moooon
Motir
Vibe the project. Bring an idea — Motir's three AI layers plan it, track it, and ship it, end to end. You're looking at Motir, built in Motir.
- Vibe Project
- Open Source
- AI Agent
- AI Loop
- 6
- 需求
- 4
- 贊同
- 524
- 已規劃
- 4,959
- 已發布
- The abandoned-plan sweep records a session as failed ("model_unavailable") while its job is still queued in motir-ai, and leaves that job queued to run later against a session already failed
- headroomScaling e2e case 6 reads the machine state before the stand-in has frozen it
- A planning machine that resumes from suspend keeps its frozen clock, so every graph-read credential it mints is born expired — re-audits and planning code reads fail with 401
- Planning bug (both planners): the gap re-lay authors only the cards it adds — the card that raised the gap, and any older sibling that gains an edge to a new card, keep bodies the re-lay made false
- Verify the resume substrate is real before anything resumes — each profile's resume form (the per-profile resumeArgs) and the session id a run records are on the default branch, and a resume really continues the same conversation instead of opening a new one
- A work-item link labelled with one key but pointing at another card passes every validation — add a `likely-mislabelled-link` shape advisory so both planners catch it before a plan closes
- A pool machine woken by a request suspends itself again a second later, so the call after the wake times out and Resume answers "Motir AI is not available"Planning capacity follows demand — a conversation turn never waits behind a planning run, the planning machines and the graph service grow and shrink on their own, and a load-tested cost ceiling is the only fixed limit
- Planning failure: validation_error — contextual planning: none of the target work items (MOTIR-9002) could be read for this user
- (motir-ai) DECIDE: a typed sections write interface over the free-text body, from one motir-core registry — an amendment to `session-model.md` AMENDMENT 3; MOTIR-4688 Q1's storage choice stands
- The CI fleet never recovers a lost runner: a job left with no runner stays queued forever, because GitHub hands each new runner to the oldest queued job and the fleet only counts its own intents
- A model served by two providers can carry only one price — provider_direct_prices.json is keyed by bare model id, so deepseek-v4-pro on the DeepSeek channel bills at Alibaba Frankfurt's $1.65 / $3.301
- A run in the person's own agent shows no agent output until the agent finishes — every profile is launched in its print-only mode, never its streaming mode, and no session id is kept
- A target added in the planning composer is dropped when the window already holds a project-wide session: the job is sent with no targetKeys, the card is still locked at Planning, and the planner asks which story to plan
- A run in the person's own agent is killed as "stalled" while it waits on its PR's CI — the stall clock counts only run events, and the CI watch writes none, so the fix loop that would repair a red PR is the thing that gets killed
- (motir-ai) Parse the per-(provider,id) upstream-prices schema version 5 — keep versions 3 and 4, refuse a duplicate (provider,id) with its path, resolve a shared id to its catalog provider for the rate guards, amend contract.md §9
- A cost guard on each pool and the graph service: a hard machine ceiling at which work queues and an alert fires instead of a machine being created, an alert as the ceiling nears, and a tripwire on an abnormal machine-creation ratePlanning capacity follows demand — a conversation turn never waits behind a planning run, the planning machines and the graph service grow and shrink on their own, and a load-tested cost ceiling is the only fixed limit
- (motir-ai) Document the cost guard for operators: the four settings per group, the four alerts and what to do on each, the deploy refusals, and the release note that an unset CODEGRAPH_SCALE_CEILING now stops graph scale-outPlanning capacity follows demand — a conversation turn never waits behind a planning run, the planning machines and the graph service grow and shrink on their own, and a load-tested cost ceiling is the only fixed limit
- (motir-ai) Cost-guard e2e on the planning-pool harness, with a planning ceiling of 2 and a tripwire of 2 creates per window: creates up to 2, then work queued and one at-ceiling alert, then a forced create-then-destroy loop trips the wire, while the conversation pool is untouchedPlanning capacity follows demand — a conversation turn never waits behind a planning run, the planning machines and the graph service grow and shrink on their own, and a load-tested cost ceiling is the only fixed limit
- (motir-ai) Story integration gate against the real database: coverage of the cost guard's changed surface to the floor, concurrent creates under each group's lock never passing the ceiling, the tripwire tripping and resetting from the record, one alert per episode across evaluators and a graph leader change, pools and graph isolatedPlanning capacity follows demand — a conversation turn never waits behind a planning run, the planning machines and the graph service grow and shrink on their own, and a load-tested cost ceiling is the only fixed limit
- (motir-ai) Graph scaleOut asks the cost guard under a per-group lock: an unset CODEGRAPH_SCALE_CEILING now means no create, the ceiling and the tripwire hold the scale-out with a log line while reads keep being served, every create and failed attempt is recorded, and each tick runs the breached checkPlanning capacity follows demand — a conversation turn never waits behind a planning run, the planning machines and the graph service grow and shrink on their own, and a load-tested cost ceiling is the only fixed limit