Skip to content

7.7.0

Choose a tag to compare

@github-actions github-actions released this 12 Oct 21:07
· 197 commits to main since this release

[7.7.0] — 2026-10-12

Changes

  • Merge pull request #82 from DailybotHQ/feat/swarm-autonomy
  • fix(addon): pin herdr-peers v0.3.2
  • fix(shared): drill round-2 hardening — packet repo, export usage, §17 rules
  • fix(test): measurement script no longer fails on sort's broken pipe
  • fix(addon): pin coding-agents-kit v0.6.1 and herdr-peers v0.3.1
  • fix(shared): hysteresis binds the packet's author; §17 drill round-1 fixes
  • fix(addon): pin coding-agents-kit v0.6.0 and herdr-peers v0.3.0
  • docs(skill): §17.2 — every herdr-peers call, check included, passes the lane's pane
  • fix(execute): never overwrite a pending packet; resume checks packets first
  • fix(shared): resolve the local review — stale packets, ordering, never-block wording
  • ci(release): fold an Unreleased block into the release section; Learnings for 7.7.0
  • fix(shared): an untouched earlier view render is machine output, not a human edit
  • feat(addon): agentkit capacity and leases; herdr notices and leadership
  • feat(execute): capacity check at each task boundary; resume reads a pending packet first
  • feat(skill): the quality bar — gates are hypotheses, verified not asserted
  • feat(skill): escalation taxonomy C1-C7 — decide and record everything else
  • feat(docs): add Guide §17 Swarm autonomy and route it on its trigger
  • feat(shared): add swarm.py — own-capacity decision and the handoff packet

Swarm autonomy

  • Guide §17 Swarm autonomy (guide/swarm.md): roles, the capacity loop,
    handoff packets, lead succession with epochs and fencing, lane-to-lane asks,
    escalation per chain, the release train, the record and how to drill it.
  • shared/swarm.py: capacity (continue / prepare / handoff over
    coding-agents-kit ak capacity) and handoff write|accept|show with the
    dwp-handoff-v1 schema.
  • AGENT_PROTOCOL §7.5 escalation taxonomy (C1–C7) and the quality bar
    (DWP_SPECIFICATION §5.1.4).
  • Pins: coding-agents-kit v0.6.1 (ak capacity, usage cache, ak lease;
    the patch fixes Codex lanes' --self, own-lease counting with --reuse
    and account resolution found by the live drill) and herdr-peers v0.3.2
    (notices, the per-plan lead record, usage_limit
    screens; the patches make lead changes succeed past closed panes, accept
    status --pane and lead show --pane, resolve a DWP_PLAN plan id, and
    keep a lead record's repository root local), each with its release
    SHA256SUMS digest.
  • Known limitation: the v7 journal has no dedicated event for a lane handoff
    or a lead succession (recorded as a cancelled delegation plus an
    observation); dedicated shapes are a follow-up.

Learnings

What 313 executed plans in 23 repositories taught the methodology (evidence
abstracted; the count is how many of six independent audits saw it, or the
measured share in the plans read):

  • Gates are hypotheses — a new gate is shown to fail before it is trusted;
    vacuous passes fail; compare with the base branch. Evidence: green gates over
    broken deliverables in 6/6 audits (self-grep gates, | tail without
    pipefail, tests that could not fail).
  • Measure first — task 1 of a fix or audit verifies the premise and the
    measuring tool. Evidence: test-bed and service audits; one gate went from
    990 s to 391 s by measuring before changing.
  • Review before the irreversible step; prefer an independent reviewer.
    Evidence: 4/6 audits; seven post-review patch releases; an independent
    reviewer found 2 critical and 4 high findings where the self-review found no
    critical one.
  • A live lead and a formal lane contract (brief = common rules + per-lane
    notes, a plan per lane, a fixed milestone vocabulary with a capacity line).
    Evidence: orchestrators that paused for a human relay stalled (4 of 8 never
    ran their finals); a live-lead swarm closed in a day.
  • Hand off before the limit; succession with epochs and fencing; one executor
    per plan.
    Evidence: account limits and lead loss were the largest swarm
    failures (most cells of one pilot invalidated by quota; asks to a dead lead
    unanswered for hours; this release's own planning lead stopped on quota).
  • Escalation taxonomy — decide and record everything else; batch the
    questions.
    Evidence: in one audit 19 stops were necessary, 8 avoidable and 2
    missed.
  • Autonomy never closes a human gate or overrides the developer's literal
    words; authority is not transitive.
    Evidence: a self-signed human review and
    a decision taken against the request under "be autonomous"; lanes that asked
    their own human after a relayed approval lost hours (3/6 audits).
  • Record the swarm in the parent journal. Evidence: one hub recorded 0
    delegation events against 31 asks.
  • Pins are data. Evidence: 3/6 audits; seven pin-only amendments and a
    two-hour refusal loop.
  • Verbatim human authority in the plan folder; relayed authority marked.
    Evidence: 56 human notes found only in scratch space (3/6 audits).
  • Generated-view reconciliation fixed (shared/views.py): an untouched
    earlier render is machine output, and a reconciliation is recorded as the
    agent's action under the named authority. Evidence: 27–50 % of journal
    events in v6 plans were reconciliations attributed to a human.
  • Scoped staging and scoped external data — stage by path, never a blanket
    git add -A; external test data by recorded ID only. Evidence: one test-bed
    audit.
  • Declared gate environment; environment failures are blockers, never
    passes or blamed on the change.
    Evidence: 5/6 audits.
  • Integrate, then tag. Evidence: 3–7 releases per ecosystem plan, about 12
    in one day, from tag → live failure → patch (3/6 audits).
  • Promote load-bearing helpers on second use. Evidence: 5/6 audits.
  • Lean evidence — reports and pointers, never products or dependency trees
    in analysis_results/. Evidence: 4/6 audits, up to 9.4 GB in one lab.
  • Records coherence at close. Evidence: records disagreed in 6/6 audits
    (9 of 13 hub ledgers, 31 of 70 product plans).
  • Merge authority declared per repository; never an admin merge. Evidence:
    5 admin merges against the stated rule.
  • Coordination hygiene — the lead notice is the single current-state
    broadcast (no ACKs), named locks for shared resources, the non-blocking gap
    protocol. Evidence: ACK storms, stale freezes and two agents on one plan
    (3/6 audits).