Skip to content

AgentWeave v2.0.0

Choose a tag to compare

@gutohuida gutohuida released this 26 Oct 02:12
· 1673 commits to master since this release

The first release since 1.1.0, and a large one: a spec-driven development loop that runs in the
Hub, flows that staff, review and integrate work on their own, runners beyond Claude, a knowledge
vault, and a performance overhaul for weaker machines.

Breaking

  • A Hub that is not told which database to open refuses to start. agentweave-hub, a bare
    uvicorn hub.main:app or any launch that skips the CLI must set DATABASE_URL; 1.1.0 fell back
    to a default file. Bare agentweave still resolves its profile and sets it, and the Docker image
    now sets DATABASE_URL=sqlite+aiosqlite:////app/data/agentweave.db itself (-e overrides it).
  • SQLite runs in WAL (synchronous=FULL, a bounded WAL file). A running Hub's database now
    lives in three files (agentweave.db, -wal, -shm): copying the .db file alone is no longer
    a backup.
    Use SQLite's backup API or VACUUM INTO. If WAL cannot be set (a network share), the
    Hub opens in the journal it has and says so in its startup line.
  • Trace spans are kept for 14 days by default, swept by the Hub off the request path. Set the
    window per project in the trace settings. Agent outputs, events, messages, runs and tasks are
    never pruned.
  • No more guessing that an agent asked a question. The detector that read a turn's trailing
    prose as a question is gone (migration 0082); an agent that needs an answer calls ask_user.
  • Upgrading runs migrations 0082 through 0146 on first start.

Specifications in the Hub

  • The spec/ corpus is owned by the Hub: capability and change documents, requirements, tasks,
    evidence and coverage, with the lifecycle exploring → proposed → approved → archived.
  • Requirements trace to tasks and evidence; approval asks that every requirement a task touches is
    covered by accepted evidence.
  • Roadmaps split large work into slices, each with its own document.

Flows

  • A document says how it will be built (delivery): a flow fires on a schedule, staffs builders
    from the project's agents, stages reviews, and stops when its queue is done.
  • Tasks have dependencies, a board, and governed transitions (pending → … → approved); work
    happens in per-task worktrees and is merged into the main branch on approval.
  • Flows advance as soon as their work moves (completed, sent back, approved), not only on their tick.
  • A stall detector raises stuck work, and the Hub's mender agent can be briefed to unstick it. It is
    never staffed as a builder or reviewer, and it can hand finished work to its reviewer.

Approval checks

  • The operator configures the project's checks; approval runs them on the merged result.
  • A passing result is reused when the merged tree is identical, so the Hub's own prerequisite
    merges and unrelated main moves no longer force a re-run.
  • A failure is re-run once before it counts against the task. One that passes the second time is
    recorded as flaky, approval goes on, and the Overview lists the flaky tests with "Create a fix
    task" and "Dismiss".
  • pytest -n auto in a check is sized to a third of the CPUs; a failing suite keeps the start of its
    tracebacks.
  • Accepted evidence survives a clean prerequisite merge into the task branch.

Runners and agents

  • Runners for Claude, GitHub Copilot (ACP) and Codex (app-server), one registry, model catalog and
    provider allowance holds; usage and spend accounting per turn.
  • Charters, agent configuration, permissions with operator approval cards, sandboxed runs, quiet
    hours, and conversations with checkpoints, hand-off and a workspace side panel.

Knowledge vault

  • Sources (notes, documents, transcripts, decisions) distilled into cited facts, with contradictions
    tracked; tracked or private per entry.

Performance

Measured on a copy of a real operator database:

  • The Hub's event loop no longer waits on git, the filesystem or process spawns. Turn start
    froze every request for 4.1 s, now 14 ms; approval for 3.1 s, now 14 ms. Unchanged git facts are
    cached by commit id, and streamed output lines are written in batches (6.2 ms of loop time per
    line, now 0.9 ms).
  • Reads under write load: p90 2.4 s, now 90 ms.
  • List routes: /agents, /conversations 0.7–1.0 s, now about 20 ms; GET /loops 2.5 s, now
    about 0.2 s; open event streams no longer hold a database connection each.
  • The dashboard: one live event refreshes each thing once; first open goes from 43 requests to
    14, the first script from 1.5 MB to 565 KB, screens load as their own chunks, and a streaming
    1,000-row conversation drops 9% of frames instead of 98% at a 6x-throttled CPU.

Fixed

  • Hundreds of defects found by driving the product end to end, recorded in
    scripts/drive/FINDINGS.md, each with a test that failed before its fix.