Repository navigation
Releases
Releases · tusharmewara/loom
v0.12.1
v0.12.0
v0.11.0
v0.10.2
v0.10.1
Loom v0.10.0 — Harness² analysis, Skill inventory + validation, Pillars 1-4 landed
Loom v0.10.0
Major Features
Harness² Analysis + Skill Inventory Overhaul
- Category analysis of Google Research's Harness² (
docs/LOOM_HARNESS2_ANALYSIS.md) — different category (method, not competitor); Rust-native adoption path scoped to a 3-day slice - Skill inventory, validation & trigger generation (
fa29056):- One scan pass over
~/.loom/skills/handling both dir (SKILL.md) and flat (.md) layouts - 7 defect codes:
missing_description(Fatal),missing_name,name_mismatch,invalid_frontmatter,empty_body,missing_script,duplicate_name loom skills install/removenow accepts both layouts; refuses Fatal defectstrigger_lines()generates agent-facing trigger lines with cap +not_triggeredtrackingloom doctorshows skill-validation row;loom skills listshows per-skill defects + triggersloom lock --generatenow produces a real lockfile (11 skills on this machine)loom-skills82 tests pass;neutral_renderer.rs(373 lines, dead, uncompilable) deleted
- One scan pass over
Pillar 1 — Harden the Core (7457c72)
- CLI bounded backoff to daemon (8 attempts, 200ms→1s) + actionable error ("run
loom daemon restart") - Embedded-pane dispatch now single owner (
App::dispatch_embedded_pane), resolves agent from full task record, falls back to PATH binary, test asserts spawn works - Backlog reconciled: daemon-lifecycle tests landed, 5/6 routing debts closed, keyword weight corrected (0.35 in code)
Pillar 2 — Security & Trust (e0edb50)
- Self-verifying audit export: recomputable chain over normalised outcome (legacy raw-result entries still verify) + Ed25519 per-machine anchor (
<loom_dir>/auth/export-signing.key); three verdicts INTACT/TAMPERED/UNANCHORED, never "intact" without pinned key - Per-run container network posture:
NetworkPolicy::NoNetwork | Restricted{allowed_hosts} | Unrestricted; Docker auto_remove restored; k8s NetworkPolicy created BEFORE Job; CNI caveat documented - Scoped revocable grants: derived per-run credential (master secret never leaves vault), expiry checked at USE, rotation invalidates,
grants.json0600, no new crypto deps
Pillars 3+4 — Protocol Spec & Usage Metering (d8eb9db, a31740d)
- Public Protocol Spec v1.0.0 (
docs/protocol/adapter-and-acp-protocol.md): Adapter contract + ACP dispatch server (TCP, Ed25519 challenge-response, daemon wiring, CLI surface, 236 body anchors + 146 index entries verified) - Usage Metering Ledger:
UsageSource{Measured,Estimated,Unavailable}; exactly ONE meter owner (loom_telemetry::UsageMeter); "unavailable" printed when provider reports nothing; cost-only runs reported not measured; ACP v1usage_updatecorrectly labelled unstable
Phase B Contrast Experiment — NEGATIVE (3f4887e)
- 5 planted defects, 10 runs (base vs candidate), graded by crate's own test suite
- Result: candidate identical to base in 5/5 → signal ZERO → probe loop does NOT ship (Phases C/D/E not built)
- Spend: $0.00 / ~41 min (1.24M input / 86k output tokens) against $10-20 cap
- Second finding: surface hint read but not injected — Loom writes nothing to agent standing context
Other
- All independent review records on
mainwith RESOLVED banners: Phase A (fa29056), Pillar 1 (7457c72), Pillar 2 (e0edb50), Pillars 3+4 (d8eb9db,a31740d) - CI: 5-job matrix (test, container, integration, benchmarks, security)
cargo test --workspacegreen (excluding network-gatedacp_livetest); clippy clean; fmt clean- 31 crates, version 0.9.6 (next: 0.10.0)
Artifacts
- Tarball:
loom-0.10.0.tar.gz - sha256:
05907f187ceca1047807cee369587418e527a0fefebf6b022177b0d759681075
Loom 0.9.6
Added
- feat(acp):
loom acp registry [--available] [--json]lists all 42 agents in the public ACP registry, marking locally registered ones with* - feat(acp): ACP agents can now be launched with an
envmap (registryenvblocks are honored, e.g.FAST_AGENT_MODELforfast-agent)
Fixed
- fix(tui): the Kanban and Orchestrator tabs no longer trap tab navigation — both tab-specific key handlers ran before the global handler and swallowed
Tab/Shift-Tab/←/→, so the Orchestrator tab could not be left at all and the Kanban tab could not be switched away from. Kanban now usesh/lfor columns andj/kfor rows; the Orchestrator uses↑/↓for field focus and←/→for tool selection - fix(acp):
loom acp run <agent>now resolves the agent's real launch command — explicit path → localconfig/agents.tomlentry → public ACP registry →PATHbinary — instead of spawning a bare binary with no arguments. Agents were previously started outside ACP mode (e.g.opencodewithout itsacpflag) and timed out oninitialize - fix(acp):
loom add-acp-agentnow picks a usable distribution deterministically (npx>uvx>binary, resolved for the current platform) instead of an arbitrary first key of the registry distribution map, so agents likekiloandsigitkeep their package runner and theiracpmode flag - fix(acp):
uvxagents are resolvable — the daemon's agent launcher previously understood onlynpmandpiprunners - fix(acp): a JSON-RPC error returned by the agent during
session/promptis now reported asACP session/prompt failed: <error>instead of being masked asACP session/prompt timed out
Changed
- fix(homebrew): bottle releases are published under a
vX.Y.Z-bottlestag in the tap repo. The tap's plainvX.Y.Ztag namespace is permanently reserved by GitHub once a tag has been used by an immutable release and can never be recreated, which broke the 0.9.5 publish - chore:
cargo fmt --allacross the workspace — 8 files that had drifted out of rustfmt compliance (loom-routingoutcomes.rs/scorer.rs, loom-clicommands/routing.rs, loom-tuiapp.rs/ui/tabs/kanban.rs, loom-workflowtask_manager.rs, two loom-integration-tests)
Loom 0.9.5
Changelog
All notable changes to the Loom project will be documented in this file.
The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.
[0.9.5] - 2026-10-19
Added
- docs: architecture diagrams — static SVG (
docs/diagrams/architecture-export.svg) and interactive HTML (docs/diagrams/architecture.html), embedded in the README intro - docs: competitive analysis refresh — Orca (stablyai) tool-by-tool entry, Orca column in the feature matrix, shipped-status annotations across the P0/P1/P2 roadmap items (worktrees, containers, PR integration, remote SSH, Kanban/Graph/Cost tabs, ACP registry, cross-session recall all marked shipped)
- docs: man pages rewritten for 0.9.x —
manpages/loom.1covers all 36 subcommands (incl. cfg-gatedtui), the 12 adapters, the 15 TUI tabs, corrected hook event types (PreToolUse/PostToolUse/SessionStart/PostPhase) and honest environment-variable docs (LOOM_AGENTS_DIR, not the never-honoredLOOM_DIR);manpages/loomd.1documents the gRPC server, TLS env vars, ZeroMQ event socket, and the WSL2/Unix-socket constraint - docs:
docs/is no longer gitignored — analysis docs and diagrams are tracked; only the archify JSON intermediates (docs/diagrams/*.json) stay ignored
Changed
- docs: README rewritten with source-verified facts (12 adapters, 15 tabs, 31 crates, CHANGELOG-driven version history, working links) replacing the 0.3.x-era drift
- feat(tui): tab bar shows per-tab shortcut labels (
1–9,0,!,@,#) and scrolls to keep the active tab visible on narrow terminals - feat(tui): Kanban tab renders workflow phases (Plan → Implement → Verify → Finish) as columns with per-task selection; the custom scrollbar widget was removed in favor of native scrolling
Fixed
- fix(workflow):
git worktree addnow retries with--orphan(git ≥ 2.42) for repositories without any commit, and worktree isolation is best-effort — a non-git project directory or missing git no longer blocks task creation (the task runs in the project directory instead)
[0.9.4] - 2026-10-18
Added
- feat(routing): persisted decision + outcome telemetry (
crates/loom-routing/src/outcomes.rs) — SQLite store at~/.loom/telemetry/routing.dbrecords one row per task-bound decision (signals JSON, numeric, level, tier, model, token estimates) and closes it with an outcome:completedwhen the task reaches Finish,overriddenwhen a manual tier override rejects the heuristic's pick; newloom routing outcomes [--limit N]reports recent decisions and outcome counts, giving the decision-model flip condition its evidence surface - feat(core):
[security] require_signatureconfig flag (default off) — when enabled, unsigned hooks are skipped and unsigned plugins are rejected at install; documents the trust model explicitly (loom dir contents are user-owned local state) - feat(config):
config/security.tomloverlay joins the merge chain alongsiderouting.toml/mcp.toml/hooks.toml/acp.toml - feat(core):
derive_level(numeric, thresholds)andderive_tier(level)— one exported source of truth for complexity level and model tier derivation - docs: track
ROADMAP.md(previously untracked, referenced by the routing debt record)
Changed
- refactor(routing): the level/tier derivation duplicated across
scorer.rs,router.rs,task_manager.rs(×2) and the CLI now routes throughderive_level/derive_tier; configurablecomplexity_thresholdsgovern level as well as tier, making the Supplement's "these thresholds are configurable" claim true - refactor: dedupe cross-crate helpers into
loom-core, split workflow machinery, dependency hygiene - chore: workspace
cargo fmtand clippy 1.99 drift lints (result_large_err,useless_format, borrow inprintln) - chore:
__pycache__/gitignored; release-bottle script untracked (artifact dirs are gitignored)
Fixed
- fix(routing):
extract_signalscounted bytes, not chars, inflating the length signal ~3× for UTF-8 CJK prompts and biasing CJK tasks toward higher tiers - fix(core):
ComplexityThresholdsderivedDefaultwas{0.0, 0.0}rather than the serde defaults (0.35/0.7), so any routing config without a[complexity_thresholds]section treated every score as High → Flagship - fix(routing): the level-vs-tier contradiction (config
low_max = 0.3+numeric = 0.32yieldedlevel = Lowbuttier = Mid) is resolved by the unified derivation - fix(cli):
init'srouting.tomltemplate claimed threshold defaults0.33/0.66; corrected to the shipped0.35/0.7 - fix(routing-hook): the Python hook imported third-party
toml(ImportError on any machine withoutpip install toml); now uses stdlibtomllib(Python ≥ 3.11), with the no-op-on-error contract unchanged - ci: install
protocfor the integration and test-matrix jobs; fail cargo steps properly (set -o pipefail);--all-featuresbuilds no longer break on duplicate cfg-gated imports
Security
- fix(core): plugin downloads were HTTPS-only for remote config but accepted plain
http://for plugins —install_from_urlandupdate()now reject non-HTTPS URLs before any network I/O, closing an on-path tool-TOML swap - feat(security):
require_signaturetrusts nothing unsigned when enabled — hooks skip and plugin installs fail rather than running unverified code
Tests
- test(routing): characterization suite pinning hook and scorer behavior before the refactor — 7
routing_hook.pycontract tests (plain asserts,tempfilerules files, never~/.loom), user-override precedence, and the adversarial keyword ceiling (keyword weight 0.35 clamps at max 0.35 contribution, so keywords alone cannot force Flagship on the static path) - test(routing): boundary and contradiction coverage for
derive_level/derive_tier, including the strict-<edges atlow_max/high_min(previously only the 0.5 midpoint was tested) - test(core): HTTPS rejection on install and update, unsigned-plugin policy follows the flag, half-signed manifests rejected,
securityconfig parse and default - test(routing): outcome-telemetry store and full workflow flow (create → advance ×3 → override → Finish = 4 rows: 3
completed, 1overridden)
[0.9.3] - 2026-10-09
Fixed
- fix(loom-ipc): sync IPC handlers now run on a bare scoped thread instead of inside the async connection task — handlers that drive tokio runtimes internally (
Handle::block_on/Runtime::block_on) panicked with "Cannot start a runtime from within a runtime", breakingloom acp status/show/stop,loom hooks list, and tool availability checks (the CLI surfaced this asEOF while parsing a value at line 1 column 0); a panicking handler now returns an error response instead of dropping the connection
Changed
- chore: workspace clippy cleanup —
cargo clippy --workspace --all-targetsreports 0 warnings (was ~110): dead code removed (loomd capture.rs never-wired tracked-spawn chain, gRPC shutdown channel, loom-remotedefault_key_paths,ForgeManager.loom_dirfield), lint fixes (matches!,sort_by_key+Reverse,while let, needless pass-by-value, complex type aliases),#[allow(dead_code)]on deserialized-but-unread serde response fields,///!→//!doc-comment headers, integration tests no longer dropsave_manifestResults - chore(loomd, loom-tui): container runtime features (docker/podman/kubernetes) forwarded via
[features]instead of forced on every dependent build
[0.9.2] - 2026-10-06
Added
- feat(loom-auth): audit trail integrity verification — SHA-256 hash-chained audit log (
prev_hash/entry_hashonAuditEntry) withverify_integrity()tamper detection; chain survives restarts and log rotation; legacy unchained entries are tolerated
Loom v0.9.4
Routing debt closed. This release pays down the seven items from the internal routing debt record, plus the security hardening they surfaced.
Added
- feat(routing): persisted decision + outcome telemetry (
crates/loom-routing/src/outcomes.rs) — SQLite store at~/.loom/telemetry/routing.dbrecords one row per task-bound decision (signals JSON, numeric, level, tier, model, token estimates) and closes it with an outcome:completedwhen the task reaches Finish,overriddenwhen a manual tier override rejects the heuristic's pick; newloom routing outcomes [--limit N]reports recent decisions and outcome counts, giving the decision-model flip condition its evidence surface - feat(core):
[security] require_signatureconfig flag (default off) — when enabled, unsigned hooks are skipped and unsigned plugins are rejected at install; documents the trust model explicitly (loom dir contents are user-owned local state) - feat(config):
config/security.tomloverlay joins the merge chain alongsiderouting.toml/mcp.toml/hooks.toml/acp.toml - feat(core):
derive_level(numeric, thresholds)andderive_tier(level)— one exported source of truth for complexity level and model tier derivation - docs: track
ROADMAP.md(previously untracked, referenced by the routing debt record)
Changed
- refactor(routing): the level/tier derivation duplicated across
scorer.rs,router.rs,task_manager.rs(×2) and the CLI now routes throughderive_level/derive_tier; configurablecomplexity_thresholdsgovern level as well as tier, making the Supplement's "these thresholds are configurable" claim true - refactor: dedupe cross-crate helpers into
loom-core, split workflow machinery, dependency hygiene - chore: workspace
cargo fmtand clippy 1.99 drift lints (result_large_err,useless_format, borrow inprintln) - chore:
__pycache__/gitignored; release-bottle script untracked (artifact dirs are gitignored)
Fixed
- fix(routing):
extract_signalscounted bytes, not chars, inflating the length signal ~3× for UTF-8 CJK prompts and biasing CJK tasks toward higher tiers - fix(core):
ComplexityThresholdsderivedDefaultwas{0.0, 0.0}rather than the serde defaults (0.35/0.7), so any routing config without a[complexity_thresholds]section treated every score as High → Flagship - fix(routing): the level-vs-tier contradiction (config
low_max = 0.3+numeric = 0.32yieldedlevel = Lowbuttier = Mid) is resolved by the unified derivation - fix(cli):
init'srouting.tomltemplate claimed threshold defaults0.33/0.66; corrected to the shipped0.35/0.7 - fix(routing-hook): the Python hook imported third-party
toml(ImportError on any machine withoutpip install toml); now uses stdlibtomllib(Python ≥ 3.11), with the no-op-on-error contract unchanged - ci: install
protocfor the integration and test-matrix jobs; fail cargo steps properly (set -o pipefail);--all-featuresbuilds no longer break on duplicate cfg-gated imports
Security
- fix(core): plugin downloads were HTTPS-only for remote config but accepted plain
http://for plugins —install_from_urlandupdate()now reject non-HTTPS URLs before any network I/O, closing an on-path tool-TOML swap - feat(security):
require_signaturetrusts nothing unsigned when enabled — hooks skip and plugin installs fail rather than running unverified code
Tests
- test(routing): characterization suite pinning hook and scorer behavior before the refactor — 7
routing_hook.pycontract tests (plain asserts,tempfilerules files, never~/.loom), user-override precedence, and the adversarial keyword ceiling (keyword weight 0.35 clamps at max 0.35 contribution, so keywords alone cannot force Flagship on the static path) - test(routing): boundary and contradiction coverage for
derive_level/derive_tier, including the strict-<edges atlow_max/high_min(previously only the 0.5 midpoint was tested) - test(core): HTTPS rejection on install and update, unsigned-plugin policy follows the flag, half-signed manifests rejected,
securityconfig parse and default - test(routing): outcome-telemetry store and full workflow flow (create → advance ×3 → override → Finish = 4 rows: 3
completed, 1overridden)
Install: brew install tusharmewara/loom/loom (bottles attach below once the release workflow finishes).