v5.41.0

Added

• The ai-content wrap step's 5-minute wait budget now scales up for a project carrying many
  build plans (the proxy for "how heavy is this repo's next-session-prime scan"), up to a
  4x ceiling, instead of a fixed cap that a heavy repository could run close to on legitimate
  work alone. A step that still takes over 80% of its own (possibly scaled) budget before
  finishing now logs a warning and marks its result output.warning, so the wrap drawer shows
  "completed with warnings" and names the plan count — the same signal whether the wrap is
  watched live in the browser or read from the server log afterward (#1782).

Changed

• priming-roll's plan resolution no longer falls back to the legacy .claude/plans directory: with the Operator's explicit authorization, all plan files across the 12 local projects that still used it (RentalClaw-Project, TangleBrain-Builder, WhitePapers, ScrapeGoat, Notse, Medusa, TangleClaw-Builder1, Volta, Kobold, PV-AI-Guidebook, UCI, Monad-1) were migrated to .tangleclaw/plans, and lib/wrap-steps/priming-roll.js's _resolvePlansDir now checks only the TC-owned location. lib/plan-docs.js still serves a plan from the legacy location for any project this migration did not cover (it is scoped to the 12 named projects, not every TangleClaw-managed one) — only wrap-time plan resolution retired the fallback (#1050).

Fixed

• A stale wrap run that lets POST /api/server/restart proceed, or lets the
  wake monitor resume nudging a project, now logs a one-shot warning naming the
  project, run id and last step — previously neither action left any trace
  (#1321).
• A timed-out wrap-step command no longer leaves its child processes running: execShell/execFileArgs
  (lib/exec.js) now spawn the command detached in its own process group and kill the whole group
  (SIGTERM, then SIGKILL after a short grace period) on timeout, instead of signalling only the
  shell Node held a handle to. The returned promise now also waits for the group to be CONFIRMED gone
  (polled, not just signalled) before resolving, so a descendant that ignores SIGTERM can no longer
  outlive the point where the caller is told the command has stopped — and if that confirmation itself
  cannot be verified within a bounded window, the result now says so explicitly (groupKillConfirmed: false, plus a caveat in the error text) instead of silently claiming a clean stop. A command is also
  now correctly reported as timed out when it traps the SIGTERM this sends and answers with its own
  exit() (0 or otherwise) — previously indistinguishable from an ordinary success or failure once the
  signal itself left no trace in the exit code. The same group kill now also applies when a command is
  cut off for producing too much output (maxBuffer): a backgrounded job sharing the group could
  otherwise outlive the result, or — if the shell itself was deferring the signal while it had an
  outstanding job of its own — turn what should be a fast overflow guard into a hang tied to that job's
  own lifetime (#1503).
• priming-roll's shipped-plan filter now also applies when a project's plans directory holds exactly one .md file: a lone plan whose cited issues are all closed is reported as an archive candidate and skipped instead of being rolled silently, matching the filter that already applied when several plans coexist (#1522).
• publishHandoff now re-validates the staged handoff file's digest inside its BEGIN IMMEDIATE
  transaction, matching recordPromotedHandoff, instead of checking it before the transaction opens —
  closing the window where a file replaced between that check and the rename would have been published
  (#1651).
• The handoff-stage wrap step now mirrors .prawduct/.handoff-notes.md into a TangleClaw-owned
  snapshot (.tangleclaw/handoff-notes-continuity.md) at the end of every wrap, and the Claude
  SessionStart prime hook recovers from that snapshot when the live file is already gone or empty by
  the time a session boots — so prawduct's forward notes survive destructive consumption of the live
  file instead of leaving the next session without them. The mirror is also kept from outliving its
  relevance: a wrap that ends with no live notes to carry clears any stale earlier mirror, and the
  SessionStart hook deletes the mirror once it has shown it, so recovered notes are replayed at most
  once rather than at every boot until the next wrap happens to overwrite them (#1733).
• priming-roll no longer blocks a wrap on a stale activePlan pointer in .tangleclaw/project.json: a pin that is missing, resolves into a plans directory's archive/, or names a now-fully-done plan is cleared and resolution falls through as if it had never been set, so a one-shot pick from the multi-plan picker stops silently poisoning later wraps. An explicit step.planPath resolving into archive/ is now rejected the same way a missing file is. A genuine ambiguity (two or more still-live in-progress plans) still blocks for the operator to pick, unchanged (#1757).
• A stale wrap run that has not yet started committing is still taken over
  automatically when a new wrap starts — but the displaced pipeline is now
  told to stop before it can branch, commit or push, closing the window where
  two pipelines ran for one project. A stale run already past that point (it
  may already be committing, branching or pushing) refuses the new wrap
  instead, with stale: true, until that pipeline actually finishes or the
  server restarts: no request can make two concurrent git mutations safe
  (#1805).
• A Medusa session listener's still-unhandled inbox is no longer lost when a wrap ends the
  session: forgetSession now captures it before the workspace id is retired, and the
  project's next listener is seeded with it, oldest first, so mail that arrived while a wrap
  was running (and so was never nudged into the pane) survives into the successor session
  instead of being stranded against an identity nobody will ever reconnect as (#1806).
• A wrap run cut off by a server restart is now durably recorded: GET /wrap/status reports interrupted: true with the step the run last reached
  and whether it had started committing, instead of silently reading as no run
  at all (#1816). The session page's own display of this is tracked
  separately (#2642).
• injectCommand now refuses to type into a pane whose project has a wrap running in
  wrap-run-registry, the same way it already refuses a pane showing an engine startup
  dialog, answering HTTP 409 WRAP_RUNNING. medusa-wake and workload-nudge already
  checked the registry themselves before calling in; this closes the gap for every other
  caller of the command route — a coordinator clearing a Builder, an Observer's restart
  prompt — which previously had no such check (#2269).

Internal

• resolveSessionRange/resolveSessionRangeAsync (lib/wrap-steps/_git-range.js) now log a warning
  naming the stale or orphaned boundary whenever a recorded lastWrapSha or launchSha was given
  but the session range still fell back to the trunk branch — previously silent, so an operator
  debugging "why didn't the changelog-coverage gate engage for my session?" had no signal that a
  downgrade happened. A project's genuine first wrap (nothing recorded at all) still logs nothing;
  the two existing callers see no change to the resolver's return shape (#670).