the runner while queued, 0 steps, then passed); independent full-suite re-run 1345 tests OK (skipped=6); remote run result.json re-read on gocc by the reviewing round.

Honest notes: the control-side ledger rows were lost when the VM swapped and wiped /tmp; the blocked-side row rests on the builder's verbatim pre-swap readback, the allowed-side run was re-verified live on gocc. No behavior change to dispatch/routing; CLI name, bin path, OPENSWARM_* env, subcommand signatures and SKILL contract untouched.

v0.4.49: M30 — SuperAgent intake covers batch + fleet

Choose a tag to compare

@surlymochan surlymochan released this 17 Nov 02:22
2d78035

M30 accepted (total #246): SuperAgent intake now dispatches through batch and fleet paths, not only single runs.

  • Slice 1 (#247, PR #248): batch-path intake core + guardrails.
  • Slice 2 (#251, PR #252): fleet-path intake core — Mandate cards dispatched in topo order across fleet runners via fleet_dispatch_once, each card health-gated (M25) before any exec attempt, Channel completion post with fleet id + per-card run ids.
  • Real-run evidence: 2 batch runs + 2 fleet runs dispatched via intake (ledger 41 lines), 1 blocked (health gate: no copilot binary on host) + 4 allowed; evidence checker 10/10 on the raw ledger. Full suite 1309 passed / 6 skipped / 212 subtests.
  • Shipped default unchanged (dispatch.mode stays run); explicit --backend stays health-gated, never exempt.

Honest notes: the live block was a pre-dispatch health-gate block, not an in-card block (in-card blocking covered by unit tests); pi health-gate verdict was unknown on this host (quota probe cannot read MiniMax quota) and dispatched under on_unknown=allow.

v0.4.48 — M29 accepted

Choose a tag to compare

@surlymochan surlymochan released this 16 Nov 14:27
8ee1d56

M29 accepted (swrm #241, slice 2 #244)

Intake controller proven against live allowlisted work on surlymochan/evalid:

  • Two real evalid issues dispatched end-to-end (intake → Mandate → real run → PR → Channel post): #201 → PR #207 (merged), #208 → PR #209 (merged, squash 28d6416). Both mains CI green after merge.
  • Validator intake_m29_slice2_evidence_check re-run by the main round against the live 55-line ledger: 8/10. The two failures trace to a single disclosed platform incident (VM reboot killed the #201 poll after dispatch_started; its PR #207 still exists and is merged). All mechanism checks pass on the live ledger: gate ignore (3), mandate withdrawn (1), no dispatch after withdraw, cross-repo duplicate skipped (1), one assignment per ticket, allowlist scope, per-thread caps respected (1–2 dispatches / 10–20 min, within 3 / 60 min).
  • Shipped in this release: PR #243 (slice 1: intake policy + per-repo allowlist) and PR #245 (slice 2: live evidence checker + acceptance record).

Acceptance comment: swrm#244 comment 6919195215.

v0.4.47

Choose a tag to compare

@surlymochan surlymochan released this 15 Nov 20:56
d3d5a28

M28 accepted: SuperAgent intake controlled rollout (shadow → active)

The intake now really dispatches on its own inside the allowlisted scope — verified end to end with live issues, not fixtures.

Evidence (all read back from the ledger, run result.json files, and GitHub):

  • 2 allowlisted issues dispatched end to end: #235 → run 20261115-193436-6f7889 → PR #238; #236 → run 20261115-194553-72ae75 (done) → PR #239. Both PRs merged; close-out by verify evidence (run-internal verify rc 0 + operator independent re-run 102/103 passed), never executor self-report.
  • Out of scope stays shadow: evalid #205, 4 triggers, 0 runs. Unauthorized trigger ignored (0 runs); stop/withdraw retires the Mandate (0 Assignments, re-poll 0 new triggers).
  • Guardrails: caps within 3 dispatches / 60 min per thread; idempotent (1 Assignment per ticket); shipped default still shadow.
  • Full suite 1187 passed, 6 skipped, 212 subtests passed (independently re-run); main CI run 43758841536 success on d3d5a28.

Merged in this release: #240 (slice-2 evidence + ledger acceptance checker), #238 (poll prints the dispatch note), #239 (thread-restricted poll leaves the repo-wide cursor untouched).

Honest notes: the #235 run reports status failed only because the executor emitted no final JSON (result null) — the work stands on the branch + verify evidence. PR opening is the operator's step (the local runner never pushes, by design). No live unhealthy-backend block was staged; that path is unit-test covered.

v0.4.46 — M27 SuperAgent 自治接入层

Choose a tag to compare

@surlymochan surlymochan released this 15 Nov 08:09
4d4628a

v0.4.46 — M27 SuperAgent 自治接入层(融合 dsh-super-agent)

M27 epic(#212)已验收。领域模型移植自 DeepSeek 官方 deepseek-ai/dsh-superagent-prototype(MIT);结单差异:swrm 不采执行器自报,Ticket 结单仍须 verify/证据通过。

  • 切片 1(PR #220):领域核心(Trigger/Thread/Ticket/Mandate/Assignment)+intake 账本+确定性 Gate 准入、Mandate 密封,shadow 模式。
  • 切片 2(PR #227):GitHub Source(gh 轮询)+Assessment 研判+accept 真派单+Channel 回贴。
  • 切片 3(PR #230):Subscription 续接(真人 follow-up 同 Mandate resume,不重走 Eval)+真人 stop 撤回 Mandate+operator intake close(凭 verify 证据)+只读 openswarm attention。

验收证据(受控仓 surlymochan/swrm,accept mode=on,探针 #229):授权 issue → accept → 真实 echo 派单完成 → Channel 回贴;真人 follow-up 续接同一 Ticket/Mandate(无新 Decision、无派单);真人 stop 撤回 Mandate;全量测试 1148 通过+6 跳过,CI 10/10 全绿。

默认 accept_mode 仍为 shadow;远程派单与既有子命令签名不变。