feat: harden consensus loop, codex error classification and orientation default

[antonbabenko]

Summary

Ten fixes from a quality, cost and performance audit of the consensus loop, the Codex provider, the analyzer and the orientation feature.

Results quality

• Codex error classification now runs on codex's own stderr lines (prompt echo removed, ANSI stripped). A review prompt that mentions "auth" or "login" no longer turns a usage-limit error into a non-retryable auth error that tells the user to log in again.
• Arbiter peer-opinion cap goes from 2000 to 3000 chars and keeps head (1800) and tail (1200). The review format puts the verdict last, so a head-only cut dropped the conclusion of every long review.
• consensus-step excerpt: a non-APPROVE voice whose reply had no [category] bullets now carries the first and last 600 chars of its text, so the host arbiter can turn the objection into a decision instead of seeing "REQUEST_CHANGES (0 critical)" every round.
• Round digests in the provider-arbiter loop list dissenting issues as addressed and report real line-count diffs, so peers stop repeating the same issue.
• Orientation auto-attach is ON by default (schema, default config, runtime resolver, docs). With it off, file-blind reviewers judged plans against a repo they could not see. Explicit enabled: false opts out.

Cost

• The per-round blind arbiter pass in runToConvergence was one extra provider call per round whose result nothing read. It now runs only when consensus.blindVote is set, the same key the single-shot path honors.
• A failed arbiter revision leg stops the loop with stopReason: "arbiter-revision-failed" instead of resubmitting the same plan and buying the same dissent again.

Performance and cleanup

• Analyzer: a 360 s absolute slow gate for models seen under consensus / consensus-step (120 s elsewhere), so a long review round does not flag the whole panel as slow.
• provider.js: removed the unused opinion-envelope schema, parser and validator (about 200 lines) and their tests.
• Comment fixes in result-cache.js (cwd is part of the cache key) and answer-floor.js.

Cross-model review

Pre-PR consensus code review (expert code-reviewer, panel: codex, gemini, grok, one OpenRouter delegate) converged in 3 rounds.

Accepted and fixed:

• R1 codex: the orientation default flip was only in the config template; the runtime resolver still turned an omitted block into off and the missing-config path had no orientation key. Fixed with a shared resolveOrientation helper plus tests ORC1-ORC3.
• R2 codex: two stale doc lines still said orientation was off by default. Fixed, host artifacts regenerated.

Dismissed, with reasons:

• grok, glm (R1): "revised starts null and the stop check sits outside the revise branch, so a unanimous APPROVE stops as arbiter-revision-failed." The loop breaks on converged before any revision code; every non-converged path assigns revised, and okText returns null for blank or errored text, so an empty plan is never submitted.
• grok, glm (R1): "accept decisions are recorded before the revision, so a failed revision records issues as addressed." Round history is written only by submitRevision, which never runs after the arbiter-revision-failed break; when a revision lands, the revision prompt is built from exactly those issues.

Verification

• npm run sync:check clean.
• npm run check: 1181 pass. Five failures (bridge M1e, openrouter O14, dashboard-server S13, dashboard-cli C-stop and C-restart) are timing tests that fail identically on untouched master in a temp worktree on the same machine; none touches files in this change.

[github-actions]

This PR is included in v3.29.0 🎉

Claude Code

/plugin marketplace update antonbabenko
/reload-plugins

If the update still shows the old version, the marketplace re-pin is a few minutes behind; run it again shortly.

Codex, Kiro, Cursor, OpenCode plugins: update deliberation from your host's plugin manager.

MCP server (any host): @antonbabenko/deliberation-mcp@3.29.0 is published. Restart your host to pick it up.