fix: session isolation + await cursor + cleanup

[marksverdhei]

Summary

• Fixes Session state not persisted across cf subcommands #2: get_session_id() now uses PPID (parent shell PID) instead of own SID when no TTY is available — session state persists across cf subcommands in non-TTY contexts (agents, subshells, bash -c)
• Adds LASTLINE cursor tracking so await doesn't return stale/already-seen messages
• Sanitizes newlines in send to preserve one-message-one-line invariant
• Improves name uniqueness check to also match [name joined] format
• Fixes ((attempts++)) crash under set -e on first collision
• Uses process substitution for await/send-await to prevent orphaned tail -f processes
• Adds delete-room command, global room support (-g), auto-resolve .Chatfile extension
• Adds test_bugs.sh regression tests (8 bug scenarios, all pass)
• Removes scratch files (sketch.py, PROMPT.md)

Test plan

• test_cf.sh — 32/32 pass
• test_bugs.sh — 0/8 bugs reproduced (all fixed)
• Manual reproduction of issue Session state not persisted across cf subcommands #2 in non-TTY context — confirmed fixed

🤖 Generated with Claude Code

[marksverdhei]

Reviewed during hivemind triage. Status:

Tests still green on the PR branch as-is:

• test_cf.sh → 32/32 pass
• test_bugs.sh → 0/8 bugs reproduced

But the PR has rotted and now conflicts with main:

• PR has CONFLICTING / DIRTY mergeable state.
• Conflict is in cf (the core shell script). PR branch's cf is 353 lines; main's is 291.
• Main has moved on with fix: persist global room sessions in $HOME/.cf_session #3 ("fix: persist global room sessions in $HOME/.cf_session, closes Session state not persisted across cf subcommands #2") and test: 6 additional tests for global room, arg validation, messaging, status #4 (more tests for global rooms) — both landed after this PR. Issue Session state not persisted across cf subcommands #2 (the headliner here) is already closed by fix: persist global room sessions in $HOME/.cf_session #3.

Unique-to-#1 work that didn't make it into #3/#4 (from the body):

• LASTLINE cursor tracking on await
• newline sanitization in send
• name-uniqueness also matches [name joined]
• ((attempts++)) set-e crash fix
• process substitution for await/send-await
• delete-room command, -g global flag, auto-resolve .Chatfile

A rebase needs hand-resolving the cf conflict against what #3 landed — couldn't do that confidently as an idle pass since it'd require deciding which side of the overlap wins. Leaving for your call: rebase + reland the unique bits, or close and re-open targeted PRs for the still-wanted pieces.

[marksverdhei]

Closing as superseded — this can no longer merge cleanly and would regress main.

Since this PR was opened (2026-03-05), main reworked the same areas via a different (and cleaner) path:

• Issue Session state not persisted across cf subcommands #2 (session persistence) was fixed by fix: persist global room sessions in $HOME/.cf_session #3 (fix: persist global room sessions in $HOME/.cf_session (#3, closes #2)) — a directory/global-room approach rather than this PR's PPID-based session id.
• Global rooms (-g), delete-room, and the double-.Chatfile fix all landed on main independently.
• A test suite + CI workflow landed via test: 6 additional tests for global room, arg validation, messaging, status #4/Add pure-function tests for presentation.py helpers (13 tests) #5/ci: add tests workflow (bash + pytest) #6.

As a result this PR's diff vs current main would delete .github/workflows/tests.yml (−21) and tests/test_presentation_helpers.py (−173) and revert test_cf.sh (−106) — i.e. it would remove CI and tests main now relies on. So it's not mergeable as-is.

Salvaged the one unambiguous, still-applicable fix → #7: the ((attempts++)) retry-loop crash under set -e (register dies on the first name collision). Verified + regression-tested there.

Still-valid follow-up candidates from this PR, for whoever picks them up (each is independent and small):

• send / send-await: sanitize embedded newlines to preserve the one-message-one-line invariant (msg="${1//$'\n'/ }").
• await / send-await: the tail -f … | head -n1 can orphan the tail -f until the next write — process substitution avoids it.
• await: a read-cursor (LASTLINE) so it doesn't return an already-seen last line.

Thanks — the diagnostics here were good; closing in favor of the rebased salvage.