TinyTitan 5.18

TinyTitan 5.18 — the launcher installs what you ask for

This release makes the first run the easy one: a model that is not on disk can be fetched from the
launcher, the menu offers the models this checkout supports but does not have, and every long
download, conversion and repack draws a real percentage instead of a silent terminal.

What is new

• A model you ask for is downloaded and converted, not refused. tools/server_launcher.sh --model katcoder --bits 4 on a checkout without it used to end in an error and a pointer at
  another tool. It now asks once (a 20–220 GB download is not something to start unannounced), runs
  tools/install_models.sh with the catalogue key the installer is addressed by, re-reads the
  catalog and starts normally. --install answers ahead of time. A piped or --dry-run
  invocation never downloads: it prints the tools/install_models.sh <key> command instead, which
  is what keeps CI and scripted use safe. Checked by benchmark/test_launcher_install.py —
  fourteen tests, including that neither a pipe nor a dry run changes models/, and that every key
  the new stem→key map returns is one the installer's CATALOGUE carries — and by a live dry run on
  this checkout.
• The menu offers what you do not have yet. Run the launcher with no arguments and it lists the
  installed models, then the supported ones that are absent, marked not installed in the same
  columns (engine, size, thinking levels) so they can be compared rather than guessed at. Picking
  one fetches it and then takes the ordinary path, so nothing downstream knows where the bytes came
  from. Checked by the same suite: the offer rows, the Rows N-M are not installed yet hint,
  choosing one, declining one, and the regression guard that Enter on the default row still launches
  an installed model.
• Downloads, conversions and repacks report a percentage. A ~70 GB MoE fetch or a ~360 GB 125B
  fetch runs for hours, and the converters used to print nothing between finished shards. One line
  now covers it — rewritten in place on a terminal (with an ETA from the observed rate), one plain
  line per 10% in a log or a pipe, TINYTITAN_NO_PROGRESS=1 to silence it — and the same format is
  used by TinyTitanRepack, whose ModelInstallProgress stream existed but was consumed by nobody:
• converting 38.7/91.6 GB 42% 4/13 shards eta 12m installing 12.3/38.1 GB 32% eta 4m10s
• Checked by benchmark/test_progress.py (twelve tests: both output modes, the ETA, the shared
  12.3/38.1 GB unit rule, a closed pipe, a disabled line), the converter suites (51 tests
  together), and RepackCLITests (six, including a new local-import case that asserts a real CLI
  run reports progress to a pipe).
• The DeepSeek Harness plugin keeps a long session going. dsh-tinytitan gains three switches a
  profile opts into: autoGoal (a direct prompt arms a goal, so the harness keeps working until the
  model completes it), autonomy (the preset policy plus the fresh-agent ralph loop) and
  handoff (an unfinished objective moves to a fresh child context when the budget is spent, with a
  wall-recovery path for a turn that ends on max-tokens). The handoff budget is derived from the
  routed model's declared window — min(0.6 × window, window − 131,072), which is 131,072 on 256K,
  600,000 on a 1,000,000-token route and 629,145 on 1,048,576 — rather than a constant. The shipped
  bundle patch writes all three as false: each starts work nobody asked for, so a profile names
  the field it wants. Checked by the plugin suite (130 tests, 129 pass and 1 skip) and the live
  chain recorded in plugins/dsh-tinytitan/README.md.

Also in this release

• Two memory defaults in the documentation were wrong and are fixed: the bootstrap cap is 60
  records (not "forty"), and the guard ships on (TINYTITAN_MEMORY_GUARD=1) rather than off.
  The wiki's agent-memory table gained the guard and consolidation rows and lost two stale cells.
• Two engineering analyses landed as repository documents: docs/clickhouse-long-session- verdict.md (a columnar store does not belong in the memory or prompt path; it belongs in the
  evaluation/telemetry plane) and docs/plan-fact-keeping-over-sessions.md (the seven-step program
  to keep facts over long sessions, each step tied to the metric it must move).
• CI runs the new suites: test_launcher_install and test_progress joined the dependency-free
  Python step.

Performance

• No timing was measured on this commit.
• Nothing here touches the engine's hot path: the Swift change is confined to the TinyTitanRepack
  tool's output, and the rest is the launcher, two Python converters and the DSH plugin.

Verification

• tools/lint.sh — all eleven gates clean.
• swift test --no-parallel — 1,528 tests in 234 suites.
• 7 of the 16 stored golden baselines compared byte-identical (qwen35-4b-4, qwen35-4b-8,
  qwen35-9b-4, qwen35-9b-8, qwen36-4, qwen36-8, qwen38-4), then a clean scratch release
  build with a clean warning scan.
• Not checked, because their install is not under models/ and nothing may be fetched to change
  that: ornith-4, ornith-8, qwen38-8, agentworld-4, agentworld-8, katcoder-4,
  katcoder-8, qwen35-2b-4, qwen35-2b-8.

Checksum

tinytitan-5.18-macos-arm64.tar.gz sha256: 0804f25f0903b76e3b32b144618f3b2b8d217c565b6d701f61f1e4f80709d2fa
tinytitan-5.18-macos-arm64.tar.gz size: 15882928 bytes
tinytitan-lib-5.18-macos-arm64.tar.gz sha256: 50a0f155ff2ad373817464cb5b0d4e075312c1a568b7e8f2562db8955b9a45ba
tinytitan-lib-5.18-macos-arm64.tar.gz size: 17033718 bytes