v0.9.0

Fixed

• An epic is no longer failed over work Tempa itself terminated. A session did 182 turns of QA
  fixes, live-verified three features, started a twelve-minute test suite in the background and
  closed its turn saying it would report back once it finished. Under a headless one-shot CLI run,
  writing that sentence is how a session ends: the CLI exited 0, Tempa's own container teardown
  reclaimed 38 processes including the suite, and the result the round was waiting for never
  existed. The next round started over with no memory of the verification, did the same thing, and
  reclaimed 35 more — and two exit-0 rounds with no feature completed is exactly what
  implement_no_progress_rounds reads as an epic blocked on something outside itself. The epic was
  marked failed and the whole runner stopped, over an epic that was blocked on nothing.
  Tempa already had a guard for this shape, but it could not fire: background_tasks_terminated_hit
  is set only by recognising the backend CLI's own "Background tasks still running after Ns;
  terminating." message, which covers background work that CLI's harness tracks — not a shell the
  session started with its Bash tool's run_in_background, for which the CLI prints nothing and
  exits 0. Tempa's own kill path set nothing at all. The teardown is now recorded on _state, and a
  round that both promised to come back with a result AND had processes reclaimed earns the epic one
  extra round before the count is allowed to matter.
  Deliberately an allowance, not an exemption. Both signals are required because either alone is
  wrong in a direction that matters: leftover processes were reclaimed on 82% of exit-0 sessions
  in the workspace this came from (62 of 76), so a grace keyed on reclaim alone would just be
  implement_no_progress_rounds: 4 wearing a mechanism costume, and would make a genuinely blocked
  epic that happens to leave a vite server running effectively unfailable. The allowance is one
  round, is dropped the moment a feature completes, is cleared by --reset-failed, and is spent
  only in the failure arm — so the two fully automatic repairs beside it (the dependency reorder and
  the QA-state desync repair) still fire on exactly the round they always have.
• A halt no longer claims a blocker Tempa has no evidence for. The failure message asserted the
  epic was "very likely blocked on something outside this epic" and then conceded, one line later
  and verbatim, that "the session didn't name a specific epic it's blocked on". The claim is the
  part a human reads first, and it sent them hunting a dependency that did not exist. When no
  blocked_by_epic was named, the halt now states only what is known; when the round was one Tempa
  cut short, it says so in Tempa's own voice and quotes the session's words underneath rather than
  instead of them. That quoting is load-bearing: the incident round's closing message was the only
  record anywhere that three features had been finished, since the round ended before writing them
  into config.json.
• The next round is no longer told to redo the thing that killed the last one. last_round_note
  is handed to the following session as a "CLAIM TO CHECK … check it against the code as it stands
  now, first". Round 2 of the incident received round 1's "I'm now waiting for the full
  Configuration.Tests suite… I'll report back once it completes." under that header and did exactly
  as instructed: it re-ran the suite, in the background, for twelve minutes, and ended the same way.
  A note now travels with a last_round_note_kind, and an unfinished_check arrives under a header
  saying its result no longer exists and pointing the round at redoing the check narrowed, reading
  config.json first. Notes written before that field existed carry no kind and still arrive as
  claims — a real blocker must — so the same warning is merged into the default frame as well.
  --reset-failed keeps the kind with the note it labels, because a note that loses its label is
  silently promoted back to a claim.
• Tempa's own cleanup notice is no longer quoted back as an epic's blocker. _log_reclaimed
  appends its message into the session log after the agent's last word, so on any session that
  left processes running it is the tail — and the log-tail fallback would hand it back as the
  session's explanation, telling a human in the Halted panel that their epic was blocked on "Turn
  off Terminate leftover processes in Settings → Runs". The existing [Done] trim could not catch
  it, since that only fires when [Done] is the last line. The message is now built from the same
  constant the outcome layer matches, so the two cannot drift apart.
• A stalled round that left nothing quotable now still explains itself. blocked_reason was
  assigned unconditionally on that path, so a round with no captured prose and an unreadable log
  stored "" — and both the dashboard's Halted panel and tempa status render their block only
  when that field is truthy, so the epic became a bare red ✗ with no explanation and, since the
  retry hint rides on the same string, no way out anywhere on the page.
• An agent's closing words no longer carry its tool-call JSON. One assistant event can carry a
  text block and a tool_use block, joined with a newline, and the capture only rejected a readable
  that started with [. So prose plus a run_in_background call was stored whole and could reach
  blocked_reason, which is the entire content of the Halted panel. Lines are now filtered
  individually.

Changed

• The prompts tell a session to narrow a long run, not to background it or ask for more time.
  The system prompt has forbidden ending a turn to wait since 0.6.6, and the incident happened
  anyway in a session that complied with it: it asked for the foreground with a 600000 ms timeout
  and was answered "Command did not complete within its 600s timeout and was moved to the
  background"; its later, compliant until [ -s … ]; do sleep 5; done foreground wait was
  backgrounded too. An instruction the harness converts into the forbidden action is not an
  instruction. The implementation and QA prompts now carry a rule naming a reachable move — narrow
  the run until it fits — plus "write output to a file rather than piping it through tail", and
  the system prompt's contradictory "raise the timeout" sentence is corrected to match.
• A feature is recorded before the wider regression run, not after it. The incident round
  finished its entire allotted batch of three features and verified all of them, then went to run a
  full suite before updating config.json — so the round was scored as having achieved nothing.
  The rule now hangs the record off finishing and checking each feature, ahead of any wider run. The
  trade is deliberate and stated: a done a later suite disproves goes back through QA as
  require_fixing, one extra round, which is the right direction because an unrecorded feature is
  invisible while a wrongly-recorded one is correctable.
• Tempa warns when a backend's background-work ceiling disagrees with backend_background_wait_sec.
  That setting reaches the CLI as an environment default only, under a variable name the vendor owns,
  so if the name is ever renamed the setting silently stops doing anything and every message telling
  you to raise it becomes advice about a disconnected knob. Nothing else in the runner could notice.
  It has something to say already: every ceiling message in the workspace this came from reported
  Claude Code's own 600s default rather than Tempa's configured value.
• The background-terminated marker is now matched against the raw output line rather than the
  narrowed failure-marker text, so it is still recognised when the CLI reports it inside a result
  event with exit code 0.
• max_session_run is documented for what it is — a backstop against a run of consecutive
  zero-progress sessions, not a lifetime session cap. It was never reached in the incident
  workspace: one epic has 20 session logs on disk and a stored total_run of 2.