Skip to content

claude-hooks v1.2.0

Choose a tag to compare

View all tags
@mann1x mann1x released this 09 May 19:42
· 861 commits to main since this release
v1.2.0
4ba3c85

[1.2.0] — 2026-05-09

MINOR bump for the caliber-grounding-proxy cloud-resilience
layer — a new opt-in retry subsystem visible to any caliber init
run against a flapping cloud Ollama. Also ships the v1.2 of the
/consultants benchmark protocol (Q3 actionability sub-rubric, first
confirmed heterogeneous PROD-READY label) and the first caliber-eval
cohort published in-repo (six labels graded against the claude-cli
reference).

Added

  • caliber-grounding-proxy cloud-resilience retry layer
    (claude_hooks/caliber_proxy/ollama.py, shared
    claude_hooks/_chat_retry.py). Two parallel retry budgets
    protect every chat completion to upstream Ollama:

    • 15-attempt HTTP/network budget with exponential backoff
      (base 1.5 s, cap 90 s, ≈ 15 min total). Catches 408 / 429 / 500 / 502 / 503 / 504 plus a curated list of retryable 4xx
      body substrings (the same set the consultants engine
      already proved against kimi-k2.6:cloud flapping).
    • 5-attempt empty-content budget for 200 OK responses
      with empty content, no tool_calls, and
      finish_reason ≠ length — the "throat-clearing" pattern
      every cloud-tagged Ollama model exhibits on heavy initial
      prompts.

    Tunable via env vars: CALIBER_PROXY_RETRY_MAX_ATTEMPTS,
    CALIBER_PROXY_RETRY_BASE_DELAY_S, CALIBER_PROXY_RETRY_MAX_DELAY_S,
    CALIBER_PROXY_EMPTY_RETRY_MAX. Defaults match the consultants
    engine, so behavior is consistent across both cloud paths.

  • FlapCounters exposed at /health.upstream_flaps. Five
    process-scoped counters surfaced in the /health JSON:
    upstream_5xx_total, upstream_retryable_4xx_total,
    upstream_empty_total, upstream_retry_succeeded_total,
    upstream_retry_exhausted_total. Lets claude-hooks-rollup
    (and any operator dashboard) detect cloud-weather degradation
    before it fails a bench.

  • Generic tool-call passthrough on the proxy round-trip.
    Provider extras like Gemini's thought_signature are now
    preserved verbatim across both legs of the round-trip instead
    of being stripped — the earlier targeted strip broke
    gemini-3-flash-preview:cloud with
    400 missing thought_signature in functionCall parts. A small
    denylist (function.index for deepseek/qwen; empty at top
    level) handles the inverse case where an upstream field would
    confuse the OpenAI-compat client. Net effect: every cloud model
    that ships a custom tool-call extra works without per-model
    patches.

  • caliber-eval cohort published at
    docs/caliber-eval-results/ —
    six labels graded against the claude-cli reference:
    gemma-native-tools-v3, gemma4-31b-cloud,
    gemini-3-flash-preview-cloud, deepseek-v4-flash-cloud,
    glm-5-1-cloud. Each label ships its score.py JSON + a
    narrative summary comparing to baseline. The workbench (full
    rsynced workspaces, run logs, fake-HOMEs) stays off-repo at
    /srv/dev-disk-by-label-opt/dev/caliber-eval/ per
    PROTOCOL.md, which documents the reproduce + publish recipe.

  • docs/caliber-eval.md — in-repo entry-point pointing at the
    off-repo workbench and the published-results dir.

  • docs/PLAN-caliber-proxy-cloud-resilience.md —
    implementation plan that drove the resilience port (marked
    "shipped 2026-05-09").

  • 27 new tests at tests/test_caliber_proxy_retry.py
    covering the decision helpers (is_retryable_status,
    is_retryable_empty_response, compute_backoff) and an
    end-to-end mocked httpx harness exercising both budgets.

  • Updated tests/test_caliber_proxy.py with passthrough
    coverage: test_assistant_tool_calls_passthrough_unknown_fields,
    test_assistant_tool_calls_function_index_stripped,
    test_response_tool_call_extras_passthrough,
    test_response_tool_call_preserves_upstream_id,
    test_round_trip_preserves_provider_extras.

Fixed

  • consultants synthesizer no longer silently synthesizes over a
    failure tombstone.     The SYNTHESIZER prompts now refuse to render
    a coherent answer when an upstream role marked the section as
    failed, surfacing the failure in the final synthesis instead.
    Caught by the v1.2 protocol's Q3 actionability sub-rubric.

Documentation

  • /consultants benchmark sweeps — 2026-05-09 cloud screening
    (7 new models), N=3 aggregate runs of the 3 PROD-READY
    candidates, Q3 actionability re-grade against protocol v1.2,
    per-role recommendation refresh, and the first confirmed
    heterogeneous PROD-READY label
    (mix-gemini-PRC-gemma4-S-2026-05-09).
  • docs/benchmarks/index.md — refreshed TL;DR (5 PROD-READY
    labels at v1.1.0 engine HEAD), per-role token/wall winner
    matrix, cross-reference to the caliber cohort with the
    caliber-init verdict (claude-cli stays default; glm-5.1:cloud
    is the recommended non-claude-cli fallback).
  • docs/caliber-eval-results/README.md — tl;dr — verdict
    section with the pick-when table and explicit disqualifications
    (deepseek-v4-flash:cloud and gemini-3-flash-preview:cloud
    both fail the references-point-to-real-files rubric — the same
    grounding-discipline weakness they show on consultants Q3).
Assets 2
Loading