mirror of
https://github.com/windmill-labs/windmill.git
synced 2026-08-25 08:00:59 +00:00
e0510fea21006a06eee7eecd4587161970d7f4d5
3 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Guilhem
|
e866b68cdf |
feat: surface execution usage in the sidebar and explain what an execution is (#10760)
* feat: surface execution usage in the sidebar and explain what an execution is Users read "executions" as a job count and are surprised by the real number, which meters a second of compute. Every place the UI prints an execution count now says so, and the sidebar carries a usage meter for the quota that will bind first. Adds SidebarUsage at the bottom of both sidebar surfaces: a ring in the collapsed rail, a labelled bar when expanded, and a modal breaking down every quota. On the free tier it meters the per-user and per-workspace 1000-execution caps; on a paid plan it meters workspace usage against the executions the workspace's seats already include. Item.tooltip was inert on disabled dropdown rows: DropdownSubmenuItem rendered the info icon inside the disabled button, which swallows hover, and the row's own title attribute shadowed any wrapper title. Both renderers now fall back to a wrapper title the way DropdownV2Inner already intended. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: keep the usage meter tied to the workspace it describes isPremiumStore held the previous workspace's tier across a switch, which no consumer noticed while it only gated affordances — the usage meter is the first surface to render a number from it, and would have shown a paid seat quota for a free workspace. It is now undefined until the active workspace's tier is known, and a superseded response no longer writes. The seat fetch had the same shape: a slow response for the workspace we left overwrote the current count and stayed wrong until the next switch. The usage wrapper also carried the padding the brand-mark row used to own, which shifted the sidebar bottom by 4px on every instance where the meter renders nothing. The component owns its own padding instead. Names the collapsed ring for assistive tech, which otherwise saw an unlabelled button whose only signal was the arc's color. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: scope the usage meter to the billing workspace and a known tier A fork's usage, tier and bill all resolve to its billing root, but its member list is deliberately a subset of the root's, so counting fork members metered root usage against a fork-sized cap and invented billed-seat overages. Seats now come from the billing root, and the paid meter stays hidden when that root is not visible from the fork. The tier was cleared only after the user-store round-trip, so the meter rendered the previous workspace's tier for the length of it — a free→paid switch showed the 1000-execution hard cap on a paid workspace, not a race but every time. The clear now happens before the first await. Workspace usage had neither guard: a superseded response overwrote the store permanently, and the meter is the first surface to print that number as its headline rather than bury it in a dropdown. The free-tier counters keyed off `!$isPremiumStore`, which reads an unknown tier as free and flashed the free-tier blocks during a paid-to-paid switch. They wait for a known tier instead. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: never render an unresolved execution count as zero The workspace-usage clear wrote 0, which is a real usage value: an in-flight or failed fetch rendered as a green "0/1,000" bar, and a rejection left it there for the session because loadUsage had no failure path. Usage is now undefined until it resolves, each endpoint is assigned on its own so one failing leaves the other's number intact, and a quota is listed only once its own usage, tier and cap are known. The legacy counters show an em dash rather than a fabricated 0. The fork gates read an unknown tier as not-premium, so clearing the tier on switch made the fork entry point disappear for the length of the fetch on a paid-to-paid switch. They hold while the tier is unknown. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: read the usage endpoints as numbers, and fall back to the free tier Both usage endpoints serve text/plain, so the client hands back a string despite the generated `number` type. Interpolation and arithmetic coerced it, which is why nothing noticed before, but `toLocaleString` on a string returns it unchanged — a five-figure count rendered without its thousands separator against a formatted cap. A failed tier fetch left the tier unknown for the session, and consumers hold premium-only affordances through the unknown window so a free workspace kept offering them. It falls back to the free tier instead. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: keep an unknown tier unknown, and refresh the seat cap on demand Falling back to the free tier on a failed tier fetch fixed the affordance gates by lying to the meter: a paid workspace's real five-figure usage rendered against the 1000 hard cap, red, under "jobs stop running for the rest of the month". The tier stays unknown instead, and the two consumers get what each needs — the meter hides, while affordances read `maybePremium`, which holds through the pending window but fails closed once the fetch has failed. Membership changes elsewhere don't reach this component, so the seat cap could show an overage against a cap that had since grown. It re-resolves when the modal opens, which is when the number is read rather than glanced at. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: let anything showing executions re-read them The counters were written in one place, the root layout, on a workspace change only — so a tab left open all day showed the count from whenever the workspace was opened, and the modal-open refresh could only reach the seat cap, leaving a freshly computed denominator over a stale numerator. Moves the fetch to lib/usage.ts, next to the stores it writes, so the meter can refresh both numbers when its modal opens. Seats follow a membership signal that WorkspaceUserSettings bumps where it already refetches after every mutation, so the cap stops lagging a role change without either side owning the other. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: order concurrent usage and seat refreshes The workspace id doesn't order two requests for the same workspace, and both refreshes can now have two in flight: usage through A→B→A or a modal-open refresh landing on one already running, seats through a membership bump arriving mid-request. An older response could win and restore the count it replaced. Each refresh takes a generation and only writes if it is still the newest. The membership signal also fired on a plain read, so opening the users tab made every consumer re-fetch a list identical to the one it held. It bumps on an observed change to the member set instead, never on the first read. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: count only billable seats, and order the tier requests The cap counted every member row, while the backend bills `NOT disabled AND NOT is_service_account` — a workspace with service accounts got an inflated included quota, which hides a real overage rather than inventing one. The seat basis matches `count_paid_seats` now, and the membership signature carries both fields so enabling or disabling a member re-resolves the cap. The tier fetch was the one refresh still ordered by workspace id alone, so a late failure for a workspace could raise the failure flag over a tier a newer request had already resolved. It takes a generation like the other two. The membership signature is keyed by workspace: this page survives a workspace switch, and comparing one workspace's members against another's reported a membership change where only the workspace had changed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: compare the member set only against the same workspace's Qualifying the signature with the workspace put the workspace inside the value being compared, so a switch made every comparison unequal and bumped the version unconditionally — the opposite of the intent, and worse than before the key. The workspace is the key now, not part of the payload: a different one has nothing to compare against and re-baselines silently. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * refactor: hold the usage and tier fetches in resources Every one of these values belongs to a workspace but lived in a bare store, so each writer and reader re-derived "does this still describe what I'm rendering?" by hand. Nine sites did, and the ones that forgot were most of this branch's review findings: three stale-workspace overwrites, three A→B→A races, and two placeholders (`0` executions, `false` tier) that read as real data because an in-band value was standing in for "not known". `resource` from runed — which frontend/AGENTS.md prescribes for async data, and which ~80 files here already use — supplies all three properties as behaviour rather than convention: a superseded fetch is discarded, the value resets when its key changes, and loading and error are states instead of magic values. The seat count keys on the billing root and the membership version, so both a workspace switch and an added member re-resolve it. That removes three generation counters, two workspace trackers, and the manual clear-and-compare around each fetch. What remains is one publish site that asserts the value still carries the active workspace before it reaches a store. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: order the resource responses that runed does not The refactor claimed `resource` discards a superseded fetch. It does not: its only ordering is an AbortController whose signal the generated client cannot consume, and `current = result` runs unconditionally once a fetch resolves. So a late answer for a workspace we had left still landed in `current`, and the publish site — which trusted `current` — cleared the value on screen for the workspace we were on. That reinstated the races the generation counters had covered. `loading` was standing in for the missing ordering, and it cannot: it is also true during a `refetch()`, when `current` is still the right value. Gating on it meant every re-read blanked the meter, and clicking it unmounted the modal that same click had opened, since both sit behind the quota it had just cleared. Values now carry the scope they describe and `scopedValue` keeps the newest one matching the active scope, so a superseded answer neither publishes nor erases, and a re-read leaves the display alone. The account-wide user counter keys on the account, so a workspace switch no longer clears it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: order responses within a scope, not just across scopes The tag carried what a value described but not when it was asked for, so two fetches for one scope — a refetch landing on an in-flight load, or a second membership invalidation — were indistinguishable and the older won if it landed last. That left the seat cap reading the pre-change number until the next bump or switch, which is the stale cap the generation counters had covered. Widening the tag to the resource key would have fixed it by blanking the bar on every membership change, so the issue order travels alongside the scope instead: `tagged` stamps each request as it is issued, and only a strictly newer answer for the current scope replaces the held one. The unit tests now cover the same-key case they missed; both new ones fail against the key-only guard. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: ignore a user list a newer read has overtaken `lastSeen` was written unconditionally after the await, so a response for a workspace already left overwrote the baseline for the workspace on screen. The next real membership change there then compared against a baseline that was never taken for it, re-baselined silently, and never bumped `workspaceMembershipVersion` — leaving the sidebar on the old seat cap. The `users` assignment had the same hole: an overtaken list could paint over a newer one. Both now go through a single check: a read whose issue order is behind the last applied one is dropped before it touches either. Also trims the two `scopedValue` docstrings and the membership rationale to the four lines AGENTS.md allows, and records there that a failed refresh keeps the last successful value rather than blanking. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: do not claim a plan before the tier resolves Widening `isPremiumStore` to `boolean | undefined` left `UserMenu`'s `{:else}` catching the unresolved state: with the tier still in flight, or after the request failed, a free workspace was told it was on the "Premium plan". Both branches under that block assert a plan, so the block now renders only once the tier is known — which also keeps the bordered divider from appearing empty while it resolves. Verified against the running instance with the tier stubbed slow: unresolved shows neither branch, `false` shows the free counters, `true` shows "Premium plan". Reverting the guard reproduces the wrong label at 300ms. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * test: pin how a late answer orders against the read that replaced it Returning to a scope whose earlier read is still in flight is the one case the guard resolves by scope rather than by sequence, and the suite only covered it with nothing outstanding. It now covers the late answer itself: it stands while it is the only value describing the scope, the read issued on returning supersedes it, and it cannot come back afterwards. Also gives the meter the explicit `type="button"` the sibling sidebar rows use. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: size the modal's plan button with unifiedSize `size` is deprecated on `Button`. `unifiedSize="sm"` renders the plan button at the same height and weight as the modal's own Close button. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: match the plan button to the modal's own action button `unifiedSize="sm"` is `h-7`, and the Cancel button `Modal` renders beside it is `px-3 py-[7px]`, i.e. 32px — so the two sat 4px apart. `md` is the unified size that lands on 32px, which pairs them without putting a deprecated prop back. Measured both boxes rather than the new one alone: 32px and 32px, same top. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * feat: instrument the execution meter, and bill-align PremiumInfo's seats The meter's only interaction is opening the modal, so that is what it counts: `usage_meter/opened`, keyed by the plan tier and the quota that was tightest — `free:user`, `free:workspace`, `paid:workspace`. The full set is a type next to the call site so the vocabulary stays readable in one place. The pair is registered in `FEATURE_USAGE_KINDS` (windmill-ee-private), without which the post is dropped with a 204 and records nothing. Verified both halves: the browser posts `{"feature":"usage_meter","kind":"opened","key":"free:user","value":1}`, the running EE image drops it because its registry predates the entry, and `is_recordable_event` accepts it once the entry is there. `PremiumInfo` computed its seats from an unfiltered user list, so the billing page counted disabled members and service accounts that `count_paid_seats` does not bill. Same filter as the sidebar's cap now. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * chore: point ee-repo-ref at the usage_meter registration Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: read the member list before the seat rows that depend on it `loadPremiumInfo` reads `users` after its own await and nothing recomputes the seat rows when the list lands, so whenever `premium_info` won the race the page rendered zero developers, zero operators and zero seats and kept them. The list is now fetched first, and a failure to read it no longer costs the rest of the page. Also refreshes the registered-action inventory in `docs/feature-telemetry.md`, which the new pair makes 21 across nine features. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs: scope the seat comment to the counter it matches The comment claimed parity with the seats actually charged, which nothing in this repo computes: `count_paid_seats` documents itself as counting provisioned members rather than billing's active-user population, and the Stripe quantity is not derived here. What the filter buys is agreement with that counter. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * chore: update ee-repo-ref to c6902ec2c51dc0ce30962afbfab3e456c5d9b831 This commit updates the EE repository reference after PR #735 was merged in windmill-ee-private. Previous ee-repo-ref: bbc48fae6b73b6d72fe2e125e6003794a4ece167 New ee-repo-ref: c6902ec2c51dc0ce30962afbfab3e456c5d9b831 Automated by sync-ee-ref workflow. --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: windmill-internal-app[bot] <windmill-internal-app[bot]@users.noreply.github.com> Co-authored-by: Ruben Fiszel <ruben@windmill.dev> |
||
 hugocasa
|
53eb94659b |
feat(telemetry): extend feature-usage tracking beyond AI features (#10681)
* feat(telemetry): extend feature-usage tracking to long-tail features Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs: describe telemetry as product feature usage rather than AI usage Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * refactor(telemetry): trim disclosure copy and drop unused pick origin Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(telemetry): count trigger fires per run and key hub picks from hub data Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(telemetry): slugify hub keys and order both writers' upserts Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(telemetry): key native trigger adoption by service so it matches fires Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * chore: update ee-repo-ref for native trigger adoption fix Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * refactor(telemetry): move feature-usage collection into the ee crate * docs: point feature-telemetry at the moved registry and rust writer * docs: correct the trigger-fire gate comment to match measured step counts * docs: put the private-build caveat on the verification step * chore: update ee-repo-ref to f079db9e7962a413b349c4ff8036080894f30771 This commit updates the EE repository reference after PR #725 was merged in windmill-ee-private. Previous ee-repo-ref: 055adb80416f9339c9a28ae7fbaeadad30d74959 New ee-repo-ref: f079db9e7962a413b349c4ff8036080894f30771 Automated by sync-ee-ref workflow. --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: Ruben Fiszel <ruben@windmill.dev> Co-authored-by: windmill-internal-app[bot] <windmill-internal-app[bot]@users.noreply.github.com> |
||
|
hugocasa
|
c3b2275864 |
docs(agents): rework agent context, fix dev-env docs, vendor skills (#10667)
* docs(agents): scope agent guidance to where it loads AGENTS.md loads in every session. Three of its sections only ever applied to one directory, and docs/autonomous-mode.md was unreferenced by anything in the repo, so none of its content was in effect. - Move "Verifying Backend Changes" to backend/CLAUDE.md, "Verifying Frontend Changes" and "Banned Patterns" to frontend/CLAUDE.md. They now load when working under those directories, which is when they apply. - Update the two cross-references that pointed at the moved sections (pr and svelte-frontend skills). - Delete docs/autonomous-mode.md. Its "don't stop early" half is already in .webmux.yaml's oneshot system prompt, which actually loads; its trigger was bypassPermissions, which does not imply an absent user; and it restated AGENTS.md and the pr skill with copies that had drifted (hardcoded ports, relative screenshot paths). Salvaged the UI traps it uniquely documented into frontend/CLAUDE.md and dropped the three stale profile references. AGENTS.md drops ~3.6k characters with no guidance lost. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs(agents): guidance for building a feature — reuse, telemetry, live verification Three recurring gaps, all cases where a pointer existed but nothing triggered on it. Component reuse. The svelte-frontend skill documented three components with props, which reads as the whole catalog; the barrel exports 23 and common/ has 34 subdirectories against those 23. So "never use raw HTML elements" was an instruction agents could not follow. Added a mandatory discovery step: read the barrel, grep the tree, and treat the documented three as examples. Brand guidelines. frontend/brand-guidelines.md is 34k characters referenced by bare path, which nothing opens speculatively. Added a table mapping what you are building to the section that governs it, entered with grep rather than a full read. Product telemetry. feature_usage has 14 registered actions across three features, and an unregistered (feature, kind) pair is dropped by valid_feature_usage_event with a bare continue — no error, still a 204 — so frontend-only instrumentation silently records nothing. New docs/feature-telemetry.md carries the criteria for when to instrument, the four-step recipe including the allowlist and the InstanceSettings disclosure, and the privacy rules. Raised in the plan for user-facing work, not as a separate question, and not at all for bugfixes or refactors. Also: validation now ends at exercising the change on the running instance, with standing permission to spin up whatever that takes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(dev): correct the worktree dev-environment guidance Several things agents were told to do did not match what the machine does. - Env discovery pointed at .env / .env.local / backend/.env. In a webmux worktree the real values are in $(git rev-parse --git-dir)/webmux/runtime.env (BACKEND_PORT, FRONTEND_PORT, DATABASE_URL, CARGO_FEATURES, WM_DB_NAME), sourced by every pane and undocumented. Reading it is also not blocked by the Read(**/.env) deny rules, which the old instruction walked straight into. - The database name rule said branch-with-underscores. worktree-common.sh uses the worktree directory basename, and Postgres truncates at 63 characters, so branch hugo/win-2340-… resolves to windmill_win_2340_…_and_eval with no hugo_ prefix and the tail chopped. A wrong DATABASE_URL guts the sqlx cache. - The restart procedure said "tmux pane 1" and sent keys to an undefined <pane1>. Pane 1 is the backend under the full profile and the frontend under frontendOnly. Replaced with finding the pane by pane_current_command, recovering the live feature set from the running process (CARGO_FEATURES in runtime.env only records what the pane started with), and restarting in place. - Added recovery for an orphaned backend holding the port: it reparents to systemd when its shell dies, so it survives anything that looks like cleanup. Three checks before killing a single pid, because pkill -f windmill takes out every sibling worktree. - Agents spawned their own servers because AGENTS.md opened by telling them to. Now it checks for the existing panes first; the spawn commands are scoped to a plain checkout. - New EE worktrees branched from the EE repo's local main, which nothing fast-forwards, so they started behind the commit pinned in backend/ee-repo-ref.txt — the one CI builds against. They now base on the pin, falling back to main only when it is unreadable. - Enabled webmux autoPull so local main stays current; new worktrees are branched from it. Documented what WM_CLONE_DB does, including that it terminates every connection to the base windmill database. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * feat(skills): vendor grilling/architecture skills; tighten PR ready and review rounds Vendors five skills from https://github.com/mattpocock/skills (MIT, pinned at 84fdeffd12f2ee307994d1eb6feb48173b6e0502). They are one dependency closure: grill-me is a stub that runs grilling, and improve-codebase-architecture draws its vocabulary from codebase-design and its CONTEXT.md upkeep from domain-modeling. .agents/skills/UPSTREAM.md records the license, the pin, and the four local deltas so a refresh stays a diff: - flattened the upstream engineering/ and productivity/ split - rewrote bundled-file links to repo-root paths, since relative links break when read through the .claude/skills symlink - dropped the upstream agents/openai.yaml packaging metadata - removed every ADR path. This repo has not adopted ADRs, and a skill that offers to create them is how the practice arrives by side effect rather than by decision. PR workflow changes, all in the pr skill: - A round that never starts is usually a conflict with main, not a CI outage. Resolve by merging, not rebasing — a rebase rewrites the head SHA that round verdicts and the clean-round marker are keyed to. If the merge advances backend/ee-repo-ref.txt, the EE worktree has to follow or cargo check --features private compiles a tree neither the author nor CI intends. - A clean round no longer means an automatic flip to ready. Wide blast radius (*_ee.rs, migrations, OpenAPI or the generated client, auth paths, shared worker infrastructure, a new public surface) asks first; self-contained changes flip. Unattended, the judgement holds and the action degrades: flip the small ones, leave the rest at a clean draft with the reason in the PR body. - Rounds that never converge are usually structural. After three without convergence, stop, name the module the findings cluster around, and suggest improve-codebase-architecture rather than burning more CI. AGENTS.local.md (gitignored, with CLAUDE.local.md importing it) holds the ready/ask calibration, recorded as dated observations rather than a rule. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs(dev): state that each worktree gets its own fresh database The per-worktree section warned which DATABASE_URL to use but never said where the database comes from: the post-create hook creates and migrates a new one per worktree, so it starts with none of the main instance's workspaces, scripts or flows. WM_CLONE_DB was documented only as a comment in .webmux.yaml, which reads as how things work rather than as a per-project opt-in. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * chore(sqlx): script the cache backup/restore instead of documenting it The update-sqlx skill spelled out a cp/comm/rm dance around `cargo sqlx prepare`, which empties backend/.sqlx before regenerating — a failed run leaves the cache gutted (observed: 2350 -> 142 entries), and a --all-targets run in a CE checkout fails that way every time. Three problems with documenting it: - The backup path was the literal /tmp/sqlx_backup, shared by every worktree. Two concurrent runs overwrite each other's backup, which is the only thing standing between a failed prepare and a gutted cache. - The restore was a copy-pasted `rm -rf .sqlx && cp -r ... && cp ...` chain. - Skipping the backup is what turns a routine failure into a lost cache, and a convention is easier to skip than a command. sqlx-cache.sh has backup / newq / restore, keeps state in a per-worktree directory, and leaves the judgement call where it belongs: `newq` prints each added entry's query field for review, and only `restore` writes them in. Also adds the general rule that scratch files belong outside the checkout — anything written into the tree has to be deleted again, and rm prompts each time. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs(agents): state why a routine cleanup prompts, and where scratch goes The guard hook already auto-allows a plain rm whose operands are under /tmp or inside a git checkout in $HOME, so deleting a temp dir or a stale .sqlx entry costs nothing. What prompts is the command shape: the hook's tokenizer defers on &&, ;, redirects, quotes and $VAR, so a chained cleanup falls through to the Bash(rm:*) ask rule. That was recorded only inside a paragraph about screenshot file paths in frontend/CLAUDE.md, where nobody looking for it would find it. Stated in Core Principles instead, alongside the rule that scratch belongs outside the tree — for the reason that actually applies, which is not committing junk rather than avoiding prompts. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * chore(security): deny agent edits to the permission hooks and project settings .claude/hooks/guard-rm-outside-tmp.sh and guard-main-branch.sh are the enforcement points for everything the permission rules are meant to catch, and nothing stopped an agent editing them. One sed -i disables the guard for every later command, silently, and the deny list in .claude/settings.json has the same exposure. Defence in depth rather than a boundary: an agent with arbitrary bash can still delete, and this may only close the Edit-tool path if Bash writes are not covered by Edit deny rules. It costs nothing and removes the cheapest way to turn the guards off. Changing them now means editing the files by hand, which is the intent. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: address review round findings on head |