v0.6.1 — gate autonomous advisors (peer-review fixes)

Follow-up to v0.6.0, driven by a consilium peer review of the new advisors (agy/Gemini + hermes/GLM, agreeing independently).

The finding

The closed-stdin guard (advisors are spawned with </dev/null) only stops an advisor that asks before acting. A fully autonomous agent that executes tools without prompting is not stopped by it — the same behaviour openhands was rejected for. So cline and goose needed explicit gates.

Fixes

• cline: pass --auto-approve false (cline defaults to auto-approve, executing tool calls without prompting).
• goose: set GOOSE_MODE=approve (goose runs tools autonomously by default), honouring a user-set value.
• pi: place -p "<prompt>" last, so a value-taking -p can't swallow --model and drop the prompt.
• bin/consult.ps1: use Process.MainModule.FileName instead of [Environment]::ProcessPath, which is null on PowerShell 7.0/7.1 and broke --panel child spawning there.
• docs: honest per-agent read-only table + a note that closed-stdin does not gate autonomous agents.

Verification

bash -n clean, pytest 16 passed. The installed advisors (grok, agy, hermes) were re-verified live and all responded.

Known limitation

cline, goose, pi, and kilo are not installed in the dev environment, so these gates are best-effort from docs and not yet verified against the actual binaries — confirm with <cli> --help + a live write-attempt test when installing them.

v0.6.0 — grok, pi, cursor, kilo, cline, goose advisors

New advisors

Six more headless AI coding CLIs are now consultable via consult <agent>:

Agent	From	Dispatch
grok	xAI (grok-cli)	grok [-m M] [-c] [--cwd DIR] -p "<prompt>"
pi	pi.dev	pi -p [--model M] [-c] "<prompt>"
cursor	Cursor	cursor-agent -p --output-format text [--model M] [--resume] "<prompt>"
kilo	Kilo Code	kilo run [-m M] [-c] [--dir DIR] "<prompt>"
cline	Cline	cline [-m M] [--cwd DIR] "<prompt>"
goose	Block	goose run -q [--model M] [-r] -t "<prompt>"

grok and cursor were verified against the installed binaries' --help (correcting several flags the published web docs got wrong, e.g. cursor's default output is stream-json — now pinned to text). The other four are wired from their official docs.

New cursor is the first roster entry where the CLI name differs from its binary, so the adapter resolves cursor → cursor-agent (or the newer agent alias) automatically.

Evaluated and left out

• kiro (kiro-cli): no headless mode — chat is always an interactive TUI, so it cannot answer prompt-in → text-out → exit.
• openhands: its headless mode forces always-approve autonomous execution with no read-only option, which would let it modify files in the shared working directory and break consilium's "the advisor never touches your files" guarantee.

Also

• Both the bash (bin/consult) and PowerShell (bin/consult.ps1) ports updated in lockstep.
• Council provider / code-access maps extended for the new agents.
• Run consult --list to see which advisors are installed and signed in.

v0.5.0 — proactive consult, end to end

v0.5.0 — proactive consult, end to end

Highlights

• Optional proactive hook (Claude Code) — clients/claude-code/hooks/consult-nudge.sh: a non-blocking PreToolUse(Bash) hook that nudges consult (1) before hard-to-reverse actions (git push/merge/rebase, DB migrations, terraform/kubectl apply) and (2) before committing finished code (doc-only / nothing-staged commits stay silent, so it fires on real code, not every commit). Opt-in: the README documents a two-step setup and deliberately does not auto-edit your settings.json.
• A 10-trigger cookbook — the consulting playbook (docs/consulting-guide.md) now lists 10 concrete when → command → what it returns recipes (migration, API change, auth audit, stuck, push/merge, verify-done, architecture choice, deploy, risky diff, cross-check).

Together with the proactive hub block from v0.4.1, consilium now tells a complete proactivity story: auto-installed instructions for every agent plus an optional deterministic hook for Claude Code.

No core or output-contract changes.

v0.4.1 — proactive consult + sharper comparison

v0.4.1 — proactive consult + sharper comparison

Highlights

• Proactive hub block — the consult-peer guidance synced into your agents (codex / opencode / gemini / hermes) now frames the checkpoints as trip-wires: "consult BEFORE you act, not after — proactively, without being asked." Agents reach for a second opinion on their own at hard-to-reverse moments instead of waiting to be told.
• Sharper comparison — the "How it compares" section (consilium vs Sakana Fugu vs OpenRouter Fusion) now lists an honest quality-vs-efficiency strength and weakness for each tool, including consilium's own (no built-in auto-routing; the hub steers).

Docs / wording only — no code or output-contract changes.

v0.4.0 — perspective-diverse panels + comparison table

v0.4.0 — perspective-diverse panels + a comparison table

Highlights

• Perspective-diverse council — in a multi-member panel, each member now leads with a distinct soft audit lens (correctness / security / concurrency / API-contracts), so the panel covers more of the defect space instead of every member flagging the same issues. A lone auditor stays full-spectrum; members remain independent (no change to the star topology).
• "How it compares" docs — a new comparison table (consilium vs Sakana Fugu vs OpenRouter Fusion) in all three READMEs (en / ru / uk), positioning consilium's local, transparent, verify-against-the-real-code approach against the hosted orchestrators.

Notes

• No breaking changes; the council output contract (COUNCIL STATUS + per-finding source verification) is unchanged.
• Regression tests continue with stdlib unittest (no runtime dependency).

v0.3.0 — the council and multi-advisor panels

v0.3.0 — the council and multi-advisor panels

The headline of this release is consult council: a multi-agent code audit that fans the same code to several advisors on different providers, reconciles their findings into one answer, and mechanically verifies every file:line citation so hallucinated findings are flagged rather than trusted.

Highlights

• consult council — multi-agent code audit with source verification (python3-only, opt-in).
• --panel a,b,c — fan a question out to several advisors in parallel; --review adversarial mode; pipe a diff in via stdin.
• COUNCIL STATUS: COMPLETE | INCOMPLETE — every council run ends with a one-line verdict, and the exit code now reflects source verification (no silent exit-0 when a citation fails to verify).
• Windows PowerShell parity — the council intercept works in the .ps1 port.
• First regression test suite — stdlib unittest, no runtime dependency.
• Advisor isolation fixes: agy reads via an isolated --gemini_dir; the hermes synth prompt-echo is removed; a non-agy synthesizer reduces drift.

Notes

• The core consult stays a zero-dependency shell tool; only the council needs python3.
• Advisors advise, they never edit your files.

v0.2.3

Robustness fixes across the installer and both adapters, plus precise safety wording in the docs.

Fixes

• install.sh — the hub-block updater now requires exactly one START/END marker pair. An unbalanced file (e.g. a duplicated START) previously slipped into the awk rewrite branch and silently dropped everything after the second marker. It now also writes the rebuilt block through the target file (cat >) instead of mv, so a dotfiles symlink keeps pointing at its real file, and removes its temp file via a RETURN trap.
• bin/consult.ps1 — the advisor's child process takes its working directory from ProviderPath, with a guard for non-FileSystem PowerShell locations (Registry::, Cert::) whose .Path is not a valid process directory.
• bin/consult — a failed tee while writing the transcript is now surfaced as a warning; PIPESTATUS is snapshotted in one step so reading the second element stays safe under set -u.

Docs

• Replaced the absolute "it never touches your files" claim with the accurate per-agent picture — codex is hard-sandboxed read-only; the others don't edit without granted permission and consilium never passes them permission-granting flags — across README (en/uk/ru), the Claude skill, and the usage guide.

Full Changelog: v0.2.2...v0.2.3

consilium v0.2.2

A bad argument order silently broke every consult agy --code DIR call — in both the bash and PowerShell clients. agy went off investigating the literal --add-dir flag instead of answering the question.

Why

agy (Gemini CLI) reads the prompt as the positional argument immediately after -p. Both clients emitted agy -p --add-dir DIR "<prompt>", slipping --add-dir between -p and the prompt — so agy parsed the token --add-dir as the request itself and asked "do you want to add --add-dir support to the CLI?" instead of answering.

This also explains the "agy has no --model flag" workaround from 0.2.0: that was the same ordering bug misdiagnosed — -p --model X "prompt" corrupts the prompt for the identical reason.

Changed

• bin/consult + bin/consult.ps1 — the agy dispatch now emits every flag (-c, --model, --add-dir) before -p, with the prompt as the final token. Verified: agy --add-dir DIR -p "..." answers correctly; the old order does not.
• --model now works for agy — agy --help and agy models confirm agy supports --model (Gemini / Claude / GPT-OSS). Removed the stale warn-and-ignore and the "agy ignores it" usage text in both clients.
• Verified end-to-end through consult (bash): correct answers with and without --model, including spaced model names like Gemini 3.5 Flash (Low).

Updating

From your consilium clone:

git pull && ./install.sh --clients

Full Changelog: v0.2.1...v0.2.2

consilium v0.2.1

Distribution fix: updates now propagate to clients cleanly, and new advisors no longer require touching any client config.

Why

Before this, install.sh only handled the consult CLI symlink. The Claude skill and the codex/opencode/agy hub blocks were manual copies that froze at install time — each with a hardcoded agent list that silently went stale on every new advisor (the same doc/code drift that hid the agy --model bug in 0.2.0).

Changed

• clients/hub-block.txt — single canonical hub block with no agent list. It points to consult --list (the runtime source of truth), so adding an advisor never requires re-syncing client blocks again.
• install.sh --clients — idempotent updater. Replaces the block between consilium:consult-peer markers in the agents you already use (never creates configs), copies the Claude skill, warns instead of silently duplicating an unmarked legacy block.
• Removed the inline agent lists from clients/README.md, docs/consulting-guide.md, and the consult-peer skill (the human-facing dispatch/provider tables in docs are kept).
• README / .uk / .ru — new "Updating" section.

Updating

From your consilium clone:

git pull && ./install.sh --clients

One command refreshes the CLI, the hub block in every agent you use, and the Claude skill.

Full Changelog: v0.2.0...v0.2.1

consilium v0.2.0

Second-opinion bus for cross-AI consultation. This release fixes a broken advisor and adds a new one.

Fixed

• agy advisor was silently corrupting prompts. The dispatch appended --model, but the agy (Gemini) CLI has no model-selection flag — the flag and its value were swallowed into the prompt, so the advisor answered about its model instead of the question. --model is now ignored for agy with a clear warning, mirroring the existing opencode/codex patterns.

Added

• hermes advisor (Nous Research Hermes Agent). One-shot dispatch via hermes chat [--model M] [-c] -q "<prompt>", verified against the official CLI docs. Supports --model and --continue; --code is not supported (no --add-dir equivalent) and warns. Wired into both bin/consult (bash) and bin/consult.ps1 (PowerShell).

Docs

• All advisor menus (architecture, consulting-guide, clients/README, consult-peer skill) now list hermes, eliminating the doc/code drift that hid the agy bug.

Notes

• openclaw was evaluated but not added: per its docs it has no headless single-prompt→stdout mode, so a dispatch branch would be guessed and break.

Full Changelog: v0.1.0...v0.2.0