Skip to content

atlas ingestion program: semi-automated updates, versioning, provenance, feedback — long-term design + tasks (2026-11-04 brief) #6018

Open

Description

@dezren39

Program (Drew, 2026-11-04)

A long-term plan for semi-automating updates to Atlas ingestion: what gets ingested, how content is versioned, a standard provenance model, delta-only updates, retrieval statistics + result feedback, classification safety, and disk/IO + plan-spend discipline. Statement inventory (S1–S40, auditable, with per-statement dispositions): ~/workspace/scratch/atlas-ingestion-brief-2026-11-04/statement-inventory.md in the sandbox workspace; this issue is the canonical tracker — children are linked in a comment below and closed into this issue as they complete.

Where Atlas stands (2026-11-04, verified)

  • Shard restoration complete: 31 TursoDB shards; internal = 28 shards / 107,001 chunks; secret = exactly 1 shard / 638 chunks (secret material only in the 2 secret DBs). Production monolith stays libSQL for now; TursoDB for all NEW databases. Cold tiers → rustfs (2026-11-04 storage ruling). Current spend anchor: Turso scaler $24.92/mo yearly — the plan is the ceiling (see also #5468, mirror storage pressure).
  • Search: fan-out over shards #5755; query front-end #5773; full parameter surface with per-run vector logging #5774; Jaccard rescore + ingest dedup ≥0.9 recorded edge #5733; staging tier with merge-on-close promotion #5669.
  • Provenance registry already designed: #3125 (chunk_occurrences, content-addressed extracts, repo@commit:path pinning). This program extends atlas: provenance registry (chunk occurrences), content-addressed extracts, multi-DB vector resolution #3125 (source-system marks, version lineage) rather than replacing it.
  • Scripts migration into code + fleet binary: #5670. Untracked Atlas material enumeration: #5981 — ingestion source enumeration overlaps it.

Topic map (children filed under this tracker)

  1. Ingestion operations design — semi-automated updates; automatic source set (reboots/boot records, session logs, chat logs, statistics, reports); usefulness filter; disk + I/O monitoring; plan-spend ceiling.
  2. Classification revalidation (task) — verify nothing sensitive sits at the wrong tier.
  3. Agent notes placement design — zettel vs repo vs publishing directly into the ID graph (ties to #5545, #5686).
  4. GitHub corpus extraction (task) — ALL issue/PR comments from ALL our repos; external repos we interact with or rely on (external half gated, see 5).
  5. External-content licensing gate (research) — may we copy external GitHub comments/discussions into Atlas at all? In the #5977/#5986 family. Must be answered before external-repo ingestion is built; design may proceed.
  6. First-pass corpus (task) — all code repos, all findable docs, dezren39/cairn.
  7. Content versioning + retention design — new versions, rename lineage, newest-resolution ("take the max"), cold rollout, deletion triggers (date / superseding increments / never accessed).
  8. Retrieval statistics + feedback + query trace design — top-K shard survival, top-1/2/3/5 after merge+rescore, thumbs up/down minimum, optional comments/essays, result↔query trace (per-stage scores, shard, exact embed state, metadata).
  9. Provenance standard — extensions filed as a comment on #3125: one standard for when/where; marks of the form repo@hash:path AND source-system marks (e.g. opencode-db @ amd @ <date>); segment search to associate unlineaged files; unassociated/duplicate as a logged state.
  10. Existing-embed provenance audit (task) — scan all current embeds for correct path/provenance details.
  11. Delta-only ingestion + offset embeds design — the 21 GB rule (20.5 GB already present → only new files + edits land; no mass retagging); span-anchored embeds (content at file offsets 17–241) with search-time joins that never double merged lines.
  12. Expansive v2–v4 design — the larger design that pushes depth/scope beyond this brief wherever research leads.

Horizons

  • v1 — first-pass corpus (own repos incl. GitHub comments, docs, cairn) + delta ingestion + provenance marking standard + embed audit + classification revalidation run.
  • v2 — file versioning/retention live (cold rollout + deletion policy) + retrieval statistics (shard survival, top-1/2/3/5).
  • v3 — feedback loop live (thumbs/comments/traces) feeding retention + ranking evolution; classification revalidation recurring/automated; offset-embed joins in production search.
  • v4 — expansive: feedback-trained ranking, provenance-complete resolution, cross-source lineage graph, ID-graph-native corpus objects, classification-aware retrieval.

Gates

  • External-repo comment ingestion is build-gated on topic 5's answer (Drew's sequencing: licensing figured out before the external build, after the design is somewhat done). Our own repos are not gated.
  • Nothing in this program raises spend past the current plan without Drew's explicit call (S3).

Activity

  1. added
    area/infraInfrastructure, deployment, and operations
    kind/trackerTracking issue: closed when its children close, never triaged by coverage
    on Nov 5, 2026
  2. developing-today commented on Nov 5, 2026

    @developing-today
    Contributor
  3. dezren39 commented on Nov 5, 2026

    @dezren39
    MemberAuthor

    Children filed (2026-11-04)

    Designs

    • #6019 — semi-automated ingestion operations: source set, usefulness filter, disk/IO + plan-spend guardrails (S1–S5)
    • #6020 — expansive v2–v4 design: feedback-driven ranking, provenance-complete resolution, self-curating corpus (S39–S40)
    • #6022 — where agent notes live: zettel vs repo vs publishing directly into the ID graph (S7–S10)
    • #6026 — content versioning + retention: versions, rename lineage, newest-resolution, cold rollout, deletion triggers (S16–S20, S29–S30)
    • #6027 — retrieval statistics, result feedback, query traces: top-K survival, top-1/2/3/5, thumbs + essays, per-stage scores (S21–S27)
    • #6028 — delta-only ingestion + offset embeds: the 21 GB rule; span-anchored (17–241) chunks with non-doubling joins (S35–S38)

    Tasks

    • #6021 — classification revalidation scan: nothing sensitive at the wrong tier (S6)
    • #6023 — GitHub corpus extraction: all our repos' issue/PR comments; external set gated (S11–S12, S14)
    • #6025 — first-pass corpus ingestion: all code repos, all findable docs, dezren39/cairn (S15)
    • #6029 — audit of existing embeds for provenance/path correctness (S34)

    Gate

    Covered by existing issues (comments posted, no duplicates filed)

    Statement inventory with per-statement dispositions: ~/workspace/scratch/atlas-ingestion-brief-2026-11-04/statement-inventory.md — 40 statements, zero unmapped.

  4. dezren39 commented on Nov 5, 2026

    @dezren39
    MemberAuthor

    Wave-2 update: research fold-in complete — 28 new children; child list extended

    The prior-art research lane (7 clusters, 46 proposed follow-ups) has been folded into this program. Dispositions: 28 newly filed (below), 16 covered by existing program issues (#6019–#6029 family — incl. join semantics → #6028, lineage schema + newest-resolution → #6026, trace schema / survival report / feedback capture / normalization → #6027, scanner execution → #6021, connector inventory + I/O politeness → #6019, manifest-first delta → #6028, agent-notes placement → #6022, first-pass plan → #6025), 2 covered by pre-existing issues (provenance record → #3125, ingest-time dedup edges → #5733 — both with substantive research comments). Six program issues also received findings comments that change their designs (licensing gate answered-provisional on #6024; CDC verdict on #6028; git-records-no-renames on #6026; log-at-query-time on #6027; three-layer scan on #6021; cost model on #6019). Full map: ~/workspace/scratch/atlas-ingestion-brief-2026-11-04/wave2-map.md.

    New children by cluster

    Delta / offset embeds (parent #6028): #6035 · #6036 · #6037 · #6038 · #6039 · #6040
    Versioning / retention (parent #6026): #6041 · #6042 · #6043
    Provenance: #6044 (parent: this tracker; home #3125) · #6045 (parent #6019)
    Stats / feedback (parent #6027): #6046 · #6047
    Classification (parent #6021): #6048 · #6049 · #6050 · #6051
    Licensing gate (parent #6024): #6052 · #6053 · #6054 · #6055 · #6056
    Ingestion ops / cost (parent #6019): #6057 · #6058 · #6059 · #6060 · #6061
    ID graph (parent #6022): #6062

    — Wave-2 research fold-in, 2026-11-04 (prior-art doc: ~/workspace/scratch/atlas-ingestion-brief-2026-11-04/prior-art-research.md)

  5. added
    queue/queuedHas a queued task in the Supabase agent queue
    task/triageQueue packet kind: verify, reproduce, and route this item
    work/assignedQueue worker lifecycle state
    queue/claimedQueue worker lifecycle state
    and removed
    queue/queuedHas a queued task in the Supabase agent queue
    on Nov 5, 2026
  6. dezren39 commented on Nov 5, 2026

    @dezren39
    MemberAuthor

    🤖 Agent note · handed off

  7. added
    queue/parkedQueue task is parked (waiting on CI, time, or an event)
    and removed
    queue/claimedQueue worker lifecycle state
    on Nov 5, 2026
  8. 48 remaining items

  9. dezren39 commented on Nov 5, 2026

    @dezren39
    MemberAuthor

    🤖 Agent note · handed off

  10. dezren39 commented on Nov 5, 2026

    @dezren39
    MemberAuthor

    🤖 Agent note · handed off

  11. dezren39 commented on Nov 5, 2026

    @dezren39
    MemberAuthor

    🤖 Agent note · handed off

  12. added
    queue/claimedQueue worker lifecycle state
    and removed
    status/parkedDeliberately paused with a recorded resume condition
    queue/parkedQueue task is parked (waiting on CI, time, or an event)
    on Nov 6, 2026
  13. dezren39 commented on Nov 6, 2026

    @dezren39
    MemberAuthor

    🤖 Agent note · handed off

  14. dezren39 commented on Nov 6, 2026

    @dezren39
    MemberAuthor

    🤖 Agent note · handed off

  15. dezren39 commented on Nov 6, 2026

    @dezren39
    MemberAuthor

    🤖 Agent note · handed off

  16. dezren39 commented on Nov 6, 2026

    @dezren39
    MemberAuthor

    Agent note — task 10384 handed off (lane code-n) — AT CAP

    • Task: task 10384 (triage, developing-today/code), attempts 6/6 — at the attempts cap. If this cycle ends without the claimant session completing the task, it will need a cap raise (fleet-monitor repair precedent) to be claimable again.
    • Execution lane: opencode per fleet queue lane — handing off to the laptop opencode claimant; this lane does not work the task directly.
    • Fresh check (2026-11-06T02:37Z): issue atlas ingestion program: semi-automated updates, versioning, provenance, feedback — long-term design + tasks (2026-11-04 brief) #6018 is OPEN, no hard-stop labels (held, user/owned, do-not-merge, do-not-auto-edit).
    • Next: claimant launches the bound opencode session; task parked pending that launch.
  17. added
    queue/parkedQueue task is parked (waiting on CI, time, or an event)
    status/parkedDeliberately paused with a recorded resume condition
    and removed
    queue/claimedQueue worker lifecycle state
    on Nov 6, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

area/infraInfrastructure, deployment, and operationskind/trackerTracking issue: closed when its children close, never triaged by coveragequeue/parkedQueue task is parked (waiting on CI, time, or an event)status/parkedDeliberately paused with a recorded resume conditiontask/triageQueue packet kind: verify, reproduce, and route this itemwork/assignedQueue worker lifecycle state

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions