Repository navigation
Routing policy alignment: map OpenChamber routing categories to the fleet's Workhorse-first tiers + POOL MAP gates #1272
Description
Problem (found by pilot #1188, 2026-10-26)
Two routing systems now coexist and nobody has reconciled them:
- The fleet's policy (Drew, 2026-10-26 + the POOL MAP in the
provider-quota registry): Workhorse-first (DeepSeek Flash high,
Luna Max, Sonnet), Frontier only on evidence or for design, Fast
for mechanical work; measurement gate (quota read ≤60 min old,
≥85% of window = no new heavy work); plan quotas only, never PAYG;
OpenAI the final reserve; Astra gated; Antigravity Argon watched. - OpenChamber's own classifier (
~/.config/openchamber/routing.json
on the laptop): categories trivial / research / implement / hard,
bound today to — trivial → antigravity gemini-3.8-flash (high),
research → opencode-go gpt-6-luna (max), implement → opencode-go
muse-spark-1.3-contributor (high), fallback → claude-code
claude-haiku-5-5.
The bindings may or may not match the policy — nobody has checked
them side by side (e.g. trivial work defaults to the Antigravity tier
the pool map says to watch; the implement default's pool isn't in the
quota-attribution picture yet). The dispatcher (#1187) will route by
one of these; it must not be by accident.
Work
- Produce the mapping table: fleet tier (Fast/Workhorse/Frontier) ×
OpenChamber category, with the pool each binding draws on and its
quota source. - Flag every binding that conflicts with the POOL MAP gates (watched
tiers as defaults, ungated Astra-class routes, pools with no quota
read) and propose the corrected binding. - Recommend which system is canonical at dispatch time (fleet policy
expressed through OpenChamber categories is the expected answer —
confirm or refute with reasons).
Acceptance criteria
- Mapping table + conflict list exist in the goal workspace and
are linked from epic epic(openchamber): session dispatch + babysitting — OpenChamber as the fleet's workhorse executor #1234. - Each conflict has a proposed corrected binding, or an explicit
"keep, because" with the reason recorded. - Pilot Pilot: Workhorse-first vs Frontier-plan split on comparable effort/m tasks (routing evidence) #1191's trial runs use the mapped categories so its
evidence lands on the real configuration.
Child of epic #1234. Filed by Muse (operator) 2026-10-26.
Activity
developing-today commentedon Oct 27, 2026
CODE-396
dezren39 commentedon Oct 27, 2026
Pool-quota steward escalation — burn-by-expiry pattern, 3rd consecutive run (2026-10-27 10:45Z)
Filing this on #1272 per the steward ledger's standing trigger: if the
opencode-go / cline-pass short-window expiry pattern repeated a third
run, escalate as a routing-config question here instead of another
dispatcher note. It has repeated. Root cause found this run — it is
deployment drift, not canonical config.
Fresh quota picture (laptop OpenChamber
/api/quota/*, 10:28–10:31Z):opencode-go 5h ~0% (reset 12:20Z today) / weekly ~0% / monthly 62.48%
(reset Fri Nov 6 23:55 CDT); cline-pass 5h 6% (reset 13:09Z) / weekly
32% / monthly 72%; codex weekly 21% (reset Mon Nov 2 00:01 CDT);
openrouter monthly 100% — CLOSED. Meta and Claude unreadable
(Unsupported provider / rate-limited read), named out of scope.
The pattern (steward snapshots): opencode-go 5h sat at ~0% at
08:31Z, 09:29Z, and 10:31Z while its window ran down to the 12:20Z
reset; cline-pass 5h went 2% → 5% → 6% against a 13:09Z reset. The
two dispatches in between (tasks 1309/1310) resolved Workhorse to
cline-pass and to codex (weekly 20% → 21%) — never to a Go leaf,
even though both Go workspaces are cancelled with unspent quota
expiring (Personal period ends Nov 2, Lab Nov 6; no rollover).
Why: the laptop's deployed config does not carry the canonical
Go leaves. Canonical
.opencode/opencode.jsoncis correct —#1328 (16e6c0a, dce7585, 2026-10-27 07:04–07:21Z) seats
opencode-go.personal/deepseek-v4.1-flash(weight 3) andopencode-go.lab/deepseek-v4.1-flash(weight 2) indeepseek-flash-plan, and weighted Go Luna leaves inluna-plan.But the laptop's deployed copy (
~/code/.opencode/opencode.jsonc,the file its
~/.config/opencoderesolves to) is a modified workingtree vs its own HEAD dce7585 —
git diff --numstat= 118 insertions/ 975 deletions — and contains zero
opencode-go.personal/deepseekor
opencode-go.lab/deepseekleaves (grep count 0, verified twice).Workhorse sessions launched on the laptop therefore cannot resolve
to a Go Flash leaf no matter what the weights say.
Proposal (bounded, one change):
path (the deployed config follows the repo file per the config(opencode): reconcile deployed config into the repo file #904
ruling) — do NOT hand-edit the deployed file; that is what caused
the 2026-10-25 drift. Whatever local modifications the laptop
working tree carries need to be reconciled into the repo or
dropped, not preserved in place.
routing.jsoncategorybindings (dated Oct 24, proposals until this issue's mapping
lands) are stale against the POOL MAP:
implement→opencode-go / muse-spark-1.3-contributor (a Meta model id on the
Go door),
hard→ claude-code / claude-opus-5-5 andfallback→claude-code / claude-haiku-5-5 (claude-code is disabled in the
fleet config on the free-tier pause, PR feat(opencode): pause claude-code (free tier) + terra-plan door correction #1316). Align them with
the Workhorse-first tiers as part of this issue's mapping.
Expected effect (measurable): after deployment, the next
Workhorse dispatch that resolves through deepseek-flash-plan shows
opencode-go 5h used% rising above ~0% on the following steward
snapshot, and Go monthly (62.48%) starts falling before the Nov 2 /
Nov 6 workspace expiries. Steward review-by: next pool-quota-steward
run after deployment lands.
— pool-quota-steward run 2026-10-27T10:45Z (issue #1339); full
snapshot + ledger event in the goal's
pool-change-ledger.md.17 remaining items
dezren39 commentedon Nov 3, 2026
🤖 Agent note · handed off
dezren39 commentedon Nov 3, 2026
agent-note:v1
event: picked-up
task: 1648 (repo-health-code-pilot, repo-worker-code; openchamber-scheduled-lane aea12c82)
state: claimed; resuming from proposals_published — this session runs on the laptop, so live routing.json + goal file write are possible directly
dezren39 commentedon Nov 3, 2026
agent-note:v1
event: done
task: 1648 (repo-health-code-pilot, repo-worker-code; openchamber-scheduled-lane aea12c82)
result: mapping proposals filed to goal workspace files/designs/code-routing-policy-alignment.md (mapping table + conflicts + 5 proposed corrected bindings + standing statements + live routing.json appendix re-read 2026-11-03T07:53Z — config now uses opencode-route pool routes). Epic #1234 linked. Proposals only; routing.json and POOL MAP untouched. Closing.