Repository navigation
feat(t27os): OpenCode free pool as a second executor, law in .t27 (Refs #12975) - #12977
Conversation
…/quota law (Refs #12975) New spec specs/t27os/opencode_free.t27 (module T27osOpencodeFree): - free means price 0 on input AND output; a missing price is not zero - provider allowlist: only the Zen provider, so a paid model cannot be picked - content gate: non-public content only to glm-5.3-flash-free and deepseek-v4.1-flash-free; every other model, and every unknown one, public only - route by task class: pool, Claude, or owner; owner-class jobs never start - lane states with a pool-wide wait on FreeUsageLimitError, no rotation, no loop - one probe per model, no repository content, only while the lane is available - card badge: free never reads as unlimited Probe receipts 2026-10-31: eight probes, one per free model, all HTTP 429 FreeUsageLimitError, pool-wide, shared reset about 2026-11-01T00:00Z. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
CI triage for head 6f20bba (Refs #12975). This PR changes one file, Fingerprints (policy: skill analog-emulation, lesson 2026-10-31). Master commit is 4cfd2f2 for every row.
pr_changes: Not proven: condition 1 for emit-bitexact. No run of that workflow exists on master at 4cfd2f2, so "red on master" rests on the five failing files all existing unchanged in the base commit. Also pending: the Corpus ratchet run on this PR (it was red on the base commit too, run 124454610439); whether the tests inside |
…probe receipts (Refs #12975) Provider nvidia is admitted next to Zen: a NVIDIA model starts UNPROVEN (a listed price of 0 is not an entitlement), gets one content-free probe, and only a 2xx with text makes it available. 401/402/403 moves it to NO_ENTITLEMENT, which a fresh listing does not lift (only the owner's re-arm does). NVIDIA models get public text or a probe only; the Zen pool limit never stops a NVIDIA lane. Header carries the 2026-10-31 receipts: 10 of 15 answered, 4 timed out (recorded OTHER), 1 invalid-request. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
NVIDIA models are now in the lane law (commit 29bae57, Refs #12975). What changed in specs/t27os/opencode_free.t27: provider nvidia is admitted next to Zen. A NVIDIA model starts UNPROVEN (a listed price of 0 is not an entitlement), gets one content-free probe, and only a 2xx with text makes it available. HTTP 401, 402 or 403 puts it in NO_ENTITLEMENT, which a fresh listing does not lift (only the owner's re-arm does). NVIDIA models get public text or a probe only, and the Zen pool limit never stops a NVIDIA lane. Probe receipts, 2026-10-31, one content-free probe per model ("Reply with the single word ok.", empty throw-away directory, 110 s limit, through OpenCode's stored credential, no repo content):
The same keys listed the catalogue on 2026-10-26 and got 403 on chat, so each result is a dated measurement of one call, not a standing entitlement. Re-probe before leaning on it. Corpus ratchet (expected-failure ledger) is red on this PR with the same verdict as on master at the base commit: 5 unexpected failures (specs/fpga/wasm_board_ax7203.t27, four specs/tri/shootout/*.t27) and 3 unexpected passes (specs/port/browseros/...). This spec is not in either list. Nothing was blessed or disabled. Tracked in #12981. Follow-ups filed: #12983 (key master NVIDIA profile for OpenCode panes, |
PR DashboardGenerated at: 2026-10-31 07:38:17 UTC
Summary
Seal Status
|
PR DashboardGenerated at: 2026-10-31 13:23:25 UTC
Summary
Seal Status
|
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using high effort and found 3 potential issues.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 9bff988. Configure here.
| pub fn may_call(state : u32, now_s : u64, until_s : u64) -> bool { | ||
| if (state == LS_AVAILABLE) { return true; } | ||
| if (state == LS_WAITING_QUOTA) { return now_s >= until_s; } | ||
| return false; |
There was a problem hiding this comment.
Unproven 429 unlocks work
High Severity
next_state moves LS_UNPROVEN to LS_WAITING_QUOTA on EV_RATE_LIMIT or EV_FREE_USAGE_LIMIT, and once the wait expires may_call allows work with no EV_OK. probe_ok then refuses another probe because the state is no longer LS_UNPROVEN. NVIDIA law requires a 200 with text before work; this path is also the 2026-10-31 Zen receipt (every free model returned FreeUsageLimitError).
Additional Locations (2)
Reviewed by Cursor Bugbot for commit 9bff988. Configure here.
| } | ||
| if (state == LS_UNAVAILABLE) { | ||
| if (ev == EV_LISTED_FREE) { return LS_AVAILABLE; } | ||
| return LS_UNAVAILABLE; |
There was a problem hiding this comment.
Relisting skips entitlement probe
High Severity
next_state sends every LS_UNAVAILABLE lane to LS_AVAILABLE on EV_LISTED_FREE, with no memory of whether the model ever got EV_OK. Combined with LS_UNPROVEN plus EV_GONE, a never-probed NVIDIA model becomes callable after a catalog blip, even though a listed price of 0 is not an entitlement.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 9bff988. Configure here.
| pub fn probe_ok(done : u32, content : u32, state : u32) -> bool { | ||
| if (done >= PROBES_PER_MODEL) { return false; } | ||
| if (content != CC_NONE) { return false; } | ||
| return state == LS_UNPROVEN; |
There was a problem hiding this comment.
Re-arm cannot send another probe
Medium Severity
EV_OWNER_REARM returns a refused model to LS_UNPROVEN so the owner can probe once more, but probe_ok still forbids it once done >= PROBES_PER_MODEL (1, “ever”). may_call is false in LS_UNPROVEN, so the re-armed lane cannot be probed or given work.
Additional Locations (2)
Reviewed by Cursor Bugbot for commit 9bff988. Configure here.
Co-authored-by: Dmitrii Vasilev <gHashTag@users.noreply.github.com>
PR DashboardGenerated at: 2026-11-09 21:13:05 UTC
Summary
Seal Status
|
…#12983) (#12986) * feat(t27os): an NVIDIA profile for OpenCode panes in the key master, outside the chain (Refs #12983) PR_NVIDIA sits above the four chain profiles; agent_profile_ok_stored admits it for OpenCode only while an NVIDIA key is stored. The inline config names provider nvidia alone (enabled_providers), holds the key by reference and no model id. Tests pin key present/absent, paid provider, unknown profile and the key-free config text. Refs #12975, #12977. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * chore(seals): re-save t27os_T27osKeys after the NVIDIA profile (Refs #12983) One seal file. Saved with the post-#12824 t27c at ~/t27-lanes/t27c-bc19b54e6 (base bc19b54), on the owner's word for exactly one save, run from a shell cell. The spec hash and the four generated-code hashes now describe keys.t27 at this head; tests 115 -> 117. Refs #12983 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com> Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Dmitrii Vasilev <gHashTag@users.noreply.github.com>


Refs #12975. Does not close it: the work-card fields (executor, model) and the live catalog reader are later lanes.
What this adds
One new spec,
specs/t27os/opencode_free.t27(moduleT27osOpencodeFree), the owner's "go" of 2026-10-31 as law:glm-5.3-flash-freeanddeepseek-v4.1-flash-free; every other model, and every model with an unknown data class, gets public content only.FreeUsageLimitErrorwaits the whole pool until its Retry-After (no model rotation, no retry loop, no probe while waiting); an ordinary 429 waits at leastRATE_COOLDOWN_S.Receipts (2026-10-31, public-safe)
Eight probes, one per free model, prompt "Reply with the single word ok.", no repository content, nothing installed or upgraded:
all eight answered HTTP 429
FreeUsageLimitError, the same Retry-After (about 62,200 s), shared reset about 2026-11-01T00:00Z. The limit is pool-wide, so rotating models does not help. Measured free capacity until the reset: zero. The per-day allowance is unpublished and still unmeasured; further probing waits for the reset and is outside the one-probe authorization.Verification
t27c check: OK.t27c(stale unstamped local binary not used).tools/dupe_scan.py:--likereports nothing in this file written elsewhere. The full scan is red on master's own tip (fix_may_startgroup copied, in no ledger) with or without this file; groups stay 689 either way. Not blessed here.Not in this PR
Executor and model fields on the work card (
specs/queen/tasks.t27,dashboard.t27), the live catalog reader, the pane profile for Zen, any coding lane. Each is its own issue after this one.🤖 Generated with Claude Code
Note
Medium Risk
The PR codifies security-sensitive routing and data-class rules that will govern executors once wired, but changes are spec-only with no runtime hooks in this diff.
Overview
Adds
specs/t27os/opencode_free.t27(T27osOpencodeFree), a new pure-decision spec that defines how a lane may use OpenCode’s Zen free pool and NVIDIA as a second executor—without implementing catalog I/O or live lanes in this PR.Selection and content: Models count as free only when both catalog input/output prices are zero (
PRICE_UNKNOWNis never free). Only providersopencodeandnvidiapassprovider_allowed. Content is gated vialane_ok/provider_may_receive: probes and public text may go to any free model; non-public text is limited toglm-5.3-flash-freeanddeepseek-v4.1-flash-freeon Zen; NVIDIA lanes accept probe or public only regardless of model name.Routing and safety:
route_tasksends low-risk work to the free pool, keeps architecture/security/multi-file/release/long-context on Claude (security to the pool only with owner approval), and never starts owner-class jobs (secrets, credentials, deletion, upstream, publish, payment) in a lane.Quota, probes, and UI: HTTP mapping includes pool-wide
FreeUsageLimitError, per-model rate limits (minimumRATE_COOLDOWN_Sfromkeys.t27),LS_UNPROVEN/LS_NO_ENTITLEMENTfor NVIDIA-style entitlement failures, one content-free probe per model while unproven, and cardbadgevalues so free capacity is never shown as unlimited.The file ships with a large embedded test suite covering the rules above; integration with agents, work cards, and a catalog reader is explicitly deferred.
Reviewed by Cursor Bugbot for commit b2078ab. Bugbot is set up for automated code reviews on this repo. Configure here.