Orchestrator Awareness — The Situation Engine

Vision (operator, 2026-08-25): AIDE itself understands its own context — hardware, models, tasks — and ADVISES: which model, which quant, what speed to expect, what to monitor, how to fine-tune better. Funding thesis: the IDE that can MEASURE and EXPLAIN its own AI assistance wins the trust the market lacks (GitKraken: 84% feel productive, 20% can prove it).

Situation Object (the orchestrator's senses)

Computed once per boot + refreshed on events; exposed at GET /api/orch/context:

{
  hardware: { ramFree, ramTotal, vramTotal, vramFree, gpu, backendCaps },
  engines: [{ id, name, backend, quant, ctxServed, tokPerSecBench, trusted }],
  profiles: { delegation, formatOnSave, lastIntent },
  activity: { egressLast24h, shipsCount, reworkCount, avgOutcomeLatencyMs },
  battery: { lastRunAt, delta }
}

Sources already shipped: probeHardware (30s cache), --list-devices capability cache, profile sidecars, served-ctx probe, egress.log, ships.log, battery JSONs. NEW code = aggregation + scoring only.

Recommendation Engine (model + quant + speed)

Scoring for "task T on machine M": candidate = installed OR catalog GGUF; score = taskFit(tags) * fitScore(RAM/VRAM via fitModel) * speedPrior(bench or quant-class prior) ; output top-3 with EXPECTED tok/s range from device-benchmark-runner data (never fabricated — "measured" vs "estimated" labeled). Surfaces:

• Cold card line: "Best for this machine: X (~N tok/s measured)"
• MODELS panel: BENCH button per engine -> runs llama-bench lane -> verdict chip
• TASK submit hint when selected engine mismatches task size ("large refactor detected — consider qwen-coder-1.5b, ~3x throughput")
• Hub search results annotated with fit verdicts (fitModel + quant class)

Monitoring Surface (what developers actually watch)

MONITOR sheet (rail badge click): live tok/s + TTFT of last reply, context fill %, RAM free, VRAM used (nvidia-smi poll 10s), egress events today, ships

• rework + outcome latency (from ships.log), diagnostics trend. All data sources exist; aggregation + render only.

Training Room v2 (fine-tune loop closure)

1. IMPORT auto-fills ctx from probeGguf metadata; registers as baseline.
2. After operator trains externally: same file path re-import -> version bump -> BATTERY reruns automatically -> delta table appended to evidence.
3. Dataset studio hooks: expose .aide/library/events.jsonl export in the fine-tune format (Loop C feeds training data).
4. Reward-design guidance surfaced in-room (combined unit-test+light-style; static-only degenerates — RLVR 2026 finding).

Funding/measurement law

Every AI-assisted change carries Assisted-by trailer; every ship logs intent->verified latency; batteries publish negatives. Public copy states: "AIDE measures its own AI assistance — outcome latency, rework rate, provenance on every reply." This is the funder differentiator (64-pt proof gap).

Threats / pitfalls

• Recommendations MUST label estimated vs measured (fabricated numbers destroy trust).
• nvidia-smi polls: 10s min interval, cached; never block request paths.
• Bench runs are GPU-heavy: P7 one-job-at-a-time + consent chip before launch.
• Catalog scoring must not phone home beyond explicit hub search (fit scores computed locally from metadata).
• Training room never auto-starts training; operator runs jobs (R2 hard rule).

Gates

1. /api/orch/context returns aggregated truth matching component endpoints.
2. Recommendation top-3 changes correctly when RAM drops below fit threshold (fixture test).
3. BENCH button produces verdict chip from real llama-bench JSON (smoke on bundled model).
4. MONITOR values match ground truth within tolerance (props/timings cross-check).
5. Battery-on-import flow triggers automatically when a re-registered artifact hash differs.