Skip to content

ci_tips.md

Latest commit

 

History

History
820 lines (619 loc) · 29.9 KB

ci_tips.md

File metadata and controls

820 lines (619 loc) · 29.9 KB
Raw

Make CI green fast

This is a playbook. Use it when a pull request is red and you want the next push to pass. It comes from current workflow files and from many CI-fix sessions in this repo.

Source of truth for status: gh pr checks, not gh run list. gh run list misses required checks that are not GitHub Actions jobs.

CI repair note, 2026-09-30: CI was already red during the Null consensus release work (PR #3206). Compare failed checks with their base-branch logs and tested SHAs before choosing a fix. Existing failures, regressions, and transient infrastructure problems need different remedies, but none is a reason to ignore a failed release check.

Always finish a local pass with:

git diff --check
git status --short

Do not run scripts/fix_rust.sh during a normal CI loop. That script creates commits.

1. The 10-minute loop

Do this in order. Fix failures as they appear and keep watching the remaining jobs, including slow checks and infrastructure retries, until CI is green.

  1. Confirm GitHub is testing the commit you think it is. Local fmt/clippy that you did not push does not count.
  2. Run scripts/preflight.sh (see §3). It scopes itself to the files you changed and stops at the cheap failures first.
  3. Push the validated fix immediately after local preflight passes; do not wait for the rest of the remote suite to finish.
  4. Watch with fail-fast. When any check fails, inspect its logs and begin the fix immediately while other jobs continue. Restart the watcher after each push and repeat until the complete check set is green.
  5. Rerun only when the log is infra (429, GHCR pull/push, warp-proof, cancelled sibling).
  6. Do not push again just to "unstick" CI. A new push cancels in-flight Runtime Checks. Clone-upgrade and try-runtime then restart from zero (often 40+ minutes).
  7. Leave clone-upgrade and e2e to CI unless you already know they are the failure. Clippy and cargo test for changed crates are part of the local gate.
gh pr view --json number,url,headRefName,headRefOid
gh pr checks --json name,bucket,state,workflow,link
gh pr checks --watch --fail-fast
# After a failure:
gh run view <run-id> --log-failed

--fail-fast stops on the first failure. Other jobs may still be green or still running. Re-read the full check set after every push. The set can change.

gh pr checks is tab-separated. Names have spaces. To list all failures:

gh pr checks <N> | awk -F'\t' '$2=="fail"'

2. What "required" means here

A required check is a GitHub branch-protection name. If that name is missing or red, merge stays blocked.

A skipped job usually counts as passing. That is why path filters and labels exist. Do not add a new required check name without a skip-success path, or docs-only PRs get stuck on "Expected".

These names matter today:

Check name Workflow Usual local match
cargo fmt Check Rust cargo fmt --check --all
cargo clippy (default) Check Rust SKIP_WASM_BUILD=1 cargo clippy --workspace --all-targets -- -D warnings
cargo clippy (all) Check Rust same + --all-features
cargo test Check Rust leave to CI
bittensor-core wasm32 seam Check Rust Linux/wasm; flake on GitHub 429
zepter feature propagation Check Rust zepter run check
SDK offline checks Runtime Checks cd sdk/python && just check (almost)
Docs and website build Runtime Checks generate.py --check then website build
spec_version newer than mainnet Runtime Checks bump or label no-spec-version-bump
Sudo-upgrade mainnet clone and test Runtime Checks clone-upgrade fan-in; see below
cargo test (eco-tests) Eco Tests cd eco-tests && cargo test
typescript-formatting Typescript E2E cd ts-tests && pnpm run fmt
typescript-e2e-zombienet_evm Typescript E2E gate name; skip when EVM not selected
typescript-e2e-zombienet_shield Typescript E2E gate name; skip when shield not selected
skeptic / auditor ai-review required; often not a compile bug

cargo fix leaves no diff and no cargo check warnings run on main / manual, not on pull requests.

3. Local gates (run before every push)

One command runs every gate below with CI's exact flags, scoped to the files changed vs origin/main, and exits nonzero on any failure:

scripts/preflight.sh          # full gate (builds the release node when runtime metadata may have changed)
scripts/preflight.sh --fast   # skip the node / wasm builds
scripts/preflight.sh --all    # every gate regardless of the diff
scripts/install-hooks.sh      # pre-push hook: a red gate aborts the push

Install the hook once per checkout. It gates the pushed commits in a detached worktree (preflight.sh --rev SHA), so uncommitted fixes do not count and git push origin other-branch validates other-branch. git push --no-verify is forbidden for agents. Pushes go out as unarbos; the gate verifies who owns the push credential for the destination remote via api.github.com/user and prints the git remote set-url fix when it is not unarbos. Build artifacts are resolved from Cargo's target dir (--print-artifacts shows the paths); scripts/test-preflight-rev.sh is the regression for the --rev path.

The rest of this section explains what each gate runs and how to fix it. Match fix-mode commands to the files you own.

Rust source, manifests, fixtures

cargo fmt --check --all

If it fails only on files you own:

cargo fmt --all

Then inspect the diff. Typical hit: import order. rustfmt wants use frame_support::storage::{TransactionOutcome, with_transaction} on one line, or use super::*; above other imports.

just clippy is not the CI job. CI is:

SKIP_WASM_BUILD=1 cargo clippy --workspace --all-targets -- -D warnings
SKIP_WASM_BUILD=1 cargo clippy --workspace --all-targets --all-features -- -D warnings

Root just clippy only denies todo and unimplemented. CI denies all warnings. Unused imports fail CI and pass just clippy.

scripts/preflight.sh runs both whenever a Rust path changed, plus zepter run check and cargo test --all-features for the changed crates (and runtime/tests/fee_baseline.rs + claim_root_weight.rs when pallets/ or runtime/ changed).

cargo clippy -p pallet-subtensor --lib misses test-only denies. A new test module usually needs the same crate-level allow as the other pallet tests:

#![allow(clippy::expect_used, clippy::unwrap_used, clippy::indexing_slicing)]

Do not touch a .rs file to "clear" a missing use. If Clippy wants WeightInfo in scope, add the import.

Python SDK

From sdk/python, when the locked env exists:

uv run --no-sync ruff check .
uv run --no-sync ruff format --check .

Use uv run ruff, not ruff on PATH. A system or stale .venv binary will report lints (for example UP045) that CI's locked ruff does not, or the reverse. CI pins ruff through uv sync --python 3.14.

Lint the whole SDK. Formatting one file (query.py) while CI scans everything will bounce.

Fix only task-owned files:

uv run --no-sync ruff check . --fix
uv run --no-sync ruff format .

CI SDK offline checks also runs (from sdk/python):

uv sync --python 3.14 --locked --all-extras --dev
uv run ruff check .
uv run ruff format --check .
uv run ty check --exit-zero-on-warning bittensor
uv run pytest
uv run python -m codegen.check --coverage
uv run python -m codegen.check --names
uv run python scripts/export_beta_baselines_rs.py --check

just check in sdk/python also runs --units and --namespaces. CI does not. A local just check failure on those two is not always a CI failure.

Generated docs

From sdk/python:

uv run --no-sync python ../../website/apps/bittensor-website/scripts/generate.py --check

If the drift is from this change:

uv run --no-sync python ../../website/apps/bittensor-website/scripts/generate.py

Repeat --check. Inspect the exit code. --check can print a long file list and still exit 0. A list is not always a fail.

Order that avoids the fmt/regen race:

  1. Finish Rust and Python source.
  2. cargo fmt --all
  3. generate.py then generate.py --check
  4. Do not fmt again after regen. Line anchors in MDX and reads.json will move and the gate goes red again.

Do not hand-edit docs/tx/, docs/query/, docs/errors/, generated hyperparameter pages, or website/apps/bittensor-website/public/catalog/.

If --check fails on tens of unrelated pages, stop. That is leftover WIP, not your one-line change. Do not dump a 50-file regen into someone else's PR.

MDX pages you do write by hand need string title and description frontmatter. Missing those fail the website build after the drift gate.

TypeScript tests

From ts-tests, when the locked env exists:

pnpm run fmt

Fix: pnpm run fmt:fix.

EVM ABI

Canonical files: precompiles/src/solidity/*.abi. Vendored copies: sdk/python/bittensor/evm/abi/*.json.

From sdk/python:

uv run --no-sync pytest tests/unit/test_evm.py::TestVendoredAbiSync -q

Update the vendored JSON only when the canonical ABI changed.

4. Failure cookbook

Each row is a failure we keep hitting. Symptom first. Then the fix.

cargo fmt

Symptom: Diff in path/to/file.rs. Often import order or a long iterator chain that rustfmt wants on one line.

Fix: cargo fmt --all. Commit only the rustfmt diff. Re-run cargo fmt --check --all.

Traps:

  • Do not edit .rustfmt.toml. Almost every option is commented on purpose. Turning options on reformats hundreds of pallet files.
  • vendor/frontier prints imports_granularity warnings. Stable rustfmt ignores them. They do not fail the job.
  • rustfmt can ICE and still exit 0. CI greps the output for panicked at / internal compiler error.
  • cargo fmt -p pallet-subtensor is not enough. CI is --all.

cargo clippy (default) / cargo clippy (all)

Symptom: unused import, clippy::expect_used, indexing, or a compile error. -D warnings turns every warning into a hard fail.

Fix: smallest code change. Then run both Clippy jobs. Default green does not mean --all-features is green.

Traps:

  • One compile error in pallet-subtensor turns Clippy, cargo test, Docker, e2e node builds, and clone-upgrade red at once. Fix the compile first. Do not chase the cascade.
  • just clippy will lie to you. Use the CI flags.
  • Build-script cargo:warning lines (missing benchmarks, freeze_struct) are not Clippy denials. Ignore them.
  • Pallet --lib Clippy misses test modules. Add the crate-level allow shown above, or rewrite the lint.

SDK offline checks

Symptom: ruff format would reformat query.py (or similar); codegen.check --names / --coverage fail; pytest fails.

Fix:

cd sdk/python
uv run ruff format .
uv run python -m codegen.check --coverage
uv run python -m codegen.check --names
# after a read-registry change:
uv run python -m codegen.emit_namespaces

Traps:

  • Do not hand-edit sdk/python/bittensor/_generated/. Codegen headers have no trailing comma after Spec version: N. A hand edit of the catalog will fail clone-upgrade even if ruff is clean.
  • Do not advertise storage that the runtime does not have. Reads for TotalAlphaStaked or leftover DeferredRootAlphaDividends failed when the chain had no such item.
  • Do not regen from a local --dev / pow-faucet node. That node adds SubtensorModule.faucet. CI's production release binary does not. You will then fail --coverage (raw-only entries no longer on chain: ['faucet']) or the reverse.
  • Do not use system ruff. Use uv run ruff.
  • Do not let a local venv's Python leak into PyO3 / binding builds. See the pyo3 trap under cargo test.

Docs and website build

Symptom: Docs drift gate (generated tx/query/errors pages) and a list of stale files. Often docs/tx/claim-root.mdx or line anchors in reads.json after a Rust edit that shifted source lines.

Fix: regenerate with generate.py, then --check. Commit the generated pages.

Traps:

  • Fmt then regen, or regen then fmt, can leave stale line anchors. Run generate after rustfmt is final. Do not fmt again.
  • --check can print many paths and still exit 0. Read the header and the exit code.
  • Broad unexpected drift that fails: stop and report. Do not "make CI green" by committing unrelated catalog churn.
  • The website build (yarn turbo run build --filter=@raofoundation/bittensor-website) is a second step. Drift can be clean and the Next build still fail on missing MDX title / description.

spec_version newer than mainnet

Symptom: local spec_version in runtime/src/lib.rs is less than or equal to mainnet.

Fix (only if the user asked): bump to mainnet + 1, or add the no-spec-version-bump label.

Check mainnet:

curl -sS -H 'Content-Type: application/json' \
  -d '{"id":1,"jsonrpc":"2.0","method":"state_getRuntimeVersion","params":[]}' \
  https://entrypoint-finney.opentensor.ai \
  | jq -r '.result.specVersion'

The file must contain exactly one spec_version: <n>, literal.

Trap: do not bump or label unless asked. A bump on a non-release PR fights every other open PR. Two in-flight PRs must not claim the same number. A missing bump on a release PR means the release train builds and then no-ops at every deploy guard.

clone-upgrade (pristine|remaining|combined) and

Sudo-upgrade mainnet clone and test

What it is: CI builds a Linux release node, sudo-upgrades a copy of real mainnet state with your runtime, then runs clone regressions and (when relevant) SDK metadata drift against that upgraded chain.

Phases:

  • Snapshot available: pristine and remaining run in parallel.
  • Live scrape or fresh-mainnet-clone label: one combined job.
  • pristine does not run the SDK drift gate. It can pass while _generated/ is stale. remaining (or the end of combined) is the metadata check.
  • remaining / end of combined runs uv run python -m codegen.check --drift ws://127.0.0.1:9944 when the path classifier sets sdk_drift.

The required name is Sudo-upgrade mainnet clone and test. That is a fan-in job. If a sibling clone job is failure, this gate dies in about 3 seconds with CLONE: failure. That is not a second product bug. Open the inner clone-upgrade (remaining) log.

If clone is skipped because the build failed, the gate is red. If clone is skipped because of skip-clone-upgrade or a docs-only PR, the gate is green. GitHub treats skipped as a pass. That label hides the check; it does not fix drift.

Common real failures:

  1. Generated SDK drift. Committed _generated/ does not match the upgraded production runtime. Regen from the CI Linux release node, not a Mac --dev node:

    gh run download <run-id> -n 'node-subtensor-release-<sha>' -D /tmp/release-node
docker run -d --name st-meta --platform linux/amd64 -p 9944:9944 \
  -v /tmp/release-node:/node:ro ubuntu:24.04 \
  bash -c 'apt-get update -qq && apt-get install -y -qq ca-certificates >/dev/null && \
    exec /node/node-subtensor --chain local --tmp --alice --validator \
    --rpc-external --rpc-cors all --rpc-methods unsafe --rpc-port 9944 \
    --unsafe-force-node-key-generation'
# Confirm specVersion == this PR before codegen. A stale
# scripts/specs/local.json can still serve an OLD genesis.
curl -sS -H 'Content-Type: application/json' \
  -d '{"id":1,"jsonrpc":"2.0","method":"state_getRuntimeVersion","params":[]}' \
  http://127.0.0.1:9944 | jq -r '.result.specVersion'
cd sdk/python
uv run python -m codegen ws://127.0.0.1:9944
uv run python scripts/record_golden.py ws://127.0.0.1:9944
uv run python -m codegen.check --drift ws://127.0.0.1:9944

    --dev is safe only if the binary is that CI release artifact (no pow-faucet). A local Mac --dev build is the faucet trap.

  2. Hand-edited _generated/. Trailing commas, extra storage items, leftover faucet, or merge-conflict leftovers. Re-run codegen. Do not patch the generated files by eye.

  3. Regression assert. Behavior on real mainnet state changed (locks conviction after transferStake, balancer, emission). Reproduce with docs/internals/mainnet-clone.mdx. Run the named script in clones/js-tests (npm run test:locks-conviction, …). A cancelled run is not a pass.

  4. Cancelled sibling / 3-second gate. --fail-fast, a cheap job, or a new push cancelled clone-upgrade. Fix the cheap job or wait for the new HEAD. Do not debug clone.

  5. Warp-sync infra. Log says Bad warp proof or Downloading finality proofs, 0.00 MiB. Nightly snapshot is stale and live scrape failed. Rerun --failed up to three times. Do not apply skip-clone-upgrade unless asked.

Traps:

  • Do not regenerate _generated/ by hand. You need a node on the new spec; the full scripts/preflight.sh builds one from this tree and regenerates for you. A Mac SKIP_WASM_BUILD=1 tree cannot produce that node.
  • Do not regen from --dev / pow-faucet. That adds faucet.
  • Do not add skip-clone-upgrade to hide a runtime/SDK change.
  • Apple clang on this machine often cannot target wasm32. Leave full runtime wasm and the release node to CI.

Local clone steps (slow; only when clone is the real failure): see docs/internals/mainnet-clone.mdx.

try-runtime (devnet / testnet / mainnet)

What it is: replay on_runtime_upgrade against a network state snapshot. Cached snapshots are the default. Label fresh-try-runtime-state to scrape live.

When it fails: a migration panics on real storage, or a new migration has no pre_upgrade / post_upgrade. Unit tests will not catch this. Put the migration in the runtime Migrations tuple and add those hooks.

If all three networks go red together, look for a compile break first (a vendored crate), not three migration bugs.

Local Mac cannot build the try-runtime wasm. Leave it to CI.

skeptic / auditor

What they are: required AI review jobs. skeptic is read-only. auditor may run scripts/fix_rust.sh and push chore: auditor auto-fix on non-fork PRs.

Read the verdict text, not the red X. Findings live in the PR comments and the skeptic-output artifact. The job log is often empty.

gh api repos/RaoFoundation/subtensor/issues/<N>/comments
gh run download <run-id> --repo RaoFoundation/subtensor -n skeptic-output

Recurring verdict: VULNERABLE on weight accounting. The bot flags dispatchables whose weight does not pay for new loops or storage growth (BasketClaimed, per-hotkey drains, dissolve cursors, root_register walking every subnet).

When to fix:

  • New unbounded walk or unpaid work this commit added.
  • Real economic loss (for example recycled credits on eviction).
  • Secrets inherited by npm install, or unbounded PR archive extract. Do not waive those. Auto-merge stays stuck until SAFE.

When not to thrash:

  • Same HIGH weight findings already on the PR, unchanged. Docs regen and rustfmt do not create them.
  • HTTP 406 / “PR exceeds 300 files” / merge-base missing: infra. The workflow reads prefetch scripts from the base branch. A prefetch.sh fix on the PR head does nothing until it is on main. Do not rewrite pallet code for a 406.
  • Fork PRs fail on auto-trigger. A nucleus member must gh workflow run ai-review.yml -f pr_number=<N>.
  • User said "green ignoring skeptic": keep fixing every other required check. Do not invent weights.
  • Auditor 👎 “missing tests” or a docs wording nit is informational unless that check itself is required. Skeptic can be SAFE while Auditor is 👎.

Do not invent WeightInfo numbers. CI measures on reference hardware (threshold cited: 40%) and uploads a bench-patch artifact. Label apply-benchmark-patch commits that patch, or download it and apply it yourself. Label run-benchmarks turns on validate-benchmarks (not required today).

bittensor-core wasm32 seam

Symptom: checkout download 429, or a real wasm compile break.

Fix: if the log is GitHub rate limit, gh run rerun <id> --failed. If the log is a compile error in bittensor-core / bittensor-core-wasm, fix the portable core. This job exists because Clippy on native does not link the wasm target.

Mac: skip this locally. CI builds it on Linux.

typescript-formatting / TS e2e gates

Formatting: cd ts-tests && pnpm run fmt. Required even when e2e shards skip.

E2E flakes (rerun once):

  • GHCR pull denied: requested access to the resource is denied
  • GHCR push unknown blob after a successful build
  • GitHub checkout 429
  • intent_multisig_cancel or stake-movement amounts that pass on siblings
gh run rerun <run-id> --failed

If the same test fails twice on this commit, it is a real regression.

Real e2e bugs (fix the test or the step):

  • NextKey is None ~17–20s after the first block → wait/retry in the test. Not a LOCALNET_IMAGE rename.
  • find-e2e-tests expected count drifted after a test was removed (for example 113 → 112). Update the count. Do not add a dummy test.
  • Workflow still calls a deleted Python e2e path (pytest tests/e2e/test_reads.py file not found). Delete or retarget the step. Keep the metadata drift gate.

Traps: ignore a LOCALNET_IMAGE / LOCALNET_IMAGE_NAME hint when 98/100 tests already hit the chain. Read the real assertion.

Bittensor E2E Test / find-e2e-tests

Label skip-bittensor-e2e-tests skips the Rust SDK e2e suite.

cargo audit

Label skip-cargo-audit skips it. Do not add advisory ignores just to go green. Do not refresh Cargo.lock unless the lockfile change is the task.

cargo test (workspace)

Leave to CI unless you know the failing test name. Then:

SKIP_WASM_BUILD=1 cargo test -p pallet-subtensor <test_name>

CI uses cargo nextest run --workspace --all-features plus cargo test --doc --workspace --all-features.

py-sp-core / PyO3 traps:

  • undefined reference to Py_* under --all-features is the crate feature extension-module. Remove that crate feature. Maturin already sets features = ["pyo3/extension-module"]. Do not exclude the crate from --all-features.

  • “Need python3-dev” is often a misread. Fix the first compile error. py-sp-core may never have been reached.

  • After source .venv/bin/activate, cargo/pyo3 can cache the venv’s python3.10. deactivate in a subshell is not enough:

    unset VIRTUAL_ENV
# take .venv off PATH, then:
hash -r
cargo clean -p pyo3-ffi

    Rebuild with system Python. This is local. CI Linux is usually fine.

5. Labels that change CI

Use a label only when the user asks, or when the failure is the label's own gate.

Label Effect
no-spec-version-bump Skip spec_version vs mainnet
skip-clone-upgrade Skip clone-upgrade; required fan-in still passes
fresh-mainnet-clone Live scrape; one combined job
fresh-try-runtime-state Live RPC instead of cached try-runtime snap
skip-cargo-audit Skip audit
skip-bittensor-e2e-tests Skip Rust SDK e2e
run-benchmarks Run validate-benchmarks on the PR
apply-benchmark-patch Apply the bot's weights.rs patch
check-node-compat Opt-in node compat harness
mainnet-clone Public tunneled clone (non-fork, base main)
auditor:run-node Auditor may start localnet (heavy)

red-team, blue-team, runtime, breaking-change do not skip checks.

6. Flake vs real

Rerun once when the log is one of these:

  • GitHub 429 on checkout (wasm32 seam, other jobs)
  • GHCR pull denied / error pulling image configuration
  • GHCR push unknown blob after a green build
  • Bad warp proof / Downloading finality proofs, 0.00 MiB
  • Job cancelled because a cheaper sibling failed, --fail-fast, or a newer push
  • Skeptic HTTP 406 / “PR exceeds 300 files” / empty verdict JSON
gh run rerun <run-id> --failed

If it fails the same way on the same commit, treat it as real.

7. Time-wasters (do not do these)

  • Run scripts/fix_rust.sh. It commits.
  • Push without running scripts/preflight.sh, or push with --no-verify.
  • Regenerate _generated/ without a node on this spec.
  • Hand-edit generated docs or _generated/.
  • Bump spec_version or apply labels without being asked.
  • Invent benchmark weights.
  • Refresh a lockfile to "make CI pass".
  • Merge latest main into the PR to hide an unrelated red check without checking that main is actually green for that check.
  • Debug clone-upgrade when cargo fmt is the first failure.
  • Treat just clippy as CI Clippy.
  • Push fmt/clippy that you never committed.
  • Push again to “unstick” CI (cancels 40+ min of clone/try-runtime).
  • Dump a broad generate.py regen when --check was already stale from other WIP.
  • Use system ruff instead of uv run ruff.
  • Regen _generated/ from a --dev / faucet node.
  • Treat pristine green as “SDK metadata is fine.”
  • Treat skeptic 406 as a pallet bug. Prefetch is read from base.
  • Add python3-dev for a compile error that never reached py-sp-core.
  • Exclude py-sp-core from --all-features instead of removing the extension-module crate feature.
  • Chase LOCALNET_IMAGE when the container already ran tests.
  • Assume an admin can force-merge main. enforce_admins: true still needs reviews and required checks.

8. Command card

# Local gate (before every push; the pre-push hook runs it for you)
scripts/preflight.sh --fast
scripts/preflight.sh          # full: node build + SDK drift + try-runtime when relevant

# Where are we?
gh pr view --json number,url,headRefOid,mergeStateStatus
gh pr checks --json name,bucket,state,link

# Watch
gh pr checks --watch --fail-fast

# Logs
gh run view <run-id> --log-failed
gh run download <run-id> -n skeptic-output
gh run download <run-id> -n 'node-subtensor-release-<sha>' -D /tmp/release-node

# Rerun infra (after the run has finished)
gh run rerun <run-id> --failed

# Fast Rust
cargo fmt --check --all
cargo fmt --all   # only after check fails on owned files

# CI Clippy (when that job is red)
SKIP_WASM_BUILD=1 cargo clippy --workspace --all-targets -- -D warnings
SKIP_WASM_BUILD=1 cargo clippy --workspace --all-targets --all-features -- -D warnings

# Fast Python (sdk/python) — use uv run, not system ruff
uv run --no-sync ruff check .
uv run --no-sync ruff format --check .
uv run --no-sync python -m codegen.check --coverage --names
uv run --no-sync python ../../website/apps/bittensor-website/scripts/generate.py --check

# Regen bindings only after a Linux release node on THIS spec is up
# uv run python -m codegen ws://127.0.0.1:9944

# Fast TS (ts-tests)
pnpm run fmt

# Zepter (when that job is red)
zepter run check
# fix: zepter run default   # inspect the diff; do not use fix_rust.sh

9. Path filters (why a job skipped)

Runtime Checks uses .github/scripts/classify-runtime-changes.sh on the trusted base SHA. Outputs: runtime, docs, python_sdk, sdk_drift, snapshot_ci.

docs=true for docs/*, website/*, sdk/python/*, and .github/workflows/runtime-checks.yml. That is why an SDK-only change still runs the docs drift gate.

Check Rust uses .github/scripts/classify-rust-changes.sh and .github/rust-ci-paths.txt. Non-Rust PRs get placeholder Clippy jobs so the required names still exist.

If the classifier is missing or the PR file list fails, the workflows enable every check. That looks like "why is clone running on my docs typo?"

10. Mac vs CI

Local preflight sets SKIP_WASM_BUILD=1 (root justfile does this too). CI builds real runtime wasm and a Linux node-subtensor on self-hosted Linux.

On this machine, Apple clang often cannot target wasm32. That is environment, not your diff. Do not spend time on a local wasm runtime build to "match CI".

Clone-upgrade and try-runtime need that Linux artifact. Reproduce them only after the cheap gates are green and the log shows a real migration, regression, or metadata-drift failure.

Null release exception (2026-10-01)

For PR #3206, the owner explicitly chose a shared Null capacity of 2,500 UIDs (divided by emission mechanism count) and accepted Skeptic as the sole allowed non-green review check. Keep its result and findings visible; do not suppress the check or describe it as passed. All other CI checks and the mandatory local preflight/pre-push gate must pass. Further reward-accounting optimizations are deferred; payouts remain atomic. This exception does not supply measured weights or establish that deployment resource limits have passed.

PoW challenge refresh validation (2026-10-02)

Competitive registration mining refreshes the latest published block challenge on a 12-second cadence. Chain admission accepts work with age 1 through 5 blocks; transaction-pool longevity reflects the remaining lifetime. The miner must not restart or poll the chain after every small GPU batch. A found proof is checked against the latest difficulty before submission. Preserve key/subnet binding, replay rejection, disjoint GPU nonce ranges and bounded cancellation.

Verify the 12-second refresh with a deterministic clock, exact accepted/rejected age boundaries, difficulty changes during mining, and live GPU registration on the rebuilt runtime. Delay a broadcast beyond the buffer to exercise bounded fresh-proof recovery; do not replay included or uncertain transactions. Raising the CLI mining timeout alone does not change on-chain proof expiry.

If the local Python-binding Rust tests fail to link -lpython3.12, check that the interpreter selected by PyO3 has its shared development library installed. Select an installed managed interpreter with that library using PYO3_PYTHON=/path/to/python3.12 scripts/preflight.sh, and preserve that setting for the pre-push hook. This is a local toolchain issue; do not disable the tests or change the binding code to hide it.