feat: ~1M context floor for role picks; docs, tests, dead-sort removal #1

Merged

crooy merged 2 commits from context-floor-and-role-picks into main

2026-09-30 06:50:16 +02:00

crooy commented

2026-09-27 13:27:54 +02:00

Owner

Adds a real context-window constraint to the analyzer and records the rules the
2026-09-11 role cut was actually made on.

Tool

context_length from the models API was being fetched and thrown away; it is
now used. New --min-context (1m / 256k / raw int) filters before the
Pareto and knee maths, a Ctx column shows every row's window, and
min_context is echoed in --json.
Models absent from the models API have unknown context and count as below the
floor, so they drop out rather than silently topping the frontier.

Tests / CI

test_openrouter_pareto.py: offline self-check for the size helpers
(suffixes, case, floats, junk rejection, inclusive floor). Wired into the
smoke workflow, which stays network-free by design.

Docs + skill

The ~1M context floor for every role except smol.
The per-role axis rule: agency ranks slow/task; intelligence ranks
plan/advisor, with tau2 only a floor for the advisor since it reads
tool-use transcripts. Rank on marginal $ per intelligence point, and compare
in-family variants before paying a tier up.
A guard for the OMP config trap: a settings block mis-indented under another
key (e.g. inside retry:) is parsed as that key's child and silently ignored,
so the file reads correct while OMP uses defaults. Verify the effective value
with omp config get, never the file.
Corrected the sample advisor pick in the skill, which contradicted the
decision it was illustrating.

Cleanup: removed an unused re-sort in compute_pareto_frontier.

Verified: py_compile, test_openrouter_pareto.py, --help, a live
--json run, and -b hle --cost-metric task --min-context 1m all pass.

Adds a real context-window constraint to the analyzer and records the rules the 2026-09-11 role cut was actually made on. **Tool** - `context_length` from the models API was being fetched and thrown away; it is now used. New `--min-context` (`1m` / `256k` / raw int) filters **before** the Pareto and knee maths, a `Ctx` column shows every row's window, and `min_context` is echoed in `--json`. - Models absent from the models API have unknown context and count as below the floor, so they drop out rather than silently topping the frontier. **Tests / CI** - `test_openrouter_pareto.py`: offline self-check for the size helpers (suffixes, case, floats, junk rejection, inclusive floor). Wired into the smoke workflow, which stays network-free by design. **Docs + skill** - The ~1M context floor for every role except `smol`. - The per-role **axis rule**: agency ranks `slow`/`task`; intelligence ranks `plan`/`advisor`, with tau2 only a floor for the advisor since it reads tool-use transcripts. Rank on marginal \$ per intelligence point, and compare in-family variants before paying a tier up. - A guard for the OMP config trap: a settings block mis-indented under another key (e.g. inside `retry:`) is parsed as that key's child and silently ignored, so the file reads correct while OMP uses defaults. Verify the effective value with `omp config get`, never the file. - Corrected the sample `advisor` pick in the skill, which contradicted the decision it was illustrating. **Cleanup**: removed an unused re-sort in `compute_pareto_frontier`. **Verified**: `py_compile`, `test_openrouter_pareto.py`, `--help`, a live `--json` run, and `-b hle --cost-metric task --min-context 1m` all pass.

crooy added 1 commit

2026-09-27 13:27:54 +02:00

feat: ~1M context floor for role picks; keep the axe, drop the dead sort

smoke / smoke (push) Successful in 11s

Details

smoke / smoke (pull_request) Successful in 7s

Details

880d7d3455

- openrouter_pareto.py: use the models API context_length (was fetched and
  discarded), add --min-context (1m/256k/raw), a Ctx column in the table and
  min_context in --json; models missing from the API count as below the floor
  and are filtered out before the Pareto/knee maths
- test_openrouter_pareto.py: offline self-check for the size helpers, wired
  into the CI smoke workflow (no network)
- docs: the ~1M floor for every role except smol, the per-role axis rule
  (agency ranks slow/task, intelligence ranks plan/advisor with tau2 as a
  floor), rank on marginal $ per point, and a guard for the OMP config trap
  where a mis-indented settings block is silently ignored
- fix the sample advisor pick in the skill (it did not match the decision)
- drop an unused re-sort in compute_pareto_frontier

crooy added 1 commit

2026-09-27 13:31:19 +02:00

docs: remove references to private setup kit

smoke / smoke (push) Successful in 7s

Details

smoke / smoke (pull_request) Successful in 8s

Details

2bfaa88a78