Main loop calls the advisor before a skill's plan, once per verify-secure run when stuck, and before a merge. plan-challenger, verifier and security-auditor may call it once before the verdict; advice only adds, never clears (amends BDR-150).
Global line and effort-shift Advisor section: the main loop calls the advisor when a stronger model is needed, sub-agents do not, advice waives no gate. Skill-runner, Limits, ask-tier and audit-row prose match BDR-150.
Fable is the advisor (BDR-150): the client-handover skill-runner children and the model-gate dispatch doctrine pin opus. A pinned child's route answer "fixed at spawn" passes the gate; a repo-wide census refuses any fable pin.
Fable is the advisor now: the best and big tiers default to opus then sonnet, the row dialog drops it, and the ask-before-Fable dialog runs only when an override heads best with fable. Fable stays ranked and hand-pickable.
User's /advisor and /model picks: advisorModel fable, model opus[1m]; skillOverrides block moved by the CLI, one tab fixed. No permission list changed.
When the guard itself fails, a main-loop code write outside a skill run is held (next.called split); tool errors replay as before. A Skill answer carrying isError no longer opens the guard.
A new ## entry must carry the next free ID, replayed in line order the way add allocates it. The rule runs only on a v2 HEAD with entries, so upgrade, first commits and template imports pass.
memory_index skipped a non-regular blockers.md and dropped the awk rc, so the digest printed a false '- open: none'. With a kind it now returns 2; without, it skips the bad registry with a stderr line. The digest prints unknown on a malformed row.
The single legacy row matched no real spelling (bench M5). 28 rows (7 shapes x 4 prefixes, bench-proven with an echo analogue) now deny them; soft_deny covers create, container run and compose up; T41 locks the rows.
alarm then exec killed one process: a child holding the captured pipe kept the 2 s bound open. perl now leads its own group, kills it at the deadline and TERMs leftovers; SD19/SD19b assert the bound and that the command ran.
profile-state.jq sanitizes printed values and shape-checks plugin ids (no forged plugin: line); the digest prints open: none or unknown instead of nothing; push-guard deny_lost checks every pipe stage, T63 split into ask and deny lists; prune-memory mktemp guarded.
Handover: rooted reads, whole-line PROJECT_ROOT read with a one-line
pre-paste rule, BDR status filter. plugin-gate: single-quoted templates,
unsafe roots printed never run, apply check against the git common dir,
$HOME refused (and in profile.sh). Advisor: minimal only from project
signals, no extras token, ctx7 wording, print-only plugin remedy. Packs
vocabulary in onboard/tour; graphify labels match the trap.
Precision, triggers and the miss share count only rows after the gate-1
fix merge (d3d5e6c, 2026-11-12T19:53:59Z); their actions wait for 20
post-fix asks, a one-off NEW line marks the sample and the reminder
keeps prompting until then. The 3rd-pick signal is gone. The reader
holds one parsed transcript at a time. labels.tsv rows and state.json
are validated on read. Exact key sets of log records pinned.
code-limits: mkdir -m 700 without -p as the atomic proof, else one
find test (ours, dir, 700). session-start: one runner, timeout then
gtimeout then perl alarm (exec, no shell). pre-commit (generator +
both copies): a git diff or grep error blocks. memory.sh line rewrap;
claude-stub reads its log path from the env. Tests T16, SD19, PC1-4.
The fable alias (head of the best tier, every best-tier phase, row and
pin that names fable) now resolves to claude-fable-5-2. Rename prepared
by another session, split from the BLK-049 diff; id probed live (claude
-p --model claude-fable-5-2 answers). Kit 640/0, validate --strict ok.