The no-effort rule was keyed on every claude-haiku id, so Haiku 5.5 ran at the engine default instead of its phase effort (mechanical: low). The guard now names Haiku 4.x; status-reporter pins effort: low.
A pick in the picker is held (PreModelSwitch deny), then the same project / conversation / everywhere question runs; only Everywhere reaches the native write to user settings.json. A model chosen in that dialog bypasses the cold-read caps but waits while held down.
This project writes model into .claude/settings.local.json and sets a sticky route; This conversation sets the sticky route only; Everywhere runs the native write. Links refused in the project tree, native never runs after an answer.
Prompt rules and the derived orchestrate route write effort-only routes, so
they never lift main above the session model; a held run keeps its model.
Fable on main comes from skill rows, the route tool or /route (BDR-123).
CLI 2.1.306 prints an advisory plugin-directory section (ICON_MISSING,
blocking: false) that the whole-output `warning` grep counted as a
failure. The verdict is now the rc of `claude plugin validate --strict`
(per-file warnings → 1, directory advisory never), through one helper
shared with a positive control on two scratch manifests.
The Sub-agents rule gains the ceiling and hands the rest to /feat or /bugfix; README and the challenge citers now say the main loop reaches Fable through best-tier rows from an opus session. /deploy's floor follows its verify row and its stale changed entry goes.
New best-tier phase audit (xhigh) carries both merge gates; their opus frontmatter stays the outage floor, since inTier never writes below a tier head. The census floors best-tier rows on the second alias, and the citers now name the audit row.
The probe trusted the plan's Supabase url, followed redirects for free,
crashed on deep JSON and stopped at the first unconfirmed map candidate.
Rules now scope hydrated state to the export chunk, silence dev configs,
UI-suffixed data attributes, GraphQL types and metadata keys, and catch
Go raw strings, triple-quoted f-strings, sql.raw, empty select() and
multi-line hidden inputs; text output is sanitised; per-variant locks.
After the report the skill proposes fixes (plan phase), stops for the
user's acceptance, applies accepted ones via bugfixer and the
verify-secure gates, then re-scans. scan.py gains --assert-absent
(contract oracle, positive control) and --baseline (no new hit).
probe.py observes a running target the user owns and names, driven by a
plan built from the static findings. Five probes: anon-read, cross-user,
no-auth route, CORS reflection, capped rate-limit burst. GET/HEAD/OPTIONS
only, --authorized gate, secrets from the environment, same-host redirect
only, bounded request budget. SKILL STEP 5, checks.md reference, docs.
Suite guardrails-probe.test.sh (16 checks) against a localhost fixture.
Eval iteration 1: the callout 'edge' variant picked up the diagram connector styles, and all three skill runs hand-wrote a template filler. fill_template.py takes values.json with @fragment files, escapes </ in the quiz; suite 21 checks.
The main loop may now move down to a cheaper model by default
(mainModelSwitch true, fail-closed picker). Both caps measure the COLD
READ a switch costs: the context minus what the target model's prompt
cache already holds, from a per-model served record written on every
main step (1 h TTL, cleared by /clear and /route reload; a smaller
context than the record counts as cold). downgradeMaxTokens defaults to
120000, upgradeMaxTokens keeps 200000; no return bypass is needed since
a warm return is cheap by construction.
The prompt classifier writes the full phase route: coding and
bookkeeping guesses move main down under the switch and the cap; a guess
never ranks above the session model; a mid-turn ultrathink floor keeps
its model; a user /effort typed between turns drops the guess for that
turn only (engine baseline is now per turn). A model-loaded helper skill
inside a best-tier run moves effort only. Main routed onto a model whose
window the context outgrew returns to the session model. The skip line
is said once per cap crossing; the toast names a cap hold; the status
line shows the switch only when off.
code-clean B5 (enable_tool split into enable_gstack/
enable_external_single/enable_pack), B7 (enable_skill into
enable_gstack_skill/enable_linked_skill/enable_plugin_skill), B9
(cmd_show into _show_entries/_show_plain/_show_grouped, `_show_plain`
ends with an explicit return 0), B18 (locals), B19 (comment).
/refactor C4: tool_kind/tool_source replace the 3 hand-copied name
lists with the MANAGED_TOOLS roster (exact-match loop), and disable_* is
split to mirror enable_*.
/refactor C5: gstack_skills/gstack_parked_names/gstack_park/
gstack_unpark move into lib/gstack-removed.sh, the only sourced file the
fixture suites copy, and profile.sh and toggle-external.sh both use
them.
496-run golden capture identical; make test 47 green, shellcheck 0.
usage() printed lines 3-26 with a hardcoded range, but the tool list
runs to line 32, so 8 of the 14 tools were missing. It now reads
`sed -n '3,/^# ====/p' | sed '$d'` and stops at the closing banner, with
no line numbers. Logged in .claude/audits/BUGS-FOUND.md.
A1 drop unused `import gzip`; B1/B2 file-level shellcheck directives
placed before `set -u`; B3 `cd "$REPO" || exit 1`; B4 hook regex split
into `pattern+=` lines (sha256 of $pattern unchanged); B11
`_build_parser`/`_dispatch` pulled out of google_seo `_cli`; B20
detect-plugins header names update-all.sh; C2 prune-memory deterministic
suite wired into SUITES (literal path); C3 `_refuse_outside_allowlist`
moved up in deploy-commit.
No behavior change except B3 (a failed cd now exits 1) and C2 (one more
suite). B5/B7/B9/B18/B19 land in the refactor commit.
Kill switch fails closed: a present-but-unreadable override at first load
or a non-boolean enabled turns the router off (logged); /route reload says
when the previous config was kept. Prompt rules with nested quantifiers
are dropped; a config path that is not a regular file reads as failed;
caught errors log kind + code only (errWord); rejected keys are masked in
logs (keyWord), alias keys match KEY_NAME, prototype names are never keys;
offers capped at 256; a models alias without a fallback list is ranked at
the end of the default chain; canonical matches a table id at a '-'
boundary only; logCall/noteMain cannot unroute a step; route-tool failure
text no longer claims nothing was routed.
Repo side: mods suite probes the 'claude plugin test' usage line, doctor
bounds plugin list to 60 s, update-all skips @skills-dir mods.
Kit 431/431, full make test green.
The text regex becomes a prefilter. The command is split into simple
commands by an awk tokenizer (quotes, escapes, ; && || | ( ) backticks,
heredoc bodies dropped) and a push is only a simple command whose argv[0]
is git (or */git) with push or send-pack as its first non-option word.
grep "git push", git log -S, git config --get push.default, git stash
push, heredocs and commit messages quoting a push are no longer denied.
bash|sh|zsh -c scripts are re-checked (depth 3); wrappers (sudo, env,
eval, xargs, timeout, git subtree...), assignments whose value reads as
a push, a missing awk or > 2000 simple commands keep the text verdict.
Prefilter also matches "git" push (quote characters stripped).
Suite 98 -> 124 checks.
Plain composer prompts with no route in force get a turn-scoped effort
from $.model.classify (design/analysis/coding/bookkeeping/chat → phase
effort, effort only, never the model). Bounded by a 3 s clock race, a
user interrupt neither counts nor writes, three failures in a row pause
it for cooldownMinutes (one toast, status marker, cleared by /route
reload). No dialog, no routing.json write on a classified route. New
Config key `classify` (default on, pickSwitch), `/route classify on|off`
session toggle, `/route show` line and argumentHint. An explicit user
/effort beats the guess (engineBase check).
Docs: README classifier paragraph, USAGE row + sentence, CHANGELOG,
effort-shift.md. Kit suite 428 tests.
Clears first-use decisions from the tracked routing.json through the
existing writer: a name is forgotten in every table (confirmed, changed
with the row restored to its recorded shipped phase under guards, project
exceptions pruned) and asked again; all and projects ask a confirmation in
the engine dialog (Cancel first) and hold the single-dialog slot; nothing
is written when there is nothing to forget; the answer names what was
restored and the frontmatter floor to realign when one was aligned; the
route tool has no forget path. Docs name the new writer. Kit suite 232 → 284.
Contract .claude/tasks/contracts/2026-10-11-model-router-w3c-forget-1457.md,
plan r3: 3 lenses + 1 confirmation, feater + 3 rounds (one real defect),
GATE 0 MET, verifier at the cap on coverage (user-accepted), security PASS.
The dialog now explains what it asks about: the skill's description (from
its SKILL.md frontmatter, five YAML forms, first sentence, cleaned), the
agent's description (from agent.offer), the phase's new 'about' line in
routing.json, and the real model id and effort the next step runs on.
Options Later / Keep / Change; Change asks the model (fable, opus, sonnet,
haiku with their tier role), then the effort among those the phases of
that model offer, then the scope; the pair maps to an existing phase (rows
stay phase names; the row's current phase wins a tie; same phase = Keep).
A main-loop phase (T3) is Later / Keep only. A main-row change toasts the
real decision and the /route switch hint when the pick is a downgrade.
Descriptions pass a hardened read (name allowlist, stat kind, size cap)
and clean(). Kit suite 190 → 232; W3-A dialog tests migrated.
Contract .claude/tasks/contracts/2026-10-11-model-router-w3b-dialog-1240.md,
plan r3: 3 lenses (2 BLOCKERs: inline rows, phase edits) + 1 confirmation,
feater + 2 rounds, GATE 0 MET, verifier CONFORME (3rd pass), security PASS.
routing.json (tracked, reached through the plugin directory) is now the
single source of the phase table and of every skill/agent row, plus the
decisions: confirmed rows/phases, changed rows (from/to) and projects
exceptions keyed by a normalized git remote (credentials never stored, no
machine paths). First use of a rowed typed skill, a rowed agent spawn or a
main-loop phase opens the engine's dialog (Later / Keep / two alternative
phases; Other = a phase name); a change asks Everywhere or This project
only. One dialog at a time, never in headless, never inside an agent,
never written by the model: only a dialog answer or /route ask writes,
serialized, size-capped, never creating the file. Layers: routing.json <
~/.claude/model-router.json; the project tree is never read. /route
pending, /route ask on|off. Census reads rows and phases from the file and
tolerates a user-changed row (WARN). Kit suite 86 → 190 tests.
Contract .claude/tasks/contracts/2026-10-10-model-router-w3a-confirm-1201.md,
plan r4: 3 lenses + 1 confirmation, feater + 4 rounds, GATE 0 MET,
verifier CONFORME then re-verify after security, security BLOCK(1) fixed
then PASS. Live: T2 dialogs answered by the user from the hot-loaded mod.
The 15 Skill(effort-*) citers now call mcp__model-router__route per phase
(orchestrate at a dispatch span, reflect/plan for the skill's own level,
apply at the bookkeeping tail, escalate at the verify-secure caps); built-in
judgment dispatches carry an explicit effort= param. lib/effort-shift.md is
the route doctrine, lib/model-gate.md the mod rule (route answer = witness,
/route on as remedy). Deleted: skills/effort-*, lib/effort-pins.txt/.sh,
lib/model-check.sh, their tests, the installers' re-apply blocks. The mod
drops its Skill(effort-*) bridge. The tracked model:/effort: frontmatter
stays as the off-state floor, census-locked equal to the rows
(lib/tests/effort-routing.test.sh rewritten, 140 checks; analyzer → xhigh).
Contract .claude/tasks/contracts/2026-10-10-model-router-w2b-1045.md, plan
r4 § W2-B: GATE 0 MET, verifier ECARTS(7) then CONFORME 10/10, security
PASS, full make test green (design-tool-gate env red only).
Tracked relative symlink skills/model-router -> ../mods/model-router: Claude
Code loads the mod in place as model-router@skills-dir wherever link.sh
links ~/.claude/skills (no CLAUDE_CODE_PLUGIN_DIRS: absolute paths in the
tracked settings.json). Engine-laid mods/*/tsconfig.json gitignored.
lib/tests/mods.test.sh: manifest name, link target, claude plugin validate
and test per mod, capability-probed, time-bounded, SKIP with reason.
doctor.sh: fail-soft Mods section (link by -ef, one guarded plugin list).
CLAUDE.md: mods/ section (loading, per-machine enabled:false switch,
dev-copy shadowing, tests).
Run D2 of manual-push mode (BDR-114).
- push-guard sources lib/gitflow.sh once (absolute path) and reads each
candidate dir through gitflow_push_mode; a missing lib denies.
- Dir tokens are extracted as whole shell words: a fully quoted token
(inner apostrophe allowed) is resolved, a backslash-escaped space is
unescaped deterministically, a token mixing quoted and unquoted parts
is refused (fail closed) instead of resolving to its parent.
- A payload jq cannot parse is scanned as raw text with its JSON escapes
folded; a push-looking one gets the static deny through the trap.
- The 20-token cap runs before any per-token classification (a flood of
20 000 tokens is refused in 0.13 s; T58 locks it under 5 s).
- `case "$mode"` has a deny default; missing core tools warn and allow.
- T42 compares the deny list against main (the last release) instead of
HEAD; literal-true, mixed-token, broken-payload, lib-missing and
banner-on-bad-value cases added (98 checks).
- session-start banner reads the mode through the verb and shows
`push : manual (autopush bad)` on an unparseable value.
- tour hints quote "<abs project>".
Run D1 of manual-push mode (BDR-114). `git config --bool --default true
gitflow.autopush` only covered a MISSING key: an unparseable value made
git die with empty output, the `= false` test failed, and every push ran
again. A typo on a work machine silently re-enabled the pushes it was
meant to stop.
- lib/gitflow.sh: `_gitflow_push_off` reads the mode through the lib
verb (`push-mode`); anything but `auto` is push-off, and the verb's
stderr line names an invalid value during start/finish.
- Emitted post-commit/post-merge hooks (POSIX sh, standalone): push only
when the key reads `true` or is unset; `false` exits quietly; any
other result prints one stderr line ("NOT pushed, treated as manual
push mode") and exits 0. Mirrors gitflow_push_mode.
- .githooks/ and githooks/ regenerated files-only through `emit-hook`
(no config read or write; .git/config hash unchanged).
- hooks/unpushed-guard.sh: mode from the lib verb (absolute lib path
resolved before any cd, no temp file); anything but auto is manual;
the SessionStart line names an invalid or unreadable value.
- Tests: gitflow-test T18q block (invalid → start, hook and finish push
nothing and say so; `true` → the hook pushes; emitted hook is
POSIX-clean), unpushed-guard T14 rewritten.
Closes the non-gap observations the gates left on runs C1/C2:
- capitalize STEP 5C/6: heading no longer says "+ push"; the --no-push
fact read is its own paragraph and scoped to that path; the
auto-persisted line requires finish rc 0 AND ahead = 0; rc 5/2/6
(merged, branch not deleted) still report the push state; the
"not on origin" line carries the once-a-remote-exists hint.
- gitflow.sh push-mode: the raw config value echoed on stderr is reduced
to printable characters (LC_ALL=C, BSD tr safe) and capped at 64.
- gitflow-test.sh exports the hermetic git config env in the file, so a
bare run on a global-manual machine stays green.
- client-handover-writer: the branch allowlist refuses a leading dash.
Run C1 of manual-push mode (BDR-111/BDR-112).
- lib/gitflow.sh: `gitflow.sh push-mode` prints auto | manual | invalid
(rc 0; an invalid value is named on stderr). It is the one reader a
skill may call: the bare `git config … gitflow.*` read is denied to
Claude since run B. Ignores GITFLOW_NO_PUSH by design (documented).
- skills/capitalize/SKILL.md STEP 5C: the explicit `git push origin
develop` is gone — `finish` has pushed develop itself since BDR-095,
mode-aware since run A. 5C is now three separate read-only calls
(finish; push-mode; `git rev-list --count origin/develop..develop`)
and prose outcomes keyed on the real ahead count: pushed / manual push
mode, you push / not on origin / push FAILED / invalid value named,
plus a finish-failure outcome (merge vs delete rc distinguished).
STEP 6 closing lines and the recap carry every outcome; the
`--no-push` line reads the branch's own ahead count ("this disk only"
only when true). Invariant: no `git push` inside any Bash call; the
user hints are prose.
- skills/close/SKILL.md, lib/gitflow-aiguillage.md: "push" claims
qualified "in auto-push mode".
- lib/gitflow-test.sh T11b: six cases for the verb (default, true,
false, non-boolean with stderr + rc 0, corrupt config, usage).
Polish items from the gates are listed in TODO.md (C1 polish).
Hardening after the security gate on a2ac018 (3 MEDIUM, all closed and
re-measured):
- dir tokens are deduplicated and capped: more than 20 distinct cd/-C
targets in one command denies before any git fork (20000 tokens: 0.15 s
against the 10 s hook timeout that used to turn a flood into an allow)
- a git or cd failure while reading gitflow.autopush denies instead of
reading as auto (git absent, usage error, unenterable dir); the key
being unset is the only "auto" answer; the decision is recorded only
after one candidate was evaluated cleanly, else the EXIT trap denies
- cd/pushd/-C targets that follow a quote or backtick (bash -c '…') are
extracted; quote characters are excluded from unquoted tokens
User decision (contract, gated): both fail-closed cases also fire in
auto mode on such pathological commands; silence in auto mode holds for
every ordinary push. Header limits list the residual misses (quotes or
backslashes inside a token, cumulative relative cd, unparseable payload)
backed by the soft_deny rule. Tests: 71 checks (T48–T50b added).