Atomic claiming
work_start / work_start_next claim and transition
in one optimistic-locked step. Do not chain a separate claim
with a status update — the two-step form races; the combined verb does not.
Filigree · work state / issue lifecycle · Python
Filigree owns issues, dependencies, and workflow state machines — giving an agent a race-free claim, a dependency-aware ready queue, and a pre-computed orientation file, so no two agents collide. Local-first: SQLite, no cloud, no accounts.
What it is
Traditional issue trackers are human-first; agents scrape CLI output or parse
API responses. Filigree flips this — agents get a pre-computed
context.md at session start, claim work with optimistic locking,
and resume sessions via event streams without re-reading history. Each project
gets a .weft/filigree/ directory (like .git/) holding
a SQLite database, config, and the auto-generated context summary; legacy
.filigree/ stores keep working and migrate on the next
filigree init.
The bug machine has a hard gate: verifying → closed requires a
recorded fix_verification. A fix is not done because someone says
so; it is done when the verification is on the record.
Key capabilities · the five things it gives you
work_start / work_start_next claim and transition
in one optimistic-locked step. Do not chain a separate claim
with a status update — the two-step form races; the combined verb does not.
Blockers, the ready-queue, and critical-path analysis — an agent asks for the next startable issue and gets one that is actually unblocked, not just open.
Agents interact natively over MCP — no parsing text. The count is the
repo's authority; see docs/mcp.md for
the live surface.
entity_association_add binds an issue to an opaque external
entity id (a loomweave:eid:… SEI or a legacy locator). The
consumer does drift detection against the hash captured at attach time.
context.md is regenerated on every mutation and injected at
session start, so an agent orients instantly instead of re-reading history.
The only Weft member with a GUI — everything else is CLI/MCP-first.
Real-time Kanban, a dependency graph, and Files/Health, local-first on
localhost:8377. See it below ↓
The dashboard · unique to Filigree
A local-first board on localhost:8377 — Kanban lanes with
drag-and-drop, a Graph v2 dependency view, Files and Health, and a slide-in
issue detail surfacing the federation's signature data vocabulary. Everything
else in the suite is CLI/MCP-first; this is the one place a human on the loop
can watch the board.
Wire post-commit hook for `warpline changed`
Wardline → Legis signed-scan handshake
elspeth judge-gate parity for the protected cell
wl-log-injection reaches preview sink, unbaselined
SEI-keyed governed sign-off binding
Loomweave + Filigree orphans issue on rename
Backfill temporal store via warpline ingest
localhost:8377 UI. Expect cosmetic divergence from the
running dashboard.
Usage snapshot · curated, not the full reference
# at session start
filigree session-context
# claim the next startable issue (atomic: claim + transition)
filigree start-next-work --assignee me
# …do the work, commit, then
filigree close <id>
# watch the board (the one Weft GUI)
filigree dashboard --port=8377 How it composes · value is the weave
Filigree's slice of the combination matrix. Every status is honest on its face:
a partial or planned pair is never rendered as
live. Each partner links across to its own site.
Findings become tracked work.
A-1: native emitter shipped; asterisk stays live until the Loomweave-absent path is proven end-to-end (currently unit/server-wiring tier).
Issues bound to live code, surviving refactors.
Drift-aware entity⇄issue binding shipped: filigree stores content_hash_at_attach; loomweave issues_for detects + reports drift, integration-tested. Asterisk until a two-real-daemon drift dogfood lands.
Governed issue lifecycle — sign-offs, RTM, verification.
The re-verification worklist becomes tracked work.
Filigree consumer (warpline_worklist_ingest) shipped in Filigree 3.0.0 — the EARNED inbound seam.
Status & honest limits
Links · pointers to the authorities
Want to see it actually run?
Watch the whole federation analyze Lacuna, the demonstration specimen — tracked work, race-free claims, and the dashboard, on a small app with catalogued flaws.