v1.8.25 — real-transcript over-block fixes

cozempic v1.8.25 — real-transcript over-block fixes

A follow-up to v1.8.24 that closes the "verify the markers against a real session" gap. Ground-truthing the reload-gate matchers against an actual captured agent transcript surfaced a class of over-blocks where completed Agent work was read as still-live — the guard would then never reload, context would bloat, and native compaction would destroy team state (the exact failure cozempic prevents).

Fixed (all verified against the real transcript)

• A foreground Agent that already returned its result (real duration_ms usage trailer) is now treated as done — previously it stayed "running" forever, so any session that used Agent subagents went inert.
• A completed background Agent now clears: the real task-notification carries <tool-use-id>/<output-file> tags that the strict parser missed; fields are now parsed independently (order- and attribute-tolerant).
• detect_in_flight no longer fabricates a phantom in-flight task from a launch marker quoted in an agent's output or echoed in a normal Bash's stdout (a background Bash is credited only when it actually ran with run_in_background).
• Bare-name completions resolve to the suffixed agent id; terminal task statuses (finished/closed/resolved/…) are unified.

Tests

Redacted real harness fixtures are now committed under tests/fixtures/harness/, and the real-fixture assertions run (previously skipped). The false-SAFE direction (never mark live work terminal) was adversarially probed clean. Full suite: 1278 passed.

Upgrade

pip install -U cozempic      # or: uv tool upgrade cozempic
npm install -g cozempic
brew upgrade cozempic