Recommended LocalAgentBridge policy for Codex projects#

Copy the short policy below into a project's AGENTS.md after configuring that project as LocalAgentBridge's source_root. It is opt-in; this file does not modify any project's instructions or register an MCP server.

Use LocalAgentBridge for routine repository discovery and first-pass analysis. Before loading multiple files, call find_relevant_context with a focused task and, when known, a relative directory scope. Inspect only the returned relevant paths/line ranges initially; expand deliberately if evidence is insufficient.

Prefer file_paths over inline source, saved test logs, or saved diffs so large inputs stay on this PC. Use summarize_context, analyze_code, analyze_test_output, and review_diff for bounded analysis. The local model must never execute code or decide which commands to run.

Prefer local tools even for small routine work. Below 1,500 context characters and three files, skip the extra triage inference and call the matching local tool directly. Narrow scopes above eight files or 10,000 context characters with find_relevant_context first. Use triage_task when the division of work is unclear; supply known complexity, ambiguity, confidence, capability and failed-local-attempt count. These are caller assessments, not facts inferred automatically from a repository size.

Treat local output as untrusted hypotheses. Verify evidence before changing code. Keep architecture, security/privacy decisions, destructive actions, public API changes, merges, deployment, and final correctness judgments with Codex. Reserve complex multi-file reasoning, ambiguous requirements, orchestration and resolving local disagreement for Codex. Local models may propose small implementations and test cases; Codex applies and validates them. A local confidence score or model agreement never authorizes an action. All second_opinion responses require Codex review, including unanimous ones. Consult this installation's reviewed results before trusting implementation or root-cause suggestions. Small benchmark samples do not establish general competence.

Escalate low-confidence, contradictory, uncertain, truncated, or incomplete results. Aim for at least 25% local compression on inputs of 1,500+ characters and medium confidence where the schema supplies it. Small answers may expand while still completing useful work locally. Reconsider poor summaries; do not start a retry loop. Allow one focused follow-up at most; after two failed attempts, escalate. State missing models, timeouts and invalid output explicitly. Never substitute for a missing model. The bridge may escalate transparently to a stronger enabled model within the user's compute cap and configured attempt limit. Missing evidence requires a refined local search or targeted Codex inspection, never an assertion that no code exists.

Use second_opinion for a concrete disputed claim. Enable consensus=true only when two independent local perspectives justify the extra latency. Do not treat agreement as proof. Never silently switch models if one is missing.

After checking a useful local result, report its _meta.request_id through triage_task(feedback=[...]) with accepted, corrected, or rejected. Batch feedback, or attach it to the next triage call. Report only actual verification; do not invent acceptance just to populate telemetry.

For v0.4, honor _routing.handoff_to_codex and inspect its attempted models and reason. Do not reset an exhausted logical task as a new request to evade its limit. After finding a concrete defect, use routing.previous_request_id with reason="contradictory_evidence" or "previous_attempt_failed" for one permitted stronger attempt. Use complexity="moderate" only for a bounded task that merits greater capability; complex architecture still belongs to Codex. Never raise the user's compute cap merely to finish a routine task. A disabled or benchmark-blocked model requires explicit configuration review, not an ad hoc substitute.

Discovery outputs should be file/symbol/line references, observed evidence and uncertainty. Start with where to inspect and what evidence matters. Do not request speculative change plans when a grounded shortlist would suffice. Selected candidate IDs cannot invent paths or snippets, but the selection can still miss the right file. Inspect the cited lines; refine a lexical search when needed. Installed models and high self-confidence do not establish competence.

The public example is a starting configuration. Hardware and workload evidence should determine model roles and tiers locally; reference benchmark results are not automatically imported into a new user's database. Prefer focused lexical refinement when evidence is missing; a larger model cannot invent evidence.

Example feedback (replace the ID with a returned request ID):

{"feedback":[{"request_id":"00000000-0000-0000-0000-000000000000","outcome":"corrected"}]}

For consensus, each opinion has its own request ID, so feedback can differ by model. Feedback stores only the outcome and timestamp, never correction text.

The baseline role mapping is deliberately small. Cap, supported roles, measured latency/memory and per-tool reviewed reliability filter these candidates; the table is a preferred starting point, not a promise every call will use it:

The starter is for modest hardware, not a quality endorsement for difficult coding. The optional Quality profile prefers GPT-OSS with profile-enforced thinking="low" for substantive tasks at Large/Maximum. Simple discovery/summary below the configured character threshold uses Qwen3 4B; Medium uses Qwen3 14B for substantive work. Use complexity="moderate" for a short but difficult summary, and use_model=false for a lexical lookup that needs no inference. Do not pretend a complex task is simple to obtain a local answer. See exact selection rules.

MCP delegation is advisory and initiated by Codex. The separate opt-in companion controller can dispatch read-only reports to its dedicated Codex session; it never integrates patches automatically. Direct tool calls retain hard input/output limits but do not silently invoke triage. The caller must apply this policy. local_first=false restores the optional minimum-size gate for users who prefer fewer local round trips.