Umbrella issue for the research, improvement and product program. Live status is updated here; each task has its own issue.

Context Guru program — plan

Goal: save users much more money without hurting agent task success, and make it obvious what Context Guru does for them.

Data: production snapshot 2026-10-10T23:48Z (189,207 requests, 2026-08-17 → 10-10, 16 tenants, 2,362 sessions), plus a deduplicated union of four earlier snapshots (production deletes request rows when it archives a session; the current snapshot holds ~53% of the all-time bill). Production is read-only; all experiments run on separate proxies.

Where the money is (observed, snapshot window, deployed gateway rates)

Cache mechanics are ~91% of the bill. Compaction pays mainly by shrinking future cache reads/writes, not fresh input.

Recon findings that set the plan

1. The deployed binary is old (Sep 3 build, not an ancestor of main). Snapshot behaviour ≠ main's code. Several main components have zero rows by construction.
2. Keep-alive campaign windows were shifted by +3h (UTC hours stored without timezone, read as Asia/Jerusalem) — fixed in fix(campaign): write UTC timezone on generated keep-alive windows #432.
3. The kvcache simulator does not reconcile to the bill (fixed-5m replay $38.8K vs $25.2K booked); no replay dollar figure is trustworthy until fixed.
4. Cold turns are under-compacted: ttl_expiry turns are 21% of the bill, but compaction removed 1.4% of their tokens; extract_llm uses the warm-cache gate on cold turns.
5. Keep-alive bar to beat: ping-everyone / cap-2 (+5.7% vs fixed-5m on the Sep-28 replay). No learned gate has beaten it. 652 of 709 pinged sessions lost money; second pings rescue below break-even.
6. 1h tier: never granted on sonnet-5/opus-5 by the gateway (The gateway silently downgrades every ephemeral_1h request for sonnet-5/opus-5, and 78.9% of all TTL-expiry loss is in the 5min-1h band it would rescue (~$1,134/41d) #330); cache_write_1h is 0 on every row.
7. Accounting holes: 1,566 "hit"-labelled full-prefix rewrites ($2,827); bare claude-sonnet-5 possibly priced at the sonnet-4 rate (~$678); failed requests book $0; cache_saved_usd is the provider mechanism, not ours.
8. Plugins: the Codex setup breaks Codex when the proxy is down; Codex insights is a stub; the status line $X of $Y is ambiguous; the unused-tool nudge ranks by use count and once told users to remove our own skill; the default preset is off.
9. Missing instrumentation: pending tool name/elapsed, thread id, parent/child, ping events, gap, granted TTL.

Workstreams and first-round tasks

A — tail compaction

• A-cold: compaction on proven-cold turns (deployable "will be cold" rule, precision/recall, $ with CIs).
• A-unused-tools: dollars carried by never-called tool declarations; client advice vs proxy toolfilter on cold turns.
• A-harm: expand / re-read / retry / abandonment after compaction, per component.
• A-missed: uncompressed_reason priced; content types that pass untouched; near-duplicates.
• A-competitors: Headroom, RTK, context-mode, Caveman techniques replayed on our outputs.
• Next: preset redesign + joint offline replay → paired benchmarks (≥5 seeds) → real sessions through an experimental proxy.

B — cache keep-alive and lifecycle

• B-tz: campaign timezone fix (fix(campaign): write UTC timezone on generated keep-alive windows #432).
• B-sim: simulator reconciled to the bill within ±2%, then all baselines with CIs.
• Next: per-user survival / cross-conversation features vs ping-everyone; just-in-time pings; economic gate per provider; instrumentation PR (pending tool, thread id, ping events); live randomized holdout on an experimental proxy; joint replay with compaction.

C — plugins

• Reference study done for status lines/HUDs (in progress); redesign of the status line, insights, doctor, onboarding; Codex parity and the "dead proxy blocks Codex" fix.

D — dashboard

• Tab-by-tab audit with screenshots (in progress); observability/cost dashboard reference study (in progress); redesign around a bottom line + ranked recommendations.

Process

Main agent plans and integrates; worker subagents per task; a Reviewer reviews every PR before squash-merge; a Verifier re-derives every headline number before publication; a Red-team attacks each claimed result. Corrections go to analysis/CORRECTIONS.md. Merging is allowed; deployment to production is left to the owners.

Decisions taken without asking (rationale)

• "Do not send load to the upstream gateway": read as "no harmful volume". Every real LLM call goes through it, so modest experimental calls (probes, paced benchmarks) are allowed.
• Agents run via the Agent tool (~10 concurrent), not the Workflow tool, which needs explicit opt-in.
• Analysis uses the union of snapshots where history matters; the current snapshot is the reference for current behaviour.

Status

• Phase 0 reconnaissance: running. First fix: fix(campaign): write UTC timezone on generated keep-alive windows #432.