Repository navigation
Map: run the agent lanes at full Claude capacity #6958
Description
Destination
A design for running the agent lanes at full capacity: every Claude subscription used close to its
weekly limit before reset, every linked Console org's monthly API credits spent before they expire,
and no lane ever stalled by a usage limit. It covers how each use case is split across local
sessions, cloud sessions, GitHub Actions runs and API-only runtimes (Managed Agents, Batches,
Messages API), the credential fallback that keeps critical lanes moving, and the metrics that
steer allocation toward work that ships. Once the decisions below close, it graduates to build
epics through /work-items:decompose.
The capacity feeds the existing end-to-end lane pipeline (intake through post-merge, see
docs/conventions/pr-pipeline/README.md and docs/conventions/loop-lane/README.md). The operator
enters only where a lane escalates with needs-human, decides, and leaves.
Notes
- Max and Team plans now carry monthly API credits in the plan's linked Console org:
help article 17154008.
Credits cover API-key traffic only (Messages, Batches, Managed Agents,claude -por Agent SDK
run with that org's key). Runs started by the Claude Code GitHub Action count as Claude Code
usage and are not covered, even with-p. Credits expire each billing cycle; one plan links to
one org, and only support can change the link. No API reads the credit balance. - Subscription usage can be read headless with
claude -p "/usage";-pand--bgruns do not
wait at a limit, so a scheduler must detect it. - ADR 0037 records "operator cannot use API keys"; the seed decision below amends it.
- Today's lanes all run claude-code-action on one subscription token; the pilot is the
pr-refinefix-docs lane (feat(pr-pipeline): turn on the pr-refine fix-docs lane #6704). - Related: feat(evals): run plugin eval suites in a CI lane and hill-climb in the background #6405 (evals in CI, credential question), feat(developer-experience): claude plugin eval suite in GitHub Actions #6649 (plugin eval in Actions), feat(planning): route new work by diff size and fog: one-shot, /planning:interview, or /planning:wayfind #6729,
Read usage through the official claude binary instead of the tool's own endpoint calls claude-code-account-rotation#220.
Decisions-so-far
- Framing: how work splits between subscriptions and API credits (Framing: how work splits between subscriptions and API credits #6959): credits fund API-only work, subscriptions stay one account per workload, lanes fall back to rotating org API keys, ADR 0037 to be amended (resolved 2026-10-10)
- After every linked org's credits are spent, keep a capped paid backstop? (After every linked org's credits are spent, keep a capped paid backstop? #6966): no; the lane pauses and alerts until a window resets or credits refill (resolved 2026-10-11)
- Inventory every agent workload with its surface, criticality and whether it can run API-only (Inventory every agent workload with its surface, criticality and whether it can run API-only #6964): 27 workloads; review lanes are the heaviest load; critical set and filler candidates listed for Which lanes are critical enough to need a never-blocked fallback? #6965 and What goes in the filler queue, and how is it fed? #6971 (resolved 2026-10-11)
- Write an ADR amending ADR 0037 to allow API keys from the operator's own linked Console orgs (Write an ADR amending ADR 0037 to allow API keys from the operator's own linked Console orgs #6961): ADR 0062 merged (docs(adr): allow API keys from the operator's own linked Console orgs #6995); fallback orgs carry no paid balance (resolved 2026-10-11)
- Managed Agents: workload fit, credit coverage of runtime, and how a session gets repo access, skills, MCP servers and network (Managed Agents: workload fit, credit coverage of runtime, and how a session gets repo access, skills, MCP servers and network #6969): fits bug hunts, sweeps, doc drift and research; plugin skills need the repo-root .claude/skills layout; runtime-fee coverage open until an invoice shows it (resolved 2026-10-11)
- Which lanes are critical enough to need a never-blocked fallback? (Which lanes are critical enough to need a never-blocked fallback? #6965): a lane is critical when a merge or the operator queue waits on it; attend-queue now, pr-merge, pr-fix-ci, pr-update and the explain-change gate once built (resolved 2026-10-11)
- What goes in the filler queue, and how is it fed? (What goes in the filler queue, and how is it fed? #6971): deferrable, bounded, unwaited work (evals, bugs:scan, improvement:find, audits, sweeps) as labeled issues picked by work-loop (resolved 2026-10-11)
- ADR 0062 amended 2026-10-16: each lane role uses one plan; decision 3 stands; see the amendment.
Child items
- loop-lane: confirm-then-record for lane posts, and research on overlapping runs (loop-lane: confirm-then-record for lane posts, and research on overlapping runs #7672): open
- loop-lane: decide an external dead-man's switch for scheduled lanes (loop-lane: decide an external dead-man's switch for scheduled lanes #7673): open
- harness-ops: probe where cloud-routine state lives between runs (harness-ops: probe where cloud-routine state lives between runs #7674): open
- managed-agents: map the docs corpus and pick a first credit-funded workload (managed-agents: map the docs corpus and pick a first credit-funded workload #7675): open
- feat(lanes): private, aggregatable store for lane transcripts (feat(lanes): private, aggregatable store for lane transcripts #7676): open
- feat(lanes): run limits by task class that save progress and resume from pushed state (feat(lanes): run limits by task class that save progress and resume from pushed state #7677): open
- loop-lane: research read-only preferences and sandbox completeness for local lanes (loop-lane: research read-only preferences and sandbox completeness for local lanes #7680): open
- feat(usage): usage panel with 70%/90% alerts and a one-click rebalance (feat(usage): usage panel with 70%/90% alerts and a one-click rebalance claude-code-account-rotation#255): open
Not-yet-specified (fog)
How the self-improvement loop closes: which findings (post-merge sweeps, /improvement:find,
review-lane misses) flow back into the filler queue and into the skills themselves, and how
"higher accuracy and confidence" is shown rather than asserted.
Out-of-scope
- Automatic rotation between subscription accounts at their limits: decided against in the seed
interview; subscriptions stay one account per workload. - The pipeline's stages, holds and merge authority: owned by the pr-pipeline and loop-lane
conventions, not this map.
Activity
3 remaining items
github-actions commentedon Oct 11, 2026
Proposed work class: work-class: structural
Basis: automated intake triage; the lanes App applied this class.
Size: Large
Complexity: High
Uncertainty: Knowable
Flags: none
Blocked by: none
Understood: this is a design map for running the agent lanes at full subscription and API-credit capacity. It covers how work splits across surfaces, the credential fallback, and the allocation metrics. Most decisions are resolved. The open child items (#7672–#7677, #7680) carry the remaining work. The latest comment says the map needs no owner input, so I'm not asking anything further. The work is too large for one change and should graduate to build epics through decomposition. I found no clear duplicates.
Automated triage, applied by the lanes App.
kyle-sexton commentedon Oct 11, 2026
Resume point (2026-10-11). Run
/planning:wayfind work #6958from any machine; every resolution lives in its item's comments. Suggested order: #6968 CI credential (quick interview), then #6970 allocation model (design session). Operator steps any time: #6963 setup-token/usageprobe (run in your own terminal, paste back only therate_limitsoutput) and #6960 link orgs (auto-reload off, no purchased credits). Then #6962 -> #6967 -> #6972, and #6973 -> #6974.kyle-sexton commentedon Oct 24, 2026
Carried from #6970 and #6972 (closing; HITL session 2026-10-24):
kyle-sexton commentedon Oct 25, 2026
More carried from #6972 and #6974 on closing them (HITL session 2026-10-24). The filler-scheduler and prompt-rule lines are in the comment above.
From #6972:
.github/lane-funding.jsonlistsrun-claude-clilanes. Amend the ADR.From #6974 (#6970, its first destination, is closed). Capacity used is an input; the goal is merged work that stays merged.
kyle-sexton commentedon Oct 26, 2026
This was generated by AI.
Released: nothing on this map needs the owner. The open lines (the late-cycle filler scheduler, which outcome signals to record and how they feed allocation, and the ADR 0062 amendment) are design inside the map's stated destination, so agents decide them from data. Open children #7675, #7676 and #7680 carry their own state.
Under the owner's 2026-10-26 stop rule, a person is needed only for a change in intent or vision, an implementation finding that changes the intent or primary design, or a high or critical security issue; host, attended and eval runs are agent work. So
needs-humancomes off.melodic-automation-lanes commentedon Oct 26, 2026
Not admitted to the issue lane yet: Size large: the split into sub-issues at intake is pending (owner Q93).