Phonton CLI v0.22.1

phonton-cli v0.22.1 — Diffs that land

Every change below was found by running benchmarks/smoke and checked with a
before/after run of the same tasks. DeepSeek flash (deepseek-flash),
Windows 11, release builds, hidden acceptance tests, 3 runs per task.

Results

All 12 smoke tasks on three fixtures, 0.22.1:

Fixture	Accepted	Tokens	Cost
todo-api (4 small files)	12/12	45,494	$0.022
ledger (one 339-line module)	12/12	126,091	$0.076
shop (one 28 KB module)	12/12	130,857	$0.088

Small n, one model, one machine; token use varies run to run. Reproduce with
node benchmarks/smoke/run.mjs --arm cloud --runs 3.

Fewer wasted retries

Models often miscount the @@ -start,count ranges in a unified diff even when
every line they quote is right. Phonton used to reject those hunks and pay for
another attempt. Now it recomputes the counts from the hunk body and, when the
stated line is wrong, moves the hunk to where its unchanged and removed lines
match the current file. A hunk that matches nowhere still fails verification.

• Direct prompts in Phonton's worker format: 6 of 13 diffs applied as sent,
  13 of 13 after recounting and relocating.
• ledger tasks, 0.22.0 vs this fix: 12/12 accepted in both; 218,139 vs
  108,591 tokens, $0.172 vs $0.063.

Large files

Files over 24 KB used to reach the worker as signatures only, so edits deep
inside a big module were guessed. Workers now get line windows around the
names the subtask mentions and the symbols retrieval picked. On the shop
tasks (measured before the planner fix below): 11/12 vs 12/12 accepted,
422,615 vs 203,489 tokens. Excerpts add about 3k tokens to an edit near the
top of a function; they remove the runaway retries on edits deeper in.

Planner

• A goal containing "release" (as in "release the stock") was treated as a
  broad, multi-agent goal and split into four or five LLM-planned subtasks.
  It is now planned like any focused change: shop-cancel-restock was split
  in 5 of 6 earlier runs (32k-214k tokens each); now its median is 4.9k.
• The LLM decomposer asks for the fewest subtasks that cover a goal and adds
  test subtasks only when the goal asks for tests.
• Removed plan preflight special cases for goals mentioning chess and for a
  src/receipt.js file; both were tuned to old benchmark fixtures.

Local models in one command

phonton models catalog                  # models that fit this machine
phonton models setup qwen2.5-coder:3b   # runtime, download, calibrate, select

setup MODEL stops at the first step that fails, for example when the machine
is too short on free memory to calibrate. Plain phonton models setup still
installs only the runtime and now prints the next step. Download progress no
longer prints ? / ? bytes for stages without a size.

Install

npm install -g phonton-cli
phonton doctor