docs(close-session): refresh classification gate post-#67 (#62)

The capture-routing in 723dab5 (#62 minimal) carried pre-#67 prose: it
warned "no populated classification.yaml yet" and steered repos_touched
away from brain/ai-sessions as if they'd escalate to confidential. #67
tagged the homelab repos internal, so that guidance is now wrong and
over-restrictive — listing the central repos is fine, and the summary's
fixed ai-sessions target no longer escalates the call. Point at
classification.yaml as the source of truth; reserve the refusal warning
for client-*/untagged material. Completes #62's veneer.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

ingestion/cmd/server/main.go

@@ -410,8 +410,12 @@ func main() {
 			os.Exit(1)
 		}
 		auditSink := buildAuditSink(ctx, brainDir, logger)
+		// The Gitea client also satisfies SummaryWriter (#66): session
+		// summaries are written to mathias/ai-sessions over the same API
+		// token. nil only if a future tracker impl lacks file writes.
+		summaryWriter, _ := tracker.(capture.SummaryWriter)
 		captureSvc := capture.NewService(
-			mcpSrv.BrainStore(), tracker, nil, classCfg, auditSink)
+			mcpSrv.BrainStore(), tracker, summaryWriter, classCfg, auditSink)
 		sovereign := splitList(os.Getenv("BRAIN_CAPTURE_SOVEREIGN_PRINCIPALS"))
 		resolver := capturehttp.NewOriginResolver(sovereign)
 		captureH := capturehttp.New(captureSvc, jwtValidator, mcpToken, "local-cli", resolver)

ingestion/internal/gitea/gitea.go

@@ -12,6 +12,7 @@ package gitea
 import (
 	"bytes"
 	"context"
+	"encoding/base64"
 	"encoding/json"
 	"fmt"
 	"io"
@@ -93,37 +94,105 @@ func (c *Client) CloseIssue(ctx context.Context, repo string, number int, commen
 	return capture.IssueRef{Repo: repo, Number: number, URL: out.HTMLURL}, nil
 }
 
-// do performs a JSON request against the Gitea API and decodes the
+// WriteFile creates or updates a file in repo at path via the Gitea
-// response into out. Errors carry the status and a truncated body for
+// contents API — the SummaryWriter port (#66). It upserts: a GET resolves
-// diagnosis but never the token.
+// the current blob sha (if any) so an existing file is updated rather than
-func (c *Client) do(ctx context.Context, method, path string, payload any, out *issueResponse) error {
+// rejected (the richer-fidelity-supersedes rule for re-captured sessions).
-	reqBody, err := json.Marshal(payload)
+// Owner is the fixed const, like every other call.
-	if err != nil {
+func (c *Client) WriteFile(ctx context.Context, repo, path, content string) error {
-		return fmt.Errorf("marshal request: %w", err)
+	cpath := fmt.Sprintf("/api/v1/repos/%s/%s/contents/%s", owner, repo, path)
-	}
+	sha, err := c.fileSHA(ctx, cpath)
-	req, err := http.NewRequestWithContext(ctx, method, c.baseURL+path, bytes.NewReader(reqBody))
 	if err != nil {
 		return err
 	}
-	req.Header.Set("Content-Type", "application/json")
+	payload := map[string]any{
-	req.Header.Set("Accept", "application/json")
+		"message": "capture: " + path,
-	// Gitea's token scheme. Held here only; never logged.
+		"content": base64.StdEncoding.EncodeToString([]byte(content)),
-	req.Header.Set("Authorization", "token "+c.token)
+	}
-
+	// Gitea contents API: POST creates a new file, PUT updates an existing
-	resp, err := c.http.Do(req)
+	// one (PUT requires the current sha). Pick by whether the file exists.
+	method := http.MethodPost
+	if sha != "" {
+		method = http.MethodPut
+		payload["sha"] = sha
+	}
+	status, body, err := c.request(ctx, method, cpath, payload)
 	if err != nil {
-		return fmt.Errorf("gitea %s %s: %w", method, path, err)
+		return err
 	}
-	defer func() { _ = resp.Body.Close() }()
+	if status < 200 || status >= 300 {
+		return fmt.Errorf("gitea %s %s: status %d: %s", method, cpath, status, strings.TrimSpace(string(body)))
+	}
+	return nil
+}
 
-	respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 4096))
+// fileSHA returns the current blob sha for a contents path, or "" when the
-	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+// file does not exist (404). Any other non-2xx is an error.
-		return fmt.Errorf("gitea %s %s: status %d: %s", method, path, resp.StatusCode, strings.TrimSpace(string(respBody)))
+func (c *Client) fileSHA(ctx context.Context, cpath string) (string, error) {
+	status, body, err := c.request(ctx, http.MethodGet, cpath, nil)
+	if err != nil {
+		return "", err
 	}
-	if out != nil && len(respBody) > 0 {
+	if status == http.StatusNotFound {
-		if err := json.Unmarshal(respBody, out); err != nil {
+		return "", nil
+	}
+	if status < 200 || status >= 300 {
+		return "", fmt.Errorf("gitea GET %s: status %d: %s", cpath, status, strings.TrimSpace(string(body)))
+	}
+	var meta struct {
+		SHA string `json:"sha"`
+	}
+	if err := json.Unmarshal(body, &meta); err != nil {
+		return "", fmt.Errorf("gitea GET %s: decode: %w", cpath, err)
+	}
+	return meta.SHA, nil
+}
+
+// do performs a JSON request against the Gitea API and decodes a 2xx
+// response into out. Errors carry the status and a truncated body for
+// diagnosis but never the token.
+func (c *Client) do(ctx context.Context, method, path string, payload any, out *issueResponse) error {
+	status, body, err := c.request(ctx, method, path, payload)
+	if err != nil {
+		return err
+	}
+	if status < 200 || status >= 300 {
+		return fmt.Errorf("gitea %s %s: status %d: %s", method, path, status, strings.TrimSpace(string(body)))
+	}
+	if out != nil && len(body) > 0 {
+		if err := json.Unmarshal(body, out); err != nil {
 			return fmt.Errorf("gitea %s %s: decode response: %w", method, path, err)
 		}
 	}
 	return nil
 }
+
+// request is the shared HTTP path: marshals an optional JSON payload,
+// attaches auth (token only ever in the header), and returns the status +
+// body so callers can branch on status (e.g. 404) without it being an
+// error. Never logs the token.
+func (c *Client) request(ctx context.Context, method, path string, payload any) (int, []byte, error) {
+	var reader io.Reader
+	if payload != nil {
+		reqBody, err := json.Marshal(payload)
+		if err != nil {
+			return 0, nil, fmt.Errorf("marshal request: %w", err)
+		}
+		reader = bytes.NewReader(reqBody)
+	}
+	req, err := http.NewRequestWithContext(ctx, method, c.baseURL+path, reader)
+	if err != nil {
+		return 0, nil, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Accept", "application/json")
+	req.Header.Set("Authorization", "token "+c.token)
+
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return 0, nil, fmt.Errorf("gitea %s %s: %w", method, path, err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+	body, _ := io.ReadAll(io.LimitReader(resp.Body, 8192))
+	return resp.StatusCode, body, nil
+}

ingestion/internal/gitea/gitea_test.go

@@ -115,3 +115,69 @@ func TestErrorPathDoesNotLeakToken(t *testing.T) {
 	assert.NotContains(t, err.Error(), testToken, "token must never appear in an error message")
 	assert.Contains(t, err.Error(), "500")
 }
+
+func TestWriteFileCreatesNewFile(t *testing.T) {
+	var getPath, postPath, postBody string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.Method {
+		case http.MethodGet:
+			getPath = r.URL.Path
+			w.WriteHeader(http.StatusNotFound) // file does not exist yet
+		case http.MethodPost: // gitea contents API: POST = create
+			postPath = r.URL.Path
+			b, _ := io.ReadAll(r.Body)
+			postBody = string(b)
+			w.WriteHeader(http.StatusCreated)
+			_ = json.NewEncoder(w).Encode(map[string]any{"content": map[string]any{"html_url": "https://git/x"}})
+		default:
+			t.Errorf("create must POST, got %s", r.Method)
+		}
+	}))
+	defer srv.Close()
+
+	err := gitea.New(srv.URL, testToken).WriteFile(context.Background(),
+		"ai-sessions", "summaries/claude-code/2026-06/2026-06-23-x-abcd1234.md", "# Summary\n\nbody\n")
+	require.NoError(t, err)
+	assert.Equal(t, "/api/v1/repos/mathias/ai-sessions/contents/summaries/claude-code/2026-06/2026-06-23-x-abcd1234.md", getPath)
+	assert.Equal(t, getPath, postPath)
+	// base64 of the content, no sha on create.
+	assert.Contains(t, postBody, "IyBTdW1tYXJ5") // base64("# Summary")
+	assert.NotContains(t, postBody, `"sha"`)
+}
+
+func TestWriteFileUpdatesExisting(t *testing.T) {
+	var putBody string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.Method {
+		case http.MethodGet:
+			w.WriteHeader(http.StatusOK)
+			_ = json.NewEncoder(w).Encode(map[string]any{"sha": "deadbeef"})
+		case http.MethodPut:
+			b, _ := io.ReadAll(r.Body)
+			putBody = string(b)
+			w.WriteHeader(http.StatusOK)
+			_ = json.NewEncoder(w).Encode(map[string]any{"content": map[string]any{"html_url": "https://git/x"}})
+		}
+	}))
+	defer srv.Close()
+
+	err := gitea.New(srv.URL, testToken).WriteFile(context.Background(), "ai-sessions", "p/x.md", "new")
+	require.NoError(t, err)
+	assert.Contains(t, putBody, `"sha":"deadbeef"`, "existing file → update with sha")
+}
+
+func TestWriteFileErrorNoTokenLeak(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Method == http.MethodGet {
+			w.WriteHeader(http.StatusNotFound)
+			return
+		}
+		w.WriteHeader(http.StatusUnprocessableEntity)
+		_, _ = w.Write([]byte("bad"))
+	}))
+	defer srv.Close()
+	err := gitea.New(srv.URL, testToken).WriteFile(context.Background(), "ai-sessions", "p/x.md", "x")
+	require.Error(t, err)
+	assert.NotContains(t, err.Error(), testToken)
+	assert.Contains(t, err.Error(), "422")
+}

skills/close-session/SKILL.md

@@ -7,14 +7,14 @@ description: Disciplined end-of-session closeout for a Claude.ai chat before arc
 
 Capture a finishing Claude.ai work session into durable storage before the chat is archived and its context is lost. The goal is simple and load-bearing: **after this runs, a fresh session (or another agent) can reconstruct what was decided, what was shipped, and what is still open — without the original chat.**
 
-This skill is **batch**: one session in, findings out, done. It does not loop or re-read its own fresh output semantically (see Phase 5). Run the phases in order. Stop at any confirmation gate that says STOP.
+This skill is **batch**: one session in, findings out, done. Run the phases in order. Stop at any confirmation gate that says STOP.
 
 ## Operating constraints (read first)
 
 - **Gitea owner is always `mathias`.** Never guess another owner.
 - **Ground-truth at HEAD before acting.** Issue bodies and doc references rot — stale hostnames, retired services, moved endpoints. Before closing/commenting on any issue, `gitea:issue_get` it fresh. Before asserting an infra fact, verify it; do not copy it from memory or from a stale issue body.
 - **Current infra truths** (verify rather than trust, but these are the known-good baseline): Gitea is `git.d-ma.be` (not `gitea.d-ma.be`). LiteLLM is `http://koala:30401/v1/` (public `https://llm-api.d-ma.be`); piguard runs NGINX Proxy Manager only — never reference `piguard:4000` or `koala:4000`. Identity provider is Authentik (Dex migration complete).
-- **Side-effects need a confirmation gate.** Closing issues, committing files, and writing to the brain are all real writes. Surface exactly what will happen and get a clear yes before doing it. Reads are free; writes are gated.
+- **Side-effects need a confirmation gate.** Closing issues and capturing to brain/Gitea/ai-sessions are real writes. Surface exactly what will happen and get a clear yes before doing it. Reads are free; writes are gated.
 - **Never fabricate.** If the session didn't produce a decision worth persisting, say so and skip that write. An empty-but-honest closeout beats an invented one.
 
 ## Phase 1 — Harvest
@@ -34,76 +34,46 @@ For every repo touched this session, get its true current state before proposing
 
 Do not write anything in this phase. This is the read pass.
 
-## Phase 3 — Confirm and act on issue changes
+## Phase 3 — Plan the issue changes
 
-Present a single consolidated plan of issue actions: which to close (with closing comment), which to file (discovered-but-deferred work — token-budget gaps, recorded limitations, v2 follow-ups), which to comment on. Include the exact title/body for any new issue and the closing rationale for any close.
+Decide the issue actions: which to close (with closing comment), which to file (discovered-but-deferred work — token-budget gaps, recorded limitations, v2 follow-ups), which to comment on. Include the exact title/body for any new issue and the closing rationale for any close.
 
-**GATE — STOP and get explicit confirmation before any issue write.** Issue closes and new issues are side-effects. Once confirmed, execute them (`gitea:issue_close`, `gitea:issue_create`, `gitea:issue_comment`, all owner `mathias`), correcting any rotted references you found in Phase 2 as you go.
+These actions are **carried into the Phase 4 capture call** as `tickets[]` rather than executed here with direct `gitea:issue_*` calls — routing them through capture puts each one into the I5 audit record. (Closing an issue that needs a separate explanatory comment first is the one case to do directly; otherwise prefer the capture path.)
 
-## Phase 4 — Commit the canonical session summary
+## Phase 4 — Capture (one uniform call)
 
-Write one summary file to `mathias/ai-sessions`, committed directly to `main` via `gitea:file_write_branch` (no PR — this repo is solo and unprotected; if branch protection is ever added, fall back to a branch + PR).
+Persist the session via a **single `capture` call** (the `brain:capture` MCP tool, live on the Claude.ai connector). Capture owns the writes server-side — insights → brain, action items → Gitea tickets, summary → ai-sessions — plus the I1 sovereignty gate, the I5 audit record, and the supersession/read-after-write discipline. The skill's job is to *assemble the payload*, not to write each store itself. Do NOT fall back to separate `gitea:file_write_branch` + `brain_write` steps unless `capture` is unreachable (see fallback below).
 
-**Path:** `summaries/claudeai/<YYYY-MM>/<YYYY-MM-DD>-<topic-slug>-<chatid8>.md`
+**Assemble one payload:**
-where `<chatid8>` is the first 8 chars of the chat's UUID if known, else a short stable slug. `claudeai` has no host segment — Claude.ai is Anthropic-side, not a homelab host.
 
-**Frontmatter — the REDUCED live-capture schema.** A live close-session capture cannot populate the batch-export telemetry (token counts, message counts, duration_ms, permission_mode) — those only exist in the account export pipeline. Write only what's truthfully known, and mark fidelity so a reader (or the batch pipeline) can tell a live capture from an export:
+- **`insights[]`** — the generalizable learnings from Phase 1 (decisions/failures worth re-reading). Each: `{text, wing, hall}`; add `supersede_slug` to revise a prior note in place instead of creating a duplicate. `hall` ∈ facts/decisions/failures/hypotheses/sources.
+- **`tickets[]`** — the issue actions from Phase 3: `{repo, action, ...}` where action ∈ create/close/comment. Owner is always `mathias` (server-forced).
+- **`summary`** — `{title, body, repos_touched}`. Capture writes it to `ai-sessions` and stamps `fidelity` in frontmatter. Body stays reconstructable: one-paragraph summary, decisions, key artifacts, open threads.
+- **`context`** — `{harness: "claudeai-chat", session_ref: <chatid8-or-slug>, fidelity: "live-capture", actor: "mathias", classification: <see gate below>}`.
 
-```yaml
+**THE CLASSIFICATION GATE (read before calling — this is where capture refuses).**
----
+Capture computes an **effective classification = the strictest across EVERY target it touches** (each insight's `wing`, each ticket's `repo`, and every entry in `summary.repos_touched`), then refuses if that effective level is `confidential` and the origin is us-nexus (claude.ai is us-nexus). Levels come from `classification.yaml` at the brain root (source of truth, #67), with the code defaults as the floor: `hyperguild`/`homelab` → internal; `client-*` → confidential; **anything untagged → confidential (fail-safe)**.
-title: "<concise session title>"
+- **Tagged `internal` today** (safe through claude.ai): wings `hyperguild`, `homelab`; repos `brain`, `ai-sessions`, `infra`, `hyperguild`, `homelab`, `tapir`, `agentsquad`, `jepa-fx-risk`, `swedsl`. Treat `classification.yaml` as authoritative — this list is a hint, not gospel.
-client: "claudeai"
+- Declare `context.classification: "internal"` for normal homelab work.
-interface: "claudeai-chat"
+- `summary.repos_touched`, insight `wing`s, and ticket `repo`s are classification INPUTS, not free-form metadata — every target must resolve `internal` or the whole capture escalates to `confidential` and the gate refuses via claude.ai. Listing the central homelab repos (incl. `brain`/`ai-sessions`) is now fine; they're tagged. The summary always lands in `ai-sessions` (internal), so the summary path itself never escalates.
-date: "<YYYY-MM-DD>"
+- If a session genuinely touched **`client-*` or otherwise-untagged** material, it cannot be captured through claude.ai — note that in the verdict rather than trying to force it.
-repos_touched: [<repo slugs>]
-topic_tags: [<tags>]
-outcome: "<shipped|in-progress|abandoned>"
-fidelity: "live-capture"   # NOT an export; reconstructed live from chat
-captured_by: "close-session-skill"
----
-```
 
-Do not invent the export-only fields. `fidelity: live-capture` is the honest signal; if the batch export later produces a richer summary for the same session, the export is source of truth and supersedes this.
+**GATE — dry-run first, then execute.**
+1. Call `capture` with `dry_run: true`. It validates the whole payload and returns the would-be receipt + `effective_classification`, writing nothing.
+2. **STOP. Show the dry-run receipt** (effective classification, the insights/tickets/summary that would land) and get explicit confirmation.
+3. On confirmation, call `capture` again with `dry_run: false`. Read the returned receipt: it is partial-aware (`errors[]`, per-item `ok`). Report exactly what landed.
 
-**Body** (keep it reconstructable, not exhaustive):
+If `capture` is **unreachable** (tool not on the connector — e.g. a session that started before a deploy; a tool-list refresh usually fixes it): say so. Only then fall back to the legacy inline path (`gitea:file_write_branch` summary + `brain_write`/`brain_update` + `brain_get` confirm), and note in the verdict that the I5 audit record was NOT produced.
-```markdown
-## One-paragraph summary
-## Decisions
-## Key artifacts
-## Open threads
-```
 
-**GATE — STOP, show the full file (path + frontmatter + body), get explicit confirmation before committing.**
+## Phase 5 — Verdict
-
-## Phase 5 — Brain orientation note (the durable "where we are" record)
-
-Write one brain note so a fresh session can orient without the chat. This uses the `brain_update`/`brain_get` verbs (live since 2026-06).
-
-**Target:** `wing: <domain>` (the project/topic domain, e.g. `hyperguild`, `jepa-fx`), `hall: decisions`. The note is a knowledge-type record (a decision/orientation), grouped by knowledge-type, not by interface surface.
-
-**Batch read-after-write discipline (important — do these in order, do not interleave):**
-
-1. **Read first, before any write.** Check whether an orientation note already exists for this wing/topic. Do your "does this already exist / what should I supersede" reads NOW, up front. BM25/keyword search and `brain_get` are immediate; semantic/vector search may lag up to ~5 min after a write, so never rely on a semantic query to find something you wrote earlier in this same run.
-2. **Write or supersede:**
-   - **New note** → `brain_write` (wing, hall: decisions). Returns `{id, path, content_hash}`.
-   - **Superseding a prior orientation note** → `brain_update` (slug or path, wing, hall, content, reason). Whole-note replace; stamps `supersedes`/`updated_at`; returns `{id, path, content_hash, superseded}`. Use this instead of a second `brain_write` to the same slug — blind re-write creates duplicates/contradictions, which is the exact failure brain_update exists to prevent.
-3. **Confirm it landed** via `brain_get(id)` and check the returned `content_hash` matches what the write returned. This is the read-after-write confirmation — do it with `brain_get`, never a semantic query.
-
-**RULE: no semantic/vector brain query after the first `brain_update` in this run.** The batch shape makes this natural — read up front, write, confirm by id. If you ever find the skill wanting to semantic-search a just-superseded note, stop and flag it (that's the signal the staleness window matters and needs the synchronous-reembed follow-up).
-
-**GATE — STOP, show the note (target wing/hall, new-vs-supersede, full content), get explicit confirmation before the brain write.**
-
-After the note lands, if it relates to a note in another wing, create the cross-link inline with `brain_tunnel(source, target)` (idempotent; both paths brain-relative, must be in different wings). Optionally append a `session_log` entry (`session_id`, `skill: close-session`, `phase`, `final_status`) for telemetry. Both are now callable directly from Claude.ai — no Claude Code/Crush handoff needed.
-
-## Phase 6 — Verdict
 
 Deliver a final "safe to archive" verdict in the chat. Either:
 
-- **SAFE TO ARCHIVE** — list what landed (issues closed/filed with numbers, summary path, brain note id, any tunnels) so the trail is auditable. Then list anything still in the user's queue (e.g. a PR awaiting their merge, a decision owed next session).
+- **SAFE TO ARCHIVE** — list what landed from the capture receipt (issues closed/filed with numbers, summary path, brain note ids/paths) so the trail is auditable. Then list anything still in the user's queue (e.g. a PR awaiting their merge, a decision owed next session).
-- **NOT YET** — name the specific gate that wasn't passed or the write that failed, and what to do about it.
+- **NOT YET** — name the specific gate that wasn't passed, the capture refusal reason, or the per-item error from the receipt, and what to do about it.
 
-Never claim safe-to-archive if any gated write was declined or errored. The verdict is the skill's contract: if it says safe, the session can be lost without losing the work.
+Never claim safe-to-archive if the capture refused, any receipt item errored, or a gated confirmation was declined. The verdict is the skill's contract: if it says safe, the session can be lost without losing the work.
 
-## Why the gates and the batch discipline matter
+## Why the gate and the single-call shape matter
 
-The whole point is durability across a context reset. Every gate is a place where a wrong write would silently corrupt the record (close the wrong issue, overwrite a good brain note, commit a half-truth). The batch read-discipline in Phase 5 exists because the brain's vector index refreshes out-of-band: write-then-semantically-reread in the same run can read stale, so the skill front-loads reads and confirms writes by id. Get those right and the skill does what it promises — nothing important is lost when the chat goes away.
+The whole point is durability across a context reset. The capture call is the one place a wrong payload would silently corrupt the record (close the wrong issue, escalate to a refusal, commit a half-truth), which is why it is dry-run-then-confirm. Routing everything through one `capture` keeps the supersession discipline, the read-after-write confirmation, and the I5 audit trail server-side — the skill never has to carry those rules itself, and every closeout is uniformly audited. Get the payload and the classification right and the skill does what it promises — nothing important is lost when the chat goes away.

specs/capture-implementation-report.md

@@ -0,0 +1,116 @@
+# Capture capability — implementation report (as-built)
+
+**Status:** Shipped 2026-06-23, tagged `v0.11.0`. Epic hyperguild #49 (sub-issues #50–#55) closed.
+**Spec:** `specs/capture-bdd-spec.md` (the design contract this implements).
+**Governed by:** `infra/docs/architecture/01-invariants.md` (I1–I5) + the I2 acceptance ledger entry in `infra/docs/security-baseline.md`.
+
+This document records what was actually built, where it lives, how it maps to the spec, and what was deferred — for onboarding and future audit. It does not restate the design rationale (see the spec and the linked brain entries).
+
+---
+
+## 1. Outcome
+
+One uniform capture capability — insights → brain, action items → Gitea tickets, optional summary → ai-sessions — reachable identically from every harness:
+
+- **In-process / direct-REST harnesses** (Claude Code CLI, Agentsquad, claude.ai Code, headless): `POST /capture` on the brain server.
+- **MCP-native harnesses** (claude.ai Chat/Cowork/Design, Crush, Pi, LLM Council): the `capture` MCP tool, reached over the existing `/mcp` OAuth connector.
+
+Both doors call the **same** `CaptureService`; only the transport and credential assembly differ. The persistence behaviour (validation, classification, I1 gate, orchestration, I5 audit, partial receipt) is written once.
+
+---
+
+## 2. Architecture (as-built)
+
+```
+                 POST /capture (REST)          capture MCP tool
+                 capturehttp.Handler           mcp.Server.brainCapture
+                          \                          /
+                           \   (auth → principal →  /
+                            \   origin; decode)    /
+                             v                    v
+                        capture.CaptureService (use-case, pure)
+            ┌───────────────┬───────────────┬──────────────┬───────────────┐
+        BrainStore     IssueTracker    SummaryWriter   ClassificationPolicy  AuditSink
+        brainstore.    gitea.Client    (nil today)     classification.Config  audit.Degrading
+        Store          (REST)                                                 Sink / SlogSink
+        │                                                                     │
+        api.WriteNote/UpdateNote/ReadNote (#45)                       LokiCentral + FileBuffer
+        + wing index + auto-tunnel + graph re-index                  + NtfyNotifier + Reconcile
+```
+
+- **`internal/capture/`** — the use-case + ports + entities. Pure; no I/O. Owns validation (fail-closed), effective-classification resolution (stricter wins), the **I1 sovereignty gate**, best-effort orchestration, the **two-phase I5 audit** (Reserve before writes / Record after), and the partial-aware receipt.
+- **`internal/brainstore/`** — concrete `BrainStore` wrapping the #45 `api` primitives + wiki upkeep (wing `_index`, auto-tunnel, graph re-index). The MCP `brain_write`/`brain_update`/`brain_get` handlers were re-pointed at it: one implementation, not two.
+- **`internal/classification/`** — `public < internal < confidential` taxonomy + per-wing/repo tags from an optional `classification.yaml`; fail-safe to confidential.
+- **`internal/gitea/`** — `IssueTracker` over the Gitea REST API; owner forced to `mathias`; token only in the Authorization header.
+- **`internal/capturehttp/`** — the REST adapter + the shared `Authenticate` / `DecodeRequest` / `OriginResolver` (also used by the MCP tool).
+- **`internal/audit/`** — `SlogSink` (default) and the `DegradingSink` (loki + durable `FileBuffer` + `NtfyNotifier` + `Reconcile`).
+- **`internal/mcp/`** — the `capture` relay tool + principal threading (re-derives the caller's principal from the Bearer header the chassis middleware discards).
+
+---
+
+## 3. Sub-issue → PR map
+
+| Sub | Issue | PR(s) | Delivered |
+|-----|-------|-------|-----------|
+| 49a | #50 | #56 | classification taxonomy + per-wing/repo tags (fail-safe to confidential) |
+| 49b | #51 | #57 | `CaptureService` use-case + ports + entities; `BrainStore` extraction (MCP re-pointed) |
+| 49c | #52 | #58 | Gitea `IssueTracker` (owner forced mathias; token never logged) |
+| 49d | #53 | #59 | `POST /capture` REST + OAuth2 + I1 sovereignty gate (server-derived origin) |
+| 49e | #54 | #60 | I5 audit path + classification-aware degradation (loki + buffer + reconcile) |
+| 49f | #55 | #61, infra #151 (ledger), #152 (deploy) | MCP `capture` relay tool + I2 ledger + I3 deploy |
+
+Predecessor: #45 (`brain_update`/`brain_get` verbs, PR #46) — the read-after-write contract capture reuses.
+
+---
+
+## 4. Invariant compliance
+
+| Inv | How satisfied |
+|-----|---------------|
+| **I1** sovereign containment | Effective classification = stricter(caller-declared, target-derived #50). Origin is **server-derived from the authenticated principal**, never `context.harness`. Confidential + us-nexus origin → refused before any write; refusal audited. Asserted-vs-derived mismatch → security event. |
+| **I2** deliberate acceptance | The relay's cross-harness reach is recorded in `infra/docs/security-baseline.md` with six containment properties + Revisit-if, **merged before relay code shipped** (infra #151). |
+| **I3** GitOps reconcilability | Env + `gitea-api-token` ExternalSecret under `infra/k3s/apps/supervisor/`, Flux-reconciled; image bumped by CD. No untracked runtime. |
+| **I5** auditability | Every capture emits a request-level audit record. Classification-aware degradation: confidential + sink-down → hard-refuse; internal/public + sink-down → durable local buffer + ntfy + reconcile-on-recovery; floor → refuse if nothing can record. |
+
+---
+
+## 5. Operational reference (env)
+
+Set on the `ingestion` deployment (`infra/k3s/apps/supervisor/ingestion-deployment.yaml`):
+
+| Env | Purpose | Notes |
+|-----|---------|-------|
+| `BRAIN_GITEA_URL` / `BRAIN_GITEA_TOKEN` | enables the `IssueTracker` → gates `/capture` + the MCP tool | token from 1P `DMABE_GITEA_API_TOKEN` via ESO; unset ⇒ capture disabled |
+| `BRAIN_LOKI_URL` | activates the `DegradingSink` | unset ⇒ `SlogSink` (audit to stdout → alloy → loki; no refuse/buffer semantics) |
+| `BRAIN_NTFY_URL` / `BRAIN_NTFY_TOKEN` | degraded-state alerts | optional |
+| `BRAIN_CAPTURE_SOVEREIGN_PRINCIPALS` | JWT subjects treated as sovereign-soil | comma-separated; static-token caller is always sovereign; unknown JWT ⇒ us-nexus (fail safe) |
+| `BRAIN_AUDIT_RECONCILE_INTERVAL` | buffer→loki replay tick | default 60s |
+
+The audit buffer lives at `<brain>/.audit-buffer/capture.jsonl` on the brain hostPath (nodeSelector-pinned to koala) — durable across restart without a separate PV.
+
+---
+
+## 6. Tests
+
+67 test functions across the six packages. Coverage maps to the spec's Gherkin: happy path, supersede-not-duplicate, fail-closed validation, partial-failure receipt, dry-run, stricter-classification-wins, I1 confidential-via-us-nexus-refused / via-sovereign-allowed / asserted-label-ignored / caller-cannot-forge-origin, I5 confidential-refuse / internal-buffer / floor-refuse / reconcile / buffer-survives-restart, and the MCP relay tool (forwards, preserves principal, unauth rejected). `task check` green.
+
+---
+
+## 7. Deferred (not in this epic)
+
+Tracked here so they aren't lost; file as issues when picked up:
+
+- **SKILL veneer** — the `close-session` SKILL becomes the claude.ai trigger/harvest layer that calls capture.
+- **Per-harness token provisioning** for Crush / Pi / LLM Council (claude.ai is done via the existing `/mcp` connector).
+- **Harvest adapters** — transcript-parse vs chat-memory-reconstruct vs agent-runlog, each assembling capture args at its own fidelity.
+- **`SummaryWriter` impl** — ai-sessions summary persistence (the port + path logic exist; the concrete writer is nil today, so a request with a summary fails that one item).
+
+---
+
+## 8. Brain learnings
+
+- `wiki/hyperguild/decisions/capture-classification-taxonomy`
+- `wiki/hyperguild/decisions/gate-on-server-derived-signals-fail-safe`
+- `wiki/hyperguild/decisions/two-phase-reserve-record-audit-gate`
+- `wiki/hyperguild/failures/mcp-bearer-middleware-discards-principal`
+- `wiki/hyperguild/facts/brain-mcp-embeddings-out-of-band-sync` (from #45, the predecessor)