voitta-ai

by voitta-ai

Skillz bundle: every skill in the catalog except those whose own plugin ships hooks (see catalog.json).

by voitta-ai

A Slack agent cannot wake itself: check whose token the posting tool holds, and post as the human instead of via the agent's own bot or webhook.

by voitta-ai

Actually rotate a leaked Slack bot or app-level token: reinstalling does NOT mint a new xoxb-, only destroying the OAuth grant does. Covers the channel-membership fallout, the incoming-webhook channel picker, and verification without printing tokens.

by voitta-ai

Drive the Slack web API as yourself via a live browser session (xoxc token + httpOnly d cookie) when you cannot install a Slack app. Ships a runnable Python client.

by voitta-ai

Capture the exact outbound request body a single-page app sends, and stop it before it reaches the server. Use when: (1) you need the real JSON payload of an action (export, purchase, submit) to compare against an expected fix, (2) triggering the action for real would cost money, credits, or create a record you cannot undo, (3) a fetch-only interceptor "did not fire" and the request went out anyway, (4) you must prove which of two endpoints an action actually posts to, (5) you are verifying a frontend fix by observing behaviour rather than reading the bundle, (6) the harness is installed but captures nothing on an Angular/zone.js app — prototype patches are bypassed and you need the constructor-level swap plus a pre-flight proof before any risky click, (7) you must prove a submission actually created something, not just that a request left the browser. Covers patching window.fetch AND XMLHttpRequest together (the XHR half is the one usually missed), the constructor swap for zone.js apps, the pre-flight gate, matcher scope (block the commit call, never the dialog's own reads), recording responses, blocking cleanly, and disarming afterwards — including why a hash-route navigation does not remove the patch.

by voitta-ai

Strip Spring/Grails profile configs down to real overrides and prove the effective config did not change. Covers flatten-and-diff verification and when a duplicate should deliberately stay.

by voitta-ai

Typecheck a TypeScript change without installing a large monorepo's node_modules: rebuild a minimal standalone repro at the repo's pinned compiler version, run orig/broken/fixed variants, and check exit codes so a crashed tool cannot read as a clean pass.

by voitta-ai

Recover the work of a spawned subagent that finished but never reported, and tell that case apart from a genuinely wedged one. Root cause: an agent type's declared toolset may omit a messaging tool, so a correct, completed agent has no channel to deliver through -- or it is a forward-only wrapper scoped not to poll the background job it dispatched. Idle notifications are runtime-generated and are NOT evidence the agent's own send worked; the reliable oracle is the transcript (tool_use with no tool_result). Covers recovering the result from the agent's final text, from the job runtime's state directory, and the one-sentence fallback brief that prevents the loss.

by voitta-ai

Tamarian mode: Claude speaks as the Children of Tama - meaning carried by metaphor and allusion (canon ST:TNG phrases plus ones coined from Earth culture), technical substance fully intact. /tamarian lite|full|ultra|off; hooks persist the level across sessions.

by voitta-ai

Diagnose and unwedge Kubernetes rollouts stuck because an AWS Load Balancer Controller TargetGroupBinding points at a target group no listener rule forwards to. Use when: (1) `helm upgrade` / a Deployment rollout times out with "context deadline exceeded" while every container in the new pods is healthy, (2) new pods show `Ready=False` with reason `ReadinessGatesNotReady` and a readiness gate named `target-health.elbv2.k8s.aws/<name>`, (3) `aws elbv2 describe-target-health` shows every target `unused` / `Target.NotInUse`, (4) a PodDisruptionBudget reports `ALLOWED DISRUPTIONS: 0` and node drains or consolidation (e.g. Karpenter) are blocked by a service that serves no traffic, (5) you staged a target group ahead of its listener rule "so it's ready for the cutover" and want to know why that wedges every future rollout. Core facts: the controller injects a target-health readiness gate into every pod a TargetGroupBinding matches; a target group attached to no load balancer listener keeps its targets permanently `Target.NotInUse`, so the gate can never pass. Unwedge with `kubectl delete targetgroupbinding`; durable fix is gating the binding separately from the target group and attaching the listener rule before the binding.

by voitta-ai

Make Terraform `check` blocks actually fail CI. Use when: (1) you wrote a check block to guard an invariant (e.g. a set of secret values must be pairwise distinct) and discovered `terraform plan` prints the failure as a Warning and exits 0, so the CI plan step stays green, (2) you are reviewing a PR that encodes a safety guard as a check block and need to know whether anything enforces it, (3) you need a guard over sensitive values whose failure message must name the offending items without leaking them. Core facts: check blocks are advisory by design - failed assertions never affect plan/apply exit status, so a guard written as one is silent by construction; the working CI gate is `set -o pipefail` + `tee` the plan output + grep for "Check block assertion failed" and exit 1; write the error_message to name offenders via nonsensitive() applied to KEYS only, never values. Proven end to end: a forced collision exits 1 and names the colliding pair with no values leaked; the passing path leaves plan behavior unchanged.

by voitta-ai

A terraform plan proposes `delete, create` (or "must be replaced") on a resource nothing in your diff touched, and the forcing attribute reads `(known after apply)`. Use when: (1) a plan shows `# forces replacement` on `name`, `bucket`, `identifier` or similar for a resource you did not edit, (2) the value is built from a `data` source such as `data.aws_region.current.id`, `data.aws_caller_identity.current.account_id` or `data.aws_availability_zones`, (3) the module declaring it has a module-level `depends_on`, (4) an apply already happened and a workload now fails with `The security token included in the request is invalid` or `The AWS Access Key Id you provided does not exist in our records` on a policy you can prove is correct. Covers why a module `depends_on` makes every data source inside it unknown at plan time, why that silently escalates to resource replacement, and why a replaced IAM role breaks running pods for up to an hour with errors that misdirect to IAM.

by voitta-ai

Tear down an ECS-on-EC2 capacity-provider stack when a single terraform apply deadlocks with ResourceInUseException: destroy order is the reverse of the dependency graph, so compute tears down before the workload; split into two sequential untargeted applies, workload first, infra second.

by voitta-ai

Apply terraform to production non-interactively when the repo's apply.sh/deploy.sh wrapper calls terraform apply without -auto-approve and hangs on the approval prompt. Covers why `echo yes | ./apply.sh prod` approves an unseen recomputed plan, the plan -out/apply <file> alternative that fails closed on stale state, reproducing a wrapper's auto-detected variables and init flags, and why a saved plan file is a secret.

by voitta-ai

Prove whether a terraform change was actually applied to an environment: census the versioned S3 tfstate objects over time (serial + resource-type counts) and cross-check CloudTrail, to tell a real apply from a reverted one, identify which variant of a stacked change ran, and separate untracked orphans from failed destroys.

by voitta-ai

Verify a pre-commit hook meant to refuse commits, without destroying the hook while testing it: the staged-vs-committed diff trap, safe test sequencing, and reflog recovery.

by voitta-ai

Verify quantitative claims about US federal grants and revenue against government open-data APIs.

by voitta-ai

A vendor console's confirm button silently does nothing and the required consent checkbox clears itself on every submit. Use when: (1) a submit button appears to work but the dialog just stays open, (2) a required checkbox visually ticks and then un-ticks on submit, (3) a click or value-set on a checkbox reports success yet the app behaves as if it is unchecked, (4) you are about to retry the same click a fourth time, (5) you need to know whether the request is even leaving the browser before blaming the UI. Two causes that imitate each other: a framework-controlled input that ignores synthetic clicks and needs a real Space keypress, and a server error the console renders nowhere. Covers telling them apart in one read-only step, reading the swallowed response body, and checking whether an earlier attempt half-succeeded.

by voitta-ai

Retrieve verbatim Truth Social and X posts, including deleted ones, for fact-checking.

by voitta-ai

Acquire a WordPress.com OAuth2 token (authorization-code flow) and publish/update posts via public-api.wordpress.com.

by voitta-ai

Author-side PR iteration loop. Watches a PR for new review comments and addresses them across rounds.