Reported issues for Agentic HIL

Pod holds 18 of 29 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to Agentic HIL.

Most discussed

setup has no non-interactive completion path for agents whose host blocks self-configuration

Every eval agent that stopped at the approval gate named the same blocker: setup writes the agent's own user-level MCP registration and skill, exactly what host permission classifiers catch. A first-class escape would convert the stop into a partially configured bench in one step: agentic-hil init already finishes the project half, so the missing piece is machine-readable output naming the exact remaining command (agentic-hil agent-install --agent <agent>), or a setup --project-only alia

Read the thread · 2026-08-12 · closed · 2 comments

Bench proof: the hot-plug read leaves a running core untouched (SysTick five-pair measurement)

#342 shipped the ST-Link typed-debug reads on the least intrusive connect the CLI documents (mode=HOTPLUG, no reset=), and deliberately claims documented, not measured. This issue tracks the measurement that upgrades the claim, on a real bench with a Nucleo attached, everything through the MCP tools:

flash_firmware an ELF whose firmware increments a RAM counter (e.g. volatile uint32_t uptime_ms in SysTick_Handler); put the symbol in debug.allowed_symbols; let the board run about

Read the thread · 2026-08-30 · closed · 1 comment

An MCP conflict names the file but not what it found there

mcp_config_conflict says an unmanaged entry exists and names the file. It does not say what it found there, and that is the one fact that settles what to do next.

A real case: an operator's ~/.claude.json held

{
  "type": "stdio",
  "command": "<a wrapper script under the operator's own ~/.claude>",
  "args": []
}

The refusal is correct: that entry points at a program the operator chose, and rewriting it would hand the hardware gate to something else. But from the report alone

Read the thread · 2026-08-18 · closed · 1 comment

setup's result should shout that the agent must restart before the tools exist

Lived through three recording takes: setup --agent claude-code succeeds, and the session that ran it then waits on MCP tools that cannot appear, because the host reads registrations at session start. The docs say it since #227, but the one place the agent is guaranteed to look is the setup RESULT itself. setup and agent-install should carry restart_required: true plus one sentence ("the session must restart before the agentic-hil tools appear; doctor works now") in their JSON res

Read the thread · 2026-08-14 · closed · 1 comment

Test plans gate dump_memory and read_symbol on the openocd backend, though both tools answer on stlink

Follow-up from #342. test_reactor.py (around line 1834) admits dump_memory and read_symbol plan steps only when the configured debugger is type: openocd, although both underlying tools have answered on the ST-Link backend since 0.18.0, and since #342 they answer non-intrusively. The coverage workflow that motivated #342 (dump RAM-resident measurement data before a run ends) therefore works through the MCP tools but still cannot be written as a plan step on an ST-Link bench.

This needs a

Read the thread · 2026-08-29 · closed · 0 comments

debuggers..type is not settable over MCP, and the lock no longer protects anything its siblings do not already open

A bench needed to switch its dut debugger from the CubeProgrammer backend to OpenOCD (to reach the typed-debug family, #342) and could not do it over MCP: debuggers.<name>.type is deliberately absent from the project_config_set keys, so the operator had to edit config.yaml by hand while executable, interface_cfg and target_cfg, the three fields that fully determine what binary runs with what scripts, are all writable behind allow_config_description_write.

The documented reason for the

Read the thread · 2026-08-29 · closed · 0 comments

upgrade_failed has no ERROR_CATALOGUE entry, so its refusal renders no What-to-do section

Noticed during #326. Refusals of type upgrade_failed have no entry in ERROR_CATALOGUE (src/agentic_hil/knowledge.py), so the rendered refusal carries Details and whatever next_steps the producing code attached, but no standing What-to-do section like other refusal types get. #326 worked around it with next_steps because knowledge.py was outside its scope.

Add the catalogue entry: likely causes (network or index unreachable, TLS-intercepting proxy per the #326 signatures, package manag