fix(tui_gateway): manual /compress keeps the session workspace in the stored prompt, so the next resume hits the prompt cache

[teknium1]

A manual /compress in the TUI/Desktop no longer persists a system prompt pointing at the backend's home directory, so the next resume reuses it instead of rebuilding it and missing the prompt cache a second time.

• tui_gateway/session_compression.py::_compress_session_history binds the session context and cwd (_set_session_context(key, cwd=_session_cwd(session))) around compress_now, then clears it. This matches the other RPC-thread sites that persist or inspect the prompt (server.py::_persist_live_session_system_prompt, methods_slash.py, model_switch.py).
• All three manual-compress routes go through this one function: the session.compress RPC, the command.dispatch compress branch, and the slash.exec mirror.
• New invariant test tests/tui_gateway/test_compress_session_cwd.py. The work done at the compaction boundary must resolve the session workspace, and the binding must not leak past the call. It is red on origin/main (it resolves the process cwd) and green with the fix.

Root cause: session.compress runs on the RPC thread, which has no session context. So the system-prompt rebuild at the compaction boundary resolved the backend's cwd (the process HOME) and stored it. Every later process (a gateway resume or hermes chat --resume) rejected that stored prompt as stale runtime and rebuilt it. That caused a second full cache miss right after the sanctioned compaction miss, and on the oneshot path the tools pin was skipped too.

Live repro: E2E tests/e2e/core/history/test_prefix_stability.py -k tui_gateway_restarts drives real python -m tui_gateway.entry processes, one durable session, with the cwd flipping between hops.

• before (origin/main 07646a7): request 9 (in hop 4 (gw in work-b)): messages[0] (system) changed; first diverging byte at 6382: '…/home\nScratch directory…' vs '…/work-a\nScratch directory…', 1 failed
• after (this branch): 1 passed. The only prefix break is the sanctioned compaction boundary.

The E2E file isn't part of this PR. It lands separately with the core E2E-suite PR. I copied it in temporarily for the red/green run above.

check	result
test_compress_session_cwd.py on origin/main	failed (/…/test_manual_compress_runs_unde0 != /…/workspace)
test_compress_session_cwd.py with fix	passed
scripts/run_tests.sh tests/tui_gateway/	1395 passed, 0 failed. test_tui_gateway_server.py hit the per-file 300s cap at load ~90, and re-run alone (with both compress files) it gave 627 passed. One unrelated timing flake in test_hosted_room_driver_runtime.py passed on retry
ruff check, check_no_tmp_literals.py, git diff --check	clean

Duplicate sweep: no open or closed PR binds the session cwd around manual compression. I searched compress cwd, session.compress, compress_now session context, compress system prompt cwd stale, compress cache miss resume, _compress_session_history, session cwd compress tui and stale runtime system prompt rebuild.

• Related, but a different command: fix(init): target the session's active directory, not the process cwd #96378 (/init targets the session's active directory instead of the process cwd). It's the same class (an RPC-thread handler resolving the process cwd), but a different call site. Neither PR supersedes the other.

Infographic

[teknium1]

PR Review — #120158 (P0 monitor pass)

Premise (verified): tui_gateway/session_compression.py::_compress_session_history ran compress_now on the RPC thread with no session context; the compaction-boundary system-prompt rebuild resolves cwd via agent.runtime_cwd.resolve_agent_cwd(), which fell through to the process cwd. The session.compress RPC (methods_session.py::_compress_live) binds nothing before calling it; slash.exec binds the key but not the cwd. Sibling RPC-thread sites already bind exactly this way: methods_slash.py:158, model_switch.py:383, server.py:1104, methods_connectors.py:231.

Change: 6 LOC, one try/finally around compress_now with _set_session_context(key, cwd=_session_cwd(session)) / _clear_session_context. ContextVar tokens, so the nested bind inside slash.exec restores the outer values on exit. All three manual-compress routes go through this one function.

Tests: tests/tui_gateway/test_compress_session_cwd.py red on origin/main 33a30fd (resolve_agent_cwd() = process cwd), green with the branch's session_compression.py; with it, test_compress_lock_skip, test_compression_config_hot_reload, test_compute_host_late_compress_ack, tests/agent/test_conversation_compression_manual.py: 41 passed. The test is a real invariant (cwd seen by the compressor + no leak past the call), not a change-detector.

Non-blocking: the nested bind passes no ui_session_id, so it reads "" for the duration of the summary call and is restored afterwards; nothing on the compress path consumes it today. If a later change routes an approval or emit through it, pass ui_session_id through from _compress_live as methods_connectors.py:231 does.

Looks good: fixed at the shared choke point, matches the established pattern, cache impact measured end to end (E2E prefix-stability run in the body). Merge decision stays with the maintainer.