v0.1.2

Published to npm as dsh-repeat-tool-breaker@0.1.2 — https://www.npmjs.com/package/dsh-repeat-tool-breaker

Changed

• Publishing now uses npm Trusted Publishing (OIDC) — no long-lived npm token.
  The release workflow authenticates with the short-lived OIDC token GitHub mints
  for the run, and provenance is generated automatically.
• The workflow upgrades the npm CLI before publishing, because Trusted Publishing
  requires npm >= 11.5.1 while Node 22 bundles an older one.
• The workflow still honours a repository NPM_TOKEN secret as a fallback.
• README: npm badge, install instructions for the published package, and corrected
  advisory-tier documentation.

Verified

• 6/6 published files are sha256-identical to this tag's tree.
• Provenance attestation records this repository and
  .github/workflows/publish.yml as the build source.
• Installing 0.1.2 from the public registry and exercising the gate reproduces
  the documented behaviour: first identical call allowed, second blocked with
  REPEAT_TOOL_BLOCKED, a different path allowed.

v0.1.1

Bug-fix release found in code review.

Fixed

• The advisory no longer fires on ordinary calls. The warnAfter notice was
  keyed only on count === warnAfter; because every fresh tool call starts a new
  run at count 1, the default warnAfter: 1 attached a misleading
  "you just ran X with identical arguments 1 time(s)" message to every tool
  call. The advisory is now inert unless 2 <= warnAfter < denyAfter, the default
  is raised to 2, and the wording describes a real repeat.
• Caller-provided config arrays are copied before being frozen.
• Removed a dead denyAfter overflow check; post-execute computes its signature
  once instead of twice.

Tests

20 -> 29 assertions (advisory reachability, no-advisory regressions on fresh
calls, fail-loud config, caller-array ownership). CI green on Node 20 and 22.

v0.1.0

Initial release.

A dependency-free DeepSeek Harness (DSH) plugin that hard-blocks an agent's
repeated identical tool calls through the synchronous monotonic
ctx.tools.guard gate. The denied call never executes: the model receives a
REPEAT_TOOL_BLOCKED tool result that quotes the previous successful result and
tells it to advance or conclude instead of retrying.

Highlights

• Denies the 2nd identical (tool name + canonically-equal arguments,
  property order ignored) call per agent by default.
• Same-path read cap (maxSamePath) so read-like tools re-reading one file with
  varying arguments are still bounded.
• Per-agent WeakMap state; a new user message clears the chain via
  agent/pre-step; excluded tools (default todo_write) neither count nor reset.
• Advisory warnAfter notice via tools/post-execute — never a deny, and it
  never increments the chain (counting lives only in the guard).
• Fail-loud configuration validation at plugin load.

Verification

• Deterministic guard-logic suite (20 assertions) runs with no model or endpoint:
  npm test.
• CI green on Node 20 and 22.
• Loads and applies on a real DSH boot against @deepseek-ai/dsh 0.1.2-rc.1:
  apply runs with ctx.tools.guard reachable after inject: ['tools'].

Not included: an end-to-end model-driven trajectory (see the README's
"Scope and verification status").