[HARNESS] Define and prove the Hermes observation/control path for local Bannerlord #722

New Issue

Timmy · 2026-03-28T17:35:45Z

Timmy commented

2026-03-28 17:35:45 +00:00

Goal:
Once Bannerlord runs locally, define the real harness path Timmy will use to perceive and act.

Acceptance criteria:

choose the observation path (screenshot/log/window title/etc.)
choose the control path (desktop input, MCP bridge, or other bounded local method)
prove at least one observe → decide → act loop on local Bannerlord
proof includes logs/command output plus screenshot evidence before/after action
keep telemetry flowing through Hermes, not through OpenClaw or ad-hoc side state

This is about the real local bridge, not a speculative architecture note.

Goal: Once Bannerlord runs locally, define the real harness path Timmy will use to perceive and act. Acceptance criteria: - choose the observation path (screenshot/log/window title/etc.) - choose the control path (desktop input, MCP bridge, or other bounded local method) - prove at least one observe → decide → act loop on local Bannerlord - proof includes logs/command output plus screenshot evidence before/after action - keep telemetry flowing through Hermes, not through OpenClaw or ad-hoc side state This is about the real local bridge, not a speculative architecture note.

Timmy self-assigned this 2026-03-28 17:35:45 +00:00

Timmy referenced this issue

2026-03-28 19:06:18 +00:00

[EPIC] Local Bannerlord on Mac — emulator, harness bridge, and portal readiness #719

Timmy referenced this issue

2026-03-29 00:12:04 +00:00

[EPIC] Local Bannerlord on Mac — emulator, harness bridge, and portal readiness #719

Timmy commented

2026-03-29 00:12:05 +00:00

Triage refinement:

This issue is the architectural gate.

Thin-adapter rule:

acceptable: launch, observe, inject bounded input, normalize logs/screenshots/telemetry
unacceptable: reconstruct game semantics from scratch, invent a second world model, or build a large interpretation layer

Success means:

one real local observe -> decide -> act loop with before/after proof
control path stays narrow enough that it could live outside Nexus cleanly

Failure condition:

if the seam only works by accreting lots of bespoke glue, that is falsework and the lane should be rejected instead of expanded.

Triage refinement: This issue is the architectural gate. Thin-adapter rule: - acceptable: launch, observe, inject bounded input, normalize logs/screenshots/telemetry - unacceptable: reconstruct game semantics from scratch, invent a second world model, or build a large interpretation layer Success means: - one real local observe -> decide -> act loop with before/after proof - control path stays narrow enough that it could live outside Nexus cleanly Failure condition: - if the seam only works by accreting lots of bespoke glue, that is falsework and the lane should be rejected instead of expanded.

Timmy referenced this issue

2026-03-29 00:12:06 +00:00

[TRAINING] Capture the first replayable local Bannerlord session trace for Timmy #723

allegro referenced this issue from a commit

2026-03-31 04:53:39 +00:00

feat: Complete Bannerlord MCP Harness implementation (Issue #722)

allegro closed this issue

2026-03-31 04:53:39 +00:00

Timmy referenced this issue

2026-04-04 17:18:39 +00:00

[TRAINING] Capture the first replayable local Bannerlord session trace for Timmy #723

Sign in to join this conversation.