Checkpoint compaction + compiled views

Compaction is the worst moment in a long-lived agent's life: the harness summarizes the window and the un-extracted detail dies with it. Cathedral 5 makes the brain the durable side of that boundary, in two halves:

Checkpoint compaction — at the compaction boundary the raw pre-compaction window is banked to disk synchronously (secret-scanned, content-addressed), facts are harvested from it promptly and asynchronously inside serve, and the post-compaction context carries brain://slug links the agent re-pulls on demand.
Compiled views — gbrain compile-context regenerates warm-file fragments (CLAUDE.md @import fragment, AGENTS.md managed block, OpenClaw workspace file) from brain pages: deterministic, token-budgeted, sensitivity-scanned, byte-stable under an unchanged brain.

The durability contract (read this before filing a "links are late" bug)

The boundary guarantee is bytes durable at the boundary; facts moments after; links at the next boundary:

gbrain hook compact (PreCompact, bootstrap-installed) writes the since-last-boundary window to <corpus>/<session>.seg-<hash24>.txt INSIDE its 3s deadline — before any IPC. The window is read from the newest 2 MiB of the transcript, and a too-tight deadline degrades to a typed skip (deadline_scan/deadline_write below) rather than an unscanned write — in both cases the session-end/sweep lane still covers the text. A hook crash after the write loses nothing: the maintenance sweep's corpus pass extracts every .txt segment.
Serve's checkpoint harvest (fire-and-forget from the same IPC round trip that banks standing entities) runs the facts pipeline over the segment, verifies each candidate link with a source-scoped page read (a link that resolves to nothing is never banked), and banks a manifest into session_context_state.checkpoint_manifest. Every extraction lane (harvest AND sweep) is capability-gated: a keyless install, or facts extraction switched off, still banks segments but defers extraction (typed keyless / extraction_disabled skips) until the gate opens.
The post-compaction SessionStart (source=compact) pack renders the banked links as a ## Compaction checkpoints section. Links that miss the immediate pack (harvest still running) surface on the next session start. Link-delivery guarantees differ by lane: the harvest FIFO is receipt-guaranteed (a failed manifest publish keeps the receipt and retries without re-extracting); the sweep backstop and the OpenClaw direct-Postgres rung publish links best-effort (facts are always at-least-once — re-extracting a whole segment to retry a link append is the worse trade). The facts themselves are recallable the moment any lane commits them, regardless of link delivery.

The three extraction lanes and the dedup contract

The same session text is reachable by three lanes. The contract keeps the happy path exactly-once and every failure path at-least-once (never loss):

Lane	Trigger	What it extracts
Checkpoint harvest	PreCompact IPC flush (serve FIFO, cap 8, 60s abort)	the boundary segment
Session-end corpus + sweep pass 3	SessionEnd hook + serve sweep	the post-last-boundary REMAINDER when segment coverage holds; the FULL transcript otherwise
`gbrain transcripts ingest --facts`	manual	whatever the operator points it at (operator's explicit choice)

Coverage is decided by exact-set hashes, not counts: session-end recomputes every boundary window's redacted hash from its own full parse and requires each in the per-session ledger (<session>.ledger.json). A missed compaction, a deadline'd segment write, or an unreadable ledger all fall back to the full-transcript write (bounded by the 10 MiB transcript parse budget — beyond it only the newest tail is read); fact-level dedup absorbs the overlap. A session where the harness never fires SessionEnd (e.g. /exit) keeps its banked segments; the post-last-boundary remainder waits for the next boundary or a manual gbrain transcripts ingest. Degraded configs with a chat key but no embedding provider may insert duplicate fence rows on that fallback (the 0.95 dedup arm needs embeddings) — segmentation makes this the exception path.

Harness wirings

Claude Code — zero setup beyond gbrain bootstrap (the PreCompact + SessionStart hooks are already installed; existing installs pick the checkpoint behavior up on upgrade with no re-install). Prompt link delivery needs PGLite + a live serve (the IPC flush lane); on a Postgres brain the hook still banks the segment and the next sweep pass extracts it and publishes the links — delayed to sweep cadence, not lost.
OpenClaw — engine-internal: compact() runs the checkpoint step before delegating (spool-first; serve IPC on PGLite, direct connection on Postgres) and assemble() injects the checkpoint block from the banked manifest. No hooks, no recipe.
Codex / opencode — not wired in this release. The pull protocol (gbrain context-pack after compaction, per docs/guides/ambient-recall.md) is the supported path; a native codex hook lane is a filed follow-up.

Operational runbook — reason vocabulary

Hook + harvest telemetry rides the hooks heartbeat JSONL (<gbrain home>/integrations/hooks/heartbeat.jsonl; read it with readHeartbeatTail — doctor's hooks check reports only the aggregate hard-error rate and deliberately ignores typed degrades). Counts and codes only — never content.

compact events carry a segment code: segment_banked / segment_dup (idempotent retry) / empty_window / deadline_scan (budget too tight to scan — a window is NEVER written unscanned) / deadline_write / scan_unavailable — plus a flush code echoing serve's schedule ack: scheduled / skip_queue_full / skip_not_found / skip_bad_basename / skip_no_session / skip_shutting_down / skip_already_queued (absent when the IPC round trip never happened).
session-end events carry the corpus mode: remainder / skip_covered / full_fallback.
checkpoint-harvest events (serve-side pump outcomes) carry inserted/duplicate/links counters and skip reasons: keyless / extraction_disabled / already_ingested / claimed_elsewhere / dream_output (dream-cycle output is never re-harvested) / aborted (retryable — nothing was written) / manifest_failed (receipt kept; the retry re-publishes without re-extracting). Hard failures carry the error name as the reason.

Single-corpus-dir invariant: the hook resolves the corpus dir from file config while serve resolves it from DB config. Keep dream.synthesize.session_corpus_dir consistent (or unset on both) — a split resolves as a typed not_found skip on the flush and the sweep extracts from wherever the hook actually wrote.

Remote thin clients: link re-pull over remote MCP sees world-grade fences (private-visibility facts are stripped by get_page's trust boundary). The trusted-local lane (CLI, hooks, IPC) sees the full fence.

Compiled views

gbrain compile-context --target claude-code --budget 3000   # .claude/gbrain-context.md
gbrain compile-context --target codex                        # AGENTS.md managed block (also serves opencode)
gbrain compile-context --target openclaw                     # .gbrain/compiled-context.md
gbrain compile-context --target claude-code --check          # exit 1 when the committed file is stale

Selection is three deterministic arms over ONE source (the command compiles exactly one source; thin clients are refused — run it on the host install): the durable prefixes (concepts/ people/ companies/ originals/, cap 50 each), pages tagged compile-context (the explicit pin: always a candidate, ranked above every unpinned page — the scan and the budget still apply), and the newest 30 days of pages (cap 500). Entries are world-visibility excerpts (600 chars max; private-visibility fences never leave the brain).

Deterministic by construction: selection anchors recency decay to the newest updated_at among candidates (never wall-clock), ordering is a total order, and an unchanged brain produces identical bytes across runs and days. The budget caps the WHOLE file, header included (a budget below the fixed header floor emits just the header plus an honest no-fit comment). Content that trips the sensitivity scan (secrets, PII shapes, operator blocklist/pattern file) drops that page's entry, and entries past the budget are dropped by the packer — both always reported, never silent — and .gbrain-scan-allow fingerprints un-drop reviewed false positives uniformly across every detector family. Engine read errors abort the run without writing. The openclaw target writes the file only — nothing auto-loads .gbrain/compiled-context.md yet; reference it from a workspace file your engine already loads.

checkpoint compaction