dsh-effort-pilot v1.0.3

Compatibility and robustness release.

Fixed

• effort_status failed with value is not lossless JSON on routes whose current effort is adapter-defaulted (reported while testing other models). Optional fields are now omitted instead of being emitted as undefined; a regression test covers it.

Added

• README "Model support": the pilot works with any route whose adapter declares two or more reasoning efforts; single-level routes and toggle-only models are left untouched.

Install

dsh plugin --profile web add dsh-effort-pilot

dsh-effort-pilot v1.0.2

Quality-focused release: measured speed, cost and task quality.

Added

• BENCHMARK.md quality A/B: a failing-test fix task and a blind-reviewed code-review task, at max / auto / low, two samples per arm.
• Result: fix pass rate identical (6/6 in all runs, test file untouched); blind judge overall scores max 9.0 / auto 9.0 / low 10.0 — no measurable quality difference at n=2.
• Auto is not a one-way ratchet: it lifted low -> high when work turned mixed.

Install

dsh plugin --profile web add github:atlasmarkz/dsh-effort-pilot

npm

Published as dsh-effort-pilot@1.0.2 — install with dsh plugin --profile web add dsh-effort-pilot.

dsh-effort-pilot v1.0.1

Measured benchmark release.

Added

• BENCHMARK.md: A/B results on a live DSH 0.2.2-alpha.2 host (OpenCode Go, DeepSeek V4.1 Flash), methodology and reproduction steps.

Changed

• README cost claim replaced with measured numbers: max → low cut output tokens ~36%, decode ~38% and wall clock ~33% on a reasoning task; auto reached low mid-task on a mechanical task (high -> low). Absolute saving is 0.2–0.6¢ per task on V4.1 Flash (more on pricier models), not a fixed 2¢.

Install

dsh plugin --profile web add github:atlasmarkz/dsh-effort-pilot

dsh-effort-pilot v1.0.0

First stable release of dsh-effort-pilot, a per-step thinking-effort pilot for DeepSeek Harness.

Highlights

• Per-step scheduling from the session tool-round history: simple rounds down, failures/heavy work up, hold window anti-flapping.
• set_effort / effort_status agent tools; the model can change its own depth with an audit reason, or hand control back with auto.
• Cache-safe: writes model/selection with the same provider/model and only a new reasoningEffort; never switches models.
• Failure-code-gated escalation (EMPTY_RESPONSE, MALFORMED_RESPONSE, SERVER, TIMEOUT); rate-limit/auth/quota/transport codes are excluded.
• 41 unit tests; zero runtime dependencies; no build step.

Compatibility

DSH 0.2.x plugin surface, verified live on 0.2.2-alpha.2: bundle composes, loader reports include:effort-pilot active, and both tool schemas pass DSH's own assertSupportedJsonSchema.

Install

dsh plugin --profile web add github:atlasmarkz/dsh-effort-pilot
# or the prebuilt tarball attached to this release