tether-sdk 0.2.14__tar.gz → 0.2.16__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (266) hide show
  1. tether_sdk-0.2.16/.claude/gru/active/run-progress.md +29 -0
  2. tether_sdk-0.2.16/.claude/gru/active/why-qwen-fails.md +83 -0
  3. tether_sdk-0.2.16/.claude/gru/why-qwen-fails/reg_ab.sh +14 -0
  4. tether_sdk-0.2.16/.claude/gru/why-qwen-fails/regcount.py +17 -0
  5. tether_sdk-0.2.16/.claude/gru/why-qwen-fails/whack.py +44 -0
  6. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/AGENTS.md +4 -2
  7. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/PKG-INFO +12 -8
  8. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/README.md +11 -7
  9. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/history.jsonl +2 -0
  10. tether_sdk-0.2.16/bench/ladder/general.py +454 -0
  11. tether_sdk-0.2.16/bench/ladder/refs.py +32 -0
  12. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/report.py +1 -4
  13. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/run.py +65 -20
  14. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/tasks.py +26 -6
  15. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/validate.py +1 -1
  16. tether_sdk-0.2.16/bench/little_coder_recipe.md +139 -0
  17. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/run.py +36 -4
  18. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/architecture.md +38 -1
  19. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/benchmarks.md +8 -6
  20. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/index.md +12 -0
  21. tether_sdk-0.2.16/docs/ladder/claude-haiku.md +124 -0
  22. tether_sdk-0.2.16/docs/ladder/general.md +39 -0
  23. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/ladder/index.md +20 -3
  24. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/llms.txt +2 -0
  25. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/cli.md +5 -2
  26. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/providers.md +2 -2
  27. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/release-notes.md +48 -1
  28. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/troubleshooting.md +8 -0
  29. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/cli-and-headless.md +1 -1
  30. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_faults.py +22 -3
  31. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_limits.py +1 -1
  32. tether_sdk-0.2.16/e2e/test_notes.py +23 -0
  33. tether_sdk-0.2.16/e2e/test_progress.py +21 -0
  34. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/mkdocs.yml +2 -0
  35. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/coding_edit_failure_modes.snap.json +5 -5
  36. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/research_web.snap.json +7 -7
  37. tether_sdk-0.2.16/tests/providers/test_ollama.py +285 -0
  38. tether_sdk-0.2.16/tests/test_bench_littlecoder.py +183 -0
  39. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_context.py +1 -1
  40. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_context_config.py +2 -2
  41. tether_sdk-0.2.16/tests/test_failing_test_source.py +120 -0
  42. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_harness.py +61 -0
  43. tether_sdk-0.2.16/tests/test_ladder_general.py +155 -0
  44. tether_sdk-0.2.16/tests/test_ladder_run.py +141 -0
  45. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_params.py +13 -2
  46. tether_sdk-0.2.16/tests/test_progress.py +138 -0
  47. tether_sdk-0.2.16/tests/test_regression_note.py +111 -0
  48. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_repeat_reads.py +21 -2
  49. tether_sdk-0.2.16/tests/test_small_rewrite.py +98 -0
  50. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_tether.py +1 -1
  51. tether_sdk-0.2.16/tests/test_unchanged_write.py +99 -0
  52. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/__init__.py +1 -1
  53. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/cli.py +23 -3
  54. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/context.py +6 -4
  55. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/harness.py +130 -7
  56. tether_sdk-0.2.16/tether/progress.py +81 -0
  57. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/providers/__init__.py +2 -1
  58. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/providers/ollama.py +60 -28
  59. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/providers/openai.py +0 -1
  60. tether_sdk-0.2.16/tether/unchanged.py +140 -0
  61. tether_sdk-0.2.14/bench/ladder/refs.py +0 -15
  62. tether_sdk-0.2.14/tests/providers/test_ollama.py +0 -144
  63. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.claude/gru/active/.gitkeep +0 -0
  64. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.claude/gru/active/north-star.md +0 -0
  65. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.claude/gru/archive/.gitkeep +0 -0
  66. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.claude/gru/archive/v0-2-bugs.md +0 -0
  67. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.claude/gru/archive/work-on-v0-2-4.md +0 -0
  68. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.claude/gru.json +0 -0
  69. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.github/workflows/ci.yml +0 -0
  70. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.github/workflows/delivery.yml +0 -0
  71. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.github/workflows/docs.yml +0 -0
  72. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/.gitignore +0 -0
  73. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/CONTRIBUTING.md +0 -0
  74. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/LICENSE +0 -0
  75. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/README.md +0 -0
  76. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/canary.py +0 -0
  77. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/canary_baseline.json +0 -0
  78. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/compare.py +0 -0
  79. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/grade.py +0 -0
  80. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/harness-bench/README.md +0 -0
  81. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/harness-bench/tether-adapter.patch +0 -0
  82. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/base.py +0 -0
  83. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/baseline/harness.jsonl +0 -0
  84. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/baseline/ladder.jsonl +0 -0
  85. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/charts.py +0 -0
  86. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/probe.py +0 -0
  87. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/ladder/snapshot.py +0 -0
  88. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/naive_agent.py +0 -0
  89. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/report.py +0 -0
  90. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/goal.txt +0 -0
  91. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/hidden/test_hidden.py +0 -0
  92. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ref/calc/__init__.py +0 -0
  93. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ref/calc/errors.py +0 -0
  94. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ref/calc/evaluator.py +0 -0
  95. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ref/calc/parser.py +0 -0
  96. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ref/calc/tokenizer.py +0 -0
  97. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ws/calc/__init__.py +0 -0
  98. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ws/calc/errors.py +0 -0
  99. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ws/calc/evaluator.py +0 -0
  100. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ws/calc/parser.py +0 -0
  101. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ws/calc/tokenizer.py +0 -0
  102. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ws/tests/__init__.py +0 -0
  103. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-hard/ws/tests/test_calc.py +0 -0
  104. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/goal.txt +0 -0
  105. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/hidden/test_hidden.py +0 -0
  106. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/ws/cli.py +0 -0
  107. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/ws/ledger/__init__.py +0 -0
  108. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/ws/ledger/parser.py +0 -0
  109. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/ws/ledger/report.py +0 -0
  110. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/ws/sample.csv +0 -0
  111. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/ws/tests/__init__.py +0 -0
  112. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-medium/ws/tests/test_report.py +0 -0
  113. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-simple/goal.txt +0 -0
  114. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-simple/hidden/test_hidden.py +0 -0
  115. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-simple/ws/tests/__init__.py +0 -0
  116. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-simple/ws/tests/test_textutils.py +0 -0
  117. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/code-simple/ws/textutils.py +0 -0
  118. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/codereview/goal.txt +0 -0
  119. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/codereview/ws/inventory.py +0 -0
  120. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/content/goal.txt +0 -0
  121. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/content/ws/.keep +0 -0
  122. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/goal.txt +0 -0
  123. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/hidden/test_hidden.py +0 -0
  124. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/AGENTS.md +0 -0
  125. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/README.md +0 -0
  126. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/docs/architecture.md +0 -0
  127. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/docs/style.md +0 -0
  128. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/mini/__init__.py +0 -0
  129. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/mini/cli.py +0 -0
  130. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/mini/fmt.py +0 -0
  131. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/mini/stats.py +0 -0
  132. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/mini/textio.py +0 -0
  133. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/tests/__init__.py +0 -0
  134. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/focus/ws/tests/test_stats.py +0 -0
  135. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/plan/goal.txt +0 -0
  136. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/plan/ws/README.md +0 -0
  137. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/plan/ws/svc/__init__.py +0 -0
  138. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/plan/ws/svc/client.py +0 -0
  139. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/plan/ws/svc/config.py +0 -0
  140. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/qa/goal.txt +0 -0
  141. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/qa/ws/.keep +0 -0
  142. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/review/goal.txt +0 -0
  143. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/bench/tasks/review/ws/paper.md +0 -0
  144. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/extending.md +0 -0
  145. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/glossary.md +0 -0
  146. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/guides/a-pipeline-with-checks.md +0 -0
  147. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/guides/add-your-own-tool.md +0 -0
  148. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/guides/fix-a-failing-test.md +0 -0
  149. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/guides/use-tether-from-your-agent.md +0 -0
  150. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/javascripts/mathjax.js +0 -0
  151. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/ladder/img/harness-cost.svg +0 -0
  152. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/ladder/img/harness-pass.svg +0 -0
  153. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/ladder/img/task-grid.svg +0 -0
  154. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/ladder/img/tiers.svg +0 -0
  155. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/ladder/results.md +0 -0
  156. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/agent-interface.md +0 -0
  157. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/context-and-store.md +0 -0
  158. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/harness.md +0 -0
  159. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/index.md +0 -0
  160. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/multiplexer-and-decorator.md +0 -0
  161. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/sandbox-and-skills.md +0 -0
  162. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/reference/tools.md +0 -0
  163. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/agent-interface.md +0 -0
  164. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/built-in-tools.md +0 -0
  165. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/index.md +0 -0
  166. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/multi-agent.md +0 -0
  167. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/sdk-and-custom-tools.md +0 -0
  168. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/security-and-sandbox.md +0 -0
  169. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/skills.md +0 -0
  170. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/docs/tutorial/workflows-and-telemetry.md +0 -0
  171. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/__init__.py +0 -0
  172. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/conftest.py +0 -0
  173. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/proxy.py +0 -0
  174. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_agents.py +0 -0
  175. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_defaults.py +0 -0
  176. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_lifecycle.py +0 -0
  177. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_quirks.py +0 -0
  178. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_restart.py +0 -0
  179. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_sandbox.py +0 -0
  180. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/e2e/test_surface.py +0 -0
  181. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/README.md +0 -0
  182. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/checks.py +0 -0
  183. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/course.yaml +0 -0
  184. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/ruff.toml +0 -0
  185. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/skills/planner/SKILL.md +0 -0
  186. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/skills/searcher/SKILL.md +0 -0
  187. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/skills/searcher/tools.py +0 -0
  188. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/skills/writer/SKILL.md +0 -0
  189. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/examples/course/skills/writer/tools.py +0 -0
  190. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/pyproject.toml +0 -0
  191. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/scripts/record_golden.py +0 -0
  192. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/skills/researcher/SKILL.md +0 -0
  193. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/__init__.py +0 -0
  194. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/conftest.py +0 -0
  195. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/fixtures/dialects.jsonl +0 -0
  196. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/coding_edit_failure_modes.json +0 -0
  197. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/done_nudge.json +0 -0
  198. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/done_nudge.snap.json +0 -0
  199. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/read_range_pingpong.json +0 -0
  200. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/read_range_pingpong.snap.json +0 -0
  201. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/research_web.json +0 -0
  202. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/shell_wrappers.json +0 -0
  203. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/shell_wrappers.snap.json +0 -0
  204. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/waiting_input_resume.json +0 -0
  205. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/golden/waiting_input_resume.snap.json +0 -0
  206. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/providers/__init__.py +0 -0
  207. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/providers/test_anthropic.py +0 -0
  208. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/providers/test_base.py +0 -0
  209. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/providers/test_declarative.py +0 -0
  210. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/providers/test_dialects.py +0 -0
  211. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/providers/test_openai.py +0 -0
  212. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/providers/test_retry.py +0 -0
  213. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_agent_interface.py +0 -0
  214. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_benchmarks_docs.py +0 -0
  215. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_cli.py +0 -0
  216. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_cli_docs.py +0 -0
  217. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_decorator.py +0 -0
  218. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_docs_structure.py +0 -0
  219. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_edit_tool.py +0 -0
  220. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_golden.py +0 -0
  221. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_guides_docs.py +0 -0
  222. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_integration.py +0 -0
  223. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_llms_txt.py +0 -0
  224. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_misses.py +0 -0
  225. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_multiplexer.py +0 -0
  226. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_pipeline.py +0 -0
  227. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_providers_docs.py +0 -0
  228. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_quiet_and_fold.py +0 -0
  229. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_readme.py +0 -0
  230. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_release_hardening.py +0 -0
  231. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_restart.py +0 -0
  232. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_retro.py +0 -0
  233. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_sandbox.py +0 -0
  234. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_security_hardening.py +0 -0
  235. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_skills.py +0 -0
  236. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_store.py +0 -0
  237. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_structured_history.py +0 -0
  238. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/test_troubleshooting_docs.py +0 -0
  239. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/tools/__init__.py +0 -0
  240. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/tools/test_agent.py +0 -0
  241. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/tools/test_bash.py +0 -0
  242. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/tools/test_done.py +0 -0
  243. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/tools/test_edit.py +0 -0
  244. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/tools/test_mem.py +0 -0
  245. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tests/tools/test_web.py +0 -0
  246. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/api.py +0 -0
  247. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/decorator.py +0 -0
  248. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/mcp.py +0 -0
  249. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/multiplexer.py +0 -0
  250. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/pipeline.py +0 -0
  251. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/providers/anthropic.py +0 -0
  252. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/providers/base.py +0 -0
  253. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/providers/declarative.py +0 -0
  254. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/retro.py +0 -0
  255. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/sandbox.py +0 -0
  256. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/schema.py +0 -0
  257. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/skills.py +0 -0
  258. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/store.py +0 -0
  259. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/tools/__init__.py +0 -0
  260. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/tools/agent.py +0 -0
  261. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/tools/bash.py +0 -0
  262. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/tools/done.py +0 -0
  263. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/tools/edit.py +0 -0
  264. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/tools/mem.py +0 -0
  265. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/tether/tools/web.py +0 -0
  266. {tether_sdk-0.2.14 → tether_sdk-0.2.16}/uv.lock +0 -0
@@ -0,0 +1,29 @@
1
+ # Run progress feedback (--verbose + activity dots)
2
+ - Slug: run-progress
3
+ - Status: active
4
+ - Started: 2026-11-06 Updated: 2026-11-06 Closed: —
5
+ - Branches / PRs: claude/tether-local-setup-error-6fb430
6
+ - Budget: warn 300k / stop 400k paid tokens
7
+
8
+ ## Goal
9
+ Foreground `tether "<goal>"` is silent while the model loads/thinks. Add `--verbose` (one stderr line per step) and an always-on animated dots indicator (stderr, TTY only). Document both.
10
+
11
+ ## Requests
12
+ - 2026-11-06 PO: feedback on model loaded / waiting; --verbose if missing; always-visible three-dots indicator even without verbose; docs.
13
+
14
+ ## Decisions
15
+ - No `--verbose` exists today (grep cli.py: none). Plan: new tether/progress.py polling SQLite turns, no harness change.
16
+
17
+ ## Plan (task cards)
18
+ see PO preview
19
+
20
+ ## Ledger
21
+ | Task | Tier | Model | Harness | Session | Attempt | Tokens | Time | Result |
22
+ |---|---|---|---|---|---|---|---|---|
23
+
24
+ ## Progress log
25
+ - 2026-11-06 plan drafted, waiting for PO accept
26
+ - 2026-11-06 T1 tests committed 0619c04. T2 attempt 1 tether MCP nemotron ses_c5e4e75b: FAILED max_steps; MCP workspace is main checkout /home/walkdead/Projects/tether, not worktree; wrote progress.py there. Salvaged into worktree, removed stray files from main.
27
+ ## Tether friction
28
+ - tether_run MCP always roots in main repo checkout, ignores the session worktree; minion burned 15 steps looking for a test file that was only in the worktree (ses_c5e4e75b).
29
+ - 2026-11-06 T2 finished by gru (PO-informed: MCP cannot reach worktree, 12-line edit). T3 docs, T4 e2e done. unit 528 pass 100% cov, e2e 85 pass (409s), real TTY run OK. PR opened.
@@ -0,0 +1,83 @@
1
+ # Why qwen3.5:9b fails Polyglot
2
+ - Slug: why-qwen-fails
3
+ - Status: active
4
+ - Started: 2026-11-04 Updated: 2026-11-04 Closed: —
5
+ - Branches / PRs: fix/model-sampling-defaults
6
+ - Budget: Haiku minions $2.42
7
+
8
+ ## Goal
9
+ One reason that covers most qwen3.5:9b Polyglot fails (v0.2.13/v0.2.14) and one general change.
10
+
11
+ ## Findings
12
+ - 50 failing traces classified by 8 Haiku minions (codebook in make_cards.py, local only). Primary: spec_misread 35, long_thinking_wall 6, loop_no_progress 5, other 2, whack_a_mole 1, test_output_unseen 1. Any: spec_misread 42, loop_no_progress 29, long_thinking_wall 18.
13
+ - Wrong design at about step 4, never revised. Repairs: 354 whole-file rewrites, 0 targeted edits. Failing count after a rewrite: better 39%, same 49%, worse 12%.
14
+ - Passes: 33/72 green at the first test run, 60/72 within 3. Fresh short attempts (best-of-k) estimated flat (0.52 against 0.53): exercises are bimodal.
15
+ - Empty turns: 94 turns where prompt+completion hit 8192 with no command (3.9% of turns, 40% of failing-run model time). Thinking runaway.
16
+ - About 15 earlier loop-envelope changes (stall, nudge, fold) were flat on full runs (R2 ledger).
17
+
18
+ ## Sampling A/B (8 exercises, v0.2.14, 1 trial each)
19
+ | Arm | Runs | Pass | Overflow turns | Model time lost |
20
+ |---|---|---|---|---|
21
+ | temperature 0.2 (v0.2.10+v0.2.13 baseline) | 32 | 7 | 12.3% | 53% |
22
+ | 0.6 / top_p 0.95 / top_k 20 | 8 | 1 | 6.2% | 29% |
23
+ | 0.8 (Ollama default) | 8 | 3 | 5.4% | 28% |
24
+ Overflow halves. Pass flat within noise (4/16 against 7/32). Out dirs: north-star-c852b8/hosted-out/samp-{vendor,default}.
25
+
26
+ ## Decisions
27
+ - 2026-11-04: drop the hardcoded temperature 0.2 in the ollama and openai adapters (PO: "commit what you've done"). Fixes the thinking runaway; does not fix the main cause (first design wrong).
28
+
29
+ ## Round 2 (2026-11-04)
30
+ - e2e on fix/model-sampling-defaults: 80 passed.
31
+ - Room theory rejected: overflow is 10% even at 2-3k prompts (5k+ free); healthy command turns use p97 about 1.6-2.2k tokens. Thinking either ends short or runs away.
32
+ - Failure output hides the inputs: 1536 of 2304 traceback assert lines show only a variable (`total(basket)`), not the values. After a failing test run, failing runs re-read the test file next 27% of the time (passing 14%) and produce no command 17% (passing 6%); passing runs write the fix next 78% (failing 47%).
33
+ - R3 (Haiku, $0.14): feedback quality is the repair bottleneck (Olausson ICLR 2024), runtime values help (LDB ACL 2024), debugging decays after 2-3 tries (arXiv 2506.18403).
34
+ - Candidate: when a test fails, append the source of the failing test function (from the traceback frame in the workspace) to that observation. Bounded, 0 static tokens, no model call.
35
+
36
+ ## Next steps
37
+ - [ ] Full 34x2 run with the change for the north star line
38
+ - [ ] New idea aimed at the first design (research card R3)
39
+
40
+ ### 2026-11-04 thinking budget (T1)
41
+ - PO said "Yes, prepare it". Branch `feat/think-budget` (stacked on `feat/failing-test-source`).
42
+ - Verified on tether-ollama: `num_predict` counts thinking tokens; a cut reply has `done_reason: "length"` and empty content.
43
+ - Acceptance tests in `tests/providers/test_ollama.py` (5 new, 3 red before the change) + `tests/test_params.py` exact-options assert loosened.
44
+ - Card: `.claude/gru/why-qwen-fails/cards/T1.md` (owns `tether/providers/ollama.py`). THINK_BUDGET 2560; one think-off re-ask when cut with no command; caller `num_predict` (e.g. -1) disables it.
45
+ - src-ab arm B at 12/16 when written; book-store t0 passed in 150 s.
46
+ - T1 done by claude:haiku (1 attempt, $0.021, 18 turns; could not run uv in sandbox, verified by gru): 481 passed, cov 100%, ruff, mypy green; tests byte-identical. Commit `6cb036d` on `feat/think-budget` with docs (release notes, providers.md, AGENTS.md §D). e2e + A/B pending (GPU busy with src-ab).
47
+
48
+ ### 2026-11-04 src-ab result (arm B = sampling fix + failing test source, 16 runs)
49
+ | metric | arm A (samp-default+samp-vendor, 16) | arm B src-ab (16) |
50
+ |---|---|---|
51
+ | pass | 4/16 | 4/16 (book-store, go-counting, scale-generator, vlq t0) |
52
+ | overflow turns | 5.4% / 6.2% | 9.2% |
53
+ | lost model time | 28-29% | 33% |
54
+ | rewrite better/same/worse | 56/26/19% (n=43) | 37/45/18% (n=38) |
55
+ | after failing test, fail runs: write / read_test / none | 48 / 25 / 14% (n=64) | 43 / 17 / 23% (n=53) |
56
+ - Read: source shown -> fewer test re-reads (25->17%), but more empty turns right after a failure (14->23%). Pass flat; n=16 cannot see < ~30 pt. Not a win yet, not a clear loss. The empty-after-failure turns are what T1 targets.
57
+ - Started arm C `hosted-out/think-ab` (feat/think-budget, 6cb036d: sampling + source + budget), e2e first. Log /tmp/claude-1000/think_ab.log.
58
+
59
+ ### 2026-11-05 think-ab result (arm C = sampling + source + thinking budget, 16 runs), e2e 80/80
60
+ | metric | arm A (16) | arm B src-ab (16) | arm C think-ab (16) |
61
+ |---|---|---|---|
62
+ | pass | 4/16 | 4/16 | 6/16 (hangman t0+t1, book-store t0, vlq t0+t1, go-counting t0) |
63
+ | turns in 16 runs | 341 | 282 | 360 |
64
+ | empty window-filling turns | 20 | 26 | 1 |
65
+ | lost model time | 28-29% | 33% | 16% |
66
+ | after failing test (fail runs): none | 14% | 23% | 6% |
67
+ | rewrite better | 56% | 37% | 38% (n=61) |
68
+ - Budget does what it should: empty turns 26 -> 1, lost time halved, empty-after-failure 23 -> 6%. Pass +2 is inside noise at n=16.
69
+ - New limit: 7 of 10 failed runs end at the 30-step cap or 600 s, mostly near-passing (0.65-0.82). Rewrites still only 38% better: spec-misread remains the main cause.
70
+
71
+ ### 2026-11-05 round 3: regression note (G1)
72
+ - PO: "keep going, one small thing at a time, general, not over-fit to qwen".
73
+ - think-ab fails: 10/10 end at 30 steps or 600 s, plateau at the same failing count. whack.py over A+B+C: edits between test runs in failing runs fix+break 25% / break only 10% (passing runs 8% / 4%).
74
+ - Literature: Huang et al. ASE'26 (2607.01855): verification gate + rollback is the only mitigation that helps all 6 models (Llama-8B 31.6 -> 47.3%); recap alone small. Wang-Lin et al. 2026 (2609.10123): iterative fixing damages correct code, add/remove cycles.
75
+ - Step 1 (no rollback yet): `[regression]` line naming broken + newly passing tests vs the last run of the same suite. Branch `feat/test-regression-note`, tests `6f0af32`, impl `d0257c8` (claude:haiku, 1 attempt, $0.019; my integration test had a pyc same-size/same-second bug, fixed by gru). 487 passed, 100%.
76
+ - Started arm D `hosted-out/reg-ab` (main + note), e2e first. Watch: whack.py fix+break/break-only rate, then pass.
77
+ - If the rate drops but pass stays flat: step 2 = opt-in rollback to the best version (the paper's gate).
78
+
79
+ ### 2026-11-05 arm D result (reg-ab, 8 ex x2)
80
+ - e2e 80 passed. Pass 6/16 (= arm C). Note fired 18x in 9 runs (regcount.py).
81
+ - whack.py failing runs, C -> D: fix+break 40% -> 15%, break only 10% -> 18%, any break 50% -> 33%, fix only 16% -> 32%. Passing runs: fix+break 9% -> 28% (n=18-22 edits, noisy).
82
+ - 3 runs FAILED as worker_error (both book-store at step 5, connect t1 at step 16): Ollama 500. Cause: #228 think-off retry (ollama.py:200) calls raise_for_status without `rejected_call`, so a malformed tool call in the retry crashes the worker instead of becoming a hint. Arm C had 0 such errors (luck). Book-store passed 1/2 in C, so D lost about 1 pass to this bug.
83
+ - Next: PR the note (honest: pass flat, regressions down); separate fix PR for the retry 500 (test first, minion types).
@@ -0,0 +1,14 @@
1
+ #!/usr/bin/env bash
2
+ # reg_ab.sh: freeze the feat/test-regression-note build, run e2e, then arm D (no params) on the 8 exercises x2.
3
+ set -u
4
+ WT=/home/walkdead/Projects/tether/.claude/worktrees/describe-failures
5
+ OUT=/home/walkdead/Projects/tether/.claude/worktrees/north-star-c852b8/hosted-out
6
+ V=/tmp/claude-1000/-home-walkdead-Projects-tether--claude-worktrees-north-star-c852b8/b1bc7023-8e89-46ef-935a-c605f5578993/scratchpad/venv-reg
7
+ TASKS=poly-hangman,poly-scale-generator,poly-transpose,poly-connect,poly-rest-api,poly-book-store,poly-variable-length-quantity,poly-go-counting
8
+ cd "$WT"
9
+ uv venv -q "$V" && uv pip install -q --python "$V/bin/python" . || exit 1
10
+ "$V/bin/python" -c "import tether.harness as h; assert hasattr(h, 'regression_note')" || exit 1
11
+ echo "== e2e"; TETHER_E2E=1 uv run pytest e2e -q 2>&1 | tail -3
12
+ echo "== bench"
13
+ uv run python bench/run.py --runners tether --models ollama/qwen3.5:9b --tasks "$TASKS" --trials 2 --poly-verify \
14
+ --tether-bin "$V/bin/tether" --out "$OUT/reg-ab"
@@ -0,0 +1,17 @@
1
+ """Count [regression] notes per run: regcount.py <run_dir>..."""
2
+
3
+ import sqlite3
4
+ import sys
5
+ from pathlib import Path
6
+
7
+ for root in sys.argv[1:]:
8
+ total = 0
9
+ for db in sorted(Path(root).glob("*/t.db")):
10
+ n = (
11
+ sqlite3.connect(db)
12
+ .execute("SELECT count(*) FROM turns WHERE content LIKE '%[regression]%'")
13
+ .fetchone()[0]
14
+ )
15
+ total += n
16
+ print(f"{db.parent.name}: {n}")
17
+ print(f"{root}: total {total}")
@@ -0,0 +1,44 @@
1
+ """whack.py <run dir>...: between consecutive test runs, how often a change
2
+ fixes some tests and breaks others."""
3
+
4
+ import json
5
+ import re
6
+ import sqlite3
7
+ import sys
8
+ from collections import Counter
9
+ from pathlib import Path
10
+
11
+ FAILNAME = re.compile(r"^(?:FAIL|ERROR): (\w+)", re.M)
12
+ SUMMARY = re.compile(r"Ran (\d+) tests?.*?(OK|FAILED)", re.S)
13
+ c = Counter()
14
+ for base in map(Path, sys.argv[1:]):
15
+ for run in sorted(base.glob("tether__*")):
16
+ ok = json.loads((run / "done.json").read_text())["pass"]
17
+ db = sqlite3.connect(run / "t.db")
18
+ sets, pend = [], False
19
+ for role, cmd, content in db.execute(
20
+ "select role, tool_cmd, content from turns order by id"
21
+ ):
22
+ if role == "assistant":
23
+ pend = bool(re.search(r"unittest|pytest", cmd or ""))
24
+ elif role == "tool" and pend and SUMMARY.search(content or ""):
25
+ sets.append(frozenset(FAILNAME.findall(content or "")))
26
+ for a, b in zip(sets, sets[1:], strict=False):
27
+ fixed, broke = a - b, b - a
28
+ k = (
29
+ "fix+break"
30
+ if fixed and broke
31
+ else "fix only"
32
+ if fixed
33
+ else "break only"
34
+ if broke
35
+ else "no change"
36
+ )
37
+ c[(ok, k)] += 1
38
+ for ok in (True, False):
39
+ n = sum(v for (o, _), v in c.items() if o == ok)
40
+ print(
41
+ "PASS" if ok else "FAIL",
42
+ n,
43
+ {k: f"{v} ({v / max(1, n):.0%})" for (o, k), v in sorted(c.items()) if o == ok},
44
+ )
@@ -53,7 +53,7 @@ Reviewers and the Team Leader cite the failing line when they reject a change.
53
53
  ### Rule 3: Explicit Termination (`act("done", stdin="...")`), Interactive Inquiry (`done --ask`), & Pre-`done` Verification Gates
54
54
  - Sessions do **not** terminate simply because the model emitted a text response without a tool call (small 9B models frequently "think out loud" before acting).
55
55
  - A task only completes when the agent explicitly calls `act("done", stdin="<final deliverable>")` (or `act("done '<short answer>'")`), pauses for human/parent input via `act("done --ask", stdin="<question>")` (transitioning the session to `WAITING_INPUT` and releasing the worker process until resumed via `send`), or hits `max_steps` (handed back as `WAITING_INPUT`, `send` resumes it with a fresh budget). A bare `done` (no answer, no stdin) completes with the result `Done.` and still runs the `verify` gates; `done --ask` with no question is an error (#169).
56
- - If the model emits plain text without calling `act`, the harness appends a brief deterministic nudge reminding it to either execute a command or call `done`.
56
+ - If the model emits plain text without calling `act`, the harness appends a brief deterministic nudge reminding it to either execute a command or call `done`. A reply of 80 or more characters may be the answer: the nudge says it was not delivered and to pass it to `done`, and a session that still ends `WAITING_INPUT` (stuck or out of steps) carries the last such reply in its result, marked "never confirmed with `done`" (at most 2,000 chars, #173). Lookups by built-in tools (`web search|fetch`, `mem search|recall`, `agent status|wait|read|query`) count as reads, not changes.
57
57
  - After 6 consecutive read-only/verification steps following a change, the harness appends a `[NUDGE]` to that step's observation reminding the model to call `done` if the task is complete, repeated every 8 more quiet steps until the next change (0 static prompt tokens, #74, #132). Once per session, at 75% of `max_steps` (limits of 8 or more), a one-line `[NUDGE]` says how many steps are left (#133). Before any change, 12 read-only steps in a row (then every 8 more) append a one-line `[NUDGE]` to make the edit or call `done` (#127). Seven reads of one file in a row with no change are a `cycle_stall`, and 4 stall warnings without a change (#92) first get one `[WRAP UP]` turn (a read-only session is asked for its final answer, a session that already changed something is asked to call `done` with what is done and what is left, unless its `verify` gates or green test already finish it). Being stuck any way (the stall guard's warnings still unanswered after the wrap-up, or the same check command failing 3 times with an edit in between) restarts the session: one `restart` turn (event_type `restart`, at most 6000 chars: the files as the model left them, the last failing output, counts of what was tried; failed commands are not replayed) and the history continues from that turn alone. After 2 restarts, being stuck again hands the task back: the session ends `WAITING_INPUT` (exit code 2) with a `[STUCK ...]` summary as its result, so the caller can refine the request and `send` it. 0 static prompt tokens, no extra model call; `tether stop <id>` terminates a detached worker (#89). After the same check command (a test run) has failed 3 times with an edit in between, the harness adds one `restart` turn (event_type `restart`, at most 6000 chars: the files as the model left them, the last failing output, and counts of what was tried) and continues from that turn alone; failed commands are not replayed, at most one restart per session, and everything before it stays in SQLite (0 static prompt tokens, no extra model call).
58
58
  - When `verify` gate commands are configured (via `Harness(verify=[...])`, CLI `--verify`, or Skill YAML `verify: [...]`), `Harness.execute_session` deterministically executes each command inside the session's `SandboxPolicy` before allowing `done` to transition to `COMPLETED` (consuming 0 static prompt tokens until a gate fails). A gate that errors instead of failing (sandbox violation, exit 124/126/127) ends the session `FAILED` at once as `gate_error` ("gate misconfigured"), without counting toward the consecutive-failure abort. When the stall guard is about to restart or hand back a session and gates are configured, it runs them first: a change was made and every gate passes, so the session ends `COMPLETED` ("Finished by the harness"), since the deterministic gate is the completion criterion the caller asked for (#163). With no gates, the model's own last passing test run (runner-recognised, nothing changed since, still green when re-run) is the implicit gate, at a stall and at the step cap; a passing run after a change also appends a one-line `[NUDGE]` to call `done` (#177).
59
59
 
@@ -72,6 +72,7 @@ Reviewers and the Team Leader cite the failing line when they reject a change.
72
72
  [obs_3 folded: `web fetch https://...` (3,420 chars) | use `mem recall obs_3` to expand]
73
73
  ```
74
74
  - Exception: the latest read of each of the last 8 files (`edit view <p>`, `bash cat/head/tail/sed -n <p>`; paths normalized against the workspace) is not folded until `<p>` is edited (`edit write|replace`, or any other bash command naming it: `sed -i`, `tee`, redirects, `python -c`) or re-read. The head+tail elision still applies (an observation may use up to a third of `max_context_tokens`, at least 2,400 chars), and under `max_context_tokens` pressure it is folded like any other (oldest first). The `mem recall` hint is shown only when `mem` is enabled (else "re-run the command").
75
+ - When a recognised test command fails, the harness appends the source of up to 3 failing test functions named in the traceback (`[failing test source]`, workspace files only, at most 1,500 chars, skipped when the output already shows it) so the model sees the test's inputs (0 static prompt tokens, no model call, #173). When the same suite ran before (same test count) and tests that passed then fail now, one `[regression]` line names them and the newly passing ones (at most 5 names each, 400 chars, #173). When a command that writes existing workspace files (`edit write|replace <p>`, `> p`, `>> p`, `tee p`) leaves all of them with the same code (identical bytes, or for `.py` an identical AST), one `[NUDGE]` line says so (`tether/unchanged.py`, at most 200 chars, 0 static prompt tokens, #173). A whole-file rewrite (`edit write p`, `> p`) of a file of 20 or more lines that changed at most 20% of them gets one `[NUDGE]` pointing to `edit replace p`, once per file per session (#173).
75
76
  - History is replayed as structured `tool_calls` + `tool` results (providers with `structured_history = True`; textual `act('...')` form only as fallback), with `stdin` replayed as a real 400-char prefix, never a `<N chars>` placeholder. A call and its result are trimmed together.
76
77
 
77
78
  ### Rule 6: Deterministic Security Envelope (Never Trust the LLM for Safety)
@@ -183,7 +184,8 @@ tether/
183
184
 
184
185
  ---
185
186
  - **Timeout/retry (#70):** `providers.yaml` `timeout` applies to every kind; HTTP providers retry timeouts/network errors/429/5xx twice (`post_with_retry` in `providers/base.py`) then raise `ProviderError` naming provider, model and timeout; each attempt is also bounded by a total `asyncio.wait_for(timeout)` deadline (#88).
186
- - **Context config (#54):** `max_context_tokens` / `fold_observations_after` resolve as explicit argument (`--context-tokens` / `--fold-after`, MCP `context_tokens` / `fold_after`) > `providers.yaml` entry keys `context_tokens` / `fold_after` > defaults 8192 / 2.
187
+ - **Context config (#54):** `max_context_tokens` / `fold_observations_after` resolve as explicit argument (`--context-tokens` / `--fold-after`, MCP `context_tokens` / `fold_after`) > `providers.yaml` entry keys `context_tokens` / `fold_after` > defaults 32768 / 2 (32k keeps `qwen3.5:9b` fully on an 8 GB GPU and passed 7 of 18 wall-capped Polyglot runs vs 2 at 8k, #173).
188
+ - **Ollama thinking (#183, #173):** `think=false` when under 3072 tokens of `num_ctx` are free; a thinking turn gets `num_predict` 2560 and, if cut there with no command or with a half tool call Ollama rejects (HTTP 500), is asked once more with `think=false` (tokens summed; a second rejection is the resend hint). A caller's `think`/`num_predict` param wins.
187
189
 
188
190
  ## 4. Development Workflow & Commands
189
191
 
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.5
2
2
  Name: tether-sdk
3
- Version: 0.2.14
3
+ Version: 0.2.16
4
4
  Summary: General-purpose, headless-first LLM-Native micro-harness for small language models
5
5
  Project-URL: Homepage, https://github.com/igormcsouza/tether
6
6
  Project-URL: Repository, https://github.com/igormcsouza/tether
@@ -45,15 +45,15 @@ Tether is built the other way round:
45
45
 
46
46
  ## How well does it work
47
47
 
48
- Tether answers to one number: how many real coding exercises a small local model finishes. On the Polyglot-Python set (Aider's 34 Python exercises, graded by each exercise's own tests), `qwen3.5:9b` through Ollama at the default 8192-token context, 2 trials each (68 runs):
48
+ Tether answers to one number: how many real coding exercises a small local model finishes. On the Polyglot-Python set (Aider's 34 Python exercises, graded by each exercise's own tests), `qwen3.5:9b` through Ollama at an 8192-token context (v0.2.15; the default is now 32768), 2 trials each (68 runs):
49
49
 
50
- | | before v0.2.10 | v0.2.10 | v0.2.13 |
51
- |---|---|---|---|
52
- | Passes | 35 of 68 (51.5%) | 35 of 68 (51.5%) | 37 of 68 (54.4%) |
53
- | Mean score | 0.718 | 0.763 | 0.783 |
54
- | Runs killed by the wall cap | 11 | 14 | 13 |
50
+ | | before v0.2.10 | v0.2.10 | v0.2.13 | v0.2.15 |
51
+ |---|---|---|---|---|
52
+ | Passes | 35 of 68 (51.5%) | 35 of 68 (51.5%) | 37 of 68 (54.4%) | 40 of 68 (58.8%) |
53
+ | Mean score | 0.718 | 0.763 | 0.783 | 0.756 |
54
+ | Runs killed by the wall cap | 11 | 14 | 13 | 12 |
55
55
 
56
- The target is 80% by v0.3.0, so this is the number every change is ranked by. v0.2.13 is not distinguishable from v0.2.10 (paired score delta +0.020, 95% interval -0.060 to +0.101). One run is noisy (about 0.1), so read the pass count and the failure breakdown, not one decimal. The second yardstick is the [Coding Ladder](https://igormcsouza.github.io/tether/ladder/): 24 tasks, 6 tiers of difficulty times write, bugfix, plan and review, which shows how far up a model climbs and what stops it. Method, how to reproduce both and fair-comparison rules: [Benchmarks](https://igormcsouza.github.io/tether/benchmarks/).
56
+ The target is 80% by v0.3.0, so this is the number every change is ranked by. v0.2.15 passes 3 more runs than v0.2.13, but that is inside the noise too (pass flips +7/-4, paired score delta -0.026, 95% interval -0.093 to +0.041). One run is noisy (about 0.1), so read the pass count and the failure breakdown, not one decimal. The second yardstick is the [Coding Ladder](https://igormcsouza.github.io/tether/ladder/): 24 tasks, 6 tiers of difficulty times write, bugfix, plan and review, which shows how far up a model climbs and what stops it. Method, how to reproduce both and fair-comparison rules: [Benchmarks](https://igormcsouza.github.io/tether/benchmarks/).
57
57
 
58
58
  ## How it works
59
59
 
@@ -143,3 +143,7 @@ docs/ the user site (MkDocs)
143
143
  5. **Gumaan, E. (2026).** *Feedback That Backfires: Why Small Language Model Agents Repeat the Call They Just Watched Fail*. Source of the restart rule: after repeated failures, start a context that describes the state instead of replaying the failed attempts. [arXiv:2608.23651](https://arxiv.org/abs/2608.23651)
144
144
  6. **Kim, J., et al. (2026).** *Scaling Test-Time Compute for Agentic Coding*. A restart carries a summary of what was tried, not the transcript. [arXiv:2604.16529](https://arxiv.org/abs/2604.16529)
145
145
  7. **Yang, Z. (2026).** *Why Retrying Fails: Context Contamination in LLM Agent Pipelines*. Why a retry should not replay the failed context. [arXiv:2605.08563](https://arxiv.org/abs/2605.08563)
146
+ 8. **Olausson, T. X., Inala, J. P., Wang, C., Gao, J., & Solar-Lezama, A. (2024).** *Is Self-Repair a Silver Bullet for Code Generation?* In **ICLR 2024**. Repair is limited by feedback quality, the reason a failing test run shows the failing test's source. [arXiv:2306.09896](https://arxiv.org/abs/2306.09896)
147
+ 9. **Zhong, L., Wang, Z., & Shang, J. (2024).** *Debug like a Human: A Large Language Model Debugger via Verifying Runtime Execution Step-by-step (LDB)*. In **Findings of ACL 2024**. Showing the values behind a failure improves repair. [arXiv:2402.16906](https://arxiv.org/abs/2402.16906)
148
+ 10. **Huang, Y., et al. (2026).** *Regression Accumulation in Multi-Turn LLM Programming Conversations*. In **ASE 2026**. Later edits break earlier passing tests; checking against them helps every model, the reason a test run names the tests the last change broke. [arXiv:2607.01855](https://arxiv.org/abs/2607.01855)
149
+ 11. **Wang-Lin, X., Isopoussu, A., & Mahon, L. (2026).** *If It's Not Buggy, Don't Fix It: On the Dynamics of Iterative Bug-fixing with LLMs*. Repeated fixing falls into cycles that undo and redo the same change, the reason a rewrite that left the code as it was is named. [arXiv:2609.10123](https://arxiv.org/abs/2609.10123)
@@ -24,15 +24,15 @@ Tether is built the other way round:
24
24
 
25
25
  ## How well does it work
26
26
 
27
- Tether answers to one number: how many real coding exercises a small local model finishes. On the Polyglot-Python set (Aider's 34 Python exercises, graded by each exercise's own tests), `qwen3.5:9b` through Ollama at the default 8192-token context, 2 trials each (68 runs):
27
+ Tether answers to one number: how many real coding exercises a small local model finishes. On the Polyglot-Python set (Aider's 34 Python exercises, graded by each exercise's own tests), `qwen3.5:9b` through Ollama at an 8192-token context (v0.2.15; the default is now 32768), 2 trials each (68 runs):
28
28
 
29
- | | before v0.2.10 | v0.2.10 | v0.2.13 |
30
- |---|---|---|---|
31
- | Passes | 35 of 68 (51.5%) | 35 of 68 (51.5%) | 37 of 68 (54.4%) |
32
- | Mean score | 0.718 | 0.763 | 0.783 |
33
- | Runs killed by the wall cap | 11 | 14 | 13 |
29
+ | | before v0.2.10 | v0.2.10 | v0.2.13 | v0.2.15 |
30
+ |---|---|---|---|---|
31
+ | Passes | 35 of 68 (51.5%) | 35 of 68 (51.5%) | 37 of 68 (54.4%) | 40 of 68 (58.8%) |
32
+ | Mean score | 0.718 | 0.763 | 0.783 | 0.756 |
33
+ | Runs killed by the wall cap | 11 | 14 | 13 | 12 |
34
34
 
35
- The target is 80% by v0.3.0, so this is the number every change is ranked by. v0.2.13 is not distinguishable from v0.2.10 (paired score delta +0.020, 95% interval -0.060 to +0.101). One run is noisy (about 0.1), so read the pass count and the failure breakdown, not one decimal. The second yardstick is the [Coding Ladder](https://igormcsouza.github.io/tether/ladder/): 24 tasks, 6 tiers of difficulty times write, bugfix, plan and review, which shows how far up a model climbs and what stops it. Method, how to reproduce both and fair-comparison rules: [Benchmarks](https://igormcsouza.github.io/tether/benchmarks/).
35
+ The target is 80% by v0.3.0, so this is the number every change is ranked by. v0.2.15 passes 3 more runs than v0.2.13, but that is inside the noise too (pass flips +7/-4, paired score delta -0.026, 95% interval -0.093 to +0.041). One run is noisy (about 0.1), so read the pass count and the failure breakdown, not one decimal. The second yardstick is the [Coding Ladder](https://igormcsouza.github.io/tether/ladder/): 24 tasks, 6 tiers of difficulty times write, bugfix, plan and review, which shows how far up a model climbs and what stops it. Method, how to reproduce both and fair-comparison rules: [Benchmarks](https://igormcsouza.github.io/tether/benchmarks/).
36
36
 
37
37
  ## How it works
38
38
 
@@ -122,3 +122,7 @@ docs/ the user site (MkDocs)
122
122
  5. **Gumaan, E. (2026).** *Feedback That Backfires: Why Small Language Model Agents Repeat the Call They Just Watched Fail*. Source of the restart rule: after repeated failures, start a context that describes the state instead of replaying the failed attempts. [arXiv:2608.23651](https://arxiv.org/abs/2608.23651)
123
123
  6. **Kim, J., et al. (2026).** *Scaling Test-Time Compute for Agentic Coding*. A restart carries a summary of what was tried, not the transcript. [arXiv:2604.16529](https://arxiv.org/abs/2604.16529)
124
124
  7. **Yang, Z. (2026).** *Why Retrying Fails: Context Contamination in LLM Agent Pipelines*. Why a retry should not replay the failed context. [arXiv:2605.08563](https://arxiv.org/abs/2605.08563)
125
+ 8. **Olausson, T. X., Inala, J. P., Wang, C., Gao, J., & Solar-Lezama, A. (2024).** *Is Self-Repair a Silver Bullet for Code Generation?* In **ICLR 2024**. Repair is limited by feedback quality, the reason a failing test run shows the failing test's source. [arXiv:2306.09896](https://arxiv.org/abs/2306.09896)
126
+ 9. **Zhong, L., Wang, Z., & Shang, J. (2024).** *Debug like a Human: A Large Language Model Debugger via Verifying Runtime Execution Step-by-step (LDB)*. In **Findings of ACL 2024**. Showing the values behind a failure improves repair. [arXiv:2402.16906](https://arxiv.org/abs/2402.16906)
127
+ 10. **Huang, Y., et al. (2026).** *Regression Accumulation in Multi-Turn LLM Programming Conversations*. In **ASE 2026**. Later edits break earlier passing tests; checking against them helps every model, the reason a test run names the tests the last change broke. [arXiv:2607.01855](https://arxiv.org/abs/2607.01855)
128
+ 11. **Wang-Lin, X., Isopoussu, A., & Mahon, L. (2026).** *If It's Not Buggy, Don't Fix It: On the Dynamics of Iterative Bug-fixing with LLMs*. Repeated fixing falls into cycles that undo and redo the same change, the reason a rewrite that left the code as it was is named. [arXiv:2609.10123](https://arxiv.org/abs/2609.10123)
@@ -3,3 +3,5 @@
3
3
  {"date": "2026-10-27", "version": "naive_agent", "model": "ollama/qwen3.5:9b", "runner": "naive", "tasks": "polyglot-fast-12", "trials": 2, "context": 8192, "note": "no-harness ceiling reference", "runs": 24, "passed": 3, "pass_rate": 0.125, "mean_score": 0.243, "fail_max_steps": 0, "fail_wall_cap": 0, "mean_steps": 4.7, "mean_prompt_tokens": 8054, "mean_wall_s": 55}
4
4
  {"date": "2026-10-28", "version": "v0.2.10", "model": "ollama/qwen3.5:9b", "runner": "tether", "tasks": "polyglot-34", "trials": 2, "context": 8192, "note": "v0.2.10 full run, --poly-verify; flat vs baseline: 14 wall-cap runs still lose to 4-6k-token empty turns", "runs": 68, "passed": 35, "pass_rate": 0.515, "mean_score": 0.763, "fail_max_steps": 18, "fail_wall_cap": 14, "mean_steps": 17.9, "mean_prompt_tokens": 77722, "mean_wall_s": 281}
5
5
  {"date": "2026-11-03", "version": "v0.2.13", "model": "ollama/qwen3.5:9b", "runner": "tether", "tasks": "polyglot-34", "trials": 2, "context": 8192, "note": "v0.2.13 full run, --poly-verify, released wheel; flat vs v0.2.10 (paired score delta +0.020, CI -0.060..+0.101, pass flips +7/-5, p=0.77)", "runs": 68, "passed": 37, "pass_rate": 0.544, "mean_score": 0.783, "fail_max_steps": 18, "fail_wall_cap": 13, "mean_steps": 17.9, "mean_prompt_tokens": 77190, "mean_wall_s": 287}
6
+ {"date": "2026-11-04", "version": "v0.2.14", "model": "haiku (Claude CLI provider)", "runner": "tether", "tasks": "polyglot-34", "trials": 2, "context": 32000, "note": "v0.2.14 released wheel, full run, --poly-verify, Claude Haiku; every exercise passed in both trials, including the 11 that qwen3.5:9b never passes; no restart or hand-back fired (max 6 steps)", "runs": 68, "passed": 68, "pass_rate": 1.0, "mean_score": 1.0, "fail_max_steps": 0, "fail_wall_cap": 0, "mean_steps": 3.6, "mean_prompt_tokens": 17325, "mean_wall_s": 27}
7
+ {"date": "2026-11-06", "version": "v0.2.15", "model": "ollama/qwen3.5:9b", "runner": "tether", "tasks": "polyglot-34", "trials": 2, "context": 8192, "note": "v0.2.15 release candidate (main + #231-#233, same code as the tag), full run, --poly-verify; +3 vs v0.2.13 but inside the noise (pass flips +7/-4, p=0.55; paired score delta -0.026, CI -0.093..+0.041); 30 min less wall time", "runs": 68, "passed": 40, "pass_rate": 0.588, "mean_score": 0.756, "fail_max_steps": 16, "fail_wall_cap": 12, "mean_steps": 17.6, "mean_prompt_tokens": 75366, "mean_wall_s": 260}