dsh-cooldown-retry
中文文档 → README.zh.md
Patient auto-retry for DeepSeek Harness. When an upstream model gateway answers with 429 Too Many Requests plus a retry_after_seconds hint — "All upstream providers are cooling down. Please retry after 28 seconds." — this plugin waits the delay the upstream actually asked for and retries, instead of letting the built-in llm-retry burn two fast attempts (500 ms → ~1 s) inside a 28-second cooldown window and then fail the turn.
It also absorbs one specific gateway bug seen in the field: during saturation a shared gateway can answer an ordinary request with a 400 whose message says "Audio modality is not supported" — although the request contained no audio at all — cache that body in a client-error circuit, and then replay the same 400 for every later request in the window, failing turn after turn. Once that provider has shown a capacity cooldown moments earlier, this plugin treats such a body as circuit replay and waits it out, on its own separate budget.
The problem
Shared LLM gateways (campus networks, self-hosted proxies, provider pools) frequently answer with a capacity 429 rather than a connectivity error: the pool is momentarily exhausted or its circuit breaker is open, and the response tells you exactly how long to wait.
DeepSeek Harness can already act on that hint — but only through a narrow path:
llm-retry does honour failure.providerRetryAfterMs when the field is present (packages/llm/llm-retry/src/index.ts). It waits that long verbatim, up to policy.maxDelayMs, and delegates to the next listener above that cap.- The field is populated by adapters that parse the HTTP
Retry-After response header. llm-deepseek does.
llm-pi-ai does not parse it at all — a full-package search finds no retry-after handling outside a single test fixture. Behind that adapter the field is never set.- So when the only hint is the prose inside the error body, nothing in DSH reads it.
llm-retry falls back to its local backoff (~500 ms → ~1 s, 2 attempts) and the turn dies inside a 28-second cooldown — even though waiting 28 seconds would have succeeded.
Reading the hint out of the message and then waiting as long as the upstream asked is the gap this plugin fills.
What it does
| built-in llm-retry (no usable hint) | dsh-cooldown-retry |
|---|
| Attempts per turn/provider | 2 | 5 |
| Backoff | fixed ~500 ms → 1 s | the upstream's own retry_after_seconds |
| Delay window | — | clamped to 1 s … 300 s |
| Abort-aware | — | yes — a cancelled turn ends the wait immediately |
| Hints parsed | — | providerRetryAfterMs, retryAfter, retry_after_seconds, retry_after_ms, prose retry after N |
It takes over only two shapes of failure: a capacity cooldown that carries a retry hint, and — with mislabeledClientError (on by default) — a 400 that claims a modality "is not supported" and arrives within mislabeledEvidenceMs of a capacity cooldown on the same provider. Everything else is handed to the next listener with next(), so auth errors, network errors, hint-less 429s, and genuine client errors keep their built-in behavior.
The budget is counted per turn and provider, not per step: a multi-step turn that trips the same cooldown in every step shares one budget instead of getting a fresh five in each. A new turn resets it.
Install
As a DSH bundle (recommended)
dsh plugin --profile web add github:dboycht/dsh-cooldown-retry
dsh plugin add installs the package into the profile, registers the bundle layer declared by dsh.bundle.patch, and the row mounts on the next config reload — no restart needed. It is permanent by design and survives upgrades. (Under the hood it forwards to pnpm in the profile directory, so it needs git on PATH for a github: specifier — the same requirement DSH itself already has.)
Verify it mounted — ask DSH for the composed tree and look for the row:
dsh --profile web --dump-config | grep -A2 'cooldown-retry'
# == dsh-cooldown-retry
- id: cooldown-retry
name: dsh-cooldown-retry
Its position matters: the row must appear after the built-in llm-retry row. agent/request-error is a waterfall and later-registered listeners bind as inner listeners, so this plugin gets first refusal on each failure instead of being short-circuited by the built-in fast retry.
Without installing a package
Point a row straight at the file. A new row requires insert, and name must be a file:/// URL — see Gotchas.
- insert:
- id: cooldown-retry
name: "file:///absolute/path/to/dsh-cooldown-retry/index.js"
Put that in your home-level $DSH_HOME/cordis.patch.yml (every profile) or a profile's own cordis.patch.yml (one profile). Both layers are watched, so saving applies it live.
Configure
Defaults are maxRetries: 5, minDelayMs: 1000, maxDelayMs: 300000, acrossSteps: true, hintlessBackoff: false, hintlessBaseDelayMs: 5000, mislabeledClientError: true, mislabeledMaxRetries: 3, mislabeledBaseDelayMs: 15000, mislabeledEvidenceMs: 600000. Override them from your own patch layer:
- id: cooldown-retry
config:
maxRetries: 10
maxDelayMs: 600000
A non-insert patch replaces the targeted row's whole config, so restate every key you want to keep. Unusable values are ignored in favour of the defaults, and an inverted window is repaired rather than accepted. maxRetries and mislabeledMaxRetries are capped at 100.
| Key | Meaning |
|---|
maxRetries | Patient retries per turn/provider before handing the failure downstream. |
minDelayMs / maxDelayMs | The window every wait is clamped into. |
acrossSteps | true (default) shares one budget across the steps of a turn; false restores a budget per turn/step. |
hintlessBackoff | Also take over a capacity cooldown that carries no hint, using hintlessBaseDelayMs * 2^(attempt-1). Off by default: hint-less throttling is indistinguishable from a permanent capacity problem, so retrying it is a policy change, not a bug fix. |
hintlessBaseDelayMs | The first hint-less backoff step. |
mislabeledClientError | true (default) also waits out a 400 that claims a modality "is not supported" while the request carried no such thing — the client-error circuit a saturated gateway creates and then replays. Set false to let every 400 reach the caller untouched. |
mislabeledMaxRetries | Retry budget for those replays, separate from maxRetries: waiting out a circuit must not compete with the 429 budget. |
mislabeledBaseDelayMs | Base delay for a replay that carries no hint of its own (retry_after_sec in a circuit body is used verbatim when present). Doubles per attempt. |
mislabeledEvidenceMs | How long a capacity cooldown keeps counting as evidence. Outside this window the same 400 is treated as a genuine client error and delegated. |
Observability
Every wait and give-up goes through ctx.logger under the cooldown-retry name, so it shows up with the rest of the app's logs instead of a bare console.log:
[cooldown-retry] upstream cooling down for nuaa (turn 3 step 2): retrying in 28000ms (1/5)
The countdown card in the UI
Each wait also writes a durable record before it starts and closes it when the wait finishes, matching the built-in llm-retry exactly:
| Record | When | Fields |
|---|
llm/retry | before the wait | retryId turn step provider policyKey retry maxRetries delayMs failure |
llm/retry-started | after a completed wait | retryId turn step retry |
The console's chat view therefore renders this wait through its own model-retry card: a live countdown plus the attempt number — "retrying in about 30 s (1/5)". That is what a patient wait looks like in the UI, with no client code at all. An aborted wait leaves only the scheduled record, because that attempt never started.
The two policies own separate chains (policyKey cooldown-retry/capacity and cooldown-retry/mislabeled), and the numbering is derived from session history rather than an in-memory counter, so a plugin remount or a resumed session continues the chain instead of restarting at 1. dsh-llm-retry's own invariant validates these records and rejects a malformed one; this repository pins the shape in test/journal.test.js and a contract-check script drives the real invariant.
In the composer, /cooldown-retry prints the counters:
cooldown-retry — budget per turn, max 5 retries, window 1000–300000ms, mislabeled-400 takeover on (max 3)
this turn: retries 3 | waited 84s | gave up 1 | no-hint capacity failures 2 | mislabeled 400 retries 1 | nuaa×4 | last: nuaa 15.0s (1/3) — mislabeled 400
lifetime: retries 11 | waited 4m 12s | gave up 2 | no-hint capacity failures 7 | mislabeled 400 retries 3 | nuaa×10 deepseek×2
The command is registered only when the commands service is present. A context without it — headless, ACP, a partial profile — still mounts and still retries; only the command is missing. That is deliberate: making commands a hard inject dependency would park the whole plugin in waiting over a cosmetic command.
Uninstall
dsh plugin --profile web remove dsh-cooldown-retry
Then drop the - id: cooldown-retry override from your patch layer if you added one.
Optional: floating countdown badge
The countdown is already a capability of the permanent row: every wait writes an llm/retry record that the console renders as its model-retry card (see "The countdown card in the UI" above). What follows is therefore a cosmetic add-on.
dynamic/ holds a two-half version for DSH's dynamic Cordis plugin path. It adds a small floating badge at the bottom of the window — "cooling down, retrying in 23 s (1/5)" — while a retry is pending.
It relies on harness.handle / host.call, which exist only for dynamic plugins, so it cannot be shipped as a bundle row and it is erased by a process restart. Treat it as a cosmetic add-on beside the permanent row, never as a replacement.
Gotchas
Both of these fail silently. Read them before wiring up any DSH plugin row by file path.
1. A new row needs insert
A non-insert patch entry means "override the row with this id". Writing - id: cooldown-retry together with name: is therefore read as an override of a row that does not exist: DSH prints patch: entry "cooldown-retry" not found to stderr and skips it. Your config looks correct, no error surfaces in the UI, and nothing is mounted.
# ✗ silently does nothing
- id: cooldown-retry
name: "file:///path/to/index.js"
# ✓
- insert:
- id: cooldown-retry
name: dsh-cooldown-retry
2. A file row's name must be a file:/// URL
name | Result |
|---|
D:/path/to/index.js | ✗ ERR_UNSUPPORTED_ESM_URL_SCHEME — Node reads D: as a URL scheme |
./path/to/index.js | ✗ resolved against the profile directory, not the patch file's directory |
file:///D:/path/to/index.js | ✓ |
The loader treats only names starting with . as relative (against its baseUrl, which is the profile directory). Everything else goes through a bare import(), which on Windows accepts only a well-formed file:// URL.
How it works
agent/request-error is a waterfall event: listeners run outermost-first in registration order, and the first one that returns { kind: 'retry' } without calling next() vetoes the rest of the chain, including the built-in behavior.
So running two copies — say the bundle row plus a dynamic plugin — does not double your retries. One of them owns each failure and the other is short-circuited. The registered-first copy wins.
Development
git clone https://github.com/dboycht/dsh-cooldown-retry && cd dsh-cooldown-retry
npm test # node --test — no dependencies, no build step needed
The suite lives in test/, which is not part of the published/installed package (the files whitelist ships only what a running DSH needs), so clone the repository rather than running this inside node_modules.
The retry-decision helpers — extractDelayMs, isCapacityFailure, clampDelay, planDelay, counterKey, resolveOptions, createStats, formatStats — are exported and unit-tested; apply() is thin Cordis wiring around them, and the suite drives it through a stub context so what it owns, what it delegates, and how it counts are all covered.
The stub context is a stand-in, not proof: npm test runs under node --test, which spawns a child process per file, so on a locked-down machine (or under a sandbox that forbids piped stdio) it can fail with spawn EPERM even though the suite is green. node test/retry.test.js runs the same tests in-process and is the fallback.
inject deliberately lists only timer. commands and logger are read through ctx.get / ctx.logger and degrade gracefully, and ctx.logger is used in its callable form — ctx.logger('cooldown-retry') — with a plain-object fallback, because Cordis's logger service is both callable and carries .info/.warn directly.
Why the hint patterns look like that
The obvious pattern, /retry-?after/, matches retry-after and nothing else — not retry_after (underscore) and not retry after (space). An implementation built on it declines the very error it was written for, and it does so silently: an unmatched hint is not an error, the plugin just calls next() and the built-in retry gives up exactly as before.
So every separator here is [_\s-]*, the unit-qualified prose pattern runs before the bare one (otherwise after 1500 ms reads as 1500 seconds), and each case has a regression test built on the real gateway message.
License
MIT © 2026 dboycht
