1. What Github Agent Does
Github Agent answers quantum analysis claims about a subnet's development activity, commits, pull requests, issues, releases, and contributors, using evidence it has actually collected rather than the model's own memory.
| Source | What it contributes |
|---|---|
| Normalized records | Commit, pull request, issue, and release counts, contributors, and timestamps, rolled up per subnet, for exact numbers such as "1,240 commits." |
| Semantic documents | Pull request descriptions, release notes, and important commit messages, capturing what changed and why. |
2. Where It Sits in the Tao Status Economy
Github Agent is a consumer of miner-contributed LLM keys. Exactly one code path uses the pool, the quantum analysis turn endpoint BT-Arena's orchestrator calls; other endpoints always use the agent's own static key and never touch miner scoring. It prefers a pooled miner key for the required provider and model, falling back to its own static key, invisible to scoring, only when contribution is disabled, the pool declines, or the vendor is unsupported. One lease, one LLM call, one report per turn, with no retry loop, so cross-hotkey misattribution is structurally impossible.
3. The Turn Pipeline
- No evidence: the agent declines and reports nothing. A success report here would wrongly credit a miner for a call that never happened, so none is filed.
- An exception during the LLM call is always reported as a real key outcome, success false, with the exception's category.
- Reporting is best-effort: a reporting failure is logged and swallowed rather than failing the turn, and fallback keys with no hotkey are never reported.
4. Quality Grading
| Grade | Meaning | When |
|---|---|---|
| 1.0 | Settled, well-evidenced | SUPPORTED/REFUTED at high confidence |
| 0.75 | Settled, decent evidence | SUPPORTED/REFUTED at medium confidence |
| 0.5 | Settled, thin evidence | SUPPORTED/REFUTED at low confidence |
| 0.6 | Reasoned uncertainty | INSUFFICIENT with real evidence attached |
| — | Not assessable | No response, exception, ungrounded, or empty evidence — neutral, never a penalty |
5. Reporting
One report is filed per turn, feeding directly into the reporting miner's scoring window:
{
"hotkey": "5F...",
"success": true,
"latency_ms": 2731.5,
"error_category": null,
"quality_score": 1.0,
"lease_token": "the token from this turn's acquire grant"
}- success means no exception escaped the turn; on failure, error_category carries the exception's class name.
- latency_ms spans the whole agent run: retrieval and the LLM call.
- lease_token is required, and each token is single-use.
6. Simulated Example
A quantum analysis turn on the claim: "Subnet 33 is actively developing."
- 1
Lease
A pooled miner key is acquired for the session's provider and model, locked for five minutes.
- 2
Retrieve
The claim maps to the subnet's repositories. Retrieval returns 1,240 commits, 87 pull requests with 61 merged, 32 contributors, and 14 releases, plus a validator-optimization pull request summary and two release notes. Eleven evidence items with metrics only 12 days old clear the high-confidence threshold.
- 3
One LLM call
The model returns a SUPPORTED verdict citing specific evidence items, an assessment, uncertainties, and counter-conditions.
- 4
Report
The turn is graded 1.0 and reported. The miner's scoring window gains one success row with top quality.
- 5
Variant: no evidence
A subnet whose repositories were never collected causes the agent to decline; the leased LLM is never called, no report is filed, and the orchestrator skips it for that round.
- 6
Variant: provider failure
A raised error is reported as success false with the exception's category; the miner's window absorbs one failure and the key goes on cooldown. A fatal authentication error instead marks the key dead immediately, zeroing that miner within one poll.
7. Operating Limits
| LLM calls per turn | Exactly 1 (no retry loop) |
| Supported pooled vendors | Gemini, Anthropic, OpenAI, Grok, Mistral, DeepSeek, Nvidia |
| Confidence thresholds | High ≥8 evidence items ≤45 days old · Medium ≥3 · else Low |
| Quantum Analysis turn budget | 180 seconds |
| Key lock per acquired key | 5 minutes |
| Cooldown after a reported failure | 60 minutes |