ROADMAP.md

roadmap

• implement toml configuration parsing
• implement markdown agent/skill parsing from .claude/ and .opencode/
• implement openai api streaming client
• implement cli main loop (stdin, api request, stdout/stderr streams)
• add unified client type config (openai/openai_responses/anthropic/etc)
• add tracing support for verbose logging
• add bash tool with wildcard permission matching
• consolidate permission checking across tools
• add --system-prompt (-s) flag for overriding the system prompt
• gitignore-style glob patterns for permissions (*, **, ?)
• implement "ask" permission with /dev/tty user confirmation
• bash metacharacter fallback to catch-all permission
• --yes flag for auto-accepting ask prompts
• --tools takes comma-separated list (exclusive override)
• --skills flag for exclusive skill override
• --list-agents, --list-skills, --list-tools, --list-providers (with descriptions)
• path mode vs command mode for glob matching
• --dry-run flag to inspect prompts without calling LLM
• --skills flag for exclusive skill override
• show usage on empty prompt (exit non-zero)
• fix find_all_in_dirs double-nesting base directories for global paths
• hook protocol (partial conformance, see the hcp-spec and DESIGN.md):
  • discover hooks from hooks/ under existing base dirs (.claude/.opencode/.agents, plus globals)
  • discover stage at startup: register hook tools (<prefix>_<tool_name>)
  • mutate_request stage: merge returned system strings into the system prompt; payload includes finalized system/user/model/tools
  • execute_tool stage: route invocations of hook-registered tools back to the owning script
  • before_tool / after_tool stages around built-in read / bash and hook-registered tools, with deny / args / result mutation responses
  • before_stop post-loop observational stage with exit_reason / final / error
  • honor permission.arg for hook tools (route through existing check_tool_permission)
  • --list-hooks flag
  • hook tools appear in --list-tools
  • typed parameters: validate args (required, type, enum) before execute_tool; normalize any type
  • dry-run inclusion of hooks (own section + tools/permissions list)
  • host capability payload ({name, version, stages}) on every hook invocation
• hook protocol, tier 1/2 follow-ups:
  • before_stop.continue re-entry into the agent loop (wire format honored; needs plumbing user-prompt re-injection back into run_agent_stream)
  • max_turns distinction in before_stop.exit_reason (today everything reports "stop" or "error"; needs rig surfacing the loop-exit cause)
  • before_turn / after_turn stages
  • on_error / on_permission stages
• --dump-request (-D) flag emits exact OpenAI chat-completions JSON request body
• make precommit target (lint + test + build)
• configurable tool-call output truncation (stderr display); full content spooled to $XDG_CACHE_HOME/hrns/<pid>/<seq>.txt with a clear warning
• read tool: optional offset (0-indexed line) and count (number of lines) parameters
• config default_system_prompt used when no -s and no agent body

hrns / hmux backend lift

lift hrns's agent loop into a reusable hrns-core library so hmux can drive it in-process as its most capable backend (../hmux ROADMAP phase 34). the CLI keeps working as a second consumer; no behavior change. the airun -> hrns rename was a SEPARATE atomic step done at the hmux 34e cutover - the repo/crate/config/binary names all became hrns (a hard cutover, no old-path fallback).

• convert to a cargo workspace: hrns-core (lib) + hrns (bin, a root package that is also the workspace). the bin keeps its src/ + tests/ + example toml in place. make targets now run --workspace so the core crate's tests + clippy run too.
• move the HCP-agnostic, tty-agnostic core into the lib: config (config.rs), glob (glob.rs), the permission model (permission.rs), agent/skill/markdown resolution (resolve.rs), the built-in read/bash tools (tools.rs), and the rig agent loop (agent.rs) -- with all their unit tests. a catalog() of providers+models is deferred to hmux 34b (its shape follows the hmux manifest; hrns's config carries no per-provider model list).
• core stays HCP-AGNOSTIC: the HCP host (src/hooks.rs, the v3 hcp-spec host) stays BIN-SIDE and imports the permission bits from hrns_core. hmux hosts HCP at the hub layer (its hcp face), so a core-embedded host would double-host an hmux-driven session.
• seam Observer (agent.rs): the loop reports text/reasoning/tool events through it instead of printing; the CLI's StdioObserver reproduces the stdout/stderr rendering (incl. spool + truncation) verbatim. hmux maps each call to a normalized hub event.
• seam PermissionResolver: the "ask" branch of check_tool_permission takes a &dyn PermissionResolver; the CLI supplies TtyResolver (/dev/tty + --yes). SYNC for now (the CLI blocks on tty, preserving behavior); async-ified when the hmux impl lands (34b).
• seam ToolMiddleware (tools.rs): before/after-tool + the gated hook-registered tool set. the CLI's HcpMiddleware adapts src/hooks.rs onto it; hmux adapts its interceptor bridge. (system- prompt mutation stays a direct main->HCP mutate_request; before_stop stays a direct main->HCP call on the exit info run returns.)
• seam: history in/out. run takes history: Vec<ChatMessage> in and returns RunOutcome { exit_reason, exit_error, history }, using rig stream_chat(prompt, &history) + FinalResponse.history() to accumulate; ChatMessage (= rig Message) + RunOutcome re-exported. the core stays sessionless; the hmux backend stores the Vec per session for follow-up turns.
• the caller-supplied seams PermissionResolver + ToolMiddleware went ASYNC (#[async_trait], new dep async-trait); check_tool_permission is async and the tools await it. the CLI's TtyResolver/HcpMiddleware are async but behavior-preserving (blocking tty/subprocess inside async, as before); the hmux InterceptorResolver awaits interceptor.raise for the native gate.
• resolve_provider (a <provider>/<model> string + Config -> rig client wiring) added to hrns-core, shared by the CLI and the hmux backend (DRY).
• seam: a CancellationToken for abort + a steer channel (for hmux steering, 34c). still pending; best built alongside the hmux abort/steer work.
• keep make precommit green: all rust tests pass -- 18 bin + 23 integration + 20 core = 61 -- and 53/53 python conformance checks pass. the HCP host was upgraded v2 -> v3 (host.version=3 + the base payload carries the workspace cwd, echoed by hooks as <hcp-cwd>), closing the two previously-failing conformance checks.