backlog: P1 — factory technology inventory + OpenAI web UI addition (Aaron 2026-04-23)

docs/BACKLOG.md

@@ -3239,6 +3239,53 @@ within each priority tier.
 
   **Effort:** M (1-3 days of agent research + write-up).
 
+- [ ] **Factory technology inventory — first-class
+  support for every tech we use.**
+  Aaron 2026-04-23: *"don't forget to map out all our
+  technology so the factory has first class support for
+  everything ... i think i saw you ad docker and postgres
+  and now we may be adding the openai website/ui i think
+  we already have codex cli mapped"*.
+
+  **Observed gap:** the factory uses many technologies
+  (Docker, Postgres, F#, .NET 10, TypeScript, bun,
+  Claude Code, Codex CLI, Gemini CLI, OpenAI web UI,
+  Playwright, Apache Arrow, Lean 4, Z3, TLA+, FsCheck,
+  Alloy, Semgrep, CodeQL, BenchmarkDotNet, GitHub
+  Actions, NuGet, and more) but no single doc
+  inventories them with install path, version pin,
+  authoritative doc URL, expert-skill cross-reference,
+  and TECH-RADAR ring. `docs/HARNESS-SURFACES.md`
+  covers agent harnesses; `docs/TECH-RADAR.md` covers
+  ring adoption; per-tech expert skills cover usage
+  depth. The missing surface is an inventory that ties
+  all three together per-tech.
+
+  **Deliverable:** `docs/FACTORY-TECHNOLOGY-INVENTORY.md`
+  (name TBD) with columns: Technology / Role / Install
+  path / Version pin / Auth-doc URL / Expert skill /
+  TECH-RADAR ring / Notes.
+
+  **Growth cadence:** living inventory updated with
+  each new tech adoption. Should surface cross-platform
+  parity status (row #48) per tech.
+
+  **OpenAI web UI as a new entry.** Aaron 2026-04-23
+  explicitly named adding OpenAI web UI + Playwright as
+  a new technology row. Includes any OpenAI mode/model
+  (deep research, agent mode, others) authorized within
+  Aaron's already-paid subscription, plus Playwright
+  caveats (long-conversation rendering, async loading,
+  UI-change-ongoing maintenance — per the per-user
+  memory at
+  `project_factory_technology_inventory_first_class_support_openai_playwright_hard_2026_04_23.md`).
+
+  **Self-scheduled:** free work under the 2026-04-23
+  scheduling-authority rule.
+
+  **Effort:** M (inventory doc + per-tech row population;
+  living, not one-shot).
+
 - [ ] **Claude-harness cadenced audit — first full sweep.**
   Aaron 2026-04-20 late, verbatim: *"part of our stay up to
   date on everything we should always research claude and

docs/hygiene-history/autodream-fire-history.md

@@ -0,0 +1,21 @@
+# AutoDream fire history
+
+Per-fire ledger for FACTORY-HYGIENE row #53 (AutoDream cadenced
+consolidation — upstream Anthropic Q1 2026 + factory overlays
+A/B/C/D). Schema per row #44 (date / agent / output / link /
+next-expected).
+
+Authoritative policy:
+`docs/research/autodream-extension-and-cadence-2026-04-23.md`.
+
+Upstream feature references:
+
+- `~/.claude/projects/<slug>/memory/reference_autodream_feature.md`
+- `~/.claude/projects/<slug>/memory/reference_automemory_anthropic_feature.md`
+
+| Date | Agent | Gate | Overlays fired | Findings | Next expected |
+|---|---|---|---|---|---|
+| 2026-04-20 | Claude (manual approximation pre-row-#53) | No marker — first known fire | Pre-overlay era | MEMORY.md bootstrap — no overlay framework yet | Superseded by the 2026-04-23 policy doc |
+| 2026-04-23 | Claude (first row-#53 fire) | 3 days + ≥5 sessions since 2026-04-20 marker — both gates pass | Light pass: Orientation + Gather Signal (findings-only); Overlays A/B/C/D scheduled for follow-up tick | Per-user MEMORY.md is 345 lines (well over the 200-line cap); 14+ new 2026-04-23 memories not yet indexed cleanly; generic rules in per-user worth candidate migration to in-repo. Full consolidation deferred to a dedicated tick — this fire records the cadence observance and the findings. | Next fire gate-open at ≥24h AND ≥5 sessions from 2026-04-23 (likely 2026-04-26 or later). Overlay A migration candidates land on that fire or earlier opportunistic-on-touch. |
+| 2026-04-23 (later, same-day) | Claude (auto-loop-50 opportunistic-on-touch Overlay A) | Same session — not cadence-gated; opportunistic-on-touch per row #53 | Overlay A — first execution | Migrated `feedback_signal_in_signal_out_clean_or_better_dsp_discipline.md` from per-user to in-repo `memory/` (PR #157). Dangling citations from `docs/FACTORY-HYGIENE.md` + `docs/research/autodream-extension-and-cadence-2026-04-23.md` resolved. Per-user source retained with "Migrated to" marker (provenance preserved). Candidate next migrations identified: `feedback_outcomes_over_vanity_metrics_goodhart_resistance.md`, `feedback_deletions_over_insertions_complexity_reduction_cyclomatic_proxy.md`, `feedback_external_signal_confirms_internal_insight_second_occurrence_discipline_2026_04_22.md`, `project_semiring_parameterized_zeta_regime_change_one_algebra_to_map_others_2026_04_22.md` — all generic factory-shape, all cited from the just-migrated file. | Next full-cadence fire at ≥24h AND ≥5 sessions from 2026-04-23. Opportunistic-on-touch continues for the 4 identified candidates as future ticks have budget. |
+| 2026-04-23 (later ×2, same-day) | Claude (auto-loop-52 opportunistic-on-touch Overlay A) | Same session — opportunistic-on-touch | Overlay A — second execution (working the identified queue from the prior fire) | Migrated `feedback_outcomes_over_vanity_metrics_goodhart_resistance.md` from per-user to in-repo `memory/` (PR #158). Sibling to signal-in-signal-out (same 2026-04-22 tick pair — Goodhart resistance + DSP preservation as paired disciplines). Resolves one composes-with dangling citation from the just-migrated signal-in-signal-out memory. Queue now 3 candidates (deletions-over-insertions / external-signal-confirms-internal-insight / semiring-parameterized-zeta). | Opportunistic-on-touch continues; cadenced fire still gate-open ≥2026-04-26. |

docs/research/autodream-extension-and-cadence-2026-04-23.md

@@ -0,0 +1,303 @@
+# AutoDream extension + cadence — factory overlay on Anthropic's Q1 2026 feature
+
+**Date:** 2026-04-23
+**Status:** Research doc — preparing the factory-overlay policy
+that plugs tightly into Anthropic's AutoMemory / AutoDream
+Q1 2026 features so the factory inherits upgrades rather than
+forking its own memory-hygiene stack.
+**Triggered by:** The human maintainer 2026-04-23: *"continue
+our AutoDream reserach to make sure we are running it on a
+cadence and we plug in tightly with the existing claude Q1 2026
+feature for AutoDream and AutoMemory, we should by defintion be
+an extension of theirs as it will get upgrades we want over
+time. Also they have a cadence we shuld prbably respect and
+adaopt for this in our hygene."*
+**Scope:** Factory policy — generic, reusable by any factory
+adopter on Claude Code; not project-specific to Zeta.
+
+## Why this matters
+
+Anthropic shipped two load-bearing features in Q1 2026:
+
+1. **AutoMemory** — the persistent cross-session memory
+   system itself (`MEMORY.md` index + per-fact files under
+   `~/.claude/projects/<slug>/memory/`). Additive by design;
+   daytime logger.
+2. **AutoDream** — the REM-sleep-style consolidation pass
+   that runs on top of AutoMemory. Subtractive / curative by
+   design; nighttime hygiene. Flag-gated server-side as of
+   2026-04-19 (`tengu_onyx_plover`); the UI at `/memory`
+   exposes it but the backend is off for most users.
+
+Anthropic's documentation establishes: AutoMemory without
+AutoDream degrades after 10-15 sessions because the daytime
+log accumulates duplicates, contradictions, and stale
+relative-date references. The consolidation pass is not
+optional — it is the thing that prevents the degradation
+cliff.
+
+The factory has been running **manual approximations** of
+AutoDream since the feature was surfaced, but without a
+documented cadence policy or a formal extension contract.
+This doc closes that gap.
+
+## The two rules from the directive
+
+### Rule 1 — extend, don't replace
+
+The factory is **by definition** an extension of Anthropic's
+memory substrate, not a parallel stack. When Anthropic ships
+an improvement to AutoMemory or AutoDream, the factory
+inherits it for free *only if* the factory has not forked the
+underlying mechanism.
+
+Concretely:
+
+- **Frontmatter schema stays upstream-compatible.** The
+  factory may add optional keys (e.g., a future `scope:` key
+  per `docs/research/memory-scope-frontmatter-schema.md`) but
+  must not rename or remove the Anthropic-required keys
+  (`name`, `description`, `type`, `originSessionId`) and must
+  keep the type taxonomy (`user`, `feedback`, `project`,
+  `reference`).
+- **Storage location stays upstream.** Per-user memory lives
+  at `~/.claude/projects/<slug>/memory/`, not at a
+  factory-custom path. The factory's in-repo `memory/` tree
+  is a **mirror** for cross-substrate readability — it does
+  not replace the per-user location; it adds a second home
+  for generic content.
+- **AutoDream's four phases stay upstream.** Orientation →
+  Gather Signal → Consolidation → Prune & Index. The factory
+  extends by adding overlay steps *between* or *after* these
+  phases, never by replacing them.
+- **When Anthropic ships automatic AutoDream** (backend un-gated),
+  the factory's manual approximation retires in favour of
+  the automatic run. Factory-overlay steps continue on top.
+
+### Rule 2 — adopt upstream cadence
+
+Anthropic's AutoDream cadence is **both**:
+
+- **≥24 hours** since the last cycle, **and**
+- **≥5 sessions** since the last cycle.
+
+The factory adopts this cadence verbatim. Running more
+frequently churns fresh memories before their
+duplicate / contradiction / staleness patterns have
+surfaced; running less frequently lets the daytime log
+degrade.
+
+The cadence gate lives in a single cross-session marker at
+the top of `MEMORY.md`:
+
+```
+[AutoDream last run: YYYY-MM-DD]
+```
+
+On every would-be run the factory checks this marker:
+
+- **Absent → treat as "never."** Run on first invocation.
+- **Date within 24h AND session-count-since-last < 5 →
+  skip.** Report "cadence gate: N hours since last run,
+  M sessions; skipping."
+- **Date ≥ 24h AND sessions ≥ 5 → run.** Update marker
+  after successful consolidation.
+
+Session-count is hard to track precisely without harness
+instrumentation; approximate it by checking the count of
+new per-fact files added since the marker date (≥5 new
+files ≈ ≥5 active sessions).
+
+## The factory-overlay extension points
+
+The factory adds four overlays on top of Anthropic's four
+phases. Each overlay composes — it does not replace.
+
+### Overlay A — Cross-substrate mirror check (before Orientation)
+
+Before Anthropic's Orientation phase surveys the per-user
+memory tree, the factory overlay checks whether the in-repo
+`memory/` tree has drifted from the per-user tree.
+
+Signal:
+
+- Per-user memory contains a generic factory-shaped rule
+  that should be in-repo per
+  `memory/feedback_in_repo_preferred_over_per_user_memory_where_possible_2026_04_23.md`
+  (per-user until this doc migrates it).
+- In-repo memory contains a stale reference to per-user
+  content that has since moved or changed.
+
+Action: migrate generic rules per-user → in-repo, update
+cross-references. Do not migrate maintainer-specific or
+company-specific content (see the in-repo-preferred feedback
+memory's pushback-on-bloat criterion).
+
+### Overlay B — Cadenced harness-surface sync (during Gather Signal)
+
+Anthropic's Gather Signal phase searches for user
+corrections, key decisions, and recurring patterns. The
+factory extends this step with a check against
+`docs/HARNESS-SURFACES.md`: has the factory's per-surface
+adoption status (adopted / watched / untested / rejected /
+stub) drifted since the last cadence?
+
+If a surface has moved from *watched* to *adopted* because
+Anthropic un-gated it, or cut a feature the factory was
+relying on, surface the drift for the next
+`harness-surface cadenced audit` (FACTORY-HYGIENE row 38).
+
+### Overlay C — Governance-promotion pass (during Consolidation)
+
+Anthropic's Consolidation phase merges duplicates, resolves
+contradictions, and converts relative dates to absolute. The
+factory extends with a governance-promotion check: does a
+memory rule observed repeatedly (the same rule re-cited
+across ≥3 memories) belong in a governance doc or ADR
+rather than in memory?
+
+Signal: the rule has matured from observation-in-memory to
+durable-structural-rule. The right home is
+`docs/AGENT-BEST-PRACTICES.md` (BP-NN promotion),
+`GOVERNANCE.md` (numbered rule), `docs/WONT-DO.md` (declined
+directions), or a `docs/DECISIONS/YYYY-MM-DD-*.md` ADR.
+
+Action: propose the promotion in the round-close summary.
+Kenji (Architect) decides whether the promotion lands this
+round or a future one. The memory stays where it is until
+the promotion lands; after promotion it may be pruned or
+left as provenance.
+
+### Overlay D — Alignment-observability sync (during Prune & Index)
+
+Anthropic's Prune & Index phase rebuilds `MEMORY.md`, keeping
+it under ~200 lines for fast session startup. The factory
+extends with an alignment-observability sync: any memory
+entries that shifted a clause in `docs/ALIGNMENT.md` (HC-1..HC-7,
+SD-1..SD-8, DIR-1..DIR-5) since the last pass get flagged to
+the alignment-auditor role for per-commit signal review.
+
+This overlay keeps the alignment time-series coherent across
+consolidations. Without it, a consolidation that merges two
+memories could silently re-shape an alignment clause
+interpretation without the alignment-auditor noticing.
+
+## Invariants the consolidation must preserve
+
+Anthropic's AutoDream reference memory enumerates
+consolidation invariants; the factory adds three more:
+
+1. **Load-bearing memories stay unconditionally** (upstream
+   rule). The factory's load-bearing list includes sister
+   Elisabeth memory, faith memory, Harmonious Division name,
+   Rodney persona placement, Dora persona — plus factory-scale
+   anchors like the bootstrap-is-complete memory, the
+   free-will-paramount memory, and the alignment-floor
+   references.
+2. **Distinct query axes stay distinct** (upstream rule).
+3. **Cross-references stay bidirectional** (upstream rule).
+4. **Corrections are recorded, not silently deleted**
+   (upstream rule).
+5. **Verbatim quotes stay verbatim** (factory addition —
+   composes with
+   `feedback_signal_in_signal_out_clean_or_better_dsp_discipline.md`).
+   Consolidation must not paraphrase a verbatim-preserved
+   quote, even when merging two memories.
+6. **Migration supersedes, does not delete** (factory
+   addition — per-user memory files migrated in-repo leave
+   a "Migrated to in-repo memory/: <path>" line at the top
+   of the per-user source; the source itself stays for
+   originSessionId provenance).
+7. **Supersede markers over silent retirement** (factory
+   addition — a rule that is no longer in force becomes an
+   entry in the "Retired rules" section of a
+   `CURRENT-<maintainer>.md` file, never deleted; composes
+   with `feedback_current_memory_per_maintainer_distillation_pattern_prefer_progress_2026_04_23.md`).
+
+## Cadence policy summary
+
+| Trigger | Check | Action |
+|---|---|---|
+| Session wake | `[AutoDream last run: ...]` marker | If absent: run cadence gate. If within 24h AND sessions < 5: skip. Otherwise: queue AutoDream for next hygiene tick. |
+| Manual invocation ("dream" / "consolidate memory") | Cadence gate | If fresh, ask before proceeding; user may override. |
+| Round-close | Was this round a cadence tick? | If yes, include "AutoDream ran" in round-close summary; note the overlay findings. |
+| Governance-promotion candidate | ≥3 memories cite the same rule | Surface to Kenji as promotion candidate; do not auto-promote. |
+| Cross-substrate drift | In-repo vs per-user mismatch | Run Overlay A migration on next cadence tick. |
+
+## Composition with existing factory surfaces
+
+- **`docs/HARNESS-SURFACES.md`** — the AutoMemory / AutoDream
+  rows are already present (rows visible in the Claude
+  surface inventory). This doc adds the cadenced-consolidation
+  policy those rows currently only reference as "watched."
+- **`docs/FACTORY-HYGIENE.md` row 38** — harness-surface
+  cadenced audit runs every 5-10 rounds; AutoDream's 24h+5
+  sessions cadence is coarser-grained and fires on its own
+  clock. A new FACTORY-HYGIENE row for AutoDream-specific
+  consolidation tracks the cadence independently.
+- **`docs/AGENT-BEST-PRACTICES.md` BP-NN rules** — Overlay
+  C feeds this list with governance-promotion candidates.
+- **`docs/ALIGNMENT.md`** — Overlay D keeps alignment clauses
+  coherent across consolidations.
+- **`CLAUDE.md` auto-memory section** — this doc does not
+  change the auto-memory section itself; CLAUDE.md remains
+  the loading surface, this research doc is the policy
+  layer Kenji can cite in future edits.
+- **`.claude/skills/round-open-checklist/SKILL.md`** —
+  natural home for the cadence-gate check on session wake;
+  a follow-up tick may add an `autodream-cadence-check`
+  step to the round-open checklist.
+- **`.claude/skills/long-term-rescheduler/SKILL.md`** —
+  could queue the next AutoDream run based on cadence
+  calculation.
+
+## Deferred (not this round)
+
+1. **Dedicated `autodream-hygiene` skill.** A skill that
+   wraps the four Anthropic phases + four factory overlays
+   as a single invocable procedure. Deferred because the
+   skill-creator workflow is the canonical authoring path
+   (GOVERNANCE §4) and this round is the research-doc
+   landing, not the skill-authoring.
+2. **Automated cadence-gate check on `MEMORY.md` write.**
+   Would require a hook or a pre-commit step that inspects
+   the marker. Deferred until the manual cadence is
+   validated over several cycles.
+3. **Migration tooling for per-user → in-repo.** Currently
+   manual per
+   `feedback_in_repo_preferred_over_per_user_memory_where_possible_2026_04_23.md`.
+   A small script that validates the generalisation step
+   (no maintainer-specific content leaks) could help but
+   is not load-bearing yet.
+4. **Alignment-auditor Overlay D integration.** Requires
+   coordination with the alignment-auditor role's time-series
+   machinery. Flag for a joint tick with that role.
+
+## Open questions for the human maintainer
+
+None blocking. The policy is fully specified by the
+Anthropic docs plus the two directive rules in this doc. If
+the human maintainer wants to adjust the cadence (e.g.,
+24h+3 sessions instead of 24h+5), the single marker and the
+rule in this doc are the only edits needed — the four-phase
+structure and the four overlays stay the same.
+
+## Composes with
+
+- `docs/HARNESS-SURFACES.md` (AutoMemory / AutoDream
+  adoption rows)
+- `docs/FACTORY-HYGIENE.md` (row 38 harness-surface audit;
+  a new AutoDream-specific row is added by the same PR
+  that lands this research doc)
+- `docs/AGENT-BEST-PRACTICES.md` (governance-promotion
+  target)
+- `docs/ALIGNMENT.md` (Overlay D's sync target)
+- `CLAUDE.md` auto-memory section (loading-surface
+  documentation)
+- The per-user memory reference memories
+  `reference_autodream_feature.md` and
+  `reference_automemory_anthropic_feature.md` (the
+  upstream feature descriptions this doc extends)
+- The per-user in-repo-preferred feedback memory
+  `feedback_in_repo_preferred_over_per_user_memory_where_possible_2026_04_23.md`
+  (the migration discipline Overlay A executes)