Files
lc-mcp/docs/feedback/2026-10-03/jev-runtime.md
T

76 lines
4.5 KiB
Markdown

# Jev deployment and live mirror controller feedback (2026-10-03)
This is a live deployment record, not a completed speedrun claim. Run
`mirror_20261003_142030` started at `2026-10-03T14:20:30.186+08:00`.
The timer includes English UI compatibility work and all interruptions.
## Deployment
- Updated main from `97eca20` to `95ae060`, preserving local AALC files.
- Restarted `LcMCPServer` and validated a fresh MCP session, screenshot/OCR,
and idle AALC task state. The desktop's old exposed transport stayed closed;
fresh subprocess MCP sessions work without another desktop restart.
- The runtime launcher now imports the repository controller and matching
modules instead of retaining an older independent script copy.
- Configured the actual `jev-1.13-free` choice endpoint at
`https://opencode.ai/zen/v1/systemone`. Credentials stay in private runtime
configuration and are excluded from this repository and feedback.
- An isolated identical route probe took 5468 ms with curl and 1859 ms with
urllib. Switched the live configuration to urllib. Actual subsequent route
calls included 1313 and 1484 ms; there are also slow requests and timeouts.
The isolated probe is not counted as a live game decision.
## Live findings and repairs
1. The restarted Steam game uses English labels. Added controller-side label
compatibility for entry, gift selection, event choices, judgement and
confirmation boundaries. Preserve the original OCR text for diagnostics.
2. `Receive` incorrectly matched the old substring check for `REC`, causing
unnecessary event advancement. Match the recording label exactly.
3. English OCR merges adjacent `Very High` labels. Derive separate explicit
rate spans; normalization is idempotent and does not invent a high rate
where OCR only says low or normal.
4. English event outcome text was offered as clickable choices. Exclude
`check`, `gain` and E.G.O Gift description lines from event buttons.
5. Repeated generic confirmation requests added API latency. Execute a sole
validated confirmation directly. Route and branch choices still use Jev.
6. Route and sinner API requests now contain current relevant observations,
rather than unrelated entry, fusion and party setup requirements.
7. Sinner candidates are constrained to the explicitly highest visible rate;
explicit OCR rates precede ambiguous colour-only candidates. The log saves
the provider's scored candidate scope. Zeros outside that scope represent
constraints, not extra probabilities returned by the API.
8. Initial candidate pruning caused a probability-set validation error. Fixed
the adaptation between scoped provider probabilities and router validation,
and added an integration regression test.
9. An enabled Commence button initially competed with sinner selection, causing
repeated selection. An enabled judgement button now takes priority.
10. Enter sometimes fails on event judgement. Added the actual observed button
as its mouse fallback. The original fallback list omitted judgement.
11. A changing event video clock reset repeat detection. Ignore both colon and
decimal clock formats when detecting unsuccessful confirmations.
12. After combat, OCR read the map heading as `Exploring Floor` or
`Exploring FloorT` without a valid digit. The old normalizer failed to mark
this as a map, so P/Enter batches continued on the map for several minutes.
Recognize the map heading independently of its floor number and retain
the previously known floor. Added a regression against continuing battle.
## Progress verified so far
- Entry blessings: base 1, 3 and 10; no enhanced blessings.
- Starter gifts: Barbed Lasso and Fluorescent Lamp.
- Actual Battery and permitted Thunderbranch obtained.
- Floor 1 pack: Nest, Workshop, and Technology, with Talisman Bundle in the
displayed clear rewards. One pack refresh was used before finding it.
- Kept the user's seven-character preset and used Enter to enter combat.
- First battle completed, then two event nodes completed. At 15:27 the
controller entered another battle using its P/Enter macro.
- No final reward or dungeon completion has been verified in this record.
The preceding interruptions do not qualify as a successful fast run.
The controller sends background game inputs. There is one game executor;
root interventions occur only after it hands off or stops. Battle batches
use P/Enter, and deterministic confirmations and event advancement avoid
per-key model decisions. Each action and timing remains in the private
runtime JSONL log; no provider key is written into it.