Milestones
List view
Follow-ups from the Aug 13 2026 re-review, filtered to items independently confirmed still true against current main (most of the review's P0 list was already fixed earlier in the same session — see #165 for the triage). Scope: EC2 resource tagging, an IAM-role concurrency question the getting-started guide's shared-role design implies but never tests, tracking App(image=)'s app-level-default gap explicitly, region/time-independent price wording, and CI/process polish (link checking, Action SHA pinning).
No due date•7/7 issues closedExternal review (2026-08-12) found calque's implementation has outpaced its docs: several design/status docs now contradict v0.3.0's actual shipped behavior (secrets injection, .spawn() driver, calque session, crossover-K retraction), and there is no dedicated first-real-AWS-run guide despite real AWS execution now being one of calque's strongest proof points. This milestone reconciles doc-vs-reality drift and adds the missing operational bridge from analyze -> real AWS run. See issue tracking the review itself for full context.
No due date•10/11 issues closedClose remaining idiom-support gaps (.starmap tuple-splat execution, spot support everywhere, multi-region capacity fallback, --entrypoint steering warm-unit selection, remaining AI-Almanac/rare-construct gaps) and fix real documentation staleness (compat-matrix legend taxonomy, .spawn()/acquisition/tenancy-vs-session drift, GPU-naming normalization, a task-oriented Modal->AWS porting guide).
No due date•10/11 issues closedRound out calque pool from create-only to a real operational primitive: slug model names before they become SQS queue names, add delete/status/list subcommands, add SQS visibility-timeout heartbeating for long-running claims, wire the occupancy sampler into pool-mode claims, support scale-up (add workers to an existing pool).
No due date•5/5 issues closedMIG/MPS/tenancy safety and policy gaps found in a source-level review: MIG PickLayout's maximize-tenant-count default, MPS Coordinator's missing synchronization, MPS NewCoordinator's construction-time validation gap, and (highest priority) tenancy TTL expiry not guaranteeing the prior holder's workload actually stopped.
No due date•4/4 issues closedRename cmd/calque/session.go->ramp.go (calque session -> calque ramp), freeing 'session' for the institutional MIG/MPS tenancy check-out/check-in verb. Bundles the tenancy lease-identity design work (#128's CheckIn auth gap) and unblocks #118/#119 (both explicitly blocked on this CLI existing).
No due date•3/3 issues closedAdopt github.com/scttfrdmn/substrate/emulator as calque's offline/deterministic AWS test tier, wired into internal/plan/acquire.go's Launcher interface (today only exercised by hand-written fakes) and internal/gate/bedrock.go's live-catalog dependency. De-risks M12/M13/M14's new fleet/tenancy/driver code, all of which was flagged by its own design as needing real-AWS verification with no offline path. Real-AWS verification tier stays AWS_PROFILE=aws-gated, opt-in like --i-understand-this-spends-money.
No due date•1/1 issues closedSupersedes #97's internal framing only — #97's goal (execute every .spawn()'d callable in parallel, block until all complete, return results) is unchanged and stays open as the umbrella issue. What changes: the claim that this needs a genuinely new persistent result-store is overstated — internal/exec/shard.go and orchestrate.go's existing S3-backed manifest/collect machinery (already shipping for .map() fan-out via fleetrun.go) covers block-and-wait durability already. Real new work: a string-keyed shard collector (offline-testable) and a .spawn()-site-to-manifest resolver plus a new spawnRun driver (needs real-AWS verification, like fleetrun.go itself).
No due date•5/5 issues closedSession-based sharing so a university's reserved/fixed GPU fleet can serve multiple concurrent interactive users per physical card, instead of one-workload-per-instance. Scope: per-card SharingMode (MIG vs MPS vs Dedicated) behind the target/plan seams; a user-facing interactive-tenancy primitive distinct from calque session's K-ramp meaning; prove one MIG-capable and one MPS-only card end to end on real AWS before broader rollout; document trust-boundary/crash-blast-radius policy explicitly (calque inventing capability Modal deliberately declines, see #96). Layers BELOW the idle-fleet milestone (M12): fleet decides which/how many whole instances are warm; tenancy decides how many users occupy slices within one such instance.
No due date•6/6 issues closedMirrors Modal's fleet-level idle-GPU buffer (not sub-GPU sharing — see #96, confirmed no such Modal behavior exists to port): a named pool of GPU instances stays warm (one loaded model per instance, long idle-timeout) and drains a shared queue of run requests across users/jobs wanting the SAME model, skipping @enter's cold-load cost on a cache hit. Built on spawn's cohort/taskcohort provisioning primitives (launch/reap/IAM), NOT spawn pool's taskpool fungible-task-per-pull contract (wipes worker state between pulls). Single-model-per-pool for v1; multi-model-per-instance sharing deferred.
No due date•5/5 issues closedClose the pedagogical gap surfaced by an external Modal-user review (verdict 8.2/10: A-grade spike, B-range public usability). Transform calque's engineering record into a deliberate user journey: clone→first-run quick start, promoted examples, capability matrix, reproducibility artifact, CI + contributor/security hygiene.
No due date•12/12 issues closedGaps B(autoscaling)/C(async) + §1/§18 non-goals. Document-only. Convert explicitly-deferred idioms into structured, attach-point-aware recognize-and-leak records (autoscaling/keep_warm kwargs, .spawn/.map.aio async futures, real card selection, MIG/NEFF/Trainium/multi-node) so the seam is documented, not missing. No product code; write entries continuously as M5–M9 land. Priority 6 (ongoing).
No due date•4/4 issues closedGap F. Serve is a documented future-architecture epic, NOT built in the spike (out of §16 batch+K success criteria). Do only the cheap slivers: detect @web_endpoint/@asgi_app/@wsgi_app/@web_server, stop run.go hard-erroring (emit a semantic_gap leak), and write a design record for the long-lived request-driven shape. Priority 5.
No due date•3/3 issues closedGap E. Add mount->S3 commit so Modal's volume.commit() round-trips, not just download-before-@enter. Shares exec/bootstrap/warmd-exit attach points with M7. Priority 4.
No due date•3/3 issues closedGap D. Headline. Shard .map across N acquired single-node instances and collect the union, making §15's 100k rung real. Embarrassingly-parallel fan-out across single-node boxes — NOT §1's forbidden multi-node/gang scheduling. Reuses Collect (index-keyed), warmd Supervisor (per-instance re-drive), Acquirer AZ-sweep. Delivers against #18. Priority 3.
No due date•7/7 issues closedGaps A/B/C(sync). Stop dropping Modal idioms the pyast parser already emits: render leaked image DSL verbs, widen the IR + readConfigKwargs for portable decorator config (cpu/memory/retries/secrets/schedule/region), and detect sync invocation idioms (.remote/.starmap/.for_each). Mechanical, low-risk, high leak-count reduction; IR-widening ripples downstream. Priority 2.
No due date•8/8 issues closedGap G. Turn the analyze-time SUGGEST BEDROCK print into a structured ReplacementOffer that actually short-circuits the run path (run/real/session), so calque won't rent a GPU for a model that's already a Bedrock API call (§11). Includes wiring the gate into run.go (currently absent) + an invoke hint + near-match caveat. Priority 1.
No due date•5/5 issues closed