josephfung · josephfung · Jul 18, 2026 · Jul 17, 2026 · Jul 17, 2026 · Jul 17, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -37,6 +37,7 @@ bus event types) are noted explicitly even in the `0.x` range.
 - **`task-completion-from-sent`** — re-check tasks are open and CEO-owned; fail closed on subtask-lookup errors. (#1419)
 - **Voice-learning capture** — a body-only draft edit refreshes the snapshot in place, preserving its original `created_at`. (#1419)
 - **`voice-learn`** — voice-guide approval loop no longer breaks after one cycle; honors dismiss cooldown. (#1419)
+- **Learning-subsystem config writes** — soft-rejected `ConfigStore` writes now hold state instead of silently advancing. (#1438)
 
 ### Added
 
@@ -70,6 +71,7 @@ bus event types) are noted explicitly even in the `0.x` range.
 - **Install docs** — README points to `releases/latest/download/install.sh` and documents `CURIA_VERSION`.
 - **`00-overview.md` Spec Index** — dropped the `Status` column and "N of M Done" counts. (#1282)
 - **Spec audit fixes** — corrected stale claims: Node 24, console UI, secrets vault, setup story, fallback-provider prose. (#1282)
+- **Learning subsystem** — queue/status/guard state moved from OKF doc bodies to config-store JSON. (#1438)
 
 ### Removed
 

diff --git a/docs/adr/029-passive-email-observation-and-counterfactual-competence.md b/docs/adr/029-passive-email-observation-and-counterfactual-competence.md
@@ -45,7 +45,10 @@ Concrete mapping:
 | Learned voice | free-form `WritingVoice.guide`, maintained by a weekly batched LLM pass and approved via the digest |
 | Todo completion | `tasks` + `task-complete` / digest confirm |
 | Capability evidence | Pre-scored rows in `autonomy_action_log` |
-| Cadence / watermarks / guard markers | Scheduler crons + `config-store` |
+| Cadence / watermarks | Scheduler crons + `config-store` (`ceo_inbox`) |
+| Queue / status / guard state (completion candidates, digest items, voice proposal, matched-draft & asked-task guards) | `config-store` JSON (`ceo_inbox`), one whole-rewritten key per store |
+
+Machine-state keys use whole-object last-write-wins (no per-doc version check); acceptable because each key has effectively a single writer per cron tick and the queue key is keyed by task id so a held-watermark retry re-adds idempotently. Prose evidence (`pending-diffs.md`), snapshots, and shadow docs remain OKF.
 
 Rationale for rejecting new stores (options 1 and 3): the destinations already exist and are versioned/auditable; a parallel schema would duplicate versioning, digest surfacing, and autonomy gating. The KG is the wrong home for raw draft prose and edit diffs (spec 21: "the workspace is the scratchpad, the KG is the record").
 

diff --git a/docs/specs/04-channels.md b/docs/specs/04-channels.md
@@ -97,7 +97,7 @@ Interactive terminal for local dev and testing. Reads from stdin, writes to stdo
 - **CEO inbox drafts (v0.35):** the CEO inbox can now find, read, and edit unsent drafts. `ceo-inbox-list` / `ceo-inbox-search` query Nylas's `/drafts` resource (the DRAFTS folder) when scoped with `folder: 'DRAFTS'`; `ceo-inbox-read` accepts a `draft_id` to return a draft's full body; and the new `ceo-inbox-draft-edit` skill updates a draft's recipients/subject/body. Draft results key on `drafts` (not `messages`) so a genuine "none" is distinguishable from a silent zero, and the inbox poll watermark is not applied to drafts.
 - **Email accounts (v0.37):** multiple agent-owned mailboxes are managed from the console under **Settings → Channels → Email → Email accounts**. Each account is a row in the `email_accounts` table (migration `064_create_email_accounts.sql`) with its Nylas grant stored in the vault at a per-account key (`channel.email.<name>.nylas_grant_id`). The shared app-level `NYLAS_API_KEY` remains an env-bootstrapped vault secret. The legacy `channel_accounts.email` YAML path and `CEO_PRIMARY_EMAIL` env var are retired.
 - **Vault credential wiring (v0.37):** at startup `applyChannelVaultSecrets()` overlays channel vault credentials (`channel.<name>.<field>`) onto the runtime config before adapters are constructed, so credentials saved via the Channels console take effect on the next restart without any env-var or YAML change.
-- **Sent-mail observation (v0.42, ADR-029):** a scheduled `ceo-inbox-sent-observe` skill polls the CEO's **Sent** folder through the existing `ceo-inbox` Nylas grant (no new credential). Scope is **all sent mail**. Each run is watermarked (`config-store` under the `ceo_inbox` namespace, same high-water pattern as the inbound email poll) and is idempotent on re-run. For each new sent message it (1) correlates to Curia-authored draft snapshots (below) by `thread_id` + recipients + send time and appends `(draft, sent)` pairs to a rolling OKF evidence doc, and (2) correlates to open `owner='ceo'` tasks for completion candidates ([spec 19](19-tasks-and-backlog.md)). It emits a `ceo.sent_observed` audit event and advances the watermark. Observe runs as a **separate daily cron** on `agents/ceo-inbox.yaml` with its own Operating Modes task text — never folded into the 15-min triage loop.
+- **Sent-mail observation (v0.42, ADR-029):** a scheduled `ceo-inbox-sent-observe` skill polls the CEO's **Sent** folder through the existing `ceo-inbox` Nylas grant (no new credential). Scope is **all sent mail**. Each run is watermarked (`config-store` under the `ceo_inbox` namespace, same high-water pattern as the inbound email poll) and is idempotent on re-run. For each new sent message it (1) correlates to Curia-authored draft snapshots (below) by `thread_id` + recipients + send time and appends `(draft, sent)` pairs to a rolling OKF evidence doc, and (2) correlates to open `owner='ceo'` tasks for completion candidates, queued in `config-store` (`ceo_inbox`, `sent_observe.completion_candidates`) rather than an OKF doc ([spec 19](19-tasks-and-backlog.md)). It emits a `ceo.sent_observed` audit event and advances the watermark. Observe runs as a **separate daily cron** on `agents/ceo-inbox.yaml` with its own Operating Modes task text — never folded into the 15-min triage loop.
 - **Draft capture for voice learning (v0.42):** the three CEO-inbox draft skills (`ceo-inbox-draft-reply`, `ceo-inbox-draft-compose`, `ceo-inbox-draft-edit`) snapshot what Curia proposed into an OKF scratch doc at `/scratch/voice-learning/<draft_id>.md` (frontmatter: `draft_id`, `thread_id`, recipients, subject, `created_at`, `linked_task_ids[]`, `agent_version`; body = drafted markdown). Capture failure logs and **must not** block draft creation. Matched `(draft, sent)` diffs accumulate in `/scratch/voice-learning/pending-diffs.md` for the weekly `voice-learn` job ([spec 13](13-office-identity.md)). That job makes a single batched LLM extraction pass over the accumulated diffs to produce an updated `WritingVoice.guide`, proposed in the digest for CEO approval.
 - **OKF evidence retention:** raw draft snapshots and pending-diff evidence are scratch-path docs. Once a pair has been folded into the writing-voice profile (or dismissed), the learning job deletes or TTL-expires the consumed evidence — default scratch TTL applies to unmatched snapshots; consumed evidence must not be retained indefinitely. Shadow drafts for punted mail follow the same scratch path conventions and never surface or send ([spec 14](14-autonomy-engine.md)).
 - Nylas abstracts away provider differences (Gmail, Outlook, IMAP) and handles OAuth

diff --git a/docs/specs/13-office-identity.md b/docs/specs/13-office-identity.md
@@ -373,11 +373,13 @@ A weekly `voice-learn` job on `ceo-inbox` (cron `0 8 * * 1`) reads accumulated
 (`ctx.infraLlm.extract()`) over that accumulated evidence to produce an updated
 free-form `WritingVoice.guide`.
 
-The job never writes the profile directly. It queues a "## Guide Proposal"
-block (`status: pending`) in the digest for CEO review. The CEO approves via
-the `resolve-learning-digest` skill's `approve_voice` action, which calls
+The job never writes the profile directly. It records a single pending proposal
+object in `config-store` (`voice_learn.proposal`), surfaced in the digest for
+CEO review. The CEO approves via the `resolve-learning-digest` skill's
+`approve_voice` action, which calls
 `executiveProfileService.update({ writingVoice: { ...current, guide } })`.
-Dismissed proposals get a guard marker and cooldown before re-propose.
+Dismissed proposals set the `voice_learn.dismissed` config cooldown key before
+re-propose.
 
 The learned guide renders into agent prompts through `compileWritingVoiceBlock`,
 under a "How the executive actually writes (learned from their edits)" heading.

diff --git a/docs/specs/19-tasks-and-backlog.md b/docs/specs/19-tasks-and-backlog.md
@@ -343,9 +343,9 @@ for these learning surfaces are explicitly deferred.
   | High-confidence match | Auto-`task-complete` with reversible digest undo note | Confirm in digest |
   | Low-confidence match | Confirm in digest | Confirm in digest |
 
-  High-risk tasks (e.g. "Plan AGM") never auto-complete. Fuzzy candidates get a guard marker
-  (`completion_asked: {date}`) so they are not re-surfaced every run. Auto-completes remain
-  reversible via reopen from the digest undo note.
+  High-risk tasks (e.g. "Plan AGM") never auto-complete. Fuzzy candidates are recorded in the
+  `sent_observe.asked_task_ids` guard set (config-store, `ceo_inbox`) so they are not
+  re-surfaced every run. Auto-completes remain reversible via reopen from the digest undo note.
 - **Autonomy** ([spec 14](14-autonomy-engine.md)). The "execute the safe stuff, escalate
   the rest" behavior is the existing autonomy engine — `action_risk` per skill against the
   live score. Drafting (low) just happens; outbound (medium) is gated; spending money