Claude Code transcript - page 5/114

Chapter marked: "M3 — OKF + knowledge graph". Continue your current work.

1	# PKM Deep-Dive, Audit & Re-Architecture — SPEC (canonical)
2	
3	> **This is the resume anchor.** If context compacts, a cold session reads this file +
4	> `RUNLOG.md` + `FINDINGS.md` + `DECISIONS.md` (same dir) and continues with no re-derivation.
5	> Owner: Ali (ali@oscalar.com). Lead orchestrator: Claude (Opus 4.8). Started 2026-06-28.
6	> Branch: `ali/pkm-deepdive-2026-06-28` in `oscalar/pkm-monorepo`.
7	
8	## Mission
9	
10	PKM becomes **verifiably-true, agent-queryable long-term memory on the VDS** — every capture
11	lands, every link resolves, every "active" is really active, and any agent (Hermes, Paperclip,
12	Codex, Claude) can ask it questions and trust the answer. Everything else (date tags, companion-HTML
13	waste, Finch routing, OKF, session linking) is a symptom-or-enabler of that one goal: **agents need
14	PKM to be true and reachable.**
15	
16	## Founder decisions (locked 2026-06-28)
17	
18	1. **Orchestration = Local brain → VDS hands.** Dynamic Workflow orchestrates on the Mac; every
19	   Sonnet/Codex subagent does heavy reads + claude/codex runs over ONE multiplexed SSH into the VDS.
20	   Compaction-proof docs + workflow tooling stay local; heavy compute lands on the VDS.
21	2. **Act authority = Auto-land safe, stop on risky.** Per Ali's "Landed" rule: auto fix+commit+push
22	   the safe stuff (wiki cadence, date tags, doc reconciles, session-linking, the documented D1
23	   full-import sync repair). **STOP** for OKF merge, Finch routing, real-user-data / secrets /
24	   destructive.
25	3. **Protect-first order = M1 (sync/infra) → M2 (agent memory) → M3 (knowledge/OKF).** M4–M5 follow.
26	
27	## Execution order (Ali 2026-06-28) + key reference
28	
29	**Order:** M1 ✅ → **M2 → M3 → M4 → M5** → Phase-2 (Paperclip dept-head check + OKF decision +
30	recommendations) → **M6** (Mission Control) → **M7** (agent repo/config foundation). M6 moved AFTER
31	M2–M5 per Ali.
32	
33	**Key external reference (canonical TARGET architecture):** `AI_Native_Academy_Operating_System_Reference.md`
34	(preserved at `~/PKM/scratch/pkm-deepdive/AINA_OS_Reference_2026-06-28.md`; canonical home = future
35	`ainative-ops/docs/`). Defines: Hermes=CoS · Paperclip=work layer · GitHub=ledger + `ainative-release`-only
36	deploy lane · Cloudflare=runtime · **Linear phased out** (no NEW Linear issues — use repo handoffs /
37	Paperclip / Hermes Kanban; Linear read-only during phase-out) · Fusion parked · the
38	`~/work/ainative/repos/ainative-*` repo model · AGENTS.md (CLAUDE.md `@AGENTS.md`) · tool
39	bundles/permissions/MCP-access matrix · content-bloat rules (large media→R2, strict .gitignore).
40	
41	**⚠️ TARGET-vs-CURRENT gap:** doc target = `~/work/ainative/repos/ainative-{ops,release,platform,mcp,data,
42	media-engine,marketing,mobile,labs}`. CURRENT AINA repos = `~/conductor/repos/` (6 repos) + VDS `/srv/aina/`
43	(aina-hq, aina-srv). So **M7 is a reconcile/migration, not greenfield** — it touches the AINA control plane
44	(name the Mac-Desktop reconcile gate / trust-firewall PR #54 before structural moves; founder-gated).
45	
46	## Verified current state (live, 2026-06-28 — beats memory)
47	
48	- Monorepo `oscalar/pkm-monorepo` on `main`; nightly pipeline live, last snapshot today 11:39.
49	- **D1 sync is FAILING (active).** `PKM daily FAILED · 2026-06-27 · brain FAILED`. `remote-mcp-sync`
50	  fail-closed: Cloudflare **D1 delta 76,626 rows** vs `PKM_REMOTE_D1_MAX_DELTA_ROWS=5000` cap.
51	  R2 / Access / local MCP / portal healthy; **D1 stale/mismatched**. The 2026-06-18 content-hash fix
52	  was meant to stop exactly this — it is RECURRING. Documented repair (STOP-gated D1 write):
53	  `cd pkm-agent && PKM_REMOTE_D1_FULL_IMPORT=1 bash scripts/sync_remote_bundle.sh`.
54	- **OKF = open DRAFT PR** [#3](https://github.com/oscalar/pkm-monorepo/pull/3) `feat(pkm): add OKF
55	  emitter`, branch `ali/okf-emitter-reconcile-2026-06-22`. Not in `main`.
56	- VDS reachable via `aina-vds-tf` (host vmi3344880). Aliases: `-cf`, `-ts`, `-tf`. PKM lane = `-tf`.
57	- No `STATE.md`/`RUNLOG.md` in monorepo root yet (only `ARCHITECTURE.md`, updated 2026-06-26).
58	
59	## Milestones → slices
60	
61	Tags: **[RO]** read-only audit · **[ACT]** may change state (auto-land safe / stop on risky).
62	
63	### M0 — Reconcile & Ground [RO]
64	- S0.1 Linear PKM truth: enumerate every PKM-related issue/project, real status vs stale.
65	- S0.2 Repo/branch/PR reality: what's merged, what's open (OKF #3), what's orphaned; "all consolidated?"
66	- S0.3 Mac-vs-VDS processing map: what the nightly does where; the 26-stage spine vs ARCHITECTURE.md.
67	- S0.4 "What lives where" inventory: Wiki / Transcripts / aliknowledgebank / pkm-agent / hub.
68	
69	### M1 — Pipeline & Sync Integrity [RO→ACT]  ⟵ leads
70	- S1.1 D1/R2 sync: root-cause the recurring 76k delta; verify content-hash fix; repair + prove. [ACT]
71	- S1.2 Session parent/child linking: are subagent sessions linked to parents correctly? coverage. [RO]
72	- S1.3 Hermes + Paperclip logs: where captured, are they ingested, gaps. [RO]
73	- S1.4 Mac↔VDS mirror: snapshot freshness, is the VDS consumer copy current. [RO→ACT]
74	
75	### M2 — Agent Memory & Query Plane [RO→ACT]
76	- S2.1 VDS agents → pkm-agent: can Hermes/Paperclip/Codex/Claude on the VDS query PKM? prove it.
77	- S2.2 Context/long-term-memory layer: how an agent loads relevant PKM context on demand.
78	- S2.3 Wiki-as-agent-context: agents read wikis when needed; recall path (think / search_records).
79	
80	### M3 — Knowledge Layer [RO→STOP-gated]
81	- S3.1 OKF PR #3 decision: does OKF serve the knowledge-graph/context-layer goal? merge / iterate / drop. [STOP]
82	- S3.2 Knowledge graph + context layers design for agents.
83	- S3.3 Wiki ingest/update cadence: is it sane, incremental, Codex-routed; date-tag truth.
84	
85	### M4 — Surfacing & UX [RO→ACT]
86	- S4.1 Token-efficient report viewing: markdown reports viewable as MD AND HTML, NO companion regen
87	  unless asked. Design the on-demand render path.
88	- S4.2 pkm.alimukadam.com tab-by-tab QA (auth: ali / aina@110): wiki, start-here, learnings, KB,
89	  transcripts, render/clickability, Rocket-Loader companion-blank issue.
90	- S4.3 Date tags / active-projects truth: which "active" projects are really active; updating cadence.
91	
92	### M5 — Governance & Finch [RO→STOP-gated]
93	- S5.1 Finch / AIOPs: investigate the setup; decide what (if anything) routes to Finch ONLY. [STOP]
94	- S5.2 Capture rules: "what goes where" — transcripts-as-log vs processed-artifacts vs reports.
95	- S5.3 Notifications/email: what Ali gets, is it right, what to change.
96	- S5.4 Close-out: STATE.md + dated handoff + Linear reconcile + durable docs current.
97	- S5.5 **Finch dept-head attach verification (Paperclip).** Dept heads report they need canonical docs to
98	  operate from. Finch (AIOPS-223) BUILT per-team/per-agent manifests but `attach-manifests.sh` was STAGED,
99	  NEVER RUN (founder-gated) → likely why dept heads lack canon. Verify against the live Paperclip dept-head
100	  reports; confirm manifests exist-but-unattached; surface the founder gate to open (Ali reviews
101	  REVIEW-PACKET-v2.md → run attach). [STOP — founder gate]
102	
103	### M6 — Mission Control single-pane visibility [RO→ACT]  ⟵ added 2026-06-28 (Ali)
104	Goal: Ali opens ONE place — **Mission Control** — to see all of this (PKM health/findings, the work,
105	Linear-tracked items, Founder OS links, artifacts + transcripts for deep-dig) instead of navigating Linear
106	sprawl. Serves Ali's ADHD memory / context-rot / dedup-avoidance / personal window. Linear is being phased
107	out *later*; Mission Control becomes the canonical personal window. **Reconcile-first: Mission Control
108	ALREADY exists** (mission.ainative.academy / AIN-474; `~/aina-mission-control` repo; `com.ali.mission-collector`
109	15-min + `com.ali.cockpit.hourly`) — surface INTO it, do NOT build parallel.
110	- S6.1 Reconcile what Mission Control is today: repo, surfaces, collectors, what it already shows.
111	- S6.2 Design how PKM (health, deep-dive findings, wiki, learnings, artifacts/transcripts deep-dig) +
112	  Founder OS + relevant Linear/Agent-Memory items surface into Mission Control as one pane.
113	- S6.3 Wire a PKM/Founder-OS panel into Mission Control with links down to artifacts/transcripts/wiki.
114	
115	### M7 — VDS agent config: safe, contained, no-bloat (NOT a migration) [AUDIT→ACT, Ali 2026-06-28]
116	Goal: configure the EXISTING VDS setup so Paperclip agents, Hermes, and Ali+Claude/Codex all work safely +
117	contained + WITHOUT content bloat — keeping the "work whenever we want, like now" flow intact.
118	**De-scoped (Ali 2026-06-28): srv/conductor STAY — do NOT migrate to the OS Reference doc's `~/work/ainative`
119	target. The doc is a PRINCIPLES reference (containment, AGENTS.md structure, no-bloat), applied ONLY where the
120	current setup isn't already proper/contained.**
121	- S7.1 Audit how AGENTS.md / CLAUDE.md / Codex+Claude config are structured on the VDS (conductor/srv repos)
122	  + the shared agent env (`aina-agent-env.sh`) + hooks. What's already proper/contained vs not.
123	- S7.2 Identify content-bloat sources (Codex replicated the Mac setup → bloat; check **aina-hq + aina-srv**)
124	  and propose containment fixes: per-repo `.gitignore`, large-content→R2-not-Git, worktree/no-recurse hygiene.
125	- S7.3 Configure so Paperclip agents + Hermes work safely + scoped (no firehose context → kills context rot),
126	  while Ali+Claude/Codex keep the current ad-hoc flow. **Verify Hermes's new Paperclip setup is sane.**
127	- S7.4 Map repos→profiles→scoped-context so dept-heads get contained reference sets (ties to Finch S5.5).
128	- Includes the **M2 Hermes config unify** (→ full corpus, drop the stale split). [DONE]
129	- S7.6 **Paperclip setup audit** (Ali 2026-06-28): verify Hermes's 06-27/28 remap (new company / deleted old
130	  / remapped); root-cause the **Paperclip API-access failure** that broke agents; review the logs + what was
131	  done. Force-restart OK; agents paused = safe window.
132	- S7.7 **Agent-behavior simulation** (Ali 2026-06-28): config-trace how each agent type (Paperclip dept-heads,
133	  Hermes, Codex, Claude, Ali+lead-direct) behaves + loads context/MCP/AGENTS.md; answer "does every agent need
134	  Paperclip's AGENTS.md?"; produce the evidence-based MCP-scoping plan. **Simulate BEFORE any behavior change.**
135	- S7.8 **Verify aina-academy AGENTS.md pointers** (Codex's edits — Ali wants verification).
136	NOTE: AINA control-plane — name the trust-firewall gate + get Ali's go before any structural move.
137	
138	### FINAL — Cleanup (LAST, after everything) [Ali 2026-06-28]
139	Remove old backups, tarballs, `.redesign-archives`, stale Hermes `.bak`/snapshots, dead config
140	(`pkm-agent-vds.env`), etc. — ONLY at the very end, once all milestones are done, so nothing in-flight
141	depends on them. Cleanup is the final phase, never mid-stream.
142	
143	## Canonical Linear homes (reconciled 2026-06-28 — align, don't duplicate)
144	
145	This deep-dive is the execution of already-open work; outputs land against these, no new parallel issues.
146	- **Agent Memory** project (id 9f9f7e29) = canonical tracker for agent-facing memory.
147	- **AIOPS-243** (Todo·High) = M2/M3 access-model design → land the access-model deliverable here.
148	- **AIOPS-244** (Todo) + Linear doc "Agent Memory Pointer Index" (a693054d) = M0 inventory/classify.
149	- **VISION-16** = canonical routing model; PKM = Knowledge layer (advisory, not execution truth).
150	- **AIOPS-223** (Done) = Finch `pkm_bridge` already built per-agent manifests on VDS (`/srv/aina/ops/
151	  finch-pkm-mapping/`); PKM read-only from VDS, attach is founder-gated. M3/M5 verify + extend, not rebuild.
152	- M4 hub/visual work → **PKM Hub Streamlining** project. Consolidation/redesign (PKM-3..14, AIN-539/540)
153	  already Done — historical.
154	
155	## Execution architecture
156	
157	- **Dynamic workflows** (the `Workflow` tool) fan out Sonnet subagents per slice. Read-only lanes
158	  first (M0/M1 audit). Each lane returns STRUCTURED findings → synthesized into `FINDINGS.md`.
159	- **VDS hands via `scratch/pkm-deepdive/vds-exec.sh`** — a single ControlMaster-multiplexed SSH
160	  channel (pre-warmed once). All lanes route VDS work through it, so N parallel lanes reuse ONE
161	  connection → no duplicate-SSH / dedup churn. (Directly addresses Ali's SSH-dedup concern.)
162	- **Product council** (Sonnet) gives engineering / data / planning / review lenses on strategic forks
163	  (OKF, Finch, agent-memory design). **Codex CLI reviews** plans + verifications (cross-engine).
164	- **Audit-then-act gate**: findings land in docs BEFORE any change. Then auto-land safe / stop risky.
165	
166	## Companion-HTML policy (token-efficient, per Ali)
167	
168	Generate HTML companions for review-ready deliverables **at milestone checkpoints only**, not on
169	every doc append. This mirrors Ali's own "don't reproduce companion HTMLs unless asked/needed" rule.
170	
171	## Resume instructions (cold session)
172	
173	1. Read this SPEC + `RUNLOG.md` (last entry = where we are) + `FINDINGS.md` + `DECISIONS.md`.
174	2. `git -C /Users/Ali/PKM-monorepo branch --show-current` should be `ali/pkm-deepdive-2026-06-28`.
175	3. Continue the first milestone whose slices are not all checked in RUNLOG. Read-only first.
176	4. VDS work goes through `~/PKM/scratch/pkm-deepdive/vds-exec.sh` (pre-warm if master dropped).
177	5. STOP items (OKF merge, Finch routing, secrets/real-user-data/destructive) need Ali's explicit go.
178	

<system-reminder>This memory is 10 days old. Memories are point-in-time observations, not live state — claims about code behavior or file:line citations may be outdated. Verify against current code before asserting as fact.</system-reminder>
1	---
2	name: project-okf-emitter-2026-06-18
3	description: "PKM adopts Google's Open Knowledge Format via a non-destructive emitter; core landed on branch, nightly-wiring + agent-impact analysis remain"
4	metadata: 
5	  node_type: memory
6	  type: project
7	  originSessionId: f9c77360-476e-4c68-8408-2a541fd4db38
8	---
9	
10	Ali decided to adopt Google's **Open Knowledge Format (OKF v0.1)** for PKM (and later AINA). Key realization: his Wiki already IS the LLM-wiki pattern OKF standardizes (md + YAML frontmatter + a link graph); the only real divergence is `[[wikilinks]]` (not OKF/GitHub-standard). Chosen approach = **Approach B (non-destructive emitter)**: keep the wikilink/Tolaria-native source, rewrite `[[wikilinks]]`→relative-path links at emit time. Engine seam so AINA's Evidence Atlas becomes producer #2 later; edges preserved so the analyzed-but-unbuilt **gbrain** graph-rerank has a substrate.
11	
12	**Landed on branch `ali/okf-emitter-2026-06-17`** (pushed, NOT merged) in `PKM-monorepo`:
13	- `transcripts/tools/okf_emit.py` (+ `test_okf_emit.py`, 9 tests), `okf_viz.py` (companion-skinned graph explorer), `docs/okf/2026-06-17-okf-emitter-design.md`. Bundle output: `~/PKM/aliknowledgebank/okf-bundle/`. 2501 concepts, 96.2% links resolved, idempotent.
14	- Codex cross-engine review (read-only) caught a **critical data-loss bug** (nested same-stem flatten) → fixed. Validated the cross-engine gate.
15	
16	**Remains** (handoff: `docs/handoff/2026-06-18-okf-emitter.md`): (1) agent-impact analysis deliverable for paperclip/fusion/multi-agent [Ali leaning do-this-first]; (2) wire `okf_emit`+`okf_viz` as one additive nightly stage + deploy viz behind auth-gated hub (don't fire full nightly w/ deploy/notify without nod); (3) merge → main.
17	
18	Confirms [[feedback-pkm-mcp-local-vs-remote-boundary]] (headless claude -p 401s) and [[reference-codex-cli-path]]. Companion answer: OKF viz + doc-companions = two view-shapes, one shared design system (companion-html-spec tokens). See [[reference-about-ali-manual]].
19	

1	# PKM Deep-Dive — RUNLOG (append-only)
2	
3	Newest entries at the bottom. Each entry: timestamp · actor · what happened · next.
4	This is the heartbeat; if context compacts, the last entry says where we are.
5	
6	---
7	
8	## 2026-06-28 — Session start (foundation)
9	
10	- **Lead**: Claude (Opus 4.8), Mac-side, branch `ali/pkm-deepdive-2026-06-28`.
11	- Brainstormed + reframed intent → mission/milestones/slices (see `SPEC.md`).
12	- Founder decisions locked (see `DECISIONS.md` D1–D3).
13	- Verified live state: D1 sync FAILING (76k delta), OKF = draft PR #3, VDS reachable via `aina-vds-tf`.
14	- Built durable docs (SPEC/RUNLOG/DECISIONS/FINDINGS) + hardened SSH wrapper `vds-exec.sh`.
15	- **NEXT**: pre-warm VDS SSH master, then launch M0+M1 read-only audit workflow (Sonnet lanes).
16	
17	## 2026-06-28 — Foundation done + big reconcile (parallel to audit workflow)
18	
19	- Built durable docs + `vds-exec.sh`; multiplexed SSH proven (1 master, N concurrent lanes, no dup).
20	- Committed+pushed scaffold (072a03a5f) to `ali/pkm-deepdive-2026-06-28`.
21	- **Launched M0+M1 read-only audit workflow** (task `wbobf55u1`, run `wf_a8b78c4c-045`, 7 Sonnet lanes,
22	  VDS reads multiplexed). Awaiting completion notification.
23	- **brain_health (live)** captured into FINDINGS M1: D1 sync delta ≈ entire dataset (76,310 changed
24	  vs 76,076 total) → baseline mismatch, content-hash fix NOT holding; VDS snapshot/mirror healthy
25	  (9h, 76,076 rec); subagents tracked separately (3,859 + 1,764); 2,513 ingest queue; date model OK.
26	- **Linear reconcile (live)** captured into FINDINGS M0: the deep-dive ≈ open **Agent Memory** project
27	  (AIOPS-243 access model = M2/M3, AIOPS-244 inventory = M0, VISION-16 routing = "what goes where").
28	  **Finch (AIOPS-223) already built** the per-agent manifest system on the VDS (read-only) — verify,
29	  don't rebuild. VDS already runs pkm-agent + fresh unified index. Consolidation/redesign all Done.
30	- **NEXT**: (a) consume audit-workflow results → fill FINDINGS, (b) Codex review of synthesized
31	  findings, (c) act on safe M1 fixes (D1 baseline repair is stop-gated D1 write — confirm w/ Ali),
32	  (d) launch M2/M3 design via product council bound to AIOPS-243 + VISION-16 + Finch prior art.
33	
34	## 2026-06-28 — Audit complete (7 lanes) + D1 root-cause FIXED
35	
36	- **Workflow `wbobf55u1` done** (7 Sonnet lanes, 454k tokens). Full result:
37	  `/private/tmp/.../tasks/wbobf55u1.output`. Synthesis ledger added to FINDINGS top (10 items).
38	- **D1 sync ROOT-CAUSED + FIXED** (commit 23bf70dfd): `recency_days` (records col 15, added by
39	  9797d6dbe on 06-26) = (export_date−record_date).days, daily-volatile, never stripped from hash →
40	  75,841/75,852 rows re-hash nightly → blows 5k cap since 06-27. Fix: blank col 15 in hash normalizer
41	  + MANIFEST_VERSION 3→4 (self-heal re-baselines on next nightly). TDD red→green, 14 tests pass.
42	  Cross-confirmed by a prior 06-18 session that found recency_days was NOT yet in the payload then.
43	- **Codex review** of commit 23bf70dfd launched (bg pid 5604 → scratch/codex-review-d1fix.txt).
44	- Key remaining (see FINDINGS ledger): #2 VDS index repoint (stale transcripts-only 5,215 vs fresh
45	  76k already on VDS) = biggest M2 win, safe; #3 Hermes/Paperclip uncaptured (design+Ali); #9 satellite
46	  repos archive (Ali); #10 OKF/Finch (Ali).
47	- **NEXT**: confirm Codex verdict on D1 fix → push; then M2 substrate repoint (Codex-gated) + present
48	  the stop-gated decisions to Ali (D1 immediate-repair? OKF? Finch? repo archive? Hermes capture?).
49	
50	## 2026-06-28 — Fixes LANDED to main + D1 repair running + Ali adds M6
51	
52	- Codex reviewed 23bf70dfd → one [P2]: blanking recency_days freezes the API value. RESOLVED:
53	  Worker now derives recency_days at read time from `date` (commit 7e5dba2a4, 26 tests). recency is
54	  display-only (ranking uses `date`), so this is correct + complete.
55	- **D1 churn fix + Worker fix FF-MERGED to main** (7e5dba2a4, pushed 60692601e..7e5dba2a4); repo restored
56	  to `main` so tonight's nightly self-heals. Reconcile clean (0 behind, all 9 files mine).
57	- **D1 immediate FULL_IMPORT repair running** (bg bl6r2opw2, log scratch/d1-repair.log): full import
58	  COMMITTED (255,566 rows / 36 tables, ~29s), now on R2 body sync. Pre-authorized per D2. Verifying +
59	  clearing failure marker on completion.
60	- **Ali added M6 — Mission Control single-pane visibility** (SPEC §M6, DECISIONS D7): surface PKM +
61	  findings + Founder OS + artifacts/transcripts into EXISTING Mission Control (mission.ainative.academy /
62	  AIN-474 / ~/aina-mission-control), not parallel build. + S5.5 Finch dept-head attach verification
63	  (manifests staged-never-attached = likely why dept heads lack canon) + D8 wiki-as-living-docs.
64	- **NEXT**: verify D1 repair + clear marker → M2 VDS index repoint (Codex-gated) → M6 Mission Control
65	  reconcile + Paperclip dept-head check → present full milestone summary + stop-gated decisions to Ali.
66	
67	## 2026-06-28 — M1 sync integrity DONE + VERIFIED GREEN
68	
69	- D1 FULL_IMPORT repair completed (`✓ remote PKM-Agent bundle synced`); D1 re-baselined to **76,081
70	  records**, published manifest **v4** → no recurrence. Stage markers reconciled (.ok written, .failed +
71	  failures digest cleared). **brain_health now `"status": "ok"`** (remote-mcp-sync ok, 0 critical
72	  failures). Memory note `reference-pkm-remote-d1-sync-repair` + MEMORY.md updated (recency_days cause).
73	- **M1 protect-first #1 (sync integrity) = fully LANDED + verified.** Code on main, D1 live, health green.
74	- Advisory (non-blocking) warnings remain: 11 raw sessions newer than data.json (this session's subagents
75	  — normal active tail); VDS snapshot 76,076 vs local 76,081 (sub-day drift); ingest_queue 2,513.
76	- **NEXT (autonomous)**: M2 VDS index repoint (point VDS pkm-agent at the fresh full 76k corpus instead
77	  of the stale 5,215 transcripts-only index) — Codex-gated. Then present stop-gated decisions menu to Ali.
78	
79	## 2026-06-28 — Ali: reorder + AINA OS Reference + M7; satellite repos archived
80	
81	- VDS disk freed by Ali (finding #5 disk → resolved). **Satellite repos ARCHIVED** (oscalar/pkm,
82	  pkm-wiki, pkm-aliknowledgebank — superseded by monorepo @06-19 cutover, reversible) → finding #9 done.
83	- Ali reordered (D9): M2→M3→M4→M5 → Paperclip dept-check + OKF + recommendations → M6 → M7.
84	- **New M7** (agent repo/config foundation, D10/D11): AGENTS.md/Codex config/hooks, content-bloat fix
85	  (Codex-replicated-Mac-setup), check aina-hq/aina-srv on VDS, verify Hermes's new Paperclip setup,
86	  repos→profiles→tool-bundles→permissions map → get Paperclip factory running. Grounded in the **AINA OS
87	  Reference doc** (preserved to scratch). Target `~/work/ainative/*` vs current `~/conductor/repos` —
88	  M7 = reconcile/migration, founder-gated, trust-firewall gate noted.
89	- M2 prep: confirmed VDS has fresh full corpus `aliknowledgebank/unified_pkm.sqlite` (101MB, 06-28) +
90	  stale partial `vds-local-index/unified_pkm.sqlite` (7.4MB, 06-26). `pkm-agent-vds.env` not at maxdepth-3
91	  — locate the actual config (systemd unit / deeper path) before repoint.
92	- **NEXT (now)**: locate VDS pkm-agent config → repoint PKM_RECORDS_DB → restart → verify full-corpus query.
93	
94	## 2026-06-28 — M2 diagnosed (verify-the-entry-point saved a wrong fix); narrow Hermes gap
95	
96	- Investigated M2 properly and CORRECTED twice: (1) `pkm-agent-vds.env` points at the stale vds-local-index
97	  BUT is **dead config** — no active launcher sources it (only README + old session log; no systemd unit);
98	  (2) the REAL live entry points — **Claude MCP** (`.claude.json`) + **CLI** (`aina-agent-env.sh`) — already
99	  resolve `PKM_RECORDS_DB`→`aliknowledgebank` = **full fresh 76k corpus** (FTS5 + artifact ledger verified
100	  on VDS). So agents querying PKM on the VDS already get the wikis/learnings/full corpus. 
101	- **Real M2 gap = Hermes split config**: active `~/.hermes/config.yaml` + `profiles/aina-chief-of-staff/
102	  config.yaml` carry BOTH full AND stale `vds-local-index` PKM_RECORDS_DB (prior half-repoint
103	  `.bak-pkm-vds-repoint-20260626`). `vds-local-index.timer` active but `.service` FAILED (5% guard).
104	- Did NOT do live Hermes surgery at the tail of a huge turn (the user flagged Hermes's Paperclip setup for
105	  careful verification — ties to M7). Captured exact files + fix. **NEXT focused pass:** unify the 2 Hermes
106	  configs → aliknowledgebank, restart Hermes, verify; retire/fix the redundant failing self-index; Codex-gated.
107	- This turn LANDED: M1 sync fixed+green+live-rebaselined; satellite repos archived; reorder + M7 + AINA OS
108	  Reference captured; M2 precisely diagnosed.
109	
110	## 2026-06-28 — M7 de-scoped (Ali): config audit, not migration; cleanup last
111	
112	- Ali: the AINA OS Reference doc is **PRINCIPLES-only reference**; **srv/conductor STAY, NO migration** to
113	  `~/work/ainative` (D12 supersedes D10's migration framing). M7 = audit the EXISTING VDS config + AGENTS.md
114	  structure → configure so Paperclip/Hermes + Ali+Claude/Codex all work **safely, contained, no-bloat**,
115	  preserving the current "work whenever, like now" flow.
116	- **Cleanup (old backups/tarballs/`.redesign-archives`/dead config) = FINAL step only (D13)**, never mid-stream.
117	- NEXT focused pass: finish M2 (Hermes config unify → full corpus, safe — also serves M7's "Hermes works
118	  safely without bloat") + read-only audit of VDS AGENTS.md/config/bloat (incl. aina-hq, aina-srv) per Ali's
119	  "look at the config of vds." Then M3 (OKF) → M4 (UX/tab-QA) → M5 (Finch dept-check + governance).
120	
121	## 2026-06-28 — M7 audit launched (Ali gave go)
122	
123	- Orienting probe: VDS disk freed (67G free, 72%). Bloat shape: `/srv/aina/aina-data-engine-room` 27G,
124	  `/srv/aina/AINA Reference Files Mac` 3.3G (Mac replica), 3× Personalization-Engine copies in /home/ali
125	  (7.6G+2G+1.7G), `migrate-in` 2.7G. Agent config: `.claude/CLAUDE.md`→`.codex/AGENTS.md` symlink (unified
126	  global contract), 24KB config.toml, 55KB .claude.json. aina-hq = `/srv/aina-hq`.
127	- **Launched VDS agent-config audit workflow** `wp6ke54n4` (run `wf_34fb91c8-a5b`, 3 read-only Sonnet lanes:
128	  config-structure / bloat / hermes-paperclip), all via the multiplexed SSH channel. Awaiting completion →
129	  synthesize → Codex review → propose safe config + do the Hermes config unify.
130	
131	## 2026-06-28 — M7 config audit DONE + M2 Hermes fix LANDED + M2 COMPLETE
132	
133	- Audit `wp6ke54n4` done (synthesis in FINDINGS M7). Verdict: VDS setup BETTER than feared — Codex lean/full
134	  MCP split good; per-repo AGENTS.md well-scoped; **Paperclip sane + contained (the "Hermes made a weird
135	  Paperclip setup" worry is UNFOUNDED)**; Finch founder-gate works correctly.
136	- **M2 Hermes fix LANDED**: disabled redundant stale `pkm-agent-vds-local` in `~/.hermes/config.yaml` +
137	  `profiles/aina-chief-of-staff/config.yaml` (`enabled: false`; backups `.bak-pkmvds-disable-20260628`; YAML
138	  valid). Hermes now uses ONLY canonical pkm-agent→aliknowledgebank. Effect on next Hermes session (no forced
139	  CoS restart). vds-local-index + failing timer now fully redundant → decommission at cleanup.
140	- **M2 COMPLETE**: all query paths (Claude MCP + CLI + Hermes) serve the full fresh 76k corpus.
141	- **S5.5 answered**: dept heads lack canon because Finch's 62 manifests are staged-not-attached, blocked on
142	  (1) PKM health (NOW healthy — D1 fix) + (2) founder token `APPROVE-AIN-47-PHASE-E`. One blocker cleared;
143	  attach is a founder-gate away (Ali approves + pause 62 agents + roster JSON).
144	- Safe-now M7 items READY (not yet applied): enable Codex durability/reconcile hooks; narrow repowise-augment
145	  matcher to Write|Edit. Needs-Ali: `.claude.json` 13-MCP firehose scoping; sandbox/trust-root; Ruflo
146	  self-modify guard; Finch attach approval. Bloat (~35-45G) HELD for FINAL cleanup (D13).
147	- NEXT: present M7 menu to Ali; then M3 (OKF) → M4 (UX/tab-QA).
148	
149	## 2026-06-28 — Ali answers + Paperclip/agent-behavior simulation launched
150	
151	- Ali decisions: Finch attach = **leave staged** (D14); Claude MCP scoping = **approved (Mac+VDS) but
152	  SIMULATE first** (D15); force-restart OK; everything PAUSED except lead session (safe window).
153	- New scope (D16): Hermes rebuilt the Paperclip company yesterday (NEW company / deleted old / remapped) →
154	  verify; **Paperclip API-access failure** broke agents → root-cause; review logs. + verify aina-academy
155	  AGENTS.md pointers (Codex's edits). + simulate-before-behavior-change.
156	- **Launched workflow `we11gqlg4`** (run `wf_edc34429-fce`, 3 read-only Sonnet lanes: paperclip-audit /
157	  aina-academy-agents / agent-behavior-sim) — config-trace only, agents stay PAUSED. Awaiting completion →
158	  synthesize → evidence-based MCP-scoping + Paperclip/API fixes → Codex review → apply (force-restart OK).
159	- M7 safe-now hardening (Codex hooks, repowise) now ALSO gated behind the sim per D15 (don't blind-change).
160	
161	## 2026-06-28 — Paperclip/agent-behavior audit DONE (we11gqlg4): myths busted + factory blocker found
162	
163	- Synthesis in FINDINGS M7b. KEY: (1) **NO Paperclip company remap ever happened** (myth busted — company
164	  `7d58fc13` intact since 06-18; Hermes 06-27 session was READ-ONLY); (2) **factory blocker = Paperclip API
165	  URL** points agents at public `paperclip.ainative.academy:3100` but server is LOOPBACK-only → agents must
166	  use `127.0.0.1:3100` (find injection point); (3) **Codex's aina-academy AGENTS.md edits VERIFIED correct**;
167	  (4) **evidence-based MCP scoping** ready: 6 specialist servers used in only 3 repos (aina-platform/
168	  aina-academy/aina-cc) → safe to scope; (5) **Hermes operational issues** (compression broken/no credit,
169	  memory full 2499/2500, Linear arg error, claude.ai usage depleted).
170	- **No live config changes applied this turn** (per simulate-first + several new needs-ali findings). Menu:
171	  (a) Paperclip API-URL fix = factory unblocker (needs injection-point find); (b) MCP scoping (approved D15,
172	  evidence ready, VDS-first); (c) Hermes credits/auth + memory = Ali account actions; (d) AINA→UUID script
173	  fix (safe-now); (e) academy think-path note (small).
174	- NEXT: Ali decides sequencing; lead recommends (a)+(b) next. Then M3 (OKF) → M4.
175	
176	## 2026-06-28 — Ali green-lit fixes + research; Paperclip API-URL diagnosed; research launched
177	
178	- Ali: fix Paperclip API URL ✓; MCP scoping but it's a DESIGN question (per-agent / per-repo-dept /
179	  Hermes-sole-coordinator — D18); research Hermes docs for memory + SQLite long-term memory + dream-cycle +
180	  **subscription auth (ChatGPT + Claude, NOT API keys — D17)**; validate Paperclip official docs; read
181	  AIOPS-262 (reports rolling-timeline — Hermes backed up erroring reports + added Linear pointers; connects
182	  to M4 report-viewing).
183	- **Paperclip API-URL root cause CONFIRMED locally (D19):** `auth.baseUrlMode:"auto"` derives the agent
184	  PAPERCLIP_API_URL from the request host → public hostname (unreachable) for some runs. Run-logs: 1910×
185	  127.0.0.1 (ok) vs 482× public (broken). PAPERCLIP_API_URL is NOT in any env file (Paperclip injects it).
186	  config.backup confirms Hermes only ADDED a hostname on 06-14 (no remap). Fix = force loopback; exact
187	  mechanism pending Paperclip-docs research (don't guess schema).
188	- **Launched research workflow `wvrswg53h`** (run `wf_710b8631-4d8`, 3 general-purpose lanes: hermes-docs /
189	  paperclip-docs / aiops262+mcp-model). Awaiting → then apply API-URL fix + Hermes auth/memory/dream setup +
190	  MCP-access-model design + AINA→UUID script fix.
191	- NEXT: consume research → apply fixes (force-restart OK) → present MCP-model options + report-view path.
192	
193	## 2026-06-28 — Research done (wvrswg53h) + Paperclip API-URL fix LANDED
194	
195	- ✅ **Paperclip API-URL FIXED** (factory unblocker): systemd `PAPERCLIP_RUNTIME_API_URL=http://127.0.0.1:3100`
196	  (backup `.bak-apiurl-20260628`), restarted, VERIFIED (listening 127.0.0.1:3100, health 200, env in process,
197	  agents stay paused). Synthesis in FINDINGS M7c.
198	- Research outcomes: Hermes Claude-subscription auth BROKEN upstream (#40014) → recommend pin aux slots
199	  (provider:auto → openai-codex) to stop Anthropic drain (Ali to confirm). Hermes long-term memory = SQLite
200	  `state.db` (the "sqlite" Ali recalled); memory wall fixable by raising char limits; "dream cycle" = Curator
201	  (skills) + proposed Dreaming (memory, not shipped). Reports viewable NOW at aina-mobile-review.pages.dev
202	  rolling index (AIOPS-262). MCP-access model: recommend dept-bundles now → per-agent Finch later (NOT
203	  Hermes-coordinator). Verified passwordless sudo on VDS.
204	- NOT applied (presenting for Ali): Hermes aux-pin (deviates from his Claude-sub ask due to upstream bug) +
205	  memory-limit raise; MCP-model pick; AINA→UUID script fix. 
206	- NEXT: Ali confirms Hermes aux-pin + MCP-model + memory tuning → apply; then M3 (OKF).
207	
208	## 2026-06-28 — Ali decisions APPLIED + verified (Paperclip + Hermes hardened)
209	
210	- Ali: clear Anthropic + pin aux to ChatGPT sub; **leave MCP as-is** (no scoping, incl. .claude.json);
211	  raise memory + enable Curator + fix AINA→UUID.
212	- ✅ **Paperclip API-URL fix LANDED + verified** (earlier this turn): systemd PAPERCLIP_RUNTIME_API_URL=loopback.
213	- ✅ **Hermes hardening LANDED + verified:** config.yaml — pinned **13 aux slots** `provider: auto`→`openai-codex`
214	  (0 auto remaining → no Anthropic fall-through drain; all on ChatGPT subscription), memory_char_limit
215	  2200→10000 + user_char_limit 1375→5000 (full-memory wall fixed); Curator already enabled; no active
216	  Anthropic creds existed (auth.json=openai-codex only). **status-sweep SKILL.md** `-C AINA`→UUID (72-error
217	  bug). Backups `.bak-hermes-harden-20260628`. YAML valid. **Gateway restarted CLEAN** (Slack connected, cron
218	  + kanban up, NRestarts=0, no config errors). Hermes Claude-subscription still blocked upstream (#40014) —
219	  parked.
220	- **MCP scoping NOT done** (Ali: leave as-is). `.claude.json` firehose untouched.
221	- NEXT: M3 (OKF decision) → M4 (UX/report-view already partly answered via aina-mobile-review index).
222	

1	# PKM Deep-Dive — FINDINGS (filled by audit lanes)
2	
3	Structured truth as lanes complete. Each finding: claim · evidence · status (✅ healthy / ⚠️ degraded
4	/ ❌ broken / ❓ unknown) · recommended action · act-tier (safe-autoland / STOP-gated).
5	
6	---
7	
8	## AUDIT SYNTHESIS — 7-lane workflow `wbobf55u1` (2026-06-28, 454k tokens)
9	
10	**Verdict:** the PKM *producer* (Mac nightly) is healthy and consolidated. The breakages are all at the
11	**edges** — remote D1 sync, the VDS query index, and uncaptured new agent logs (Hermes/Paperclip).
12	Prioritized ledger (tier: safe=auto-land · design=needs design+Ali · account/strategy=Ali gate):
13	
14	| # | Finding | Tier | Status |
15	|---|---------|------|--------|
16	| 1 | D1 sync: `recency_days` volatile col not stripped from hash → 76k churn/night | safe | ✅ **DONE + VERIFIED GREEN** (23bf70dfd + Worker 7e5dba2a4 on `main`; D1 re-baselined live to 76,081 rec; manifest v4; markers reconciled; brain_health `ok`) |
17	| 2 | M2 query plane: Claude MCP + CLI already full+fresh; Hermes carried a redundant stale `pkm-agent-vds-local` server | safe | ✅ **DONE** — disabled `pkm-agent-vds-local` in both Hermes configs (backups + YAML valid); all paths now on full 76k corpus. Self-index retire = cleanup |
18	| 3 | Hermes + Paperclip logs **entirely uncaptured** by PKM (new agent systems running now) | design | ⛔ propose capture adapter (VISION-16 routing decision) |
19	| 4 | Session deep-link tier degraded: 88.6% subagent md lack turn/prompt attribution (missing `.tasks.json` sidecars) | safe | ⏳ backfill sidecars; add `is_subagent` flag |
20	| 5 | VDS `vds-local-index` rebuild failing (5% count-delta guard too tight); VDS disk 98% | safe / account | ⏳ relax guard or repoint (ties to #2); disk cleanup ⛔ Ali |
21	| 6 | ARCHITECTURE.md drift: 27→28 stages, missing `vds-html-backfill` + delta/topic-refresh cadence | safe | ⏳ doc fix |
22	| 7 | `~/PKM` is NOT a symlink facade (independent data tree); CLAUDE.md says otherwise | safe | ⏳ correct doc |
23	| 8 | `all_records.json` + `daily_briefs.json` 40d stale (pipeline bypasses them) | safe | ⏳ confirm dead → retire/document |
24	| 9 | 3 satellite repos (oscalar/pkm, -wiki, -aliknowledgebank) unarchived sprawl | account | ⛔ archive needs Ali |
25	| 10 | OKF PR #3 draft; Finch live attach | strategy | ⛔ M3 decision + founder gate |
26	
27	---
28	
29	## M0 — Reconcile & Ground
30	- ✅ **Your deep-dive ≈ the open "Agent Memory" Linear project — ALIGN, don't duplicate.** Canonical
31	  tracker = **Agent Memory** project (id 9f9f7e29). Live open issues that ARE this work:
32	  - **AIOPS-243** (Todo·High·2026-06-28) "Design PKM-backed Agent Memory access model for Paperclip/
33	    RunFusion/Hermes" = our **M2/M3**. Has design questions + initial policy + acceptance criteria.
34	    → canonical home for the access-model deliverable (land output here).
35	  - **AIOPS-244** (Todo·Med) "Inventory + cross-link PKM/legacy memory into Agent Memory" = our **M0**;
36	    backed by Linear doc "Agent Memory Pointer Index" (a693054d). Rule: classify (move/keep-pointer/
37	    archive), don't bulk-move, preserve history.
38	  - **VISION-16** (Backlog) "AINA Memory Architecture v1" = canonical "what goes where" routing:
39	    Soul=Identity · Constitution=Governance · **Linear=Operational truth** · Hermes=Executive ·
40	    Fusion=Execution · Beads=Working · GitHub=Impl-ledger · **PKM(ali-mcp)=Knowledge layer** (advisory,
41	    NOT execution truth). This is the answer to "how should things be organized/captured."
42	  - Related research: **AIN-581** (Agent Memory Hygiene — CLAUDE/AGENTS.md audit), **AIN-583** (push/
43	    pull split), **PKM-2** (retrieval/reranking). act_tier: no-action (alignment); M2/M3/M5 land here.
44	- ✅ **Finch↔PKM already built (AIOPS-223, Done) — VERIFY, don't rebuild.** Finch = "Knowledge & Canon
45	  Steward / pkm_bridge". On the VDS (2026-06-26) it gathered ~36,885 docs (agentops 526, brigade
46	  7,457, pkm-daily 19,433, pkm-wiki 5,819, repo-doc 3,650), classified canon/not-canon/historical,
47	  built per-agent (59) + per-team (10) CONTEXT-MANIFESTs + MUST-NEVER / DECISION-CONTRACTS /
48	  FAILURE-MODES + a STAGED (un-run) `attach-manifests.sh`, under `/srv/aina/ops/finch-pkm-mapping/`.
49	  Founder rules baked in: PKM **read-only from VDS, no write path ever**; Finch stages for review, does
50	  NOT auto-attach. Review loop (Gimli adversarial + Calibrator 58/100) caught stale-as-canon over-
51	  marking (526 refs); Finch v2 fixed it. → answers "give it all to Finch only?" **NO** — Finch builds
52	  SCOPED manifests; broad recall mediated by Donna/Hermes; only librarian/research agents get direct
53	  retrieval; dept heads get scoped packets (tiered, per AIOPS-243). act_tier: stop-gated (live attach
54	  = founder gate).
55	- ✅ **VDS already has the agent-query substrate.** Per AIOPS-223 receipts: VDS has `pkm-agent`
56	  (`/usr/local/bin/pkm-agent`) + fresh unified index (`/home/ali/PKM/vds-local-index/unified_pkm.sqlite`)
57	  + Wiki/Transcripts trees; read-only `pkm-agent search` returns hits. So M2 "can VDS agents query PKM"
58	  is largely YES — verify freshness + access tiers live. _[S1.4/S2.1 lanes]_
59	- ✅ **Consolidation + redesign DONE.** PKM-3/4/6/9/10/11/12/14 redesign + AIN-539/540 + INFRA-21/23
60	  all Done (2026-06-26); one-repo consolidation tracked done (PKM-14). Legacy "PKM - Personal Knowledge
61	  Management" project = historical pointer. No PKM issue stuck "In Progress". Other live projects:
62	  "PKM Hub Streamlining" (hub/visual = our M4), "PKM Compounding Learnings: Codex Route" (done).
63	- ⚠️ **AINA Reconcile Gate:** the Mac-Desktop gate targets ACAD/DAT/PLAT trust-firewall (PR #54) — NOT
64	  implicated by this PKM/AgentOps work. Proceeding.
65	- ✅ **Mac = canonical producer; VDS = guarded consumer (S0.3).** Nightly `transcripts-daily.sh` (launchd
66	  `com.ali.transcripts`, 11:15) + delta runs (`com.ali.pkm-delta`, 17:00/23:00/05:00) + watchdog (12:30) +
67	  weekly topic-refresh (Mon 09:00). VDS runs ONE independent job: `vds-local-index.timer` (2h) — currently
68	  failing (see M1). **ARCHITECTURE.md DRIFT**: doc says 27 stages, reality is 28 (+ inline `vds-html-backfill`
69	  not in STAGES array); delta cadence + topic-refresh undocumented. act_tier: safe (doc fix).
70	- ⚠️ **"What lives where" + a doc correction (S0.4).** Monorepo = code/scripts/hub/pkm-agent. **`~/PKM` is
71	  NOT a symlink facade into the monorepo — it's an independent data tree** (`realpath ~/PKM/Wiki` ≠ monorepo);
72	  CLAUDE.md's "live contents are symlinked" is misleading → correct it. Live wiki = monorepo/wiki (2,863
73	  pages). `unified_pkm.sqlite` (76,081 rec, FTS5) fresh, is the agent store. **`all_records.json` (94MB) +
74	  `daily_briefs.json` 40d STALE (last built May 19)** — pipeline feeds unified_pkm.sqlite directly, bypassing
75	  `classify_and_merge.py`/`extract_briefs.py` → confirm dead and retire/document. act_tier: safe.
76	- ⚠️ **Repo sprawl (S0.2):** 3 satellite repos `oscalar/pkm`, `oscalar/pkm-wiki`, `oscalar/pkm-aliknowledgebank`
77	  are NOT archived (last push 06-15, "offsite backup"). `oscalar/pkm-agent` archived ✅. Recommend archiving
78	  the 3 to kill canonical-confusion. act_tier: **account** (GitHub repo state — Ali's go). Also a stale
79	  merged branch `ali/pkm-remote-provenance-2026-06-26` (no PR trail) — deletable for hygiene.
80	
81	## M1 — Pipeline & Sync Integrity
82	- ✅ **D1 remote sync — ROOT-CAUSED + FIXED (commit 23bf70dfd, MANIFEST_VERSION 3→4).** Definitive root
83	  cause (S1.1 lane): commit `9797d6dbe` (2026-06-26) added `recency_days` (records col 15) =
84	  `(export_date − record_date).days`, which increments by 1 daily for every dated record but was never
85	  added to the hash normalizer's strip list. Result: 75,841 of 75,852 rows re-hashed every night →
86	  always > the 5,000 cap → `too_large` since 2026-06-27. Fix blanks col 15 for hashing only (stored row
87	  keeps the value) + bumps MANIFEST_VERSION so `sync_remote_bundle.sh` self-heals via a one-time full
88	  re-baseline on the next nightly. TDD: 14 tests pass incl. new recency-only-diff→same-hash regression.
89	  **Open**: optionally run `PKM_REMOTE_D1_FULL_IMPORT=1` now to fix immediately vs waiting for tonight's
90	  self-heal (stop-gated D1 write — Ali's call). Original evidence (brain_health
91	  2026-06-29T00:48Z + failure digest): D1 delta `too_large`, delta_rows 76,332, changed_or_new
92	  76,310, deleted 22; counts records 76,076 / packages 920 / artifacts 2,219 / body_refs 10,048.
93	  `changed_or_new ≈ total records` → the D1 baseline is effectively empty/mismatched so nearly every
94	  row recomputes as "new", then the `PKM_REMOTE_D1_MAX_DELTA_ROWS=5000` cap fail-closes. Stage
95	  `remote-mcp-sync` failed 2026-06-28 16:30; watchdog rerun also failed. **The 2026-06-18 content-hash
96	  fix is NOT holding** (baseline likely lost/reset, or manifest invalidating). Repair
97	  `PKM_REMOTE_D1_FULL_IMPORT=1 bash scripts/sync_remote_bundle.sh` republishes the baseline; if churn
98	  is structural it recurs (then raise cap w/ justification OR fix source churn). act_tier: **stop-gated**
99	  (D1 write) — on auto-land-safe list per D2 but confirm reversibility first. _[S1.1 lane confirming]_
100	- ✅ **VDS snapshot + mirror healthy & fresh.** vds_snapshot ok (age 9.1h, remote_records 76,076, max
101	  2026-06-28, host aina-vds-tf, backups kept); vds_mirror ok (age 9.0h, 5 mirrors). Drift: local
102	  76,081 vs VDS 76,076 = 5 records, sub-day. "Mirror on both surfaces" works. _[S1.4]_
103	- ✅ **Sessions: subagents tracked separately.** data.json: 3,859 sessions + 1,764 subagents = 5,623;
104	  raw 4,178 (1 active-tail = this session). Old "subagents-as-top-level" issue appears resolved.
105	  _[S1.2 lane verifying parent/child link representation + render coverage]_
106	- ⚠️ **ingest_queue: 2,513 queued files.** Large backlog — confirm whether normal queue depth or a
107	  stuck/slow ingest. _[investigate]_
108	- ❌ **Hermes + Paperclip logs ENTIRELY uncaptured (S1.3).** Hermes writes `~/.hermes/logs/` (Mac) +
109	  `/home/ali/.hermes/profiles/aina-chief-of-staff/logs/` (VDS, active 2026-06-28, ~16MB history).
110	  Paperclip writes Codex rollout JSONLs at `~/.paperclip/instances/default/companies/<id>/codex-home/
111	  sessions/` (23 today, ZERO overlap with `~/.codex/sessions`) + a 54MB `server.log`. `vds_session_pull.py`
112	  SOURCE_SPECS only covers `.claude/projects` + `.codex/sessions` + `.codex/archived_sessions` →
113	  Hermes/Paperclip never enter `data.json`, and `check_capture_coverage.py` has no dimension for them so
114	  the gap is INVISIBLE to nightly monitoring. Paperclip JSONLs are Codex-format (easy: add path to
115	  SOURCE_SPECS or symlink). Hermes `agent.log` needs a new adapter or raw-archive. act_tier: design/
116	  stop-gated (new capture surface — VISION-16 routing: do these become PKM Knowledge, or stay ops logs?).
117	- ⚠️ **Session deep-link tier degraded (S1.2).** Structural `parent_session_id` HEALTHY (1,764 Mac
118	  subagents linked, all 104 parents resolve; HTML render 99.3%). But the rich deep-link (turn position/
119	  prompt via `link_subagents.py`) is written for only **219 / ~1,923 subagent md (11.4%)** — 1,658 blocked
120	  by missing `<short_id>.tasks.json` sidecars for their parent. Codex sessions carry no parent link (by
121	  design — no `subagents/` path layout). Subagents counted FLAT in `data.json` (no `is_subagent` field) →
122	  old "subagents as top-level" persists at the data layer. act_tier: safe (backfill sidecars + add
123	  `is_subagent` flag; doesn't affect search/render).
124	- ⚠️ **VDS query index is stale AND wrong-scope (S1.4 — the M2 substrate gap).** VDS pkm-agent's
125	  `PKM_RECORDS_DB` points at `/home/ali/PKM/vds-local-index/unified_pkm.sqlite` = **5,215 records,
126	  transcripts-only, 06-26 stale**; meanwhile the FULL fresh corpus (76,076 rec, today) already sits on
127	  the VDS at `/home/ali/PKM/aliknowledgebank/unified_pkm.sqlite` (kept current by rsync). The self-index
128	  rebuild also keeps FAILING (15.1% count-delta > 5% guard). Cleanest fix: repoint `PKM_RECORDS_DB` to the
129	  always-current aliknowledgebank db, retiring the redundant stale self-index. act_tier: safe-autoland
130	  (VDS read-only consumer repoint, Codex-gated, reversible). **VDS disk 98% (5.8GB free)** — flag to Ali.
131	
132	## M2 — Agent Memory & Query Plane
133	- ⚠️ **CORRECTED (2026-06-29): CLI agents already get the full fresh corpus; only the MCP service is
134	  stale-partial.** Two configs resolve `PKM_RECORDS_DB` differently:
135	  - **CLI / interactive** (Codex/Claude/SSH) source `/home/ali/.config/aina-agent-env.sh` →
136	    `PKM_RECORDS_DB=…/aliknowledgebank/unified_pkm.sqlite` = **fresh full 76k corpus ✅** (this is why a
137	    live `pkm-agent search` returned X Bookmarks/Recall — it queried the FULL corpus, not the partial).
138	  - **`pkm-agent-mcp` service** sources `pkm-agent-vds.env` → `…/vds-local-index/unified_pkm.sqlite` =
139	    **stale transcripts-only self-index (5,215 rec @06-26) ❌** — MCP-connected agents (e.g. Hermes) get a
140	    narrow frozen view (no Wiki/Learnings/full corpus; the self-index rebuild is failing on the 5% guard).
141	  **Design tension (U7, deliberate — config.py:120-165):** the self-index was chosen for *immediacy* of the
142	  VDS's own same-day sessions (full Mac corpus lags ~24h via the daily snapshot) at the cost of the full
143	  knowledge base. For Ali's goal (agents query PKM as long-term memory + read the wikis), the **full corpus
144	  wins**; the ~24h VDS-session lag is acceptable (daily mirror + CLI already points at full).
145	  **M2 fix = point `pkm-agent-mcp` at the full snapshot** (`vds_profile_env` index_dir → aliknowledgebank):
146	  unifies MCP with CLI, adds Wiki/Learnings + the artifact ledger (which the self-index lacked), retires the
147	  failing self-index. Change: `config.py` + `deploy/vds/pkm-agent-vds.env` + `tests/test_vds_profile.py`,
148	  Codex-reviewed, then redeploy/restart MCP on VDS + verify. act_tier: safe (reversible consumer repoint)
149	  but **overrides a deliberate design → Codex-gated**. _[implementing]_
150	- ✅ **M2 FINAL (verified REAL entry points, 2026-06-29): the query plane is MOSTLY already full+fresh.**
151	  Claude MCP (`.claude.json` mcpServers/pkm-agent `env`) + CLI (`aina-agent-env.sh`) both resolve
152	  `PKM_RECORDS_DB`→`aliknowledgebank/unified_pkm.sqlite` (full 76k, FTS5 + artifact ledger) ✅.
153	  `pkm-agent-vds.env` (→vds-local-index) is **DEAD** — no active launcher sources it (only a README + an
154	  old session log; no systemd unit) → `config.py vds_profile_env()` is NOT the live path, do NOT "fix" it.
155	  **Real gap = Hermes split config:** active `/home/ali/.hermes/config.yaml` + `profiles/aina-chief-of-staff/
156	  config.yaml` carry BOTH the full AND the stale `vds-local-index` `PKM_RECORDS_DB` → some Hermes
157	  pkm-agent-mcp procs serve the stale 5,215 index (a `.bak-pkm-vds-repoint-20260626` shows a prior
158	  half-repoint). Plus `vds-local-index.timer` active but `.service` FAILED (5% guard). **Fix (focused, live
159	  Hermes surgery — ties to M7 "verify Hermes setup"):** unify the 2 Hermes configs → aliknowledgebank,
160	  restart Hermes, verify; then retire/fix the redundant failing self-index. Codex-gated. _Supersedes the
161	  "MCP service stale" framing above — that env file is dead; Hermes is the live gap._
162	- ✅ **Finch reference outputs intact on VDS** (`/srv/aina/ops/finch-pkm-mapping/`: CLASSIFIED-INDEX.yml
163	  14MB, MASTER-REFERENCE-INDEX.yml 10MB, REVIEW-PACKET-v2.md, MUST-NEVER / DECISION-CONTRACTS /
164	  FAILURE-MODES) but dated 2026-06-26 → 2d stale, no freshness loop running yet (AIOPS-223 U6 pending).
165	
166	## M3 — Knowledge Layer
167	- ⚠️ **OKF = open draft PR #3**, not merged. Decision pending council + Ali (STOP-gated). _pending_
168	
169	## M4 — Surfacing & UX
170	- _pending_
171	
172	## M5 — Governance & Finch
173	- ✅ **WHY dept heads lack canonical docs = Finch manifests never attached (answered).** Finch built 62
174	  per-agent + 11 per-team manifests (`/srv/aina/ops/finch-pkm-mapping/`), all `attach_allowed=false`,
175	  triple-gated: `FOUNDER_APPROVAL_TOKEN=APPROVE-AIN-47-PHASE-E` + PKM health=healthy + live roster JSON;
176	  `attach-manifests.sh` defaults DRY_RUN, exits 10 (no token) / 12 (degraded PKM) before any mutation. Gate
177	  working correctly. **PKM health is NOW healthy (D1 fix) → blocker (2) cleared; remaining = your approval
178	  token + pause the 62 agents + export roster JSON, then `attach-manifests.sh --apply`.** act_tier:
179	  stop-gated (founder token). _This is the path to give dept heads their canon — STOP for Ali._
180	
181	## M7 — VDS agent config audit (workflow `wp6ke54n4`, 2026-06-28)
182	
183	**Verdict: the VDS agent setup is BETTER than feared — mostly well-designed + contained.** Codex has a clean
184	lean/full MCP split; per-repo AGENTS.md (DER 171L, aina-hq 81L) are well-scoped; Paperclip is sane; Finch's
185	founder-gate works. Real issues = a few config-hardening items + bloat (cleanup = final step).
186	
187	### Safe-now (my authority)
188	- **M2 Hermes fix:** Hermes loads TWO pkm-agent MCP servers — `pkm-agent`→aliknowledgebank (canonical full
189	  76k ✅, enabled) AND `pkm-agent-vds-local`→vds-local-index (stale 7.4MB partial, enabled) — in both
190	  `~/.hermes/config.yaml` + `profiles/aina-chief-of-staff/config.yaml`. Fix = **disable the redundant stale
191	  `pkm-agent-vds-local`** (keep canonical). Removes source-ambiguity + a wasted slot, zero loss. Then the
192	  failing `vds-local-index.timer` is fully redundant (decommission at cleanup).
193	- **Codex hooks ALL disabled** (enabled=false) → no durability-autopush / reconcile-prime for Codex sessions
194	  (containment gap). Fix = enable durability-autopush + reconcile-prime-codex in `.codex/hooks.json`.
195	- **repowise-augment fires on every Read/Grep/Glob/Bash** (PostToolUse) → context injection per tool call
196	  (bloat vector). Fix = narrow matcher to Write|Edit|MultiEdit.
197	
198	### Needs-Ali
199	- **Claude `.claude.json` = firehose: 13 MCP servers load globally for EVERY session** (Clerk, PostHog,
200	  Sentry, 5× Cloudflare, Playwright …) — the primary context-rot vector. Move specialist servers to per-repo
201	  `.claude/settings.json`; keep pkm-agent/codegraph/linear/github global. **Shared Mac+VDS → your call.**
202	- `approval_policy=never` + `sandbox_mode=danger-full-access` + `/`-root in trust list (trust=no-op) → no
203	  sandbox boundary for any subagent (intentional-for-speed; risk-tolerance call).
204	- **Ruflo can update Claude config/hooks/agents** (self-modify risk); guard is prose-only. Consider a
205	  PreToolUse block on writes to `~/.codex/AGENTS.md` + `config.toml`.
206	- Global AGENTS.md has a stale Mac path (`/Users/Ali/Projects/pkm-agent`→ VDS path) in wiki-recall (confirm
207	  Mac/VDS sharing before editing the shared contract).
208	
209	### Paperclip ✅
210	- Sane + contained (1 company, 1 instance, loopback, auth-gated, 1.5GB). The "Hermes made a new Paperclip
211	  setup" concern is UNFOUNDED — clean single configure session 2026-06-14, not a confused duplicate.
212	
213	### Bloat (cleanup = FINAL step, D13 — captured, NOT removed)
214	~35–45G reclaimable: Mac-replica "AINA Reference Files Mac" 3.3G; dup PE checkouts (personalizationv1.2 2G,
215	migrate-in 2.7G, ALIPE 1.5G, maybe aina-core 8.5G); DER artifacts→R2 (rollback snaps 2.5G, vector
216	parquet/duckdb); node_modules in idle worktrees ~2G; bloated .git (PE 1.4G, DER 505M); RunFusion triple-snap
217	230M. Full map in `wp6ke54n4` output. HOLD for final cleanup phase.
218	
219	## M7b — Paperclip + agent-behavior audit (workflow `we11gqlg4`, 2026-06-28; agents stayed PAUSED)
220	
221	### Myths busted
222	- ✅ **NO Paperclip company delete/recreate/remap ever happened.** `7d58fc13` is the ONLY company (born
223	  2026-06-18). The old 24-agent scaffold was retired IN PLACE (agents swapped inside the same company,
224	  06-22); 54→62 = +8 agents (Legal/Compliance + Mobile stubs). Hermes's 06-27 session was READ-ONLY (a
225	  status check "was the agent mapping configured?"). Hermes 06-28 config backups = MCP/provider tuning, not
226	  Paperclip. → discard the "Hermes rebuilt the company" belief.
227	- ✅ **Codex's aina-academy AGENTS.md edits (3, 06-19/20) VERIFIED correct + well-scoped** — all pointers
228	  resolve (in-repo docs + /srv/repos doctrine + pkm-agent MCP). Minor: pkm-agent `think` CLI line is Mac-only
229	  (`/Users/Ali/...`) → add VDS variant (`pkm-agent think`). _[needs-ali, small]_
230	
231	### The factory blocker (root-caused)
232	- ❌ **Paperclip API unreachable: agents use `http://paperclip.ainative.academy:3100` but the server binds
233	  LOOPBACK-only `127.0.0.1:3100`.** Public hostname routes Cloudflare HTTPS on 443; port 3100 is never
234	  public → agent heartbeats time out (Jun 25-27). **Fix = agents must use `http://127.0.0.1:3100`** (find
235	  the PAPERCLIP_API_URL injection point — codex-home config.toml doesn't set it; likely the paperclipai
236	  run-agent wrapper). Tracked AIOPS-204/206. act_tier: needs-ali (clear fix, find injection point).
237	- ⚠️ **Companion bug:** sweep/heartbeat scripts call `paperclipai ... -C AINA` (prefix/slug) where the API
238	  needs the UUID `7d58fc13...` → 72× `invalid input syntax for type uuid: AINA`. Fix = use the UUID.
239	  act_tier: safe-now.
240	
241	### Agent-behavior simulation (the evidence-based scoping answer)
242	- **Claude Code (Ali+lead): 13 MCP servers global** (effective for every session; per-repo settings.json
243	  exist but empty). The firehose. **Codex CLI: lean 4 / full-vds 10** (well-designed, no change). **Hermes:
244	  4 scoped** (linear/github/pkm-agent/playwright) — correct.
245	- **Paperclip dept-heads: per-agent role `AGENTS.md`** at `.../agents/<uuid>/instructions/AGENTS.md`. →
246	  **answer to "does every agent need Paperclip's AGENTS.md?": NO — there is no single shared Paperclip
247	  AGENTS.md; each agent reads only its own role instructions.** 56/62 are **mapped-not-applied** (cwd=
248	  /srv/repos/aina-academy = repo-root FIREHOSE, no scoped sources because Finch manifests aren't attached);
249	  only 6 content agents have scoped cwd. → the context-rot is the un-applied scoped context (Finch), which
250	  Ali chose to leave staged (D14).
251	- ✅ **Evidence-based MCP scoping:** grep of package.json shows the 6 specialist servers (clerk, posthog,
252	  sentry, cloudflare-bindings, cloudflare-builds, cloudflare-observability) are used in ONLY 3 repos
253	  (aina-platform, aina-academy, aina-cc). **Plan: remove those 6 from global `.claude.json` → add to
254	  per-repo `.claude/settings.json` in those 3 repos. KEEP cloudflare-docs + pkm-agent + codegraph + linear +
255	  playwright + github global.** Zero functional loss for ~95% of work. act_tier: needs-ali (approved D15;
256	  apply VDS-first→verify→Mac).
257	- ⚠️ **Hermes operational issues (needs-ali):** (1) auxiliary context-compression BROKEN — openrouter no
258	  credit + nous no auth → context summaries fail silently at ~207k tokens (sessions split); (2) Hermes
259	  memory FULL (2,499/2,500 chars) → writes failing; (3) Linear MCP `unrecognized_keys: includeResources`
260	  (deprecated arg in Hermes scripts); (4) claude.ai extra-usage depleted (Anthropic API 400 on aux path).
261	  These matter for "factory running" — fix before heavy Hermes work.
262	
263	## M7c — Hermes/Paperclip research + Paperclip API fix (workflow `wvrswg53h`, 2026-06-28)
264	
265	- ✅ **Paperclip API-URL FIXED + verified.** Added `Environment=PAPERCLIP_RUNTIME_API_URL=http://127.0.0.1:3100`
266	  to `/etc/systemd/system/paperclip.service` (highest-priority override per Paperclip's `execution-target.ts`;
267	  backup `.bak-apiurl-20260628`), daemon-reload + restart. Verified: server listening 127.0.0.1:3100, health
268	  200, env in live process. New agent runs use loopback → the factory API blocker is cleared. Agents stay
269	  paused. (Companion: sweep scripts using `-C AINA` should use the UUID — safe-now script fix.)
270	- ⚠️ **Hermes auth — reality (deviates from Ali's ask):** Hermes is ALREADY on ChatGPT-subscription OAuth
271	  (`openai-codex`, no API key). The Anthropic "extra usage" drain came from its **11 auxiliary slots set to
272	  `provider: auto`** (config.yaml ~L169-255) falling through to Anthropic. **Claude-subscription auth is
273	  BROKEN upstream** (Hermes GitHub issue #40014 — OAuth routes to pay-per-token, not subscription; no fix
274	  shipped). → can't give Hermes Claude-subscription yet. **Recommended fix: pin the aux slots to
275	  `provider: openai-codex`** (stops the drain, keeps everything on ChatGPT subscription). Ali to confirm.
276	- ⚠️ **Hermes memory:** `~/.hermes/config.yaml` `memory_char_limit: 2200` / `user_char_limit: 1375` (full →
277	  writes fail; no auto-compaction). **Long-term memory IS SQLite** — `~/.hermes/state.db` (FTS5), already
278	  enabled, queried via `session_search` (this is the "sqlite something" Ali recalled). Fix the wall: raise
279	  `memory_char_limit`/`user_char_limit`. "Dream cycle" = the **Curator** (shipped, SKILLS-focused, `hermes
280	  curator run --consolidate`); true memory "Dreaming" (3-phase consolidation) is proposed-not-shipped.
281	- ✅ **Reports are viewable NOW (answers M4 + AIOPS-262):** Hermes built a rolling mobile index —
282	  **https://aina-mobile-review.pages.dev/docs/reports/paperclip-review-index.html** (20 backfilled report
283	  companions; source .md at `/srv/repos/aina-academy/docs/reports/paperclip-agent-runs/`). AIOPS-262 (Backlog)
284	  has 3 open decisions for Ali (stable index URL, Slack per-report vs batch, Finch enrichment scope).
285	- 📐 **MCP-access model (the design Ali flagged):** Paperclip MCP is per-COMPANY (shared codex-home config.toml)
286	  — correct, no change. For AINA agents, 3 options: **(A) per-agent manifests** (Finch-style, max least-
287	  privilege, fits the 62-manifest scaffold + AIN-47 fence); **(B) per-repo/department bundles** (~6-8, simpler,
288	  faster); **(C) Hermes-only-full-access coordinator** (NOT recommended — Hermes becomes a bottleneck, fights
289	  Paperclip/Fusion parallel dispatch). **Recommendation: B now (fast, unblocks activation) → A later** (graduate
290	  to Finch per-agent manifests). Specialist Claude MCP servers still scope to the 3 repos that use them.
291	

=== current branch / status ===
main
## main...origin/main
 M transcripts/state/reality-annotations.jsonl
 M wiki/.wiki-failure-counts.json
 M wiki/agent-queue.md
 M wiki/daily/2026-06-28.md
 M wiki/log.md
?? wiki/sessions/2026-06-28-019f0f15-users-ali.md
?? wiki/sessions/2026-06-28-019f0f1f-users-ali.md
?? wiki/sessions/2026-06-28-b8c6cf02-users-ali.md
?? wiki/sessions/2026-06-29-019f10e9-pkm-monorepo.md
?? wiki/sessions/2026-06-29-019f1117-aina-aina-data-engine-room.md
?? wiki/sessions/2026-06-29-019f1118-aina-aina-data-engine-room.md
?? wiki/sessions/2026-06-29-e4e94815-aina-aina-data-engine-room.md
?? wiki/sessions/2026-06-29-f3622214-aina-aina-data-engine-room.md

=== OKF-related branches (local + remote) ===
  ali/okf-emitter-reconcile-2026-06-22
  remotes/origin/ali/okf-emitter-reconcile-2026-06-22

=== PR #3 state ===
{"additions":1583,"baseRefName":"main","changedFiles":8,"deletions":0,"headRefName":"ali/okf-emitter-reconcile-2026-06-22","isDraft":true,"mergeStateStatus":"UNKNOWN","mergeable":"UNKNOWN","number":3,"state":"OPEN","title":"feat(pkm): add OKF emitter","updatedAt":"2026-06-22T16:03:27Z","url":"https://github.com/oscalar/pkm-monorepo/pull/3"}

=== any other open PRs ===
[{"headRefName":"ali/okf-emitter-reconcile-2026-06-22","isDraft":true,"number":3,"title":"feat(pkm): add OKF emitter"}]

=== PR #3 files ===
51+	docs/handoff/2026-06-18-okf-emitter.md
81+	docs/okf/2026-06-17-okf-emitter-design.md
502+	docs/reports/2026-06-22-okf-agent-impact.html
153+	docs/reports/2026-06-22-okf-agent-impact.md
33+	ops/dispatch/01-okf-emit.md
386+	transcripts/tools/okf_emit.py
252+	transcripts/tools/okf_viz.py
125+	transcripts/tools/test_okf_emit.py

=== merge-base / behind-main check ===
commits on PR branch not in main:
4
commits on main not in PR branch (branch is behind by):
55

=== does main already contain OKF emitter files? ===
docs/reports/2026-06-22-okf-manual-canary.html
docs/reports/2026-06-22-okf-manual-canary.md

=== OKF design doc + handoff present in main? ===
ls: docs/okf/: No such file or directory
---

=== last commit on PR branch ===
4c0aae4f8 2026-06-22 12:01:57 -0400 docs(pkm): add OKF agent impact report

Workflow launched in background. Task ID: wj0d3r6mo
Summary: M3 audit: OKF PR #3 decision evidence + knowledge-graph/context-layer design + wiki ingest/date-tag truth
Transcript dir: /Users/Ali/.claude/projects/-Users-Ali-PKM/ea21dc9e-a684-460b-b5d8-244fa739715e/subagents/workflows/wf_69e6974a-066
Script file: /Users/Ali/.claude/projects/-Users-Ali-PKM-monorepo/ea21dc9e-a684-460b-b5d8-244fa739715e/workflows/scripts/pkm-m3-knowledge-layer-wf_69e6974a-066.js
(Edit this file with Write/Edit and re-invoke Workflow with {scriptPath: "/Users/Ali/.claude/projects/-Users-Ali-PKM-monorepo/ea21dc9e-a684-460b-b5d8-244fa739715e/workflows/scripts/pkm-m3-knowledge-layer-wf_69e6974a-066.js"} to iterate without resending the script.)
Run ID: wf_69e6974a-066
To resume after editing the script: Workflow({scriptPath: "/Users/Ali/.claude/projects/-Users-Ali-PKM-monorepo/ea21dc9e-a684-460b-b5d8-244fa739715e/workflows/scripts/pkm-m3-knowledge-layer-wf_69e6974a-066.js", resumeFromRunId: "wf_69e6974a-066"}) — completed agents return cached results.

You will be notified when it completes. Use /workflows to watch live progress.

prewarm_ok
vmi3344880
2026-06-28T23:35:48-0400
Master running (pid=65294)

The file /Users/Ali/PKM-monorepo/docs/pkm-deepdive/RUNLOG.md has been updated successfully. (file state is current in your context — no need to Read it back)