LLMRig v0.11.0: Rescue for Local Coding Agents

LLMRig v0.11.0 | Rescue: Diagnose Local Coding-Agent Failures

LLMRig v0.11.0 adds the experimental Rescue workflow for local Ollama
models used through OpenCode. It distinguishes a model producing an expected
correction from the agent actually completing its turn.

What's new

• llmrig rescue: read-only preflight for a named Ollama model and OpenCode.
  No model inference or OpenCode execution without the explicit test flags.
• --agent-test: opt-in, disposable read-only debugging fixture with
  exact file-integrity and tool-permission checks.
• --probe: optional synthetic protocol canary, explicitly not a full
  coding-agent benchmark.
• --ab-test: fixed three-task comparison of declared OpenCode contexts.
• --compaction-ab: alternating baseline/disabled automatic compaction
  experiment across three fixed tasks, same declared context and timeouts.
• --agent-no-auto-compact and --agent-steps: isolated, opt-in
  hypotheses for a single disposable test, never persistent user changes.
• llmrig rescue-evidence: offline, privacy-filtered receipt comparison
  with explicit unknowns, including support for older baseline receipts.
• Privacy-preserving event sequencing that differentiates correction
  evidence, agent exits, timeouts and explanation-rubric misses.

Measured experimental case

Environment: One user-reported Apple M4 Max (48 GiB), OpenCode
1.19.3, Ollama 0.33.3 and qwen3.8:27b-mlx.

At an OpenCode-declared 4096-token context, a matched,
alternating-order three-task comparison recorded:

Compaction policy	Strict completed tasks	Timeouts	Correct-fix evidence
OpenCode default	0/3	3	3/3
Explicitly disabled in disposable fixture	3/3	0	3/3

Earlier independent repetitions also had 3/3 successful runs with
compaction explicitly disabled (28.43 s, 22.36 s and 10.21 s).
All results were self-reported by one local machine, with small synthetic
tasks. The matched experiment is exploratory, not statistical proof
of a broad improvement or an independently diagnosed root cause.
No direct OpenCode compaction events or effective Ollama per-request
context were measured.

Rescue labels the narrow failure pattern suspected
compaction-associated post-answer continuation, but its default
recommendation remains no automatic or global compaction change.
Disabling compaction in long coding sessions may cause context exhaustion.

Install and try

pipx upgrade llmrig
llmrig --version
llmrig rescue --agent opencode --model YOUR_OLLAMA_MODEL

# Only if you trust your installed OpenCode and want the explicit real-agent test
llmrig rescue --agent opencode --model YOUR_OLLAMA_MODEL \
  --context 4096 --agent-test --agent-timeout 90

Details: Rescue guide.

Security and limits

• OpenCode is third-party code running with your OS user privileges.
  Temporary HOME, restricted tools and isolated workspaces are not
  an OS-level sandbox or network firewall.
• No user project is opened, no permanent configuration is updated, no
  raw agent messages are included in Rescue reports, and reports stay local.
• Do not interpret text-matched expected fixes as verified code execution,
  or infer general agent reliability from the tiny included tasks.
• Version v0.11.0 keeps the same Python 3.9+ baseline and dependency contract.

Full changelog:
compare v0.10.0...v0.11.0.

LLMRig v0.10.0: Verified Autopilot and Rig Reports

LLMRig v0.10.0: Verified Autopilot and Rig Reports

End-to-end execution validated

• Completed local Autopilot verification with oMLX 0.6.4 and mlx-community/Qwen3-0.6B-4bit on an Apple M4 Max (48 GiB): 2 measured race-v2 runs, 0.29 s reported latency; throughput not recorded.
• Completed local verification with Ollama 0.33.3 and installed qwen3.8:27b-mlx: 2 measured race-v2 runs, 49.93 generation tok/s, 2.5861 s reported latency. These different models are not a controlled runtime comparison.
• Added reproducible field-run notes and an illustrative evidence card.

New workflows

• Model-specific read-only llmrig doctor MODEL, actionable guidance, offline JSON Plan Explorer.
• Privacy-filtered llmrig report output in terminal, Markdown, JSON, plus an offline local Rig Report viewer.
• Safer Hugging Face artifact acquisition, provenance, and informative runtime failure receipts.
• Suppress potentially misleading Ollama prompt-throughput estimates from sub-millisecond phases.
• Exact, sanitized Ollama artifact tag in shared reports; catalog format clearly distinguished from verified runtime metadata.
• Improved clean installation documentation and independent wheel/pipx smoke coverage.

Boundaries

• Mutating actions remain approval-gated; read-only planning never downloads weights.
• Verification numbers are local, self-reported observations, not independent benchmarks or hardware attestations.
• Ongoing Ollama catalog metadata and oMLX HTTP 401 classification limitations should not be mistaken for fixed bugs.
• Python 3.9+ and the existing optional local runtime requirements are unchanged.

See README and Installation.

LLMRig v0.9.3 — Sharded GGUF planning

Release LLMRig v0.9.3 (#51)

* release: bump version to 0.9.3

* docs: add v0.9.3 release notes

* release: update citation metadata for 0.9.3

* docs: describe sharded GGUF fit planning

* test: enforce release metadata version consistency

* test: update Autopilot version expectation for 0.9.3

* test: update project version expectation for 0.9.3

LLMRig v0.9.2 — README and measured evidence refresh

What's Changed

• docs: sharpen README positioning and quickstart by @sunilteja93 in #48
• release: prepare LLMRig v0.9.2 README refresh by @sunilteja93 in #49

Full Changelog: v0.9.1...v0.9.2

LLMRig v0.9.1 — Hugging Face launch metadata

What's Changed

• release: fix v0.9 PyPI publish validation by @sunilteja93 in #45
• docs: add Hugging Face launch links by @sunilteja93 in #46
• release: LLMRig v0.9.1 documentation and metadata patch by @sunilteja93 in #47

Full Changelog: v0.9.0...v0.9.1

LLMRig v0.9.0 — Autopilot for local AI

What's Changed

• v0.9: Autopilot for local AI by @sunilteja93 in #43
• assets: finalize monochrome v0.9 hero by @sunilteja93 in #44

Full Changelog: v0.8.1...v0.9.0

LLMRig v0.8.1

LLMRig v0.8.1

Documentation and package-metadata patch release.

Fixed

• Fixed broken documentation links in the PyPI-rendered README.
• Converted repository-relative links to absolute GitHub URLs for:
  • CHANGELOG.md
  • ROADMAP_V1.md
  • RELEASING.md
  • SECURITY.md
  • CONTRIBUTING.md
• Added regression coverage to prevent repository-relative README links from breaking on PyPI again.

Scope

No runtime, solve, benchmark, model-resolution, evidence, or Autopilot behavior changes.

LLMRig v0.8.0

LLMRig v0.8.0 — Runtime Intelligence

LLMRig v0.8.0 expands the Autopilot foundation into an evidence-driven runtime intelligence layer for local AI.

Highlights

• Added llmrig runtimes and llmrig runtimes --json for read-only inspection of oMLX, Ollama, MLX-LM, and llama.cpp.
• Unified runtime readiness, capability, execution, and verification routing behind the runtime-adapter registry.
• Added first-class oMLX discovery and authenticated local execution through its OpenAI-compatible API.
• Added provenance-safe oMLX local inventory matching with exact Hugging Face repository identity.
• Extended solve --verify to measure already-local oMLX configurations using the existing race-v2 methodology.
• Hardened Hugging Face artifact classification for MLX, GGUF, and generic Safetensors.
• Preserved quantization and context evidence from exact repository metadata without downloading model weights.
• Added fail-closed handling for ambiguous provenance, quantization, context metadata, and incomplete shard sets.
• Refreshed the README, runtime architecture, terminal hero, release metadata, and packaging validation.

Evidence first

v0.8 continues to enforce LLMRig's core evidence rules:

unknown != false
compatible != local
local != executable
executable != measurable
measurable != measured
measured performance != model quality

Default discovery and solve workflows remain read-only. LLMRig does not automatically install runtimes, start services, download model weights, or scan arbitrary filesystem locations.

Compatibility

• Python 3.9+
• macOS
• Linux
• Windows
• No third-party Python runtime dependencies

This release was validated across macOS, Linux, and Windows on Python 3.9 and Python 3.14, including wheel/sdist installation tests and CodeQL.

LLMRig v0.7.0

LLMRig v0.7.0

Know what your rig can run.

v0.7.0 introduces LLMRig's first Autopilot solve workflow and strengthens LLMRig's compatibility and performance intelligence for local AI.

Autopilot solve

New read-only planning: llmrig solve MODEL

Solve evaluates the evidence LLMRig currently has across model/artifact resolution, runtime compatibility, local availability, execution support, measurement support, and recommendation state.

Unknown information remains explicit. Solve does not invent a winner when the evidence is insufficient.

Explicit measured verification

Use llmrig solve MODEL --verify to measure compatible, already-local candidates through the existing race-v2 execution path and balanced Pareto decision logic.

Verification preserves:

• the existing 5% comparison threshold
• at least two comparable executable configurations
• failure invalidation for intended competitors
• Pareto and inconclusive outcomes
• no universal performance score
• no model-quality inference from throughput

Verification does not install runtimes, pull models, download artifacts, or scan arbitrary filesystem locations.

Python SDK

A minimal Python SDK is now available through llmrig.solve(...).

The deliberate top-level solve surface includes:

• llmrig.solve
• SolveResult
• SolveCandidate
• SolveError
• SolveInputError
• SolveEngineError

Internal orchestration types remain private.

Evidence and privacy

v0.7.0 formalizes orthogonal state instead of collapsing unknown facts into false conclusions.

In particular:

• compatible does not mean local
• local does not mean executable
• executable does not mean measurable
• measurable does not mean measured
• measured does not automatically mean recommended

Explicit native artifact paths remain private to the inventory-to-execution boundary and are not intended to enter public solve JSON, human output, result structures, or errors.

Release and packaging hardening

This release also adds:

• actual wheel and sdist validation in CI
• clean-environment artifact installation tests
• Python 3.9 and 3.14 CI across Linux, macOS, and Windows
• release tag/source-version enforcement
• release-commit provenance enforcement against main
• PyPI Trusted Publishing through OIDC
• CodeQL analysis
• GitHub Actions Dependabot monitoring
• CITATION.cff
• documented release procedures
• refreshed security and contribution guidance

LLMRig continues to have no third-party Python runtime dependencies.

Validation

• 237 tests pass
• wheel and sdist install validation passes
• Linux, macOS, and Windows CI passes
• Python 3.9 and 3.14 CI passes
• CodeQL passes

LLMRig remains Qwen-first today. Broader model-family intelligence, prediction/calibration, RigGraph, and stable adapter/plugin interfaces remain future work.

LLMRig v0.6.0

What's Changed

• Add cross-runtime local execution by @sunilteja93 in #18
• Add explainable decision intelligence and Pareto optimization by @sunilteja93 in #19
• Prepare LLMRig v0.6.0 by @sunilteja93 in #20

Full Changelog: v0.5.1...v0.6.0