How long does the integration take?

Change one base URL in an OpenAI-compatible client, or install the Python or TypeScript SDK; the quickstart routes a first request in a few minutes.

Will this lock us into another vendor?

No. The router is Apache-2.0 open source, your provider keys stay yours, and the API is OpenAI-compatible, so leaving is a change of base URL.

What happens when a provider fails?

The router falls back to the next model your policy allows and records the fallback in the trace, so an outage costs a retry, not an incident.

wow-viewer Training 03 v7.6 Paired Training Alignment - Prompt

Name: wow-viewer Training 03 v7.6 Paired Training Alignment
Author: akspa0

Prompt file imported from akspa0/parp-tools (.codex/prompts/wow-viewer-training/03-v76-paired-training-alignment.md). Copyright stays with the author.

Plan or implement the v7.6 alignment slice: make the paired reconstruction branch consume the same structured dataset truth discipline instead of living off legacy cache assumptions.

Read First

gillijimproject_refactor/memory-bank/activeContext.md
gillijimproject_refactor/memory-bank/progress.md
gillijimproject_refactor/docs/VLM_Training_Guide.md
gillijimproject_refactor/docs/v76-model-architecture-guide.md
gillijimproject_refactor/docs/v76-output-dataset-spec.md
gillijimproject_refactor/src/WoWMapConverter/scripts/cache_v7_6_data.py
gillijimproject_refactor/src/WoWMapConverter/scripts/train_v7_6.py
gillijimproject_refactor/src/WoWMapConverter/scripts/inference_v7_6.py
gillijimproject_refactor/src/WoWMapConverter/scripts/stitch_full_map.py
.codex/prompts/wow-viewer-training/01-shared-dataset-truth-gates.md
AGENTS.md

Goal

Realign v7.6 with the structured dataset workflow so its cache building, training, inference, and stitched outputs all agree on one explicit input and output contract.

Current Concrete Problem

v7.6 is a real paired-output branch, but it is not yet cleanly aligned with the structured datasets/ workflow.
the cache builder still hardcodes a legacy dataset root under test_data/vlm-datasets/053_Azeroth_v30.
inference and stitched outputs still lean on loose-file assumptions instead of the documented output dataset spec.
until those contracts are explicit, v7.6 runs are hard to compare, reproduce, or trust.

Non-Negotiable Constraints

Keep v7.6 as a separate paired branch; do not force it into v7.5.1 supervision semantics.
Remove hardcoded legacy dataset roots and make the real dataset contract explicit.
Align cache inputs, train inputs, inference outputs, and stitched outputs to v76-output-dataset-spec.md.
Prefer manifest-driven or config-driven paths over one-off local edits.
Do not claim v7.6 recovery from a cache-only patch if the inference and output contract are still loose.
Keep the scope on data-flow and reproducibility; do not turn this into a new architecture branch unless the current guide is clearly wrong.

What The Work Must Produce

the exact root and manifest contract v7.6 should consume
the exact script or config changes needed to remove legacy-root drift
the exact output contract for train, inference, and stitch steps
the minimum reproducible command sequence for cache, train, inference, and stitch proof
the pass or fail gate for saying v7.6 is aligned enough to iterate seriously
the follow-on backlog after the first aligned run is proven

Deliverables

Return all items:

current drift map
required script or config alignment work
first reproducible runbook
proof expectations for outputs
residual risks after the first slice
what should stay out of scope for this slice

First Output

Start with:

the exact reason v7.6 is still off the structured workflow
the single highest-leverage alignment slice to do first
the first concrete proof sequence after that slice lands
what you are explicitly not claiming yet

Install wow-viewer Training 03 v7.6 Paired Training Alignment and it becomes one more candidate the router can pick - when it fits.

A listing is a routing target with a manifest: what it does, which domains it covers, what it costs and who publishes it. Once installed it sits beside your own models and tools, is scored like any other candidate for each request, and shows up in the trace when it wins. Ratings come from workspaces that installed it, one per account.

Plan or implement the `v7.6` alignment slice: make the paired reconstruction branch consume the same structured dataset truth discipline instead of living off legacy cache assumptions. ## Read First 1. `gillijimproject_refactor/memory-bank/activeContext.md` 2. `gillijimproject_refactor/memory-bank/progress.md` 3. `gillijimproject_refactor/docs/VLM_Training_Guide.md` 4. `gillijimproject_refactor/docs/v76-model-architecture-guide.md` 5. `gillijimproject_refactor/docs/v76-output-dataset-spec.md` 6. `gillijimproject_refactor/src/WoWMapConverter/scripts/cache_v7_6_data.py` 7. `gillijimproject_refactor/src/WoWMapConverter/scripts/train_v7_6.py` 8. `gillijimproject_refactor/src/WoWMapConverter/scripts/inference_v7_6.py` 9. `gillijimproject_refactor/src/WoWMapConverter/scripts/stitch_full_map.py` 10. `.codex/prompts/wow-viewer-training/01-shared-dataset-truth-gates.md` 11. `AGENTS.md` ## Goal Realign `v7.6` with the structured dataset workflow so its cache building, training, inference, and stitched outputs all agree on one explicit input and output contract. ## Current Concrete Problem - `v7.6` is a real paired-output branch, but it is not yet cleanly aligned with the structured `datasets/` workflow. - the cache builder still hardcodes a legacy dataset root under `test_data/vlm-datasets/053_Azeroth_v30`. - inference and stitched outputs still lean on loose-file assumptions instead of the documented output dataset spec. - until those contracts are explicit, `v7.6` runs are hard to compare, reproduce, or trust. ## Non-Negotiable Constraints - Keep `v7.6` as a separate paired branch; do not force it into `v7.5.1` supervision semantics. - Remove hardcoded legacy dataset roots and make the real dataset contract explicit. - Align cache inputs, train inputs, inference outputs, and stitched outputs to `v76-output-dataset-spec.md`. - Prefer manifest-driven or config-driven paths over one-off local edits. - Do not claim `v7.6` recovery from a cache-only patch if the inference and output contract are still loose. - Keep the scope on data-flow and reproducibility; do not turn this into a new architecture branch unless the current guide is clearly wrong. ## What The Work Must Produce 1. the exact root and manifest contract `v7.6` should consume 2. the exact script or config changes needed to remove legacy-root drift 3. the exact output contract for train, inference, and stitch steps 4. the minimum reproducible command sequence for cache, train, inference, and stitch proof 5. the pass or fail gate for saying `v7.6` is aligned enough to iterate seriously 6. the follow-on backlog after the first aligned run is proven ## Deliverables Return all items: 1. current drift map 2. required script or config alignment work 3. first reproducible runbook 4. proof expectations for outputs 5. residual risks after the first slice 6. what should stay out of scope for this slice ## First Output Start with: 1. the exact reason `v7.6` is still off the structured workflow 2. the single highest-leverage alignment slice to do first 3. the first concrete proof sequence after that slice lands 4. what you are explicitly not claiming yet

{ "prompt": "Plan or implement the `v7.6` alignment slice: make the paired reconstruction branch consume the same structured dataset truth discipline instead of living off legacy cache assumptions.\n\n## Read First\n\n1. `gillijimproject_refactor/memory-bank/activeContext.md`\n2. `gillijimproject_refactor/memory-bank/progress.md`\n3. `gillijimproject_refactor/docs/VLM_Training_Guide.md`\n4. `gillijimproject_refactor/docs/v76-model-architecture-guide.md`\n5. `gillijimproject_refactor/docs/v76-output-dataset-spec.md`\n6. `gillijimproject_refactor/src/WoWMapConverter/scripts/cache_v7_6_data.py`\n7. `gillijimproject_refactor/src/WoWMapConverter/scripts/train_v7_6.py`\n8. `gillijimproject_refactor/src/WoWMapConverter/scripts/inference_v7_6.py`\n9. `gillijimproject_refactor/src/WoWMapConverter/scripts/stitch_full_map.py`\n10. `.codex/prompts/wow-viewer-training/01-shared-dataset-truth-gates.md`\n11. `AGENTS.md`\n\n## Goal\n\nRealign `v7.6` with the structured dataset workflow so its cache building, training, inference, and stitched outputs all agree on one explicit input and output contract.\n\n## Current Concrete Problem\n\n- `v7.6` is a real paired-output branch, but it is not yet cleanly aligned with the structured `datasets/` workflow.\n- the cache builder still hardcodes a legacy dataset root under `test_data/vlm-datasets/053_Azeroth_v30`.\n- inference and stitched outputs still lean on loose-file assumptions instead of the documented output dataset spec.\n- until those contracts are explicit, `v7.6` runs are hard to compare, reproduce, or trust.\n\n## Non-Negotiable Constraints\n\n- Keep `v7.6` as a separate paired branch; do not force it into `v7.5.1` supervision semantics.\n- Remove hardcoded legacy dataset roots and make the real dataset contract explicit.\n- Align cache inputs, train inputs, inference outputs, and stitched outputs to `v76-output-dataset-spec.md`.\n- Prefer manifest-driven or config-driven paths over one-off local edits.\n- Do not claim `v7.6` recovery from a cache-only patch if the inference and output contract are still loose.\n- Keep the scope on data-flow and reproducibility; do not turn this into a new architecture branch unless the current guide is clearly wrong.\n\n## What The Work Must Produce\n\n1. the exact root and manifest contract `v7.6` should consume\n2. the exact script or config changes needed to remove legacy-root drift\n3. the exact output contract for train, inference, and stitch steps\n4. the minimum reproducible command sequence for cache, train, inference, and stitch proof\n5. the pass or fail gate for saying `v7.6` is aligned enough to iterate seriously\n6. the follow-on backlog after the first aligned run is proven\n\n## Deliverables\n\nReturn all items:\n\n1. current drift map\n2. required script or config alignment work\n3. first reproducible runbook\n4. proof expectations for outputs\n5. residual risks after the first slice\n6. what should stay out of scope for this slice\n\n## First Output\n\nStart with:\n\n1. the exact reason `v7.6` is still off the structured workflow\n2. the single highest-leverage alignment slice to do first\n3. the first concrete proof sequence after that slice lands\n4. what you are explicitly not claiming yet", "variables": [], "notes": "Plan or implement the v7.6 paired training alignment slice. Use when the goal is to bring the paired reconstruction branch onto the structured datasets workflow, remove hardcoded legacy cache roots, and make train, inference, and stitch behavior follow the documented output contract. Arguments: Optional v7.6 blocker, cache-root issue, inference contract, or proof target to prioritize.", "metadata": { "source": { "provider": "github-codex-prompts", "repository": "https://github.com/akspa0/parp-tools", "path": ".codex/prompts/wow-viewer-training/03-v76-paired-training-alignment.md", "ref": "4f7fb94cfad6bf6348c55c28ae9361c2d2fd9cb7", "url": "https://github.com/akspa0/parp-tools/blob/4f7fb94cfad6bf6348c55c28ae9361c2d2fd9cb7/.codex/prompts/wow-viewer-training/03-v76-paired-training-alignment.md", "key": "akspa0/parp-tools/.codex/prompts/wow-viewer-training/03-v76-paired-training-alignment.md" } } }

wow-viewer Training 03 v7.6 Paired Training Alignment

About

Read First

Goal

Current Concrete Problem

Non-Negotiable Constraints

What The Work Must Produce

Deliverables

First Output

Install wow-viewer Training 03 v7.6 Paired Training Alignment and it becomes one more candidate the router can pick - when it fits.

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Use it

Manifest

Reviews