Conversation
Every dispatched request now logs a local prompt-size estimate in the audit log: total chars, estimated tokens (ceil(chars / 4)), a section breakdown (relay instructions, conversation, tool results, tool schemas, payload JSON), retention limits, and forwarded tool counts. Hyperagent does not report authoritative usage, so figures are labeled as local estimates everywhere. New CLI surface: - hacb cost-report [--last N]: per-request table plus totals/averages - hacb explain-prompt [--verbose]: section breakdown of the last prompt - hacb audit --cost [count]: receipts with estimated token columns Prompt excerpt capture is strictly opt-in via debugPromptExcerpts; excerpts stay bounded (240 chars) in the private 0600 state directory and never reach the sanitized audit or gateway logs.
Classify each turn locally into final_answer_only, read_or_search, edit_code, run_command, debug_failure, or unknown, then forward only the client tool categories that task type needs. Function schemas are minimised to argument names, types, and required lists with capped descriptions unless forwardFullToolSchemas is enabled. Safety rules enforced by tests: exact tool names preserved, no tools invented, mid-task conversations keep their tools, unknown turns keep the full inventory, and disabling enableSmartToolSelection restores unconditional forwarding. Defaults: on, minimal schemas, 160-char descriptions.
Add reduceToolOutput() applied to function_call_output and custom_tool_call_output items during input normalisation: ANSI and progress noise stripped, successful output capped to its last meaningful lines, failures keep the first error block plus tail lines, and search output is capped per file with remainder notes. Reductions are marked inline; user prose is never reduced. Defaults: on, 20 successful lines, 80 failure lines, 5 matches/file. Set enableToolResultReducer false for byte-identical prior behaviour.
Load bounded .hacb/CODEX_STATE-style checkpoints and inject them into the relay payload ahead of the conversation so long sessions can rely on cheap repo-local state instead of huge chat history. Only exact configured filenames inside checkpointDir are read; traversal entries are ignored; missing files are skipped; content is capped at maxCheckpointChars with an explicit truncation marker. Defaults: on, .hacb directory, four standard checkpoint files, 4,000 chars. enableCheckpointMemory false disables all filesystem access.
Add hacb task start/status/stop managing durable named tasks with their own request, prompt-char, and tool-loop limits. The bridge reserves from the active task after the daily reservation succeeds; exhaustion releases the unused daily slot and fails closed with an explicit operator message. No active task means no per-task enforcement; daily caps always remain authoritative. Defaults: 8 requests, 180k prompt chars, warn at 120k chars.
Store completed successful responses by sha256(canonical JSON body) in the private state directory and replay identical un-keyed requests locally until the TTL expires (bounded by maxEntries). Errors, cancellations, and indeterminate dispatches are never stored; explicit Idempotency-Key flows take precedence and skip the cache entirely. Disabled by default: enableResponseCache must be explicitly true. Replay is marked with X-Response-Cache-Replayed and a sanitized gateway log event.
Accept type:function_calls relay output when enableMultiToolCalls is true: every call is validated against forwarded function tools and the whole response fails safely to a final answer on any unavailable name or when maxToolCallsPerResponse is exceeded. Valid multi-calls render as multiple Responses function_call items with distinct suffixed call_ids over SSE and non-streaming output. Default remains disabled; single-tool behaviour is unchanged.
Collapse the relay header from 21 lines of prose to the mechanical minimum that survives imperfect relay-agent setup: ownership boundary, forwarded tool list, four JSON shapes, never-invent-names, and one-JSON-object rules. Trivial-prompt size drops from 1,978 to 1,370 chars (-31%), applied to every sampling request. Documented the recommended relay agent setup in README and marked RELAY_AGENT_PROMPT.md as the durable home for behavioural rules.
Add pickAgentRoute() classifying each request into debug_failure, large_refactor, planning, tool_selection, or final_answer, and bridge support for substituting a configured agentRoutes target before dispatch when enableAgentRouting is true. Route targets resolve via the standard model resolution rules; unknown targets log agent_route_fallback and keep the requested model; selected routes and reasons are recorded in audit receipts. Disabled by default: no hidden model substitution.
Add runLocalPreprocessor() executing a configured local command with sanitised request metadata on stdin (requestId, model, toolCount, inputChars; never prompt content) and honouring an explicit reject decision as HTTP 400 preprocessor_rejected. Timeouts kill the child; spawn errors, non-zero exits, and invalid JSON follow the configured failure mode: fallback allows with a logged reason, fail_closed returns HTTP 503 preprocessor_failed. Disabled by default: enableLocalPreprocessor plus an explicit localPreprocessorCommand are both required.
# Conflicts: # src/protocol.mjs
# Conflicts: # README.md # src/config.mjs # test/protocol.test.mjs
# Conflicts: # README.md # src/config.mjs # src/protocol.mjs # test/protocol.test.mjs
# Conflicts: # README.md # src/bridge.mjs # src/config.mjs # src/protocol.mjs # test/protocol.test.mjs
# Conflicts: # README.md # src/config.mjs # test/protocol.test.mjs
# Conflicts: # README.md # src/bridge.mjs # src/config.mjs # test/bridge.test.mjs # test/protocol.test.mjs
# Conflicts: # README.md # src/bridge.mjs # src/config.mjs # test/bridge.test.mjs
# Conflicts: # README.md # src/bridge.mjs # src/config.mjs # test/bridge.test.mjs
# Conflicts: # README.md # src/bridge.mjs # src/config.mjs # test/bridge.test.mjs
Cut mid-turn tool dumps, scrub Codex plugin chrome, coerce/sanitize relay JSON, and make profile/doctor OpenRouter-aware after local QA.
Keep tool_search MCP discoveries eligible, reserve a bounded other-tool slot, tighten chrome scrub / apply_patch / coercion / doctor key trim, and count the OpenRouter system prompt in budgets.
feat: tighten OpenRouter harness efficiency and relay reliability
OpenRouter often emits apply_patch as a function_call or bare type; Codex only accepts the freeform custom tool, so remap those shapes before rejecting.
gpt-4.1-mini sometimes emits `@@` plus a following `-1,9 +1,31 @@` line; Codex treats that as search context and the patch loops forever.
budget --set was clamped at 100 in both CLI validation and dailyLimit(); allow higher explicit caps (used for OpenRouter harness QA at 500).
Codex freeform apply_patch treats unified ranges (split or joined) as search context, so rewrite them to a bare @@ marker before Codex sees them.
High-default models like Claude Sonnet reserve ~64k completion tokens on OpenRouter; without an explicit max_tokens the staging key returns 402 and the bridge surfaces a generic upstream_error.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.