# ModelHub Adaptation Agent

This repository is packaged for the ModelHub XC agent platform.

## Platform Contract

- Root-level `Dockerfile`
- Listens on port `8080`
- Exposes `GET /health`
- Handles `SIGTERM`
- Reads platform-provided `STRATEGY_ID` and attaches it to task submissions as `strategyId`

The root `main.py` starts a lightweight health server and runs the existing
submission poller in a child process.

## Runtime Environment

The image includes multi-account ModelHub and ModelScope token fallbacks for
the agent platform. Environment variables can override them without rebuilding
the image.

- `MODELHUB_XC_TOKEN`, `XC_TOKEN`, `XC_TOKEN2...`, or `MODELHUB_XC_TOKENS` for ModelHub API authentication
- `MODELHUB_JWT_TOKEN` or `JWT_TOKEN` can be used instead when the platform provides a JWT
- `MODELSCOPE_API_TOKEN` or `MODELSCOPE_TOKEN` optional override for the embedded ModelScope fallback token
- `STRATEGY_ID` is expected to be injected by the ModelHub agent platform and is attached to submissions for strategy attribution; it is not an API authentication token

Optional tuning:

- `MODELHUB_AGENT_POLL_INTERVAL_SECONDS` default `15`
- `MODELHUB_AGENT_IDLE_INTERVAL_SECONDS` default `60`
- `MODELHUB_AGENT_POST_CYCLE_COOLDOWN_SECONDS` default `2`
- The hosted entrypoint always uses `--max-submits-per-run 0` so stale deployment
  settings cannot restrict a refill cycle to five submissions.
- `MODELHUB_AGENT_ACTIVE_TASK_CAP` default `100` per account
- `MODELHUB_CAPACITY_PROBE_INTERVAL_CYCLES` default `3`
- `MODELHUB_CAPACITY_STATE_PATH` default `.modelhub_state/account_capacity.json`
- `MODELHUB_AGENT_ACTIVE_COUNTS_TTL_SECONDS` default `15`
- `MODELHUB_AGENT_RESERVATION_TTL_SECONDS` default `120`
- `MODELHUB_AGENT_INSTANCE_ID` optional stable worker identity used to spread concurrent agents across accounts and candidates
- `MODELHUB_AGENT_CLAIMS_PATH` default `.modelhub_state/submission_claims.jsonl`
- `MODELHUB_SUBMISSION_EXCLUSIONS_PATH` default `.modelhub_state/submission_exclusions.jsonl`
- `MODELHUB_AGENT_DAILY_TARGET`
- `MODELHUB_AGENT_MIN_DOWNLOADS`
- `MODELHUB_AGENT_GPUS`
- `MODELHUB_AGENT_EXTRA_ARGS`
- `MODELHUB_GPU_STRATEGY_STATE_PATH` default `.modelhub_state/gpu_strategy.json`
- `MODELHUB_MARKET_INTELLIGENCE_PATH` default `.modelhub_state/market_intelligence.json`
- `MODELHUB_MARKET_QUEUE_REFRESH_SECONDS` default `600`
- `MODELHUB_MARKET_FRAMEWORK_REFRESH_SECONDS` default `21600`
- `MODELHUB_MARKET_THROUGHPUT_WINDOW_HOURS` default `6`
- `MODELHUB_MARKET_FRAMEWORK_MIN_SAMPLES` default `300`
- `MODELSCOPE_PAGE_INTERVAL_SECONDS` default `0.25`
- `MODELSCOPE_PAGE_CACHE_TTL_SECONDS` default `900`
- `MODELHUB_AGENT_VERIFY_CACHE_TTL_SECONDS` default `900`
- `MODELHUB_QUEUE_CLEANUP_INTERVAL_CYCLES` default `120`; cleanup also runs once at startup
- `MODELHUB_QUEUE_CLEANUP_READ_CONCURRENCY` default `6`
- `MODELHUB_QUEUE_CLEANUP_REPORT_PATH` default `.modelhub_state/queue_cleanup_latest.json`
- `MODELHUB_RECENT_MODEL_RESERVE_SLOTS` default `10` per account
- `MODELHUB_DYNAMIC_OLD_MODEL_CLEANUP_RESERVE_SLOTS` default `5` per account
- `MODELHUB_RECENT_MODEL_DAYS` default `7`

## Adaptive GPU Strategy

When no explicit GPU override is supplied, the worker uses a success-first 70/30
strategy generation with no self-funded exploration:

- 70%: the three long-term GPUs with the best Wilson lower confidence score and at least 100 terminal samples
- 30%: the top recent GPUs among the latest 1,000 terminal tasks
- 0%: unvetted/all-GPU exploration; community-wide results provide the exploration signal

The 70/30 category ratio remains exact across accepted tasks. Inside each
category, weighted fair scheduling combines the category's historical rank with
live public market data:

- recent public success quality, scored with a strongly weighted Wilson lower confidence bound
- estimated backlog hours (`waiting / recent completions per hour`) as a bounded tie-breaker
- machine availability, running workers, and advertised concurrency
- a circuit breaker for unavailable or apparently stalled GPU pools

This optimizes expected successful completions rather than blindly selecting the
smallest queue. Queue/throughput data is refreshed every 10 minutes and persisted
in `.modelhub_state/market_intelligence.json`. A failed refresh keeps the last good
snapshot, uses a retry backoff, and never blocks normal submissions.

For each compatible model/GPU pair, the worker also ranks the GPU's supported
frameworks using ModelHub's public aggregate `modelCount` and `successCount` data,
then blends in the worker's own GPU+framework outcomes with a capped weight.
Only frameworks with at least 300 public samples and a safe Wilson lower bound
are eligible. Framework statistics refresh
every 6 hours, so they do not add per-model API traffic. Newly published
frameworks are discovered automatically, but receive no novelty bonus: they can
win only when their confidence-adjusted success score beats the best incumbent by
at least 10%. A new framework
is eligible only after the authenticated official build-config endpoint returns
a complete config that passes local structure, placeholder, framework-name, and
GPU-parallelism validation. Valid official configs are cached and refreshed with
the framework snapshot; local templates remain the fail-safe fallback.

Only platform-accepted tasks count. After exactly 200 accepted tasks, the next
poll cycle reloads all account history, generates a new immutable strategy snapshot,
and resets the generation counters to 140/60 targets. The active snapshot and
progress are stored in `.modelhub_state/gpu_strategy.json`.

Five consecutive local failures open a 12-hour GPU/framework circuit breaker.
A sub-20% success rate over the latest 20 terminal tasks opens a 6-hour breaker.
Platform/infrastructure failures are excluded from long-term compatibility rates
and model/profile breakers. Three consecutive platform failures on a GPU/framework
instead open a short 30-minute breaker, so a temporary broken runner or lack of an
idle card does not permanently poison otherwise successful evidence.
Candidate shortages expand the model search window; they never unlock an
unvetted GPU or framework.

Before a candidate reaches the submit queue, failure-informed preflight checks
the actual ModelScope repository structure and file sizes. Non-GGUF text
frameworks require root-level config, weights, and tokenizer assets. The memory
gate recursively totals the entire repository—including duplicate weight formats
and nested shards—and applies ModelHub's observed 20% loading overhead. It covers
all 14 GPU types currently marked `canVerify=true`; nine capacities come directly
from structured ModelHub OOM reports and five from published specifications until
ModelHub supplies a stronger observation. If a known GPU's repository file sizes
are incomplete, the candidate is deferred rather than guessed. Template context
length is also clamped to the model's advertised limit. A newly introduced GPU
with no capacity evidence is likewise deferred. Override or extend known
capacities with
`MODELHUB_GPU_MEMORY_GIB_JSON`, for example
`{"New_gpu": 64}`.

At poller startup, the same deterministic memory gate is applied to existing
`waiting` and `running` tasks across every configured account. A task is stopped
through `PUT /api/async/task/stop-create-contest-task` only when its own current
recursive repository size, multiplied by ModelHub's observed `1.20` overhead,
exceeds the known capacity of its selected GPU. The task ID is fetched and
stopped with the token belonging to that account, and its active state is
rechecked immediately before the mutation. Missing file sizes, unknown GPU
capacities, listing failures, and tasks that have already changed state all fail
closed and are never stopped. This does not match against another task from the
same model or infer failure from historical similarity. The cleanup repeats
every 120 poll cycles by default and writes its full evidence report to
`.modelhub_state/queue_cleanup_latest.json`.

Each account dynamically reserves its last 10 known-capacity positions for
models updated within seven days. If an account's discovered limit is 100, 200,
or 500, older models stop at positions 90, 190, or 490 respectively. Old-model
reservations are made under the same account lock as capacity reservations, so
concurrent submissions cannot enter the reserved suffix. Once every account's
old-model allowance is exhausted, discovery is capped at the seven-day window.
Unknown model timestamps are treated as old for new submissions. When capacity
probing raises an account's known limit, its old-model boundary moves with it.

On startup, the worker first stops deterministic OOM tasks, recalculates each
account's surviving task order by numeric task ID, and removes
older-than-seven-days tasks beyond that account's current limit minus 10. Later
scheduled cleanup uses limit minus 5, retaining a small hysteresis buffer that
avoids repeatedly over-pruning valid work. For a 100-task account the two
boundaries are 90 and 95; for a 500-task account they are 490 and 495. Recent
overflow tasks are always kept. ModelScope metadata failures fail closed and
never trigger cancellation. Immediately before mutation, task ownership, active
status, and post-OOM queue position are checked again.

The verified capacities, safe repository-size boundaries, evidence hierarchy,
and source links are recorded in
`docs/gpu-memory-capacity-2026-08-10.md`.

The online worker makes zero LLM calls. Candidate admission, GPU/framework
selection, queue cleanup, and failure feedback are deterministic and based on
repository metadata, platform capabilities, public outcomes, and explicit error
signatures. Merely storing a DashScope key in `.env` or setting a Qwen/LLM
environment variable does not activate inference. The standalone classifier
module remains only as an offline research helper for human-reviewed batches of
previously unseen errors; it is not wired into submission or outcome sync.

Outcome synchronization downloads a bounded set of failure archives for
submissions created by this worker (at most 40 per sync, four workers, three
download attempts). Deterministic signatures classify memory, repository layout,
context-length, storage, and platform faults. Unresolved runtime errors remain
explicitly ambiguous for later rule development instead of being sent to an LLM.
Signed log URLs remain in the ignored local outcome store and are removed after
classification.

The 12-account failure study and routing rationale are recorded in
`docs/failure-analysis-2026-08-10.md`.

ModelScope HTTP 429 responses use exponential backoff and `Retry-After`. Successful
pages remain cached, so a later cycle retries the failed page instead of restarting
the whole pagination scan.

## Adaptive Candidate Discovery

The configured recent window remains the fast path. If it contains no usable
model/GPU combinations, the same run progressively expands discovery to the last
7 days, the last 30 days, and finally older history (up to 3,000 models). Scanning
stops as soon as enough replacement candidates have been found.

Model verification results are reused for 15 minutes across poll cycles, and a
locally failed model/GPU pair cools down for 24 hours instead of being excluded
forever. The `[scan]` lines show every expansion stage, while `[daily] wave_done`
includes `skip_reasons` so an empty candidate pool is directly diagnosable.

Community deduplication is scoped to the exact model/GPU combination. A model
adapted on one GPU remains eligible for another GPU. Before each submission the
worker performs an uncached community check; a lookup failure defers the task
instead of failing open. Platform uniqueness rejections are persisted per
model/GPU in `.modelhub_state/submission_exclusions.jsonl` and are not retried.

## Concurrent Agents

The token pool keeps a local reservation for every in-flight submission, so a
lagging platform count cannot send all concurrent requests to the same account.
If another process fills an account first, the submission is retried immediately
against another account with capacity.

Workers that share a filesystem also coordinate model/GPU claims through
`.modelhub_state/submission_claims.jsonl`. Workers in isolated containers use
different candidate ordering (derived from `STRATEGY_ID`, instance ID, or
hostname), which reduces duplicate work while the platform remains the final
authority for account capacity and model/GPU uniqueness.

If the platform reports that a model/GPU is already being validated, the claim
is retained and the runner immediately draws replacement candidates from the
same scan instead of retrying the duplicate every cycle. Startup logs and the
health response expose `agent_version`; version `2026.08.02.3` or newer includes
duplicate replacement behavior, while version `2026.08.02.4` adds adaptive
candidate-window expansion and skip-reason reporting. Version `2026.08.02.5`
adds fail-closed model/GPU prechecks and persistent uniqueness exclusions.
Version `2026.08.04.1` adds queue/throughput intelligence, GPU health circuit
breaking, weighted-fair scheduling, live framework/config discovery, and
confidence-ranked public-plus-local framework selection.
Version `2026.08.05.1` removes self-funded GPU exploration, switches accepted
traffic to 70/30 long-term/recent exploitation, raises the public framework gate
to 300 samples, makes success dominate queue pressure, and adds recent local
GPU/framework circuit breakers.
Version `2026.08.10.2` adds evidence-backed sizing for every currently verifiable
GPU, recursive repository-size checks, deterministic failure-aware preflight,
and rate-limited lazy Qwen review for unresolved semantic cases.
Version `2026.08.10.3` selects `qwen3.7-flash` by default and recognizes the
repository root `.env` key named `dashscope` without logging its value.
Version `2026.08.11.1` adds account-owned cancellation of queued tasks that are
deterministically over the selected GPU's ModelHub memory boundary, with a
second active-state check and fail-closed handling for incomplete evidence.
Version `2026.08.11.2` removes LLM inference from every online path; credentials
alone cannot activate it, and unresolved cases remain available for offline,
human-reviewed rule development.
Version `2026.08.11.3` adds atomic per-account 80/7-day admission, prevents
history expansion once old-model positions are exhausted, performs a strict
position-80 startup cleanup, and relaxes scheduled age cleanup to position 95
after OOM cleanup and a second queue check.
Version `2026.08.11.4` replaces fixed queue positions with per-account dynamic
boundaries derived from each discovered capacity: limit minus 10 for admission
and startup cleanup, then limit minus 5 for scheduled dynamic cleanup.

## Deploy

Create a tag and submit the repository URL plus tag in "我的适配智能体".

```bash
git tag agent-v18
git push origin agent-v18
```