The complete harness-quality queue from docs/coding-harness-roadmap.md — eight improvements distilled from source-verified analysis of pi, Cline, Codex CLI, and Gemini CLI, targeting Gemini 3.x first and open-weight OpenRouter/NVIDIA models second.
P0 — wasted-turn killers
- edit_file repair ladder (#338): exact → whitespace → indentation (re-indented replacement) → punctuation-folding → block-anchor matching, over-escape repair, BOM/CRLF preservation, updated-lines echo, instructive errors, lazy-placeholder guard.
- Structured output truncation (#339): head+tail shell capture with omitted counts and a named full-output spool file; read_file offset/max_lines paging with "Use offset=N to continue".
- History hardening (#340):
loop.HardenHistorysynthesizes results for dangling tool calls, drops orphans/empty turns, normalizes call IDs, strips foreign-model artifacts — before every request.
P1 — reliability and cost
- Retry/overflow recovery (#341, ADR-0017): classified loop-level retries with backoff and server retry hints; typed
OverflowError→ compact-once-and-retry in sessions. - Compaction upgrade (#342): structured state snapshot with injection firewall, handoff framing, cumulative file ledger, tool-pair-safe splits, inflation guard.
- Capability registry + tool-owned prompts (#345):
providers.Capabilitiesper provider;coding.SystemPromptassembled from the active toolset (the--codingagent previously ran with no system prompt at all).
P2 — polish
- Next-speaker stall recovery (#343): bounded "Please continue." nudges for Gemini's narrate-then-stop stall, auto-enabled per the registry.
- Loop guardrails (#344): graduated identical-call and consecutive-mistake detectors with typed halt errors.
Default-behavior notes: loop-level retries and guardrails are on by default with opt-outs (RetryPolicy.Disabled, GuardrailPolicy.Disabled); a 429 now retries up to 3× before failing. All changes additive otherwise; full changelog in CHANGELOG.md.