Retry Desk / API
Get a token

Driving Retry Desk from code

Everything the web app does is available over HTTP. The base URL is https://api.skillsafe.ai/v1/app-api, every request carries Authorization: Bearer <token>, and every response is the same envelope.

The task field comes first

This app has five lanes behind one endpoint. Every run body must carry a task field naming the lane - it is what the system prompt routes on. Send the wrong one and you get a valid package of the wrong kind; omit it and the model picks the closest lane and tells you which it chose.

One more shape trap: the run body is the input object. Do not wrap it in an {"input": ...} envelope - that returns 200 while hiding task from the model, which is the most confusing way this API can fail.

taskLaneFieldsSections returned
planDesign a retry policy the budget can pay forbrief, knownSummary, The Chain, The Numbers, Reasoning, Next Step
readWhat this chain actually costs, end to endsheet, worrySummary, Verdict, Findings, Corrected Chain, Next Step
budgetThe deadline: who spends it, and whose retries are decorationsheetSummary, Every Hop’s Worst Case, How The Budget Is Spent, Which Attempts Are Pure Load, Next Step
loadThe load: what one client call becomes, and when it arrivessheetSummary, The Count At Each Depth, When The Herd Arrives, What Jitter Would Change, Next Step
decideDecide what changes: an attempt count, the deadline, or the callsheet, fixedSummary, One Attempt Count Fixes, Only A Deadline Or A Key Fixes, Nothing Fixes, Next Step

Only task and the lane's own required fields are mandatory: sheet on read, budget, load and decide; brief on plan. Every field is a string - there are no number fields on this app. sheet is the chain sheet itself: a header of KEY: value lines and a CHAIN: block with one hop per line, outermost first.

BUDGET is the client's deadline - the point at which the caller gives up and everything after it is unpaid work; assumed 30s. PROPAGATE says whether that deadline shrinks as it passes down the chain; assumed no, which is what almost every real chain does, and it is the single largest lever on the sheet. CLIENTS is the concurrent client count, which turns a per-call fan-out into the load the innermost dependency actually feels; assumed 1000. BREAKER takes window=, threshold= and probes=.

A hop needs a timeout= and nothing else. Durations parse as 10s, 250ms, 1.5s or 2m. attempts= is the total number of tries, so attempts=3 is one call and two retries; retries= is accepted too and is cross-checked against it, because the two disagree by one and that off-by-one is the most common mistake on a real policy. backoff= is the first wait, factor= multiplies each subsequent one, cap= bounds them, jitter= takes none, full or equal, and idempotent= takes yes or no.

Order is the chain. The first hop is the outermost, so it spends the budget first and the last one is the one that runs out. A hop's worst case is attempts × timeout + sum(backoff), which is not its timeout, and the chain's is the sum of those - while the leaf-request count is the product of the attempt counts. Getting the order wrong changes every remaining-budget figure in the response at once.

Anything the reader cannot place is listed as a problem rather than skipped, and so is a hop named twice or one whose attempts= and retries= disagree. A hop that quietly vanished would change the fan-out, the worst case and every downstream hop's remaining budget in the response at once. A hop with no timeout= is an error: it has no worst case, so there is nothing to add.

A duration is reported as 10s, 250ms or 1.65m, a request count with thousands separators, a share as a percentage and a fan-out as the product written out. Every figure is the worst case rather than the typical one, computed from the pasted policy alone - no request is made to any hop, nothing is measured, and the failure rate of a dependency is not knowable from here.

Add $model to any body to choose the model for that run: gpt-5.6-luna, gpt-5.6-terra (the default) or gpt-5.6-sol. Luna caps output at 4,096 tokens and will fail the read, budget and load lanes rather than shorten them - a findings table, a corrected chain, or three tables with a row per hop or per retry, is several thousand characters before the reasoning starts.

The response envelope

Success and failure have the same outer shape, so one check covers both.

{
  "ok": true,
  "data": {
    "...": "the result"
  }
}
{
  "ok": false,
  "error": {
    "code": "VALIDATION_ERROR",
    "message": "seconds should be number, got string",
    "details": {}
  }
}
HTTPerror.codeWhat it means
400VALIDATION_ERRORThe body was not a JSON object, or a declared field had the wrong type. A number field sent as a string is the usual cause.
401UNAUTHORIZEDNo token, or a token that has expired or been revoked. Mint a new one.
402INSUFFICIENT_CREDITSThe balance is below the run's minimum. Call /estimate first and compare hold_credits against /me.
404NOT_FOUNDWrong path, or a job id that does not belong to this token.
409CONFLICTAn Idempotency-Key replay whose body differs from the original request.
429RATE_LIMITEDToo many requests. Back off; do not tight-loop.
503UPSTREAM_UNAVAILABLEThe model provider is unavailable. Retry with backoff.

1. Get a token

Open /tokens.html in a browser and copy the token this app already holds - no developer console needed. A guest token is minted automatically and is enough for /me and /estimate; writing a package is metered and needs a personal token, which comes from signing in on that page.

Keep it in an environment variable rather than in source:

export SKILLSAFE_TOKEN="YOUR_TOKEN"

2. Check the session and the balance

GET /me is free. It returns only three fields: subject_type, subject_id and credits. Signed-in means subject_type == "user" - there is no username or email to test.

curl -sS -X GET "https://api.skillsafe.ai/v1/app-api/me" \
  -H "Authorization: Bearer $SKILLSAFE_TOKEN"

3. Price the run before making it

POST /estimate costs nothing, creates no job, and returns the worst-case cost. Compare hold_credits against the balance from step 2 before you submit: a 402 after the fact is avoidable. hold_credits is a reservation priced at the full output cap - the actual charge is usually far lower.

It also echoes model, model_alias and markup_bps, which is the authoritative check that a run is bound to the model you think it is. Estimate each lane separately: their prompts and caps differ, so their holds do.

curl -sS -X POST "https://api.skillsafe.ai/v1/app-api/estimate" \
  -H "Authorization: Bearer $SKILLSAFE_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
  "task": "read",
  "sheet": "<BUDGET, PROPAGATE and a CHAIN block, one hop per line; the grammar is in /llms.txt>",
  "worry": "inventory sees a load spike every time anything upstream has a bad minute",
  "rules": "<the working rules for this lane, sent by the app>"
}'

4. Write a package

POST /run submits the job. Always send an Idempotency-Key: a network blip that replays the same request must not bill twice. A replay with the same key returns the stored result and is not charged again; a replay with the same key but a different body is a 409.

The response carries output.output (the Markdown package), charged_credits and truncated. If truncated is true the balance sat between min_credits and hold_credits and the output was cut short - render what arrived and say so rather than presenting it as complete.

curl -sS -X POST "https://api.skillsafe.ai/v1/app-api/run" \
  -H "Authorization: Bearer $SKILLSAFE_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
  "task": "read",
  "sheet": "<BUDGET, PROPAGATE and a CHAIN block, one hop per line; the grammar is in /llms.txt>",
  "worry": "inventory sees a load spike every time anything upstream has a bad minute",
  "rules": "<the working rules for this lane, sent by the app>"
}'

5. Stream a run

POST /run-stream is the same call with a text/event-stream response. Worth knowing before you build on it: from a server or from cURL you get event: delta frames carrying the output token by token; from a browser you get event: tick heartbeats and then one event: done with the whole output. Handle both, and treat ticks as liveness rather than progress.

Frame types are job (the job id), delta ({"text": "..."}), tick ({"t": seconds}), done, and error. An idempotent replay returns plain JSON with no stream at all, so check the content type before you start reading frames.

curl -sS -N -X POST "https://api.skillsafe.ai/v1/app-api/run-stream" \
  -H "Authorization: Bearer $SKILLSAFE_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -H "Idempotency-Key: cbd-$(date +%s)" \
  -d '{
  "task": "read",
  "sheet": "<BUDGET, PROPAGATE and a CHAIN block, one hop per line; the grammar is in /llms.txt>",
  "worry": "inventory sees a load spike every time anything upstream has a bad minute",
  "rules": "<the working rules for this lane, sent by the app>"
}'

6. Read the result

output.output is Markdown in the envelope this app's system prompt guarantees: every section is a level-two heading spelled exactly as listed in the lane table above, in that order; tables are GitHub pipe tables with the declared columns; prompts are in fenced blocks opened with three backticks and the word text; checklists are - [x] lines.

So parsing is a split on /^## / - but do it fence-aware, because a prompt block can legitimately contain a line starting with ##. Count the sections you got against the ones the lane declares: a short list means the run was truncated, not that the contract changed.

def sections(md):
    out, name, buf, fence = {}, None, [], False
    for line in md.split("\n"):
        if line.lstrip().startswith("```"):
            fence = not fence
        if not fence and line.startswith("## "):
            if name:
                out[name] = "\n".join(buf).strip()
            name, buf = line[3:].strip(), []
            continue
        if name:
            buf.append(line)
    if name:
        out[name] = "\n".join(buf).strip()
    return out

The artifact most callers want is the fenced text block inside ## The Sheet or ## Corrected Sheet - that is a complete sheet in the grammar above, so it can be fed straight back into another lane with nothing carried alongside it. Every other section is prose and tables meant to be read.

Rate limits and good manners