Skip to content

Bound worker launch cost by tier and branch #964

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
thomasluizon merged 6 commits into from
Sep 15, 2026
Merged

Bound worker launch cost by tier and branch #964

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
c709391
chore: bound worker launch cost
thomasluizon Sep 14, 2026
849a2df
docs: catalog worker budget controls
thomasluizon Sep 14, 2026
7d4f0e5
fix(tools): reserve worker launches atomically
thomasluizon Sep 14, 2026
8f1b730
fix worker launch reservation ordering
thomasluizon Sep 14, 2026
e0e0fe2
Merge remote-tracking branch 'origin/redesign/main' into chore/ticket…
thomasluizon Sep 14, 2026
94701d5
fix worker launch slot reservation
thomasluizon Sep 14, 2026
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
142 changes: 80 additions & 62 deletions .claude/calibration.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,236 +1,254 @@
{
"calibratedAt": "2026-09-13",
"calibratedAt": "2026-09-14",
"workerEngine": "codex",
"workerCommand": "codex",
"workerModel": "gpt-5.6-sol",
"workerArgs": [
"exec",
"-c",
"windows.sandbox=\"unelevated\"",
"--dangerously-bypass-approvals-and-sandbox",
"-c",
"model_reasoning_effort=\"high\"",
"--model",
"gpt-5.6-sol"
],
"workerModelSource": "resolveWorkerInvocation(config.worker, ..., \"default\") in tools/lib/orchestrator-config.mjs, the exact vector launch-worker.mjs launches: engine args, then models.default args, then the model",
"workerTiers": {
"default": {
"model": "gpt-5.6-sol",
"args": [
"exec",
"-c",
"windows.sandbox=\"unelevated\"",
"--dangerously-bypass-approvals-and-sandbox",
"-c",
"model_reasoning_effort=\"high\"",
"--model",
"gpt-5.6-sol"
]
},
"mechanical": {
"model": "gpt-5.6-sol",
"args": [
"exec",
"-c",
"windows.sandbox=\"unelevated\"",
"--dangerously-bypass-approvals-and-sandbox",
"-c",
"model_reasoning_effort=\"medium\"",
"--model",
"gpt-5.6-sol"
]
}
},
"workerTierVerdict": "default handles product, design, architecture, and ambiguous orders requiring judgment; mechanical handles merge-forward work, known conflict lists, and reviewer-directed test experiments",
"workerModelSource": "resolveWorkerInvocation(config.worker, ..., tier) in tools/lib/orchestrator-config.mjs for every configured tier, with each exact launch vector: engine args, then selected profile args, then the model",
"entries": {
".agents/skills/merge-prs/SKILL.md": {
"model": null,
"effort": null,
"digest": "fa4a10a59a391992",
"calibratedAt": "2026-09-09",
"verdict": "current: a pointer with no behaviour, so it declares no model and no effort and inherits whatever the Codex host runs. Its digest is the whole verdict: the frontmatter name and description decide whether Codex finds this skill at all, and the body names the one canonical definition both hosts read."
},
".agents/skills/orchestrate/SKILL.md": {
"model": null,
"effort": null,
"digest": "d55e8bc268758fcb",
"calibratedAt": "2026-09-09",
"verdict": "current: a pointer with no behaviour, so it declares no model and no effort and inherits whatever the Codex host runs. Its digest is the whole verdict, and it matters most here: orchestrate is the entry point every other piece of work passes through, so a pointer that stops resolving takes the whole queue with it."
},
".agents/skills/ticket/SKILL.md": {
"model": null,
"effort": null,
"digest": "e1f13a268ec17004",
"calibratedAt": "2026-09-09",
"verdict": "current: a pointer with no behaviour, so it declares no model and no effort and inherits whatever the Codex host runs. Its digest is the whole verdict, because a ticket is the prompt (D2) and a host that reads a forked definition writes a forked ticket."
},
".claude/agents/Explore.md": {
"model": "haiku",
"effort": null,
"digest": "74e69f2e9ecb53a7",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: locates code and returns excerpts, never reviews it, so haiku is the right floor and effort is undeclared because breadth is passed in the prompt."
},
".claude/agents/audit-readonly.md": {
"model": "haiku",
"effort": null,
"digest": "ccf83adcaf2a327b",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: a read-only fan-out finder over Read/Grep/Glob, so haiku is the right floor and no effort is declared because the work is enumeration rather than judgement."
},
".claude/agents/completeness-critic.md": {
"model": "sonnet",
"effort": "high",
"digest": "18fa9e64efcdb3a1",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: its whole job is to falsify a completion claim across a surface inventory, which is judgement, so sonnet at high effort is right and must not fall to haiku."
},
".claude/agents/design-reviewer.md": {
"model": "sonnet",
"effort": "medium",
"digest": "2948735644b388c8",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: judges a diff against DESIGN.md with the spec in front of it, so sonnet at medium effort is right; the rules are written down, which is what keeps it off high."
},
".claude/agents/design-specialist.md": {
"model": "inherit",
"effort": "high",
"digest": "0238593f7564142c",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: shapes the UI half of a ticket and must refuse to improvise where the design system cannot meet a need, so it inherits the session model at high effort deliberately."
},
".claude/agents/product-manager.md": {
"model": "inherit",
"effort": "high",
"digest": "06e63d937fc86cb1",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: runs the eight-category edge-case pass and decides how many tickets a request becomes, so it inherits the session model at high effort."
},
".claude/agents/web-researcher.md": {
"model": "sonnet",
"effort": "medium",
"digest": "b0eba80af5cdd41f",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: verifies load-bearing facts against live pages for one narrow slice, so sonnet at medium effort is right; it has no Agent tool, which is the structural cap that matters more than the model."
},
".claude/skills/android-generate/SKILL.md": {
"model": null,
"effort": "low",
"digest": "4aab9a05b8f27d13",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: a mechanical gradle build and emulator install, so low effort is right."
},
".claude/skills/android-release/SKILL.md": {
"model": null,
"effort": "low",
"digest": "511b9e409e86a2d9",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: dispatches one workflow with computed version numbers, so low effort is right."
},
".claude/skills/audit-code-quality/SKILL.md": {
"model": null,
"effort": null,
"digest": "2b278674dc8f8764",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: a judgement-level debt audit that opens tickets, which argues for an explicit high. Left undeclared in this first pass because declaring it changes behaviour and cost, and eleven skills are in the same position; that is the follow-up this pass names rather than a change smuggled into the mechanism."
},
".claude/skills/audit-performance/SKILL.md": {
"model": null,
"effort": null,
"digest": "97dae7a146ee0fc6",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: same shape as audit-code-quality and the same follow-up."
},
".claude/skills/audit-security/SKILL.md": {
"model": null,
"effort": null,
"digest": "e92ff1f4ff0b8b21",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: same shape as audit-code-quality and the same follow-up. This is the one where an inherited low effort would cost the most, because a missed authz hole is not visible in the output."
},
".claude/skills/audit-tests/SKILL.md": {
"model": null,
"effort": null,
"digest": "291bf85e61e31045",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: same shape as audit-code-quality and the same follow-up."
},
".claude/skills/deep-research/SKILL.md": {
"model": null,
"effort": null,
"digest": "a7bbf3c505b096e8",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: it fans out web-researcher subagents that carry their own sonnet/medium tuning, so the orchestrating half inheriting the session is defensible today."
},
".claude/skills/dev-server/SKILL.md": {
"model": null,
"effort": "low",
"digest": "b95013bea459e4fc",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: brings up Docker, the API and the web server in dependency order with readiness gates, which is mechanical, so low effort is right."
},
".claude/skills/drift-review/SKILL.md": {
"model": null,
"effort": null,
"digest": "1fae48b525fe2fe2",
"calibratedAt": "2026-09-12",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: it judges repeated evidence against the current workflow files, but every result remains a staged candidate for human review."
},
".claude/skills/handoff/SKILL.md": {
"model": null,
"effort": "high",
"digest": "3f7239924dc41147",
"calibratedAt": "2026-09-13",
"calibratedAt": "2026-09-14",
"verdict": "current: high effort, and it earns it. It decides what survives into a spec that outlives every session, and under-thinking it is how a rule Thomas set in week one disappears by week four."
},
".claude/skills/investigate/SKILL.md": {
"model": null,
"effort": null,
"digest": "9e633595abd89253",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: root-causing a production incident across Sentry, Render, Postgres and the LSP is judgement, so this is a follow-up candidate."
},
".claude/skills/lesson/SKILL.md": {
"model": null,
"effort": null,
"digest": "f3af59784bce4557",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: it graduates a correction into a rule, which is judgement, but it always runs with Thomas present, so an inherited effort is checked by a human in the moment."
},
".claude/skills/merge-prs/SKILL.md": {
"model": null,
"effort": null,
"digest": "30a5ce2efa3fb3ea",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: the dangerous half of this skill is mechanical (an exact-head preflight, an ordered admin squash), and its safety comes from the preflight rather than from reasoning depth."
},
".claude/skills/orchestrate/SKILL.md": {
"model": null,
"effort": "high",
"digest": "4df860787a7b5091",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: high effort, and it earns it: it plans the queue, verifies delivery from artifacts and clears the review, and it is the entry point every other piece of work passes through."
},
".claude/skills/prod-readiness/SKILL.md": {
"model": null,
"effort": null,
"digest": "24aa498f98faf0c3",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: it consolidates four child audits into one honest launch verdict, which is judgement, so this is a follow-up candidate."
},
".claude/skills/progress/SKILL.md": {
"model": null,
"effort": "medium",
"digest": "976dcf419b3253d8",
"calibratedAt": "2026-09-13",
"calibratedAt": "2026-09-14",
"verdict": "current: medium effort, because it reads live git and ticket state and must judge whether a part-built screen is honestly described, which low effort gets wrong by rounding up."
},
".claude/skills/questions/SKILL.md": {
"model": null,
"effort": "high",
"digest": "430ce323de282429",
"calibratedAt": "2026-09-13",
"calibratedAt": "2026-09-14",
"verdict": "current: high effort, because the filter decides what NOT to ask, and a wrong call either wastes his attention or ships a guess as a decision."
},
".claude/skills/second-opinion/SKILL.md": {
"model": null,
"effort": null,
"digest": "8977e10e62d65f75",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current with nothing to declare: the reasoning happens in the other model, by construction. Declaring an effort here would tune the wrong side of the call."
},
".claude/skills/sleep/SKILL.md": {
"model": null,
"effort": null,
"digest": "3acd8155b2528682",
"calibratedAt": "2026-09-11",
"calibratedAt": "2026-09-14",
"verdict": "undeclared, inherits the session: it takes every decision alone overnight, which is the strongest argument for an explicit high in the follow-up, and the weakest place to guess it in this pass."
},
".claude/skills/ticket/SKILL.md": {
"model": null,
"effort": "high",
"digest": "ce7376bb28eca9c8",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: high effort, and it earns it: a ticket is the prompt (D2), so a shallow ticket is a shallow implementation, and the cost lands on whoever executes it."
},
".claude/skills/validate/SKILL.md": {
"model": null,
"effort": "low",
"digest": "e1ef5a36050d37f1",
"calibratedAt": "2026-09-09",
"calibratedAt": "2026-09-14",
"verdict": "current: runs lint, type-check and tests across both repos, so low effort is right."
},
".agents/skills/merge-prs/SKILL.md": {
"model": null,
"effort": null,
"digest": "fa4a10a59a391992",
"calibratedAt": "2026-09-14",
"verdict": "current: a pointer with no behaviour, so it declares no model and no effort and inherits whatever the Codex host runs. Its digest is the whole verdict: the frontmatter name and description decide whether Codex finds this skill at all, and the body names the one canonical definition both hosts read."
},
".agents/skills/orchestrate/SKILL.md": {
"model": null,
"effort": null,
"digest": "d55e8bc268758fcb",
"calibratedAt": "2026-09-14",
"verdict": "current: a pointer with no behaviour, so it declares no model and no effort and inherits whatever the Codex host runs. Its digest is the whole verdict, and it matters most here: orchestrate is the entry point every other piece of work passes through, so a pointer that stops resolving takes the whole queue with it."
},
".agents/skills/ticket/SKILL.md": {
"model": null,
"effort": null,
"digest": "e1f13a268ec17004",
"calibratedAt": "2026-09-14",
"verdict": "current: a pointer with no behaviour, so it declares no model and no effort and inherits whatever the Codex host runs. Its digest is the whole verdict, because a ticket is the prompt (D2) and a host that reads a forked definition writes a forked ticket."
}
}
}
5 changes: 5 additions & 0 deletions .claude/orchestrator.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,10 @@
"default": {
"model": "gpt-5.6-sol",
"args": ["-c", "model_reasoning_effort=\"high\""]
},
"mechanical": {
"model": "gpt-5.6-sol",
"args": ["-c", "model_reasoning_effort=\"medium\""]
}
}
}
Expand All @@ -31,6 +35,7 @@
"parallelTickets": 3,
"cloudParallelTasks": 8,
"reviewFixAttempts": 3,
"workerLaunchesPerBranch": 2,
"workerLogMegabytes": 512
},
"cloud": {
Expand Down
Loading
Loading