Lukker sløyfen pluginen manglet: en rettelse operatøren gjør i utgave N
håndheves i N+1 i stedet for å bli gjenoppdaget. Maskineriet fantes
(foldIns-skjema + promote→ratify siden fix#1); det som manglet var wiring —
ingenting skrev til køen og ingen fase tømte den.
- Fangst (A2-F7): Steps 2.5/3a/5.5/6.5 appender rettelsen ordrett med trigger,
decision "pending". Ubetinget og bevisst dum — å avgjøre ved fangst om noe
"fortjener" en regel er nettopp slik køen holder seg tom og sløyfen dør.
- Step 11 retro (A2-F8), ≤5 min ETTER scheduling: promoter køen med eksplisitt
JA/NEI (mekanisk → atomisk contract-gate-promotering, teller kun med grønn
--ratify; dømmekraft → operatørens språkregelfil; NEI → rejected, beholdes),
effort-oppsummering fra MÅLT phaseLog (aldri re-estimert), og ÉN
friksjons-spørring som besvares tilbake til operatøren. Faser 18 → 19;
resumption-tabellen ruter scheduling → Step 11, retro → complete.
- articles.NN.retro: additiv-valgfri (default null), schemaVersion forblir 1.
- Background-headless (A2-F6): --background kjører pakken i en bakgrunnsagent
som skriver rapporten til disk; drafting-sesjonen leser fila. Samme isolasjon
som fersk sesjon, uten copy-paste-sømmen. Inline fan-out = eksplisitt fallback.
- Språkregler (C-10): ${DATA}/language-rules/<lang>.md (opt-in, template).
Leses TO ganger — av language-reviewer (fanger) og av Step 4 (forebygger).
Shippet banliste forblir baseline; brukerfila utvider.
- references/fold-in-loop.md (A2-F9): loopen dokumentert domene-generelt in-tree,
så en adopter uten ekstern skrivekontrakt har hele sløyfen. Refs 28 → 29.
Suiter (alle grønne): test-runner 197 → 217 (Section 16u: 18 ubetingede greps +
non-vacuity self-test; fase-sveip 16 → 17 faser; floor 179 → 198) · hooks 174 ·
trends 300 · brain 134 · editions 72 · specifics-bank 45 · contract-gate 33 ·
tests 35 · render 60.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QxvWAjte7vPcF79QeSRvRJ
A1-10 + D-5. The queue answers "what is scheduled"; nothing answered "is the week
covered, and what is still being written". N13 makes both readable at session start.
- Publishing slots are operator config: profile/publishing-slots.json in the per-user
data dir, schema + OPT-IN template in config/publishing-slots.template.json. The
plugin ships no publishing times and writes the grid for nobody; absent config means
every slot surface stays silent (same never-nag discipline as the trend/brain nudges).
Data dir rather than the state file, which is machine-written and whose frontmatter
reader does not parse lists.
- hooks/scripts/slots.mjs (new): one implementation behind four surfaces. Zero-dep (a
SessionStart hook must not spawn tsx), pure core (dates in as strings, no clock),
imported by the commands exactly as queue-manager.mjs already is — so session-start's
vacancy warning and Step 10's slot default cannot drift apart.
- Coverage counts the short-form queue (scheduled/published; cancelled frees the slot)
AND editions-register rows claiming that date, counted per date, so long-form and
short-form cannot be double-booked and a two-slot day needs two posts.
- Session start gains "## Editions in Flight" (series, phase, next action, claimed slot,
days in flight) and "## Publishing Slots" (next open slot, open count in 14 days, and
how to fill it when the gap is <=3 days). The existing discovery nudge is untouched.
- /linkedin router + /linkedin:calendar surface the same roll-up; calendar documents the
opt-in copy. /linkedin:newsletter Step 10 defaults to the next open slot (one
confirmation, not an interview) and writes --slot onto the register row.
- D-5: references/scheduling-strategy.md marked "confidence: low / directional" per the
algorithm reference's own standard, deferring to the operator's grid and their own
analytics. Neutral example time replaces the operator-specific one in the N12 docs.
TDD: red proven first (module absent; 5 session-start behaviours failing).
hooks 140 -> 174 · test-runner 184 -> 197 (Section 16t: 13 unconditional greps incl. a
de-niche check that no publishing time is hardcoded in the slot path and no operator grid
is committed, + non-vacuity self-test; anti-erosion floor 166 -> 179).
Suites green: trends 300/0 · brain 134/0 · editions 72/0 · specifics-bank 45/0 ·
tests 35/0 · render 60/0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QxvWAjte7vPcF79QeSRvRJ
Two things were structurally invisible: WHICH editions are in flight (readable
only by opening every series folder's edition-state.json by hand) and HOW LONG an
edition takes (not recorded anywhere). At two editions a week both are
load-bearing — a production line cannot be dimensioned on numbers never collected.
Register (scripts/editions, new module beside distillate.ts): one row per edition
in ${LINKEDIN_STUDIO_DATA}/editions/register.json — series, edition, title, series
path, current phase, next action, slot, startedAt/completedAt. Data-dir placement
(M0) because the register spans ALL series; the distillate is per-series and stays
in the series root. Verbs: register-upsert / register-list / register-complete.
phaseLog: articles.NN.phaseLog[{phase, completedAt}] in edition-state, additive so
schemaVersion stays 1 — pre-N12 editions load unchanged and their absence reads as
"not measured", never as zero. Per-article, mirroring articles.NN.phase: lead time
is a property of an edition.
One call per transition, both writes: newsletter.md gains a phase-transition
protocol defined once beside the resumption table, and all 16 canonical phases
invoke it. register-upsert appends the phase-log entry AND mirrors the register
row. Deliberately one command, not two — telemetry the command layer must remember
to write separately is incomplete inside a week, and an incomplete log measures
nothing. Step 10 closes the row and prints the measured lead time.
Mirror discipline: resumption still reads edition-state.json and only that. Delete
the register and the next transition rebuilds it; a failed upsert is reported, never
a reason to stop the pipeline. startedAt is the one unrecoverable value, so it never
moves — re-upserting a completed edition reactivates the same row (what
/linkedin:pivot does), keeping the clock on real elapsed production time.
Deterministic: no clock in the core (now passed in, --at/--now at the edge, as the
distillate takes lockedAt). Idempotent where a re-run is legitimate (repeated
transition logs once; completing twice keeps the first completedAt), not where it is
real work (a phase recurring after another one is logged again — a pivot back through
cleared gates is production time that happened). Missing facts are refused at the
edge, not defaulted: a row naming the wrong phase is worse than no row.
TDD (Iron Law): all three test files written first and verified red before any
implementation existed (register.test.ts + editionState.test.ts failed on missing
modules, cli-register.test.ts on "unknown command: register-upsert").
Suites: editions 27 -> 72 (floor raised) · test-runner 173 -> 184 (Section 16s: 11
unconditional greps incl. a 16-phase coverage sweep + non-vacuity self-test;
anti-erosion floor 155 -> 166) · trends 300/0 · brain 134/0 · specifics-bank 45/0 ·
hooks 140/0 · tests 35/0 · render 60/0. tsc --noEmit clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QxvWAjte7vPcF79QeSRvRJ
Every long-form gate agent sees ONE edition, so «the reader has heard this
before» was structurally invisible — and the specifics-bank dedupe actively
encouraged re-surfacing the same material. At series cadence that is the
fastest-growing defect class. N11 makes it visible BEFORE the skeleton is
approved, at two grains:
- Series grain — new scripts/editions package: each locked edition's spent
anecdotes/arguments/hooks are written to <serie>/linkedin/series-distillate.json
at Step 8 lock (distil-append), and the next skeleton is checked against it at
Step 2.5 (distil-check) with the finding folded into the annotation gate the
operator already reads. Advisory, never blocking: a deliberate callback is a
legitimate move, an unnoticed retread is not.
- Material grain — specifics-bank usedIn log: record-usage stamps «used in
edition NN» on the specifics an edition actually consumed (read from the bound
slot-map, so abstrakt/ekstern stamp nothing). At lock, so it means published;
idempotent under a pivot re-lock. Additive-optional, schema stays v1.
Placement deviates from the plan text (${DATA}) after premise-verification:
per-series state belongs in the series root beside edition-state.json, where
Step 0 already resolves the path and no slug→path map is needed. Operator
approved. The distillate module also lands in scripts/editions now rather than
at N12, since the AC required a testable roundtrip and N12 planned that package
anyway — N12 extends it instead of creating it.
Similarity is character-trigram Jaccard, not word overlap: the plugin is
language-general and inflection (migrere/migreringen) breaks word tokens.
Calibrated on real paraphrase pairs — retellings 0.44-0.55, unrelated 0.04-0.06,
same-topic-different-story 0.24 — so the default threshold sits in the gap at
0.40. Word-Jaccard scored a shortened hook paraphrase at 0.11.
TDD (Iron Law): tests written first and verified red before implementation, for
both the new package and the bank logging.
Suites: editions 27/0 (new suite line + guard, floor 27) · specifics-bank
28 -> 45 · test-runner 163 -> 173 (Section 16r: 9 unconditional greps +
non-vacuity self-test; anti-erosion floor 146 -> 155) · trends 300/0 ·
brain 134/0 · hooks 140/0 · tests 35/0 · render 60/0. tsc --noEmit clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QxvWAjte7vPcF79QeSRvRJ
C-8: gjør «praktisk anvendbar på alle nivå» til en konfigurerbar gate-akse og
fjerner den hardkodede primær-personaen. 6-fils scope (KTG-valgt) — gjør
targetLevel til en EKTE ende-til-ende-dimensjon i stedet for et fantomfelt:
- references/longform-quality-rules.md: «non-technical line manager» → brukerens
konfigurerte target-level span (domene-generelt).
- config/personas.template.md: «primær trumfer» krever nå obligatorisk sekundær i
MOTSATT ende av spennet; sekundær-NEI kan kun avskrives som «signal, ikke svikt»
med eksplisitt ceiling-begrunnelse. Ny tekniker-ende-seed (Persona 4,
Løsningsutvikler/AI-ingeniør) — tidligere kun 3 leder-personas.
- config/edition-state.template.json: additivt articles.NN.targetLevel-slot
(default null, INGEN schemaVersion-bump — samme mønster som sourceTrendId/language).
- commands/newsletter.md: persister resolvert targetLevel ved Step 1.5-checkpointet.
- agents/persona-reviewer.md: leser articles.NN.targetLevel og plasserer personaene
på spennet (fallback til persona-ekspertise om feltet er null).
Premiss-korreksjon: planens 4-fils scope antok targetLevel fantes i edition-state
via N7-broen; det stemte ikke (fantes kun i trend-store + fritekst-mapping). +2 filer
(edition-state-slot + newsletter-persist) gjør konsumet reelt.
TDD: ny Section 16q i test-runner (9 UNCOND. .md/.json-greps + non-vacuity self-test);
en vacuity-bug i egen edition-state-check fanget og skjerpet (grep slot-literal
"targetLevel": null, ikke bare ordet — sourceTrendId-docstringen inneholder allerede
"targetLevel"). test-runner 154→163/0, anti-erosion floor 137→146. Øvrige fem suiter
uendret grønne (trends 300 · brain 134 · hooks 140 · tests 35 · render 60).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014bE7VbkmR3cqHFEeGfzgwb
C-4/C-9: Step 9 (post-lock) gets a scoped fact-reviewer (F1/F3) + language-reviewer
(L1/L2) pass over the four feed-facing surfaces — delingstekst hook, samle post,
carousel slide text, edition title — plus a title-strength gate, before the hook
conversion binary. Placed at Step 9 (first phase where all four exist; operates on
distribution copy only, never the locked body — body/canonical-title defect = pivot).
C-6: deterministic post-factcheck fact-invariant after Step 5 — any later fold-in
whose diff touches a sentence with numbers, names, or quotes fires a mandatory scoped
fact-reviewer mini-recheck (not a judgment call), closing the pivot-heuristic blind
spot; enforced at the 6.5 + 7 fold-in sites.
C-7: WebFetch added to fact-checker + fact-reviewer (verbatim source text for quote
precision / exact-figure attribution).
TDD RED→GREEN: test-runner Section 16p (8 unconditional greps + non-vacuity
self-test), anti-erosion floor 129 → 137. Suite 146 → 154/0. Domain-general.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014bE7VbkmR3cqHFEeGfzgwb
Bind the long-form VOICE before prose — the discipline the line already
applies to SOURCES (validate-binding), its strongest link.
- C-1/A2-F1: voice-trainer distils a chronicle-voice-profile from the
published gold corpus (per language/format); newsletter Step 0/3a/3b bind
drafting to it, with an explicit never-silent fallback to the short-form
corpus (which the engine itself calls degrading for chronicle voice).
- C-2: Step 3a Pass 0 READS the chronicle-voice-drift-log as a negative list
(it was append-only, never consumed); Step 3b re-reads it in expansion.
- C-3: Step 1 resolves + persists the top-level edition-state.language, making
the template doc-string's "Resolved at Step 1" claim true; the Norwegian
review axis (language-reviewer / voice-scrubber) now fires when the edition
IS Norwegian instead of being silently off by default. Step 6.5 already
consumed the field — resolving it closes the axis end-to-end.
- A2-F2: Step 8 lock appends the locked edition to ingest/published
(provenance=published), so the gold corpus the profile distils from grows
one edition at a time — never the engine's drafts.
Domain-general throughout (languages/formats from the user's own corpus).
test-runner 140 -> 146 (Section 16o: 6 unconditional voice-chain greps +
non-vacuity self-test; anti-erosion floor 123 -> 129). All six suites green
(trends 300, brain 134, hooks 140, tests 35, render 60).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014bE7VbkmR3cqHFEeGfzgwb
Del 2 (broen): commands/newsletter.md Step 1 trend-intake pre-fills brief
(angle/targetLevel/key-points/source-URLs) from a /linkedin:trends candidate via
the trends CLI; sourceTrendId persisted at Step 1.5; Step 10 auto-act flips the
source trend to acted (closes the discovery->production loop deterministically);
Step 2 external fact-package intake path (kilde-så-draft with finished research).
edition-state.template.json: per-article sourceTrendId (additive-optional, no
schemaVersion bump). A1-8: grep -ci trend commands/newsletter.md 0 -> 18.
Del 2.5 (F7 gate, TDD): score.ts capForActionability — a candidate whose reader-grip
is explicitly not formulated (actionability.formulated=false) caps to at most High
regardless of composite. Pure; overrides the composite->band derivation downward only
(composite/dimensions/mode preserved, score stays etterprøvbar); absent actionability
left uncapped (legacy-safe). Wired into the item.ts capture path so the PERSISTED
priority is gated; store carries the envelope verbatim (re-capture safe). No KTG values
hardcoded — the N6 field carries the verdict, the gate only enforces it.
Suites (all green): trends 276/0 (was 266, +10), test-runner 139/0, brain 134/0,
hooks 140/0, tests 35/0, render 60/0; tsc --noEmit clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014bE7VbkmR3cqHFEeGfzgwb
Ukalibrert modell-dømmekraft inflaterte topp-båndet (observert 13/20 → Immediate).
Legger til en eksplisitt fordelings-forventning som scoring-SSOT: de fleste kandidatene
hører hjemme i Medium–High (4.0–7.9); Immediate (≥8.0) er unntaket (arbeidsmål ≤~3/20).
Score relativt over batchen, ikke sjenerøst i isolasjon.
Fordelings-formen er mekanisme; terskelen for hva som teller som eksepsjonelt bor i
brukerprofilen/dimensjons-rubrikkene, aldri hardkodet (domene-generelt). Etterprøvbart nå
fordi composite+band persisteres (RE-R3a); MR-F7-bånd-cap-gaten (N7) bygger på samme signal.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S7SQpXJpBSNvpWaTNq1kWZ
OKF-runden lukket ved trinn F (konsensus). Vår trinn E-svarfil bodde i den
interne postkassen (~/repos/_okf-interim/), som nå rives. Løfter den varige
substansen som kun fantes der inn i planen (ny §10): vår §7-registerrad
(status `planned`, ingen dør-A-stempling, uforfalskbarhet utenfor fila),
D1-posisjon (ratifisert D1 matcher vår innvending), D3 `partial`-posisjon
(ratifisert def = samme dør; vi er `planned`, ikke `partial`), samt tre
fremadrettede tekniske krav som ellers gikk tapt: tomt-vokabular-ikke-snap
(+ okrs egen RECOMMENDED-konstant som støtte), index preserve-unowned-grensen,
og lag-1-som-sett (delte golden-fixtures over fire resolve_link-akser).
Alt verifisert mot kode 2026-07-20. Ingen kodeendring i pluginen.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S7SQpXJpBSNvpWaTNq1kWZ
To selvkorreksjoner verifisert mot kode:
- type-vokabularet er tre verdier, ikke fem (JournalEntry/TributarySummary
finnes kun i designnotat, aldri i kode)
- tilslutningen til lukket vokabular ble gitt uten aa sjekke innholdet;
okf-check --strict-ingest exit 1 paa vaart konforme bundle, og safe
default kollapser alle tre typene til Dokument
Nytt empirisk argument mot obligatorisk noekkel-prefiks: referanse-checkerens
egen konstant heter RECOMMENDED, ikke REQUIRED (exit 0 med advarsler).
Q3 (fase 4-omfang) besvart av llm-ingestion-okf: lag 1-primitivene, ikke
materialize_bundle-paritet. Status forblir planned.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019uzhvLjCm39mNQXVFwG7w2
Forrige commit hevdet at delt fixture-korpus MÅ bære et verbatim,
ikke-normalisert tilfelle. Kryssjekk mot claude-playlist-corpus' trinn
D-svar viser at det var overdrevet: vår newline-frie serializer gjelder
ingest/published/, som er ekskludert fra OKF-bundelen ved design.
Filene write_concept faktisk ville skrevet (brain/) far trailing newline
fra var egen serializer. Pa konseptstien er vi enige med dem.
Det som overlever: write_concept skal ogsa tjene som generell verbatim-
writer utenfor en bundle, sa ramme-normalisering bor vaere en dokumentert
parameter, ikke en ubetinget garanti.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019uzhvLjCm39mNQXVFwG7w2
Trinn D-review besvart i _okf-interim (postkasse, ikke tracked her).
Verifisering mot egen kode avkreftet to påstander i plan §4 R5:
- Ingen stabile err.code i noen av de fire CLI-ene (fritekst-Error +
numerisk process.exit). err.code er et krav TIL biblioteket, ikke en
disiplin vi speiler.
- "Nøyaktig én avsluttende newline" er ikke vår invariant;
serializePublishedRecord utelater den bevisst for byte-eksakt
round-trip (SC2). Delt fixture-korpus må bære et verbatim,
ikke-normalisert tilfelle.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019uzhvLjCm39mNQXVFwG7w2
Legger det varige innholdet fra adopsjonsrunden i eiende repo (postkassens
regel 2: beslutninger bor i eiende repos docs, kopien i _okf-interim er
varselet).
Nytt par. 6: bundle-inventar + plassering. Vi har EN OKF-bundle (brain/),
og alt bruker-eid ligger allerede utenfor repo-treet via data-dir-seamen
(M0/v0.6.0) - beslutningen krever ingen migrasjon hos oss.
Innsigelse mot taksonomien: beslutningstabellen diskriminerer paa hvem som
skriver. Vaar brain/ skrives av pluginen og ville derfor klassifisert som
plugin-eid og hoert hjemme i repoet - som er akkurat feil for det mest
personlige vi har, i et offentlig distribuert repo. Riktig akse er hvilken
livssyklus innholdet foelger. Maskinskrevet + bruker-eid er ikke et hjoerne,
det er hva ethvert laerende system produserer.
Par. 7-9 renummerert.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019uzhvLjCm39mNQXVFwG7w2
Kartlegger repoets inntaksflater mot dor A i llm-ingestion-okf v0.3.1 og
leverer kravgrunnlaget som avgjor F1 for dette repoet.
Kjernefunn: F1 (manglende fritekst-connector) er ikke var blokker. All
ekstern henting i pluginen skjer i en modell-tur (trend-spotter WebSearch/
WebFetch, /linkedin:react WebFetch) - det finnes ingen rah HTTP-fetch i
plugin-kode. Dor A forutsetter at kode kan hente bytene; for var storste
og minst betrodde flate kan den ikke det. Kravet vart er derfor R1: en
dor som tar imot allerede-hentet, modellprodusert payload og
materialiserer den deterministisk. trends capture er den kontrakten
allerede, og donerbar som presedens.
Fire verifiserte kollisjoner mot dor A: fast 7-nokkels frontmatter uten
extension keys; rewrite-on-run-eierskap mot inkrementell akkresjon;
slug-ids mot content-adresserte ids; tabell-render som kollapser
linjeskift.
ingest/published/ forblir plugin-lokal ved design - na med teknisk gulv,
ikke bare beslutning: byte-eksakt round-trip (SC2), un-normalisert
mintContentId(body), YAML-fri grammatikk (samme grunn som Stage-1-bundlen
ekskluderte ingest/), og model-collapse-guard-semantikk.
Markorlinje satt i STATE.md (local-only): planned.
Separat fra docs/ingestion-guard/plan.md - grensen er beskrevet i par. 6.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019uzhvLjCm39mNQXVFwG7w2
Plugin eier hele render/ (KTG-go 16.07, eierskapsvalg a). Portet fra
maskinrommets tools/build-html.mjs (read-only): flerlinjers blockquote
(avsnitt i quote), [tekst](url)-lenker med http/https/mailto-whitelist,
FIGUR-markoer -> inline SVG med figcaption + fallback til blockquote.
CSS for de tre konstruktene. 7 nye paritetstester (repoets testtilfeller
kopiert + med-SVG-case); paritetsbevis: samme fixture gjennom begge
motorer = byte-identisk parser-HTML. Render-floor 53 -> 60.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: c61b308e-e4fa-4a49-9ff0-a3ce725cf703
B-F1 (prod-bug, prune de facto no-op): fjernet `m`-flagget fra Recent Posts-
regexen i `state-updater.mjs`. Med `/m` matchet `$`-alternativet i lookahead
slutten av hver linje, så den late capturen stoppet etter FØRSTE entry-linje;
gamle entries under en fersk (nye prepender øverst) ble aldri skannet/prunet.
Ny regresjonstest (old-under-fresh) rød->grønn beviser mekanismen.
B-F2 (kalender-flake): `pruneContentHistory` fikk valgfri `today = new Date()`-
param (deterministisk rotårsak-fix; CLI/hook-kall uendret via default). Prune-
testene injiserer fast syntetisk today og asserter kun på Recent Posts-seksjonen
(ikke helinnhold/frontmatter), så today-100d aldri kolliderer med SAMPLE_STATE-
datoer. Samme grep i `scripts/check-replace-safety.mjs`.
Suiter (alle grønne): test-runner 138/0 · hooks 140/0 (+1 regresjonstest) ·
trends 245/0 · brain 134/0 · tests 35/0 · render 20/0.
Utført av Fable 5 (high) subagent, orkestrert + verifisert av Opus-hovedkontekst.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: a814039f-da8b-41b1-8d7c-60abe1e03c5a
Read `llm-ingestion-guard` v0.2 adoption brief; mapped this repo's untrusted-ingest
surface (brief §7 checklist) against 882f6ee via two independent read-only surveys.
- New `docs/ingestion-guard/plan.md`: when/where to wire the guard at the deterministic
persist gates (`screen_output` at trends `capture`, `brain ingest`, specifics-bank
`ekstern` bindings, analytics CSV). Ranked by automated-reinjection risk; the trends ->
`session-start.mjs` reinjection is the one live poison->trusted-read loop (priority 1).
- Honest scope: only the `screen_output` half maps cleanly (fetch/transform happen inside
the model turn — no `your_model()` code seam); `prepare_input` has no clean wiring point.
- Python<->Node interop is the blocker (plan §7); no code wired (brief = plan-only).
- OKF `import_bundle` has no seam today (brain is export-only); relevant only when SB-S4
connector or a cross-plugin shared skill lands.
Machine-readable marker line for the guard repo's roll-up lives in STATE.md
(LOCAL-ONLY / gitignored, so the roll-up is machine-local).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: 57adea42-c8f1-4b88-acb6-2453e1239c79
Independent two-lens cold review of the 6 Measure-journey surfaces (import,
report, analyze, audit, ab-test, measure) on frozen HEAD 69f37ba. Largest
batch + only analytics-class batch; analytics-honesty predicate carried
alongside the standard intent/correctness lenses.
Verdict: REWORK (0 BLOCKER · 1 MAJOR · 4 MINOR · 6 SUGGESTION).
- report.md MAJOR: heatmap report type (:72) routes to a nonexistent "Step 6c";
real handler is Step 2c (:106) — provably-wrong cross-ref on a primary menu
branch. Caught by the correctness lens alone (intent lens never traced
step-jump arithmetic); confirmed by main grounding the step inventory.
- 5 of 6 surfaces ALLOW (measure notably clean — delegate-only enforced by the
allowed-tools whitelist, not just asserted).
Analytics-class predicate PASSES on all 6: saves framed as manual/count-only/
no-API and never folded into engagementRate; dwell called unmeasurable;
parseOptionalCount (csv-parser.ts:71) + getAnalyticsRoot seam described
accurately wherever quoted; graceful degradation present everywhere.
Independence: 2 convergences (import Step 6a invalid trends flags; ab-test
ER-omits-clicks) + 3 divergences resolved by main grounding in both directions
(intent over-rated import 6a MAJOR->MINOR; correctness uniquely caught report
6c + analyze twin severity scales). Two-lens method earned its keep again.
New R4 finding clusters (operator-gated fix, not done here): sibling
interface/metric drift (import stale trends flags vs report; ab-test ER vs CLI
engagementRate), one true under-declaration (report Step 8b Write). No code
changed; cold review finds only.
Cumulative cold-review coverage: 23/29 (S1 + R2a + R2b + R3 + R4). Remaining:
R5 (Grow+Router, 6 surfaces) -> 29/29. test-runner 138/0 unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012GqEHp4uDiivfrAUjw4BdE
Reproduces the S1/R2a/R2b non-fabricating method: 2 blind cold Opus lenses for
the round (intent + correctness), each covering all 4 surfaces, no cross-feed,
every mechanical claim tool-grounded. Surfaces: firsthour, calendar,
headless-review, pivot.
Verdict REWORK — 1 of 4 surfaces, 1 MAJOR; 0 BLOCKER:
- calendar: the queue load (queueFormatSummary) surfaces none of the
id/draft_path/character_count that publish/reschedule/cancel require, and the
reschedule step's "carry from the entry shown in Step 2" is a direct
contradiction (those fields are never shown) -> MAJOR
- firsthour: ALLOW (1 MINOR bare ref paths; clipboard printf pointer)
- headless-review: ALLOW (2 MINOR: SendUserFile absent from allowed-tools on the
primary surfacing path; dead v3.1.0 reload anchor post version-reset)
- pivot: ALLOW (clean — heuristic, worked example, off-by-one phase map all
reconcile)
Independence axis earned its keep again: convergence on headless SendUserFile
(both lenses) + one divergence resolved by main grounding (calendar -
intent-lens flagged the data-gap MAJOR, correctness-lens passed it on structural
arithmetic; main grounded queueFormatSummary's output and confirmed the MAJOR,
same shape as R2b's newsletter resumption table).
Connects to existing systemic findings (no new cross-cutting): clipboard printf
(firsthour confirmed, folds into R2a's 10-file finding) and bare reference paths
(firsthour adds 3 sites to R2b's pattern). New SUGGESTION-class pattern:
allowed-tools over-declaration on 3 of 4 surfaces.
Counts: 0 BLOCKER, 1 MAJOR, 3 MINOR, 5 SUGGESTION. Cumulative cold-review
coverage 17/29. Review finds; changes no code. test-runner 138/0 unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012GqEHp4uDiivfrAUjw4BdE
Reproduces the S1/R2a non-fabricating method: 2 blind cold Opus lenses per
surface (intent + correctness), no cross-feed, every mechanical claim
tool-grounded. Surfaces: create, batch, pipeline, newsletter.
Verdict REWORK — 3 of 4 surfaces, each 1 MAJOR; 0 BLOCKER:
- batch: 3a/3b component scaffold (format-blind) contradicts the format-aware
band gate
- pipeline: Step 2 scaffold (960-1640) cannot satisfy Step 3 total band
(1200-1800)
- newsletter: resumption table omits the contract-gate phase (Step 4.5) ->
breaks deterministic resume between Step 4 and Step 5
- create: ALLOW (1 MINOR, 8-option AskUserQuestion vs documented 2-4 cap)
Two systemic patterns surfaced (main-grounded, not single-reviewer):
- 5-component draft scaffold 960-1640 != 1200-1800 band, in exactly 3 files
(post.md [R2a], batch.md, pipeline.md)
- bare reference paths vs ${CLAUDE_PLUGIN_ROOT}/ in batch + pipeline
Independence axis earned its keep: on newsletter the intent-lens asserted the
resumption table complete; the correctness-lens counted the gap; main confirmed
the correctness-lens (a divergence resolved by grounding, not just convergence).
Review finds; changes no code. Fixes are separate operator-gated decisions.
test-runner 138/0 unchanged. Hardening-class artifact.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012GqEHp4uDiivfrAUjw4BdE
R2a of the cold-review sweep: post/react/carousel/video/multiplatform, 2 independent cold
Opus reviewers per surface (intent + correctness lenses), no cross-feed, every mechanical
claim tool-grounded (anti-fabrication mandate). Verdict REWORK — 2 MAJOR (systemic
clipboard-printf apostrophe corruption across all 10 content commands; post personal-stories
band 1,000-1,400 vs its own Step 5 gate + canonical SSOT 1,200-1,800), 0 BLOCKER. Findings
only — no code changed. Local-only hardening artifact, not pushed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012GqEHp4uDiivfrAUjw4BdE
CLAUDE.md is loaded every turn while working in this repo (measured 4,846
always-loaded tokens — the entire per-repo delta, since the repo has no
.claude/rules or .mcp.json). The bulk was a version-history narrative: the
2026-05-31 re-baseline note plus a single paragraph recounting the full
v2.0.0 → v4.1.0 journey (per-version motivation, gate-by-gate evolution).
That is CHANGELOG material, not an invariant for working on the plugin —
CHANGELOG.md (18 version headings) already owns it.
Trimmed to what is invariant: a terse current-maturity intro (v0.5.3, M0
done, v1.0.0 remainder) pointing to CHANGELOG/docs; the architecture,
hooks, command (29) and agent (19) tables, and content-quality rules. The
verbose per-row prose (newsletter phase-list, agent (vX.Y)/Step-tag
motivation) is compressed to one-liners. Agent name/model/color cells,
all counts, and the version stay byte-exact so the structure lint holds.
Verified with the repo's own gates: scripts/test-runner.sh 138/0 ("All
structural checks passed!" — counts, version-consistency, stat-consistency,
model-consistency, render-chain all green); scripts/check-model-consistency.mjs
OK (19 agents, all surface declarations match frontmatter). CLAUDE.md
127→109 lines, 19,572→10,457 B, 4,846→2,580 tok (−47%). Docs-only — no
version bump, no catalog ref change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01683eAqVecv9VZfQzL8CQ9h
Completes the linkedin-studio in-scope Stage 1 (docs/okf-convergence-brief.md):
- serializeProfile + operations.md seed gain the cheap recommended OKF fields
`title` + `description` (timestamp/resource intentionally omitted — a timestamp
would break the pure/deterministic serializer; resource is N/A internally).
- renderDiffMd leads the transient pending-diff.md with `type: PendingDiff`, so
the brain/ bundle passes okf-check even mid-propose.
Verified: 2 new tests (okf-conform + consolidate-cli); full brain suite 134/134
(0 regressions); cross-tool — okr/scripts/okf-check.mjs validates brain/ exit 0
WITH a pending-diff present.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
Cross-plugin second-brain convergence on OKF-compatible form. The brain is the
reference design (most mature of the three); OKF is a thin interop veneer.
Records: three verified premise corrections (mdcode != OKF; OKF has no
document-folder ingest; classify/convert is build-yourself and the sibling
docs never asked for it); the three-consumer landscape (okr built, architect
designed, linkedin-studio richest); the staged plan (shared spec -> measure ->
conditional shared skill); per-repo scope boundaries (each its own go); and the
landed Stage-1 outcome with its premise refinement (bundle=brain/, ingest/
excluded as round-trip-critical tributary).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
Cross-plugin OKF convergence, Stage 1 (docs/okf-convergence-brief.md): the
second-brain hub now conforms to OKF-compatible form so a shared retrieval
skill (and a sibling agent) can traverse it.
- serializeProfile leads with a constant `type: Profile` frontmatter block
(round-trip-safe: parseProfile skips it, parse-serialize identity holds).
- operations.md seed -> `type: Operations`; index.md seed -> `okf_version: 0.1`
marker (markdown text; index files carry no frontmatter per OKF); new
brain/journal/index.md (per-level index).
Premise correction: the brain is deliberately YAML-free and ingest/published
has a byte-exact round-trip invariant (SC2) a frontmatter block would break,
so the concept-bundle is scoped to brain/ ONLY; the ingest/ tributary is
excluded and pointed to from the hub. We emit frontmatter, add no parser.
Verified: new tests/okf-conform.test.ts (5/5); full brain suite 132/132 (0
regressions); cross-tool — okr/scripts/okf-check.mjs validates brain/ (exit 0).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
CLAUDE.md:3 was stale (it predated the hardening campaign). Corrected to the
tool-verified reality:
- Hardening complete (29/29): every command through the interactive quality-gate
(docs/hardening/log.md: HARDENED/PASS/FIXED) + the S27-S30 reference/terminology/
magnitude scrubs.
- Command testing is effectively that campaign (persona-sim + 4-axis eval + lint),
backed by the script-level suites.
- Independent cold /trekreview adjudication persists for S1 only (4/29); S2-S26 were
gated by the operator-in-the-loop v2 method after the reviewer swarm was dropped
following the S2 fabrication incident -> brief SC-H not met as written.
- GUI is the one workstream not yet begun.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
Closes hull #7 ("ingen brief-historikk"). Each morning brief now records the
trend ids it showed into its own YAML frontmatter (surfaced: <id-csv> =
surfacedIds(ranking)) and renders a day-over-day diff against the most recent
prior brief — a "## Nytt siden sist (<prior-date>)" section that leads the
ranked list, plus a " N nye siden sist." marker on the one-line summary the
SessionStart hook surfaces (no hook change).
- brief.ts: BRIEF_SCHEMA_VERSION 1->2 (artifact frontmatter gained surfaced:;
the store's SCHEMA_VERSION stays 4 — no store field). Three PURE helpers
(diffSurfaced / parseSurfacedFrontmatter / selectPriorBriefFile) + the
surfaced: emit + the section + the summary marker. No fs/clock in brief.ts.
- cli.ts: the brief handler discovers the prior dated file (existsSync-guarded
readdirSync -> selectPriorBriefFile, strict < today so a same-day re-run is
byte-identical), parses its surfaced: line, computes the diff, threads it into
renderBrief AND the shared briefSummary(ranking, diff) (one-source: file
frontmatter == --json summary, cli.test one-source invariant). --json gains a
diff:{priorDate,added,carried,dropped} counts object; the console line appends
the delta. Any fs error degrades to the empty-prior (first-brief) path.
TDD two-phase: stubs -> 17 value-RED (no module-not-found) -> GREEN. Trends suite
216 -> 245 (brief +27, cli +2), 0 fail. New unconditional gate Section 16n (6
checks); ASSERT_BASELINE_FLOOR 117 -> 123; TRENDS_TESTS_FLOOR -> 245. Full gate
FAIL=0; hook suite 139/139 + R3c schedule/run-daily green untouched. Behavioural:
real two-day rename-real-write diff + same-day byte-identity confirmed. Counts
29/19/27 unchanged; no version bump (additive, v0.5.2 dev).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
The shipped render dates the section header (## 🆕 Nytt siden sist (<prior-date>))
when a prior brief exists — the form SC9, Phase-B, and the behavioural step already
specify. The plan Step 3 snippet showed a bare header and the brief S-history prose
was silent on the date; both are corrected to match the shipped code.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
Closes hull #7 ("ingen brief-historikk"): each morning brief records the trend
ids it showed (surfaced: frontmatter, BRIEF_SCHEMA_VERSION 1->2; store schema
stays 4) and renders "Nytt siden sist" against the most recent prior brief.
Pure render-time diff (brief.ts stays store/fs-free; the dir+file reads live at
the cli.ts edge). Zero new source/test files — all EDITs.
Light-Voyage (3 Opus reviewers — scope-guardian MIXED, brief-reviewer
PROCEED_WITH_RISKS, plan-critic REWORK 0.88) folded into brief #9 / plan
Plan-critic. Converged on 2 MAJOR + 4 MINOR, all re-verified against live code:
- MAJOR-1: brief.test.ts:574 assert.equal(BRIEF_SCHEMA_VERSION, 1) is a hard
literal outside the frontmatter set -> Step 3 flips it to 2 with the bump.
- MAJOR-2: cli.ts:350 const summary = briefSummary(ranking) left unthreaded ->
day-2 --json.summary would lose the marker the file carries (breaks the
cli.test.ts:268 one-source invariant); Step 4 threads diff -> briefSummary.
- M1 import type for the BriefDiff interface; M2 SC9 rename-real-write (no
hand-fixture); M3 SC1 cross-partition disjointness wording; M4 empty
surfaced: contradiction reworded.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
R3 slice (b): the rest of hull #3. The morning brief now reads the temporal axis
the R3b seen-log records but the ranking ignored. Two DERIVED signals, computed at
brief time from already-persisted fields (publishedAt/capturedAt -> ageDays,
surfacedCount), never stored:
- first-mover: recent (ageDays <= --first-mover-days, default 2) AND never surfaced
on a prior day -> ranked up, badge "first ute". Future-dated (ageDays<0) excluded.
- saturation: surfaced on >= --saturation-at (default 3) prior days -> ranked down,
badge "mettet (Nx)". Self-surfacing (our seen-log), not market coverage.
- warming (1..at-1) keeps the R3b "sett Nx" badge but only at >=2 (contract intact);
neutral carries no badge.
SB1 derived (no schema bump: SCHEMA_VERSION 4 / BRIEF_SCHEMA_VERSION 1 untouched).
SB2 the R3a relevance composite stays the PRIMARY sort key; the temporal rank is a
new cmp key after pillar-overlap, before effectiveDate -> re-orders only WITHIN a
(composite, overlap) tier. temporalSignal is pure (saturationAt clamped >=1).
Prior-day surfacings exclude today (via lastSurfacedAt), so a same-day re-render is
byte-identical (caught by the R3c run-daily SC7 regression; fixes a latent R3b
prior-day imprecision too). brief CLI gains --first-mover-days / --saturation-at;
schedule untouched (nightly uses defaults).
Wiring: trend-spotter.md (prose), trend-scoring-modes.md (one-line consumer note),
README (## Temporal overlay), gate Section 16m (+6 unconditional -> ASSERT floor
111->117), TRENDS_TESTS_FLOOR 192->216. Counts 29/19/27 unchanged. Zero new files.
Gate: Passed 132 / Failed 0; trends 216/216; hook suite 139/139 untouched.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
Slice (c) of the R3 build-out: a `schedule` CLI verb (print-first launchd
plist / cron-table line) + `run-daily.sh`, a bash-3.2 headless wrapper that
runs the DETERMINISTIC morning brief from a profile-less scheduler env.
Closes hulls #1 (no autonomous trigger) + #6 (no headless entry point).
Operator-confirmed (AskUserQuestion 2026-06-26): C1 deterministic brief-only
(no AI capture — that is slice e, which plugs into the documented pre-brief
seam); C2 print-first installer (the tool emits the artifact + the install
command; `--install` writes only the inert launchd plist file; never runs the
scheduler activation itself).
Light-Voyage hardened — three Opus reviewers, each verifying against live
code: scope-guardian ALIGNED (0 creep/0 gaps), brief-reviewer
PROCEED_WITH_RISKS, plan-critic REVISE. All findings folded, incl. the
pretty-printed `brief --json` log-line compaction, the `cd "$DIR"` cron fix,
the logPath base pinned to `dirname(defaultStorePath())`, the canonical
`ScheduleSpec.env`, and the `ASSERT_BASELINE_FLOOR` :1259->:1329 cite. No
schema/count change (29/19/27, store v4). Tracked feature-design.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
The lifecycle layer over the trend store: what happens to a trend AFTER first capture.
- re-score on re-capture (last-wins; addTrend duplicate branch, score the one mutable
field; provenance + lifecycle untouched; no false-merge via JSON compare). Reverses
R3a's first-sight D3 — that R3a test reconciled to the new behaviour.
- status new/acted/skipped (effectiveStatus/setStatus + act/skip/reset CLI verbs);
rankForBrief EXCLUDES handled trends (a work queue, not an archive).
- seen-log surfacedCount/lastSurfacedAt (markSurfaced, per-day idempotent); the brief
CLI records surfacing on the store AFTER the pure render, unless --no-mark.
- render: entry id in backticks (copy-paste for act/skip) + · sett Nx prior-day hint.
- schema v3→v4 (additive lossless); the R3a migration block reconciled to the bump,
the new R3b block committed against SCHEMA_VERSION (breaks the reconcile cycle).
score.ts + item.ts untouched (re-score reuses the R3a capture path). RED-first (two
phase: 16 logic-RED + 4 stub-RED). Gate: Section 16k (6 emitters), TRENDS_TESTS_FLOOR
146→171, ASSERT_BASELINE_FLOOR 99→105. trends 171/171, gate 120/0/0, hook suite 139/139.
Plan: docs/research-engine/{brief,plan}-re-r3b.md (light-Voyage hardened @ c40b937).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vmzxpsFpc8q19LaogAWLD
R3 slice 1 (research-deepening): persist a 4-field TrendScore {mode, dimensions,
composite, priority} on the store record (schema v2->v3, additive lossless), computed
by the already-built score.ts (composite+band, one owner), threaded item->store, and
rank rankForBrief on composite first + surface band+mode in renderBrief.
Go-gate confirmed (operator "Go"): D1 4-field envelope · D2 composite primary within
bucket · D3 first-sight only · D4 one slice · D6 mode shown per body entry.
Light-Voyage: scope-guardian ALIGNED (0) / brief-reviewer PROCEED_WITH_RISKS (6 MINOR)
/ plan-critic REVISE (1 BLOCKER, 4 MAJOR, 4 MINOR) — all folded. Headline fold: the RED
proof is now explicitly two-phase (logic-RED for store/brief/cli; stub-first then
assertion-RED for score/item, since a missing named import throws at module-load under
Node16 ESM, not on assertion).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VmHCQjJHUyWwxGAVVjNLgp
