pacphi · pacphi · Sep 27, 2026 · Sep 27, 2026 · Sep 27, 2026 · Sep 27, 2026
diff --git a/docs/DASHBOARD.md b/docs/DASHBOARD.md
@@ -639,7 +639,9 @@ session."
 A session cwd is only a discovery candidate; it does not confer project scope. If that cwd makes a
 candidate surface resolve to the same host, kind, and path as a user surface, the user occurrence
 wins and no project-pressure row is fabricated. One shared user surface carried by two hosts remains
-visible on both hosts.
+visible on both hosts. A Codex session that the ChatGPT desktop app imported from a Claude Code
+transcript is not a candidate at all: it adds no project, host or Desktop origin, and the Projects
+note under the System KPIs says how many were set aside.
 
 The measurement counts a physical artifact once and retains each host that discovers it as a
 separate consumer binding. It fingerprints bounded skill definitions and individual MCP

diff --git a/docs/MAINTENANCE.md b/docs/MAINTENANCE.md
@@ -108,8 +108,9 @@ Project choices group beneath repositories only when shared Git metadata or the
 contract establishes the association. A worktree remains its own selectable project. Folder/name
 headers, a Git/Worktree/Folder designation, and all detected language icons identify each choice.
 An independent **Session origin** facet can match Claude Desktop, Codex Desktop, or unclassified
-observations; Desktop origin never replaces repository membership. Facet counts are installations
-in matching projects, not numbers of sessions. Selecting several origins does not duplicate a
+observations; Desktop origin never replaces repository membership. Codex copies of Claude Code
+sessions imported by the ChatGPT desktop app never count as a Desktop origin. Facet counts are
+installations in matching projects, not numbers of sessions. Selecting several origins does not duplicate a
 placement. Exact root paths stay private; the public grouping uses opaque repository identities
 and bounded display labels.
 

diff --git a/docs/TRANSCRIPTS.md b/docs/TRANSCRIPTS.md
@@ -75,7 +75,7 @@ Each line has a top-level `type`. The parser (`parseClaude`,
 | "ai-title" | The model-written session title (`usage-parsers.mjs:768`) — preferred over the first-prompt fallback |
 | `user` | A user-**role** turn — which is *not* the same as "the human"; see §3. On a turn that passes isHumanPrompt, also its `permissionMode` — the session's permission posture, read on the person's own turn only (`usage-parsers.mjs:593-609`) — and the opening of the response-latency window |
 | `assistant` | One content block of a model message: `model` id, the message's `usage` token counts (repeated on every block's line, so counted once per `message.id`, else `requestId`, last line winning), `tool_use` blocks (`usage-parsers.mjs:661-743`) |
-| any | Side-band fields read regardless of type: `attributionSkill`/`attributionPlugin` (`usage-parsers.mjs:758-759`), `isSidechain` (`usage-parsers.mjs:760-763`), `cwd` for project derivation |
+| any | Side-band fields read regardless of type: `attributionSkill`/`attributionPlugin` (`usage-parsers.mjs:770-771`), `isSidechain` (`usage-parsers.mjs:772-773`), `cwd` for project derivation |
 
 A real assistant completion also closes two pieces of per-entry evidence the
 transcript does not state outright. It **closes the latency window** the
@@ -139,7 +139,10 @@ Three rollout shapes are handled before any of that:
   no `thread_source`, `input_tokens: 0`). The parser stops at the first such line and
   returns an identity-only record marked `imported`; the scan keeps it out of
   Codex sessions, prompts, responses and yield diagnostics and counts it in
-  `importedExcluded`. The real conversation is in the Claude transcript.
+  `importedExcluded`. The real conversation is in the Claude transcript. Project
+  discovery reads the same marker in each rollout's head and leaves the rollout out
+  of projects, hosts and Desktop origins, counting it in the discovery scan's
+  `importedExcluded`.
 * **Forked subagents.** A forked subagent's rollout replays its parent's history
   before its own turns. Every envelope carries an `ordinal`; the replay ends at
   `subagent_history_start_ordinal` when some event lies at or beyond it, else at the

diff --git a/docs/UPGRADING.md b/docs/UPGRADING.md
@@ -39,6 +39,15 @@ and supported `claude mcp serve` tool exposure are preserved. See
 [ADR-0051](adr/0051-supported-peer-delegation-and-host-realignment.md) for the policy,
 official source citations, authority boundaries and verification limits.
 
+## 2026-09-27: System snapshot v8 (imported Codex copies)
+
+When the ChatGPT desktop app imports a Claude Code transcript, it saves a copy as a Codex session.
+Project discovery no longer counts these copies: they give a folder no Codex host and no Desktop
+origin, and System says how many it set aside. A snapshot taken before this change still holds the
+old hosts and origins, so the Footprint snapshot schema advances to v8. This build reports a v7
+snapshot as unreadable until you run **Full scan** in System or `ak system --deep`. It is never
+shown under the new rule. See [ADR-0060](adr/0060-session-surface-initiator-and-product-names.md) §3.
+
 ## 2026-09-26: `ak sync`'s exit code ignores fixes you do by hand
 
 `ak sync` now exits 0 when everything it can repair has converged, even if a row whose fix you

diff --git a/docs/USAGE-SCORECARD-METRICS.md b/docs/USAGE-SCORECARD-METRICS.md
@@ -199,7 +199,7 @@ responses = Σ over included sessions of session.responses
 - `responses` accumulation: Claude increments once per API message id — every
   transcript line of one message counts once, the last line's usage winning
   (`usage-parsers.mjs:661-696`); Codex increments per `agent_message` event
-  (`usage-parsers.mjs:1008-1013`).
+  (`usage-parsers.mjs:1021-1025`).
 - Totals: `totals.responses += s.responses` per included session
 (`usage-aggregate.mjs:928`).
 - Render: `kpi("sessions", fmtNum(t.sessions), fmtNum(t.responses)+" assistant
@@ -836,7 +836,7 @@ byDay[day].sessionsActive = count of distinct sessions with any usage row that d
 
 **Source:** the day key is the row's own `row.day`, computed once at parse
 time as **local calendar day**, not UTC
-(`usage-parsers.mjs:35`/`usage-parsers.mjs:1185` call `localDay(at)`) — so a
+(`usage-parsers.mjs:35`/`usage-parsers.mjs:1174` call `localDay(at)`) — so a
 session that runs from 23:58 local to 00:05 local has its session count attributed
 to the day its *first* usage row landed on (test:
 `tests/kit/usage-index.test.mjs:738`, "a session that opens before midnight
@@ -926,7 +926,7 @@ punchcard[dow + "-" + hour] += 1   per assistant/agent_message response, at its
 **Source:** incremented once per Claude API message (all of a message's
 transcript lines are one hit)
 (`usage-parsers.mjs:42`, keyed by this call: `punchKey(at)`) and once per Codex
-`agent_message` (`usage-parsers.mjs:1008-1013`), merged into the window-level
+`agent_message` (`usage-parsers.mjs:1021-1025`), merged into the window-level
 `punchcard` object per session (`usage-aggregate.mjs:943-1007`). Cell intensity is
 linear against the single busiest cell in the window:
 `v = pcMax ? n/pcMax : 0` (`dashboard/client.mjs`) — this is a

diff --git a/docs/adr/0050-dashboard-project-identity-and-context-reporting.md b/docs/adr/0050-dashboard-project-identity-and-context-reporting.md
@@ -2,9 +2,13 @@
 
 - **Status:** Implemented
 - **Date:** 2026-09-09
-- **Updated:** 2026-09-26 — the `project-concentration` Finding ranks `gitProjects`, not the
-  folder-label `byProject`, so it agrees with Score → Projects (#238 item 7)
-- **Earlier update:** 2026-09-09 — scoped context-card network claims for issue #211
+- **Updated:** 2026-09-27 — a Codex rollout the ChatGPT desktop app imported from a Claude Code
+  transcript gives no session origin, even though it declares `Codex Desktop`
+  ([ADR-0052](0052-codex-usage-attribution.md) §3,
+  [ADR-0060](0060-session-surface-initiator-and-product-names.md) §3)
+- **Earlier updates:** 2026-09-26 — the `project-concentration` Finding ranks `gitProjects`, not
+  the folder-label `byProject`, so it agrees with Score → Projects (#238 item 7); 2026-09-09 —
+  scoped context-card network claims for issue #211
 - **Related:** [ADR-0036](0036-dashboard-client-modularization-and-shared-loopback-server.md),
   [ADR-0048](0048-inventory-led-maintenance-resource-management.md)
 
@@ -35,6 +39,9 @@ Session origin comes from bounded transcript metadata, separate from the existin
 - Codex `session_meta.originator`: `Codex Desktop`, `codex_work_desktop`.
 - Everything else, including ambiguous SDK and VS Code markers: unknown.
 
+A Codex rollout whose turns carry the `external-import-turn` marker is a copy of a Claude Code
+transcript, not a Codex session. It supplies no origin, whatever its `originator` says.
+
 These are source declarations, not attestation of the initiating application.
 Claude's installed 2.1.266 runtime maps those entrypoints to Claude Desktop;
 Codex values were observed in local bounded session metadata. No prompt content,

diff --git a/docs/adr/0052-codex-usage-attribution.md b/docs/adr/0052-codex-usage-attribution.md
@@ -2,6 +2,9 @@
 
 - **Status:** Accepted
 - **Date:** 2026-09-19
+- **Updated:** 2026-09-27 — imported copies are also excluded from project discovery: they give no
+  project, host or Desktop origin, and the discovery scan counts them in `importedExcluded`
+  (ADR-0060 §3). The marker now lives in one leaf module shared by usage and discovery.
 - **Deciders:** agentic-kit maintainers
 - **Related:** [ADR-0009](0009-usage-scorecard-local-transcript-analytics.md),
   [ADR-0038](0038-consistent-cross-host-session-metrics.md),
@@ -117,6 +120,11 @@ out of aggregation, out of every yield statistic, and counted in
 `diagnostics.importedExcluded` (796 on the reference machine). Nothing is dropped
 silently. The record itself is still cached, so a rescan is cheap.
 
+Project discovery applies the same marker to each rollout's bounded head (256 KiB, 40 lines): an
+imported copy names no project, host or Desktop origin, and the scan reports how many it set aside
+(`importedExcluded`, 924 on the reference machine on 2026-09-27, every marker on the rollout's
+second line).
+
 ### 4. Cumulative counter restarts are summed, per event
 
 `total_token_usage` restarts from zero mid-file: 48 restarts across 33 files
@@ -196,6 +204,15 @@ subagent and previously dropped usage is now priced.
 - A subagent with no ordinals still reports no usage (decision 2).
 - One rollout carries `token_count`s but no agent message, so the pre-existing
   `partial-response-yield` warning remains.
+- Whole-rollout exclusion may drop real usage (open, plausible, 2026-09-27). On the reference
+  machine 6 of 924 imported rollouts carry a later turn that is not an import: one `task_started`
+  whose `turn_id` starts with `rollout-`, no `user_message` event, `role: user` response items in
+  five of the six (2 to 76 per file) and non-zero `token_count` usage (the per-file sum of
+  `last_token_usage.total_tokens` is about 8k to 449k). Both usage and discovery set the whole file
+  aside at the marker, so this usage is not counted. With no `user_message`, the turn may be
+  automatic (a compaction or title pass). Measured from counts only. Decided 2026-09-27 (audit
+  decision 12): Branch 8 excludes per turn instead of per file, so imported turns are never counted
+  and later turns are, after it establishes whether they are the user's work or an automatic pass.
 
 ## Verification
 

diff --git a/docs/adr/0060-session-surface-initiator-and-product-names.md b/docs/adr/0060-session-surface-initiator-and-product-names.md
@@ -1,7 +1,10 @@
 # ADR-0060 — Session surface, initiator and official product names
 
-- **Status:** Proposed (staged follow-on; nothing implemented)
+- **Status:** Proposed; §3 implemented for project discovery (2026-09-27), the rest staged follow-on
 - **Date:** 2026-09-26
+- **Updated:** 2026-09-27 — §3 implemented for project discovery and the System projects note:
+  imported copies give no project, host or origin and are counted. The ledger-derived source labels
+  (Cursor, Cowork) and the other views remain proposed.
 - **Deciders:** agentic-kit maintainers
 - **Related:** [ADR-0050](0050-dashboard-project-identity-and-context-reporting.md) (session origin
   rule, superseded in part by this record once accepted),
@@ -160,8 +163,9 @@ in full.
 ## Consequences
 
 - Usage, System → Projects, Maintenance facets, Intelligence designation (which today mixes Git
-  scope, origin and host in one enum) and the Runtime table change labels and counts. Twenty-three
-  project folders lose a false Desktop origin on this machine.
+  scope, origin and host in one enum) and the Runtime table change labels and counts. On this
+  machine 32 project folders lost a false Desktop origin when discovery began setting imports aside
+  (re-measured 2026-09-27; 23 on 2026-09-26).
 - The usage cache schema changes (new session fields); a rebuild is expected.
 - Tests that pin current names change together (inventory in the audit record, Addendum 3).
 - `CLAUDE_CODE_ENTRYPOINT` and the transcript format are internal to Claude Code and may change;
@@ -173,6 +177,10 @@ in full.
 - Whether "Cloud session" should appear at all in local views, given none was observed locally.
 - Whether the "on 3P" attribute is worth showing.
 - How ADR-0057's role lenses consume surface and initiator.
+- Whether a later turn inside an imported copy that is not itself an import (6 of 924 rollouts on
+  2026-09-27, with real token usage) counts as the importing app's own session. Decided 2026-09-27
+  (audit decision 12): it counts, excluded per turn in Branch 8
+  ([ADR-0052](0052-codex-usage-attribution.md), "Not done").
 
 ## Verification (when implemented)
 
@@ -182,4 +190,17 @@ fixture.
 
 ## Implementation status
 
-Not started. Staged as follow-on work in the audit record's Addendum 3.
+§3 is implemented for project discovery and the System projects note (2026-09-27): an imported copy
+gives no project, host or origin, and discovery counts it in `importedExcluded`. The per-source
+labels from the imports ledger, Runtime attribution and §1, §2 and §4–§6 remain follow-on work
+(the audit record's Addendum 3).
+
+Three views already show the smaller project counts but do not yet say how many imported copies were
+set aside; §3's "each view reports how many it excluded" is still owed for them:
+
+- the Intelligence census line (`src/lib/dashboard/client/intelligence.mjs`, which prints
+  `everSeen`; the server's `readCensus` in `src/lib/dashboard-server.mjs` drops `importedExcluded`);
+- the System → Projects liner (`sysProjectsLinerHtml` in
+  `src/lib/dashboard/client/system-projects.mjs`);
+- the `ak system` text output (`renderProjects` in `src/commands/system.mjs`, which prints only the
+  count; `ak system --json` carries `importedExcluded`).
diff --git a/docs/adr/README.md b/docs/adr/README.md
@@ -391,7 +391,8 @@ component out.
 
 ## ADR-0060 — Session surface, initiator and official product names
 
-[ADR-0060](0060-session-surface-initiator-and-product-names.md) (Proposed) derives a session's
-surface and initiator from the hosts' declared log fields, keeps every raw value, uses official
-product names (Claude Desktop, ChatGPT desktop app, Codex CLI, and others), and excludes imported
-session copies from every origin view.
+[ADR-0060](0060-session-surface-initiator-and-product-names.md) (Proposed; §3 implemented for
+project discovery) derives a session's surface and initiator from the hosts' declared log fields,
+keeps every raw value, uses official product names (Claude Desktop, ChatGPT desktop app, Codex CLI,
+and others), and excludes imported session copies from every origin view. Project discovery already
+sets imported copies aside and counts them; the other decisions remain proposed.