claude-advisor

An advisor for Claude Code sessions that run on Amazon Bedrock.

Claude Code's built-in advisor tool lets the working model consult a stronger model at decision points. It is a server-side tool on Anthropic's API, so it is not available on Bedrock. This project re-creates it locally:

• bedrock-advisor, a Claude Code plugin (a mod), registers an advisor tool. When an agent calls it, the plugin finds that agent's own transcript on disk: a subagent's lives at ~/.claude/projects/<project>/<session>/subagents/agent-<id>.jsonl.
• advisor-engine, a small Go binary, filters the transcript down to the task, messages, tool calls and results, and asks the advisor model on Bedrock (Converse API) for guidance. The guidance comes back as the tool result.
• When the same tool error happens twice in a row, the plugin consults the advisor on its own and attaches the advice to the tool result.

It fits an orchestrator setup where Opus delegates to Sonnet or Haiku subagents: the subagents can escalate mid-task to Opus, with their full context, instead of guessing or handing back.

Install

macOS / Linux:

curl -fsSL https://model-harness.github.io/claude-advisor/install.sh | bash

Windows (PowerShell):

irm https://model-harness.github.io/claude-advisor/install.ps1 | iex

The installer downloads the engine for your platform from the latest release, verifies its SHA-256 against the release's checksums.txt, adds this repository as the model-harness plugin marketplace pinned to the same release tag, installs bedrock-advisor@model-harness, and checks your AWS credentials. Start a new Claude Code session afterwards and run /bedrock-advisor to see the setup. Re-run the installer to upgrade; it reinstalls cleanly and keeps the options you set (on macOS and Linux this needs jq).

Each release binary also carries a GitHub build provenance attestation. When a GitHub CLI (gh) new enough to pin the tag is installed and signed in to github.com, the installer checks that the binary was built by this repository's release workflow for the tag you install, and refuses one that fails; otherwise it warns that it skipped the check. To check by hand:

gh attestation verify advisor-engine_linux_amd64 --repo model-harness/claude-advisor \
  --source-ref refs/tags/v0.1.3 --signer-workflow model-harness/claude-advisor/.github/workflows/release.yml

Options: --version v0.1.0 / -Version v0.1.0 (or CLAUDE_ADVISOR_VERSION) pins a release; --from-source <checkout> / -FromSource <checkout> builds the engine with Go instead of downloading it; --uninstall / -Uninstall removes everything; CLAUDE_ADVISOR_SKIP_ATTESTATION=1 skips the build provenance check (with a warning).

Requirements

• Claude Code v2.1.293 or later, configured for Bedrock.
• For the engine's direct Bedrock call: AWS credentials the AWS SDK default chain can find, AWS_REGION, and bedrock:InvokeModel on the advisor model. Without them, the default auto backend sends the advisor request through Claude Code's own model call, with the engine still building the prompt.

Use

Claude calls the advisor when it judges it useful: before committing to an approach, when stuck, or before declaring a multi-step task done. You can also ask for it in a prompt or a subagent brief ("consult the advisor before you edit anything"). The command /bedrock-advisor shows the status; /bedrock-advisor doctor checks the engine and Bedrock access; /bedrock-advisor ask <question> consults the advisor about the main session.

Configure

Set options in /config inside a session, or from your shell:

echo '{"advisor_model":"sonnet"}' | claude plugin configure bedrock-advisor@model-harness --values-stdin

claude plugin configure bedrock-advisor@model-harness without --values-stdin shows the current values.

Option	Default	Meaning
advisor_model	opus	Alias resolved through ANTHROPIC_DEFAULT_OPUS_MODEL (and the Sonnet and Haiku equivalents), or a full Bedrock model or inference profile id
backend	auto	auto: engine calls Bedrock, falls back to Claude Code's model call on credential or access errors. bedrock: engine only. session: Claude Code's model call only
auto_consult	true	Consult automatically when the same tool error happens twice in a row (at most 5 times per session)
max_tokens	2048	Cap on one advisor reply
max_context_chars	400000	Characters of transcript sent to the advisor; the task and the newest entries are kept
engine_path	set by the installer	Path to advisor-engine

Pick an advisor at least as strong as the model that calls it. The plugin does not enforce the built-in advisor's pairing rules.

Compared with the built-in advisor

	Built-in advisor	claude-advisor
Where it runs	Inside one API request, on Anthropic's servers	A normal tool call: one extra round trip
Providers	Anthropic API only	Bedrock, or any provider through the session backend
Context	The exact request, server-side	The calling agent's transcript, filtered
Forced consultations	None	On a repeated tool error
Advisor prompt	Anthropic's, not published	Open source
Billing	Advisor model rates on the Anthropic API	Advisor model rates on Bedrock

How it works

Opus (main) --Agent(model: sonnet)--> Sonnet subagent
                                        | calls mcp__bedrock-advisor__advisor
                                        v
                         plugin tool.call hook  (knows the caller's agentId)
                                        | finds <session>/subagents/agent-<id>.jsonl
                                        v
                         advisor-engine ask  (JSON on stdin)
                                        | filters transcript, builds prompt
                                        v
                         Bedrock Converse (advisor model)
                                        | guidance
                                        v
                         tool result returned to the subagent

The engine keeps user and assistant messages, tool calls and tool results, and the context hooks attached after tool results (including its own earlier advice), and drops other attachments, request snapshots, meta messages and thinking blocks, which are most of a transcript's bytes. The engine opens only regular .jsonl files inside the Claude Code projects directory, which it derives itself, through an os.Root, so neither .. nor a symlink can lead it elsewhere. See ADR 0001 for why this is a plugin plus a local engine rather than a gateway.

Develop

cd engine && go test ./...                 # engine
cd plugin && claude plugin test            # plugin hooks, with stubs
claude plugin validate --strict ./plugin   # manifest and hooks module
./install.sh --from-source .               # install from this checkout

See CONTRIBUTING.md. Not affiliated with Anthropic or AWS.

License

MIT