Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion .claude/guides/spend-and-inference.md
Original file line number Diff line number Diff line change
Expand Up @@ -109,6 +109,12 @@ inference API credential this deployment holds. It is bound by all of these:

- **One model:** `claude-opus-5-5`, fixed in the lane's kagent `ModelConfig`. No fallback model,
child, advisor, model switch or automatic routing, and No-Fable binds it like every other run.
A child here is a routed run with its own model or credential. The subagents the run loop
dispatches, such as the portfolio surveyor (definition and plugin guide), run inside the parent
run on its model and credential, as on the host Claude lane. They are part of the run and its
cost: the run prompt runs them in the foreground and uploads their transcripts
([platform#6262](https://github.com/devantler-tech/platform/pull/6262),
[platform#6310](https://github.com/devantler-tech/platform/pull/6310)).
- **One run a day, with no time limit:** at most one scheduled run a day, and a run goes until the
agent ends it. Cutting a run off mid-task wastes what it has already spent (maintainer decision
2026-11-09, which replaced the earlier cap of about $6 a run). An API workspace spend limit equal to
Expand All @@ -124,7 +130,8 @@ inference API credential this deployment holds. It is bound by all of these:
- **In the run, a backstop:** the run stops on what it observes: a model ID other than
`claude-opus-5-5`, or a billing route other than the credit.

The schedule stays suspended until every sub-issue of #4866 is done.
The schedule has run daily since 2026-11-15. The maintainer turned it on for a trial of about half
a month before every sub-issue of #4866 was done (ADR 0014, amendment of 2026-11-15).

**No-Fable is a protected limit:** the entire Fable model family is prohibited in recurring parent
runs, children, advisors, fallback, retries and experiments, even if a provider describes it as
Expand Down
20 changes: 16 additions & 4 deletions .claude/plugin-consumption/inference-routing-runtime.md
Original file line number Diff line number Diff line change
Expand Up @@ -64,10 +64,22 @@ The `kagent-cluster` instance is not an accepted subscription route. It is the o
bound by the *kagent credit lane* rules in `AGENTS.md` **Inference routing** (spend and inference
guide), maintainer decision 2026-11-07, monorepo#4866. Unlike the host lanes it has pre-inference
controls by construction: kagent compiles the model fixed in the lane's `ModelConfig` into the
runtime before any turn, and the API workspace spend limit is enforced by the provider. Neither is
verified yet. The lane stays suspended until every sub-issue of monorepo#4866 is done and the
canary run (monorepo#4869) records the effective model, the per-run cost and the maintainer's
confirmation of the spend limit here.
runtime before any turn, and the API workspace spend limit is enforced by the provider.

The canary runs (monorepo#4869) record:

- **Effective model:** `claude-opus-5-5` on every assistant message in the main-session transcripts
of the 2026-11-14 and 2026-11-15 runs and in the 2026-11-15 run's survey-subagent transcript. The
subagent ran inside the parent run on its model and credential, so it is not a routed child (the
spend and inference guide's *kagent credit lane*).
- **Per-run cost:** about $3.80 for the 2026-11-15 run, the first whose subagent transcript reached
the lane store. The main session cost about $2.13 and the portfolio-surveyor subagent about
$1.67. Each is counted from its transcript's token usage at Opus 5.5 prices, taking each
message's largest count, because the first streamed line of a message carries only part of its
output tokens.
- **Spend limit:** the workspace's monthly limit equal to the credit. It has not been reached, so
the provider's enforcement of it is not yet observed. The maintainer turned the daily schedule on
2026-11-15 for a trial of about half a month (ADR 0014, amendment of 2026-11-15).

## Native verification procedure

Expand Down
2 changes: 1 addition & 1 deletion AGENTS.md
Original file line number Diff line number Diff line change
Expand Up @@ -165,7 +165,7 @@ disagree, the scheduler is the defect — reconcile it there, per *Agent definit
|---|---|---|
| **Claude** — `claude/*`, hourly at `:50` | Every hour at `:20` and `:50` | 00:00, 06:00, 12:00, 18:00 |
| **Codex** — `codex/*`, hourly at `:10` | Every hour at `:10` and `:40` | 03:00, 09:00, 15:00, 21:00 |
| **kagent** — `kagent/*` | 1 run/day; suspended | — |
| **kagent** — `kagent/*` | 1 run/day | — |

Work [*The work-selection ladder*](#the-work-selection-ladder--one-ordering-checked-top-down-every-run)
top-down every run and **keep going while actionable work remains** — the floor is at least one
Expand Down
10 changes: 10 additions & 0 deletions docs/adr/0014-kagent-claude-api-credit-lane.md
Original file line number Diff line number Diff line change
Expand Up @@ -109,3 +109,13 @@ The maintainer decided on 2026-11-07:
the deadline is set to ten years to keep the 2026-11-09 amendment. The schedule stays paused until
every sub-issue of #4866 is done
([platform#5999](https://github.com/devantler-tech/platform/pull/5999)).
- **2026-11-15: the daily run is on for a trial.** Decision 6 kept the schedule suspended until
every sub-issue of #4866 was done. Seven manual runs came first, and the last two finished every
step on their own. The 2026-11-15 run measured the first full per-run cost, about $3.80 with its
survey subagent, so a trial of about half a month costs about $60 at that rate. The maintainer
turned the daily run on for that trial while sub-issues stay open, among them the GitHub App's
egress injection, the harness tooling, whole-plugin loading, the checkout-based hooks and
overlays, Go modules, claim refs and review-request markers. The maintainer would rather the lane
did the work it judges important and stopped when that was done. The lane still runs at most once a day,
and the workspace spend limit stays its only spend cap
([platform#6340](https://github.com/devantler-tech/platform/pull/6340)).
Loading