stage-11-agentics/glideslope/glideslope

Glideslope: position check

Run this, then relay the result:

glideslope

If glideslope is not on PATH, run it from the clone: python3 <path-to-clone>/glideslope.py.

Relay stdout verbatim as markdown. Never put it inside a code fence. The script already emits GitHub-flavored markdown. Fenced, it shows raw | pipes |. Unfenced, it renders as clean tables. Add nothing and reformat nothing: the script owns the format. Keep the meter badges and the honesty markers exactly as printed.

Flags

How to read it

• ◆ mark: where an even burn to the reset would put used right now. Used above it is ahead of budget. Used below it is banking capacity, which on a flat-fee plan means capacity that expires unused at the reset.
• Meter badges: the letter is the account, the outline is the window. Circle ⃝ weekly all models, triangle ⃤ 5h session, diamond ⃟ weekly Fable.
• Fable is Anthropic's model tier with its own weekly meter. It is often the binding weekly quota. none on a Claude row means the plan does not include it.
• ● name marks the machine a Claude account is logged into. ◦ name is a login held elsewhere. Only Claude rows carry it.
• Honesty markers: stale (last-known read), presumed (rolled over, nothing could have spent it), unread (unknowable), floor (the truth is this or higher), undecided (a pooled floor below the mark). Never restate one of these as a plain number.
• On remote seats, when present, names the coding agents running on cloud sandboxes and the account each bills: 6 Grok (Grok) · 1 Claude (Alpha). Mention it in a position check. That burn is already inside each account's meters, so it says who is spending, never extra usage; never add it to a number. A stale part is an old file, not the current fleet: say so.
• API-equivalent, when present, is a valuation at list API prices, never money spent. Say so if you summarize it.
• 💸 amount on a 5-hour or 7-day all-models cell is usage-credit spend during that window. Real money. The percent and the ◆ stay. The 5-hour amount is inside the 7-day amount. Absence means no known burn. floor after the amount means it understates. Relay it exactly. It is not the API-equivalent panel, and it is not on Fable.

Output order: login banner, On remote seats (only when seats are live), Weekly status, All windows, OpenRouter, API-equivalent (only if the user keeps a ledger), Codex reset banks (only when credits are banked), Recent switches. Relay all of it, in order.

API-equivalent spend

When asked for a spend estimate from local coding-agent records, run:

~bash glideslope-spend --print ~

If the command is not on PATH, run python3 <path-to-clone>/spend.py --print. It updates <store>/spend.json and prints combined 24-hour, 7-day, 30-day and all-time totals plus 7-day, 30-day and all-time rows for each Claude account, Codex, Grok and unattributed spend. Claude and Codex dollars use API list rates. Grok Build dollars use the costUsdTicks supplied by its local CLI, not the LiteLLM rate table. The totals are estimates, not subscription charges or an invoice. It covers local Claude Code, Codex and Grok Build only.

Where the config sets [spend] producer = false, another program owns <store>/spend.json, and this command would overwrite that ledger with this machine's records. Read the existing file instead, or point GLIDESLOPE_HOME at a scratch store.

Claude account totals use the strongest evidence each request has, and --print shows how much spend used each basis:

• transcript: the latest credential_org in the request's own transcript, joined to the

roster. A subagent inherits its parent's.

• unowned org: the transcript names an organization no single roster seat owns. It stays

unattributed; the record proves the machine's login did not pay.

• token: the satellite beacon saw the session running under CLAUDE_CODE_OAUTH_TOKEN and the

token's account is known. The beacon stores a fingerprint, never the token. A warning naming a fingerprint means claude-account whose <token-file> needs to run once where the token is.

• unseen token: no organization from Claude Code 2.1.281 or later, which always records one for

a login. A token nothing identified paid, so it stays unattributed.

• timeline: active-login samples and switch-log events, resolved through

glideslope.call_sign, for older versions only. It is the only basis that can name the wrong account.

• none: no earlier evidence, or evidence more than 24 hours old. It stays unattributed.

--print also shows what woke the agent over 7 days: dollars, share, turns and requests per trigger, with the top session. spend.json triggers has every horizon and the top five sessions. A c11 send older than about 14 days reads typed prompt, because c11 keeps its events that long.

--no-collect rebuilds spend from cached per-file records, writes <store>/spend.json, and does not reread transcripts; add --print to also display totals. It can miss transcript changes since the last normal collection and may still refresh the price table.

A cold first collection took about 100 seconds and reached 1,059 MiB maximum resident memory in one benchmark. A separate warm full-producer run took about 18 seconds and reached 653 MiB maximum resident memory. It reuses unchanged files from the store's per-file cache. Normal collection checks a SHA-256 hash of the collector source, file size and modification time; --no-collect checks the cached parser hash and payload without checking transcript files. These measurements are approximate and depend on history size and machine. Normal runs refresh the public price table when it is missing or more than 24 hours old. Use glideslope-spend prices to inspect rates or glideslope-spend prices --refresh to force a refresh. From a clone, run python3 <path-to-clone>/spend.py prices or add --refresh to force a refresh.

The pricing object in spend.json contains counters, not pricing status labels. Run glideslope-spend prices to see rate labels in its status column. The standard fallback status is omitted from that column; affected tokens appear only under underpriced_tokens.

Read these pricing labels and counters alongside the totals:

• list uses an available list rate; there is no separate list_tokens counter.
• standard means a tiered speed used the standard rate because no separate rate was available. This includes flex; those tokens count in underpriced_tokens. For a cheaper tier such as flex, the counter name does not mean the estimate is below the actual cost.
• derived means a missing tier-specific input, output or cache rate was derived by applying the tier-to-base input-rate ratio to the base rate; its tokens appear in derived_tokens.
• estimated means an explicit model-and-speed multiplier was used for a premium rate that is not published; its tokens appear in estimated_tokens.
• unpriced_tokens lists up to ten model entries with no usable rate. Those tokens remain counted but have no dollar value.
• speed_evidence counts requests, or Grok model calls, by provider, speed and evidence source. unknown means the request has no speed evidence.
• fast_mode_armed lists positive detections of premium speed defaults in local Claude Code or Codex settings. An absent provider is inconclusive because settings may be unarmed, missing or unreadable. It does not show which requests used premium speed.
• litellm_age_h is the cached LiteLLM price table's age in hours; null means no table is available. Old or missing rates can make list-rate totals a floor. This age does not describe Grok's cost ticks.

Switching Claude accounts

claude-account use <account> routes new Claude Code sessions to that account's login home. claude-account use auto lets Glideslope pick: it stays put while the current account is under the alert line, then moves to the account with the most slack against its own ◆ mark. Running sessions are never moved. Suggest a switch. Do not run it unless the user asks.

claude-account whose <token-file> --expect <account> --json says which account a claude setup-token token really bills, from the organization ID the API returns, never from the file's name. Run it before handing a token to remote agents; exit 0 means it bills the expected account. An account nobody is logged into still reads live when a meter token for it sits in ~/.claude/accounts/meter-tokens/ on a satellite (rows marked source: meter-token).

The views

Offer a view when the user wants to see the position rather than read it. One command rebuilds the page from the position it just read and opens it in the default browser:

glideslope --open            # the Detail view: weekly status, all windows, the approach plot with sampler history, reset horizon, switch log
glideslope --open popup      # the approach plot and register alone
glideslope --open history    # the approach plot on a real clock, from the sample store only

The pages are self-contained HTML files under views/ in the clone, so --open needs the clone, not a tool install. They reload themselves every minute. Never upload or host one.

When a row looks wrong

Read the provider's section in PROVIDERS.md. tools/kimi_probe.py and tools/grok_probe.py make one request and print the raw payload beside Glideslope's normalization. A Grok probe may refresh the Grok Build access token in ~/.grok/auth.json in place. Never debug a provider with repeated ad-hoc requests.

Installing this skill

From the clone:

mkdir -p ~/.claude/skills
ln -s "$PWD/skills/glideslope" ~/.claude/skills/glideslope

A symlink keeps the skill current with git pull.