@hasna/loops 0.4.10 → 0.4.12

package/CHANGELOG.md

@@ -5,6 +5,99 @@ documented in this file. Version entries are generated from the
 conventional-commit git history; one commit maps to one released patch version
 unless noted.
 
+## 0.4.12 (2026-07-05)
+
+Drain throughput: `--max-active` is now a per-route ceiling instead of a
+store-wide one, drain dispatch spreads across subscription accounts by live load,
+and every step records the account it ran on.
+
+### Fixed
+
+- **Route admission — per-route `--max-active`:** the global admission count was
+  store-wide, so the busiest router's `--max-active` became a shared ceiling that
+  throttled every other router (average concurrency ~1.9 while the agent lane
+  allowed 12; ~253 items/24h deferred as `global active workflow limit reached`).
+  The global count is now scoped to the route/drain that set the limit; scope
+  precedence is an explicit `--max-active-scope <key>` first, then the running
+  loop's `LOOPS_LOOP_NAME`, then the route key as the final fallback — so each
+  router's `--max-active` is its own ceiling.
+  Existing routers get correct scoping with no config change. `project` and
+  `project_group` counts are unchanged: they remain cross-route anti-hog caps.
+- **Per-account attribution:** Codewith agent steps carry their subscription
+  account in `authProfile`, not an `AccountRef`, so `workflow_step_runs.account_profile`
+  was NULL for every drain worker and per-account burn could not be attributed.
+  The resolved auth profile (and provider as `account_tool`) is now recorded on
+  each step at run creation, and the per-role assignment is surfaced in drain
+  reports.
+
+### Added
+
+- **Least-loaded auth-profile pool selection + `--max-per-profile`:** pooled
+  Codewith dispatch picked a member by a pure hash of the work item, which stacked
+  several concurrent workers on one account at high concurrency and tripped
+  provider-side 429/stream-drop limits. Live drains now count running steps per
+  account and place each route on the least-loaded pool member (verifier/planner
+  still land on a different account than the worker; the hash is the deterministic
+  tie-break, so a cold store reproduces the prior assignment). New
+  `--max-per-profile <K>` (default 2 for pools of two or more, `0` disables) defers
+  a route when every pool account already has `K` running steps.
+
+Schema: additive sqlite migration `0008_work_item_route_scope` (nullable
+`workflow_work_items.route_scope`; its index is created only by the migration —
+never in the baseline DDL, which re-runs on every open and would crash pre-0008
+databases). `SCHEMA_USER_VERSION` is intentionally not bumped so an older binary
+can still open a database this migration touched (rollback-safe). Postgres gets
+the equivalent additive migration `0004_work_item_route_scope`; the released,
+checksummed `0002_workflows_goals` block is untouched so existing postgres
+ledgers keep verifying. Upgrade coverage: a real pre-0008 (0.4.11 schema)
+fixture database must open cleanly, gain the column + index, and keep its data;
+released postgres migration checksums are pinned in tests.
+
+## 0.4.11 (2026-07-05)
+
+Scheduler fairness and PR-route hygiene: fast command loops no longer starve
+behind long agent workers, PR tasks dedupe by GitHub identity, and merged/closed
+PR routes close themselves out of the queue.
+
+### Fixed
+
+- **Daemon scheduler — separated concurrency lanes:** the daemon shared a single
+  concurrency pool, so long agent/workflow workers (minutes to over an hour) could
+  occupy every slot and starve fast command loops (monitors, digests, syncs) —
+  even the merge router starved. Command-target and agent/workflow-target loops
+  now draw from independent claim budgets. New env `LOOPS_DAEMON_COMMAND_CONCURRENCY`
+  (default 4) and `LOOPS_DAEMON_AGENT_CONCURRENCY` (default 8); the legacy
+  `LOOPS_DAEMON_CONCURRENCY` / `--concurrency` still work as the agent-lane knob for
+  back-compat, and new `--command-concurrency` / `--agent-concurrency` daemon flags
+  are added. A saturated lane no longer consumes the other lane's budget.
+- **Todos-task routing — PR fingerprint dedupe:** the repos registry maps several
+  local checkouts to one GitHub repo, so a single PR was minted as 2-3 duplicate
+  todos tasks (distinct ids + checkout paths) that each dispatched a full worker.
+  PR-subject tasks now dedupe by a stable `owner/repo#number` fingerprint (from an
+  explicit PR fingerprint field or a canonical PR URL / `github-pr:` handle,
+  lowercased) instead of the `(source-path, task-id)` key, collapsing the
+  duplicates to one work item. Non-PR tasks keep the path/id key (no false dedupe).
+- **Todos-task routing — freshness skip closes the task:** 0.4.10's freshness gate
+  stopped dispatching a worker for a merged/closed PR but left the task pending +
+  route-opted-in, so every drain tick re-skipped it forever. On a definitive
+  MERGED/CLOSED freshness skip the drain now marks the source task done and strips
+  its `auto:route`/`route:enabled` opt-in so it leaves the queue. Only fires when
+  gh/metadata definitively reports MERGED/CLOSED; never on dry-run.
+
+## 0.4.10 (2026-07-05)
+
+Hotfix bundle for the PR-merge drain pipeline.
+
+### Fixed
+
+- **Todos-task drain:** re-admit a terminal work item whose todos task is still
+  actionable (bounded by a redispatch cap + per-attempt backoff) instead of
+  deduping it away forever, so real task work keeps dispatching.
+- **PR review gate:** freshness gate skips merge/review routes for already
+  merged/closed PRs; normalize `app/<slug>` and `<slug>[bot]` bot logins so
+  bot-authored PRs resolve an author instead of hard-failing the format check.
+- **Runtime/store:** cap persisted run stdout/stderr; project path resolution fix.
+
 ## 0.4.9 (2026-07-04)
 
 Unblock the PR-merge pipeline: task-lifecycle/route workers now dispatch real

package/README.md

@@ -594,8 +594,15 @@ This is the intended deterministic-to-agentic path: a producer creates a todos
 task, `@hasna/events` delivers `task.created`, OpenLoops records the invocation
 and admission item, OpenLoops creates a worker/verifier workflow when admitted,
 and the workflow updates todos with evidence. Use account pools so worker and
-verifier steps do not burn the same profile; OpenLoops picks deterministically
-and uses a different verifier profile when the pool has at least two entries.
+verifier steps do not burn the same profile. For live Codewith drains, OpenLoops
+spreads each route to the **least-loaded** pool account (counting running steps
+per account) and still keeps the verifier on a different profile than the worker;
+the deterministic hash is the tie-break, so a cold store reproduces the previous
+assignment exactly. `--max-per-profile <K>` (default 2 for pools of two or more,
+`0` to disable) defers a route when every pool account already has `K` running
+steps — the guardrail that keeps higher concurrency from stacking on one
+subscription and tripping provider-side rate limits. The chosen account per role
+is recorded on each step (`account_profile`) and surfaced in drain reports.
 Use `--dry-run` to inspect the rendered invocation, work item, workflow, and
 loop input without storing anything.
 
@@ -645,11 +652,21 @@ loops routes schedule todos-task oss-task-route-drain \
   --max-active-per-project 1 \
   --max-active-per-project-group 4 \
   --max-active 12 \
+  --max-per-profile 2 \
   --worktree-mode required \
   --evidence-dir "$HOME/.hasna/loops/reports/oss-task-route-drain" \
   --compact
 ```
 
+`--max-active` counts active routed workflows **per route**. Scope precedence:
+an explicit `--max-active-scope <key>` wins, else the running loop's
+`LOOPS_LOOP_NAME`, else the route key — so each router's limit is its own
+ceiling rather than a store-wide total shared by every router.
+`--max-active-per-project` and `--max-active-per-project-group` remain cross-route
+anti-hog caps counted over all routes. Raise a router's `--max-active`
+deliberately once counting is per-route; keep `--max-per-profile` set so the
+extra concurrency spreads across subscription accounts.
+
 Only route tasks that explicitly opt in with `auto:route`, `route_enabled=true`,
 or `automation.allowed=true`. Use Codewith account pools, required worktrees,
 max-active throttles, and bounded evidence directories. Do not dispatch or paste

package/dist/api/index.js

@@ -3,7 +3,7 @@
 // package.json
 var package_default = {
   name: "@hasna/loops",
-  version: "0.4.10",
+  version: "0.4.12",
   description: "Persistent local loop and workflow runner for deterministic commands and headless AI coding agents",
   type: "module",
   main: "dist/index.js",