Island-Dev-Crew
- 78 skills
- 0 followers
- 11 hours ago last updated
- ▌ Arch Lens · island-dev-crew bundleHave the agents build the repo its own drill-down architecture viewer - a small repo-local static tool that shows the modular structure as a diagram with dependency arrows, drills from module into submodules on click, and lands on the code itself. Use when a codebase has outgrown what one head holds, when an architecture conversation needs a shared picture, or when the user says "show me the architecture", "build an architecture viewer", "I want to click into the modules", or "map the dependencies visually". Differentiator - viewing and navigating only, built by the agent for this repo; a look renders out-of-tree and only an authorized instrument build writes into the repo, never committed into a repo you were asked to visualise; ranking refactor candidates and the deletion test live in arch-survey, the module vocabulary in deep-modules.
- ▌ Crap Gate · island-dev-crew bundlePer-function CRAP ceiling as a fix-until-green gate on agent-written code — the Savoia & Evans 2007 coverage-weighted complexity score, with threshold regimes for humans vs agents. Reach for it when wiring a quality gate over freshly generated code, setting or tuning a CRAP threshold, or when the user says "crap gate", "CRAP score", "coverage-weighted complexity", or "run crap over what you just wrote". Differentiator - this island owns the metric content (formula, regimes, input contract, the coverage hole); hook plumbing, gate infrastructure, and evidence format live on neighboring islands.
- ▌ Leak Scan · island-dev-crew bundleDetect information leakage - one design decision expressed in two or more modules - and report it as a duplicate-knowledge finding rather than a style complaint. Reach for it when a diff touches a client and a server together, when a magic constant or wire format turns up in a second file, or when the user says "leak scan", "duplicate knowledge", "this fact lives in two places", or "why did that one change need edits in two files". Differentiator - this island detects and reports leakage sites and adjudicates each one; the depth vocabulary and the churn-mined ranking of a refactor programme belong to neighbouring islands.
- ▌ Seat Relay · island-dev-crew bundleStage one story through the serial five-seat specialist relay - specifier, coder, cleaner, hardener, QA - one fresh-context agent per seat, each handing a typed artifact to the next. Use inside spec-pipeline's implement stage when a story warrants gated quality, or on "run the seat relay", "five-seat relay", "stage this story through the seats". Differentiator - serial role specialization with born-do-die contexts; gauntlet-loop fans out parallel builders against one falsifiable bar, this island hands one baton down five seats in order.
- ▌ Spec Mulch · island-dev-crew bundleTreat a story spec as scaffolding that mulches on merge. It exists to launch one story through the relay and comes down once that story has merged and a human confirms; the durable specification is the end result plus its standing gates (gherkin, tests, the dependency fence), never the plan doc. Reach for it on a repair-shaped ask - "can I delete the spec now", or a merged story's spec file still sitting in the repo - and on an audit-shaped ask, where the scan's verdict is the whole deliverable. Retain-shaped asks are answered with an argument and never a deletion, so "keep the spec updated", "archive the specs", or a specs folder proposed as living documentation get the durable-home table and the drift case instead. Differentiator - it mulches plan documents, not code (throwaway code is prototype's seat), and it is a named divergence from spec-pipeline's publish-the-spec rule.
- ▌ Do It Twice · island-dev-crew bundleRoyce 1970 read forwards - the paper that drew the waterfall also called the single-pass model risky and prescribed building it twice, a pilot version first. Walk one thin end-to-end slice through the entire pipeline and capture the evidence before any parallel fan-out, because dispatching N agents down an unvalidated path multiplies one wrong assumption by N. Reach for it before spawning parallel agents or worktrees, when someone says 'let us parallelize this', 'spin up three coders', 'fan out the remaining stories', or 'the plan is ready, dispatch the fleet'. Differentiator - this island owns only the pre-dispatch evidence requirement and the record that carries it; ticket machinery, batch size, and fleet sizing belong to neighboring seats.
- ▌ Mutant Hunt · island-dev-crew bundleThe merciless hardener gate - prove a story's tests can fail by injecting single-operator mutants into the diff's covered lines, then demanding zero non-equivalent survivors on the diff, never a global score, under a runtime budget cap. Reach for it after a story's suite goes green, when the question becomes whether the tests would actually catch a break - "run mutation testing on this diff", "harden this story", "any surviving mutants", "would these tests catch anything". On a diagnostic ask it reports the surviving mutants and stops there; writing the kill-tests is the repair mode a human asks for. Differentiator - this island owns only the metric and its budget; excusing an equivalent survivor belongs to mutant-excusal-ledger, and the gate's loop plumbing belongs to agent-guardrails and archipelago.
- ▌ Gherkin Gate · island-dev-crew bundleThe red-before-green gate on a story's Gherkin scenarios - the acceptance bar as an executable file bound to its story by content hash, proven failing before implementation and passing after, so it cannot decay the way a prompt rule decays. Reach for it when a story's .feature file exists and its acceptance evidence needs checking, when turning acceptance criteria into a gate instead of a steer, or when the user says "gate the Gherkin", "prove the scenarios red first", or "check the acceptance ledger". Differentiator - this island judges evidence only, and a check-shaped ask gets the verdict rather than an edit to the team's .feature file; authoring the scenarios belongs to the specifier seat and the UI-level QA procedure belongs to the qa-script seat.
- ▌ Thrash Watch · island-dev-crew bundleLive recognition of agent thrash - the struggle Uncle Bob says novices fail to see - read from a running agent's behavior (break-one-fix-another chains, circling, the outright give-up), plus the ordered intervention ladder that clears it. Reach for it while supervising a working agent that seems stuck, or when the user says "the agent keeps breaking things", "it's going in circles", "is this agent thrashing", or "it just apologized and changed two lines again". It reports by default - the destructive rungs of its ladder need an explicit repair ask. Differentiator - it watches the agent's behavior only; debugging the code is diagnose's seat, and the respawn rung executes via trajectory-hygiene.
- ▌ Egoless Fleet · island-dev-crew bundleReview posture for agent-written work - a review starts from the assumption that defects exist and goes hunting for them, and the record it leaves must be a found-or-falsified list rather than a bare approval. Reach for it when a diff an agent wrote is about to be approved, when a reviewer reports no issues at all, or when someone says 'looks good to me', 'LGTM, ship it', 'the agent says it is correct', 'nothing to flag'. Differentiator - this island owns the posture and what a review record must contain; who may review and what a verdict binds belongs to the cross-family-review ceremony, and durable findings graduate to a finding register.
- ▌ Margin Ledger · island-dev-crew bundleContinuous productivity accounting for a gate stack. Keeps a per-story ledger of agent-with-gates wall clock against an honest human baseline, defends Bob's observed 2-4x margin, and reports any gate that would push the margin below 1x for a human to cut. Reach for it when stacking or tuning quality gates over agent work, when a gated pipeline starts feeling slower than doing the work yourself, or on "margin ledger", "are the gates worth it", "productivity margin", "have we lost the game". Differentiator - it prices gates already in flight; whether to build at all is job-to-be-done's triage, and per-step model cost picks are model-routing's.
- ▌ Priority Zone · island-dev-crew bundleThe head-of-context budget for the standing prompt - the primacy/recency U-curve (Liu et al., TACL 2024), instruction-density decay (IFScale), and Horthy's smart zone, encoded as a hard size gate plus a placement lint that keeps must-follow rules where attention actually lands. Reach for it when a CLAUDE.md or AGENTS.md keeps growing, when agents obey rules only sometimes, or on "trim my system prompt", "why does my agent ignore its rules", "context budget for CLAUDE.md". Differentiator - this island owns the position and budget evidence and its enforcement; how many directives at once is instruction-density-cap's seat, the wording levers live in writing-for-agents, and prompt-vs-gate rule sorting is steering-audit's seat.
- ▌ Story Cadence · island-dev-crew bundleThe comparative planning doctrine for the agent era - heavy upfront specification is the 1970s waterfall temptation replayed, and it decays mid-run for the same reason it always did, so where change is near-free, small batches with feedback beat plan-perfection. Reach for it when sizing how much plan to write before agents run, when someone proposes a spec-first pipeline or 2025-26 spec-driven-development tooling, or says "write the full plan first", "spec everything before the agents start", "how much should we plan upfront". Differentiator - doctrine only, the argument and its economics; ticket machinery, decision maps, and mid-run divergence detection live on neighboring islands.
- ▌ Tests As Spec · island-dev-crew bundleThe unit suite written as the teaching document for the next fresh-context agent, carrying names that state behaviour, arrangement that shows the module's shape, one behaviour per test, and the paths a newcomer guesses wrong. Reach for it when the reader of your tests is an agent that will not open the implementation, when a suite has to survive being handed to a context that never met its author, or when someone says "tests as documentation", "my test names say nothing", "make this suite readable for the next agent", "the tests should explain the system". Differentiator - this island owns the unit suite as a document for a reader; the story-level acceptance contract and the proof that the suite bites belong to neighbouring islands.
- ▌ Gate Toolchain · island-dev-crew bundlePer-language tool map for the pack's two flagship gates - which CRAP scorer and which mutation runner exist for Java, Go, Clojure, JS/TS, Python, PHP, Rust and .NET, what coverage artifact each one eats, and whether it has an incremental or diff mode. Reach for it when standing a CRAP or mutation gate up in a language the harness has not gated before, or when a gate run got too slow to sit inside the agent loop - "which mutation tool for Rust", "is there a CRAP scorer for PHP", "our mutation run takes an hour", "what does crap4go read". Differentiator - this island picks the tool and states its input and mode facts only; the CRAP formula and thresholds belong to crap-gate, the mutation gate contract and diff scoping to mutant-hunt.
- ▌ Human Subagent · island-dev-crew bundleCurriculum design for Uncle Bob's education inversion - a junior joining an agent-heavy team is run AS an agent, given the same task briefs and held to the same deterministic gates, spending months deliberately unproductive, until they can be trusted to direct agents of their own. Reach for it when onboarding a junior or a career-changer into an agent fleet, when planning what a new hire actually does for their first six months, or on "how do juniors learn anything now", "what do we do with new grads", "how should I train someone to run agents", "is there still an entry-level path". Differentiator - it designs the drill and its scoring only; the reading list is strategy-shelf, live struggle-spotting is thrash-watch, and the machinery that delivers lessons and remembers them belongs to the Forge.
- ▌ QA Script Seat · island-dev-crew bundleThe relay's exit seat - turn the written human-viewpoint QA procedure into a deterministic executable script that drives the real system through its UI and returns a binary pass/fail verdict bound to its story. Use when a story's QA document exists and needs its executable gate, when the user says 'turn this QA doc into a script', 'make the QA procedure executable', or 'wire the QA verdict gate'. Differentiator - this seat generates and binds the executable; runtime assertion mechanics belong to computer-use-smoke, and a live agent's judgment is never the verdict.
- ▌ Specifier Seat · island-dev-crew bundleThe relay's intake seat - transform ONE settled human intent doc into EXACTLY TWO artifacts, a Gherkin acceptance spec and a human-viewpoint QA procedure, each with a defined contract that the coder seat and qa-script-seat consume. Use at the start of a relay pass, when a story is settled and ready for the gauntlet, or when the user says "specify this story", "write the Gherkin and the QA doc", or "run the specifier". Differentiator - this seat transforms, it never elicits; interviewing the human is grill's seat, and the specifier starts only after understanding is settled.
- ▌ Steering Audit · island-dev-crew bundleAudit a standing prompt (CLAUDE.md, AGENTS.md, a system prompt) into a before/after rule inventory. Every rule is classified generative (must shape generation, stays in the priority zone) or checkable (has a mechanical test, migrates out into a deterministic gate), because steering decays and gates do not (C3, C4). Use when a rules file has grown past trust, when hardening an agent fleet's prompts, or when the user says "audit my CLAUDE.md", "my agent ignores its rules", or "move rules into hooks". Differentiator - it owns only the classification and the migration; hook and denylist mechanics belong to agent-guardrails, tool-vs-model routing to model-routing, prompt wording to writing-for-agents.
- ▌ Strategy Shelf · island-dev-crew bundleThe old-books curriculum for engineers who direct agents - the eight-entry 1968-1999 shelf behind Bob Martin's old-books prescription - the three he names, plus the five the research brief sources around them, each entry cut down to the handful of ideas that still carry weight when a model writes the code, plus what to leave on the page in a book written in 1979. Reach for it when someone asks where strategic judgment comes from now that agents hold the tactical seat, when assembling a reading path behind the general's job, or when someone says "what should I read to get better at architecture", "are the old software books still worth it", "how do I learn design if I never write the code myself". Differentiator - curriculum content only, every entry traceable to one research brief; the teaching engine, the drills, and the argument for why fundamentals persist all live elsewhere.
- ▌ Threshold Port · island-dev-crew bundlePort a human engineering practice into an agent lane and recalibrate its numeric gate parameters (complexity caps, coverage floors, CRAP cutoffs) by controlled experiment. Use when adopting a human discipline for agents or picking a gate level, e.g. "should agents do strict TDD", "what CRAP threshold for agent code", "is our coverage floor right for the agent lane", or when a threshold was set by vibe or by polling the agents. Differentiator - this island owns the value/ritual/threshold split for numbers; markdown files and the measure-edit-keep loop belong to skill-tune.
- ▌ Comment As Spec · island-dev-crew bundleInterface comments treated as the machine-readable spec an agent acts from - every exported symbol states what the abstraction is and what a caller must honor, with no implementation showing through. Reach for it when a model calls a module it never opened, when reviewing or generating a public API, or when the user says "interface comment", "document the exports", "docstring lint", "the comment leaks implementation", or "can you use this without reading the body". Differentiator - this island owns the content of one symbol's comment and the leak test on it; token accounting for interface-first loading and duplicate knowledge across modules belong to neighboring islands.
- ▌ Coupling Budget · island-dev-crew bundleTreats every cross-module edge a change adds to the edge sets it is handed as spend against a declared budget - each added edge carries a written reason or the gate goes red, and the count itself is capped so prose cannot buy unlimited coupling. Reach for it when reviewing an agent diff that reaches into a module it never touched before, before fanning parallel agents across module lanes, or on trigger phrases like 'coupling budget', 'the agent added a new dependency', 'why does this import cross modules', 'how much coupling did that change add'. Differentiator - this island owns the budget on coupling a change ADDS; declared layering direction belongs to the sibling dependency-fence, computed instability and abstractness to stability-order, and cohesion sizing to component-cohesion.
- ▌ Essence Pointer · island-dev-crew bundleSpecify a tool by exemplar - a live working reference (unclebob/swarm-forge, crap4java/crap4go/crap4clj) is the whole spec for a customized local equivalent, with parity proven against the exemplar's own test suite where one exists. Use when adopting one of Uncle Bob's swarm or gate tools, porting a checker to your own stack, or the user says "build us our own crap tool", "point the agents at that repo and make one for us", or "study it, don't vendor it". Differentiator - the exemplar stays upstream as spec, never as dependency; trust of the built copy is a separate red-fixture acceptance, not this island's claim.
- ▌ Stability Order · island-dev-crew bundleComponent stability computed as a check instead of argued as an opinion - instability I = Ce/(Ca+Ce), abstractness A = abstract-types/total-types, distance from the main sequence D = |A+I-1|, with Stable Dependencies and Stable Abstractions violations failing loudly and naming the offending component. Reach for it when components are argued about rather than measured, when deciding which component owes you abstractions, or on trigger phrases like 'stable dependencies principle', 'instability metric', 'distance from the main sequence', 'which component should be abstract'. Differentiator - this island owns the computed NUMBERS only; the declared layering direction belongs to the sibling dependency-fence, and hook plumbing, gate infrastructure and evidence format live on neighbouring islands.
- ▌ Boredom Dividend · island-dev-crew bundleThe mining method behind the whole pack - systematically revive practices that were shelved only because they bored humans, not because they were wrong, and turn each one into an agent-run gate. Reach for it when hunting for new quality gates, when an old practice (a metric, an exhaustive check, an overnight analysis) feels worth a second look, or when someone says "what gates should we add", "revive that old practice", or "we stopped doing that because it was too tedious". Differentiator - this island owns the mining move only (shelf inventory, wrong-vs-tedious triage, agent-feasibility pass); each revived practice ships as its own gate island, and the plumbing every gate plugs into lives on the Forge islands it names.
- ▌ Boyscout Ratchet · island-dev-crew bundleBaseline-and-ratchet on touched files - every file an agent touches must leave measurably no worse than it arrived (CRAP score, cyclomatic complexity, coverage), with crap4j's project-allowance pattern so a legacy repo can adopt a gate it cannot pass today. Reach for it when an absolute quality gate would fail the whole repo on day one, when adding a no-regression check to a diff, or when the user says "boy scout rule", "ratchet the metrics", "no worse than baseline", or "we can't turn the gate on yet". Differentiator - this island owns the relative no-regression discipline and the legacy-adoption path; the CRAP formula, its threshold regimes, and the absolute per-function ceiling belong to the crap-gate island, whose numbers this one consumes.
- ▌ Dependency Fence · island-dev-crew bundleDeclare which module may depend on which and how the flow runs - a tight machine-checkable fence spec of layered direction, a checker run at the end that goes red on any outward edge, and exactly three sanctioned repairs when it does (invert the dependency, insert an interface, split the module in half). Reach for it when locking a codebase's dependency direction before agents build, or on trigger phrases like 'enforce clean architecture layers', 'which module may depend on which', 'the imports flow the wrong way'. Differentiator - it owns layering DIRECTION only, as a named extension of deep-modules' dependency-cruiser lane, and ships the Clean Architecture inward-only Dependency Rule as its flagship preset.
- ▌ Interface Budget · island-dev-crew bundlePrices a module in context tokens and holds the agent to interface plus interface comment plus tests before any implementation is opened, with every implementation load logged against a closed vocabulary of reasons; the ledger is a self-report and the shipped gate scores that accounting, not the behaviour. Reach for it when a task is about to pull whole source files into the window, when auditing why a session's context filled up, or when the user says "interface budget", "why did it read the implementation", "log the files you loaded", or "what did this module cost us". Differentiator - this island owns the token accounting and the load ledger only; depth vocabulary and the deletion test, interface-comment content, and cohesion sizing all live on neighbouring islands.
- ▌ Parnas Partition · island-dev-crew bundleThe decomposition criterion from Parnas 1972 - a module boundary is judged by the design decision it HIDES, never by the processing step it performs, so every proposed module must name its secret or the split is rejected. Reach for it when an agent proposes a new module, package, or file split, when a repartition plan comes back from interrogating the agents about structure, or when someone says "split this module", "where should the boundary go", "decompose by secrets", "what does this module hide", or "why does one change touch six files". Differentiator - this island owns only the question that decides WHERE a boundary goes; depth vocabulary and the deletion test, interface token economics, cohesion sizing, and duplicate-fact detection all live on neighbouring islands.
- ▌ Strategic Ledger · island-dev-crew bundleA tactical/strategic effort ledger for a repo worked by agents. Every unit of work is tagged as shipping behaviour or improving shape, the evidenced strategic share is held inside a target band, and each strategic claim is backed by a captured before/after structural metric. Reach for it when a repo gets harder to work in cycle after cycle, when deciding how much of the next cycle to book against design, or on "strategic ledger", "tactical debt", "how much refactoring should we book", "the 10-20% investment". Differentiator - it accounts for the SPLIT of effort between behaviour and shape, where margin-ledger accounts for agent-versus-human productivity on a story and job-to-be-done decides whether a thing is worth building at all.
- ▌ Tornado Detector · island-dev-crew bundleTrend alarm on files-touched-per-change - Ousterhout's tactical tornado caught while it is still blowing, now that a fast agent can smear one design decision across a codebase in an afternoon. Reach for it when a small feature keeps touching many files, when reviewing a burst of agent-written changes, or when the user says "why does this one-line change touch nine files", "are we accumulating change amplification", "run a tornado check", or "files per feature is creeping up". Differentiator - it raises one threshold alarm on one measured trend and hands anything actionable to the survey island; it never mines churn hot spots, ranks refactors, or judges whether an agent is stuck.
- ▌ Change Cost Probe · island-dev-crew bundleMeasure what a change actually costs in your repo - minutes or tokens per story through the pipeline - and gate the trend, because a rising cost curve is what accumulating mess looks like from outside the code, and because the measured price is what sets planning depth honestly. Reach for it before asserting that change is cheap, when stories that used to take an hour keep taking three, or on "is our cost of change rising", "measure cost per story", "how expensive is change here", "how much should we plan upfront". Differentiator - it trends one absolute cost in the agent lane across finished stories; the agent-versus-human ratio belongs to margin-ledger and live struggle to thrash-watch.
- ▌ Define Errors Out · island-dev-crew bundleOusterhout's define-errors-out review pass. Delete error cases by redesigning the API instead of adding another handler for them, with the count of raised and propagated error paths before vs after as the verdict. Reach for it when an interface is accumulating exceptions, when an agent keeps answering a failure with one more try/except block, or when someone says "define errors out of existence", "too many error paths", "this API throws too much", or "stop adding handlers". A complaint enters DIAGNOSE, which reports and edits nothing; REPAIR needs explicit authority to change a named module. Differentiator - this island removes branches by redesign; measuring complexity and enforcing its ceiling belong to crap-gate, and debugging a failure that actually happened belongs to the Forge.
- ▌ Abstraction Ladder · island-dev-crew bundleThe closing doctrine of the conversation - binary, then assembly, then the compiler, then the models, and at each rung the people below predicted ruin for the same reason and were wrong the same way, because a rung removes labor and never removes the job of organizing complexity. Reach for it when a new rung tempts you to drop a practice, when deciding whether a discipline is genuinely obsolete or merely inconvenient at this altitude, or when someone says "we do not need that now that agents write the code", "why would anyone learn the layer below", "that rule is legacy, cut it". Differentiator - this island owns the ladder argument, the rung-below weekend, and the discipline test a practice must survive before it is cut; which practices survive the port to agents and what number their gate gets belongs to threshold-port.
- ▌ Component Cohesion · island-dev-crew bundleComponent cohesion for the agent era - the REP/CCP/CRP tension triangle plus the constraint that a component is also a context-window unit, so what changes together must fit in one agent's context together. Reach for it when drawing or redrawing component boundaries, when commits keep fanning across many components, or when someone says "component cohesion", "common closure principle", "what changes together ships together", or "this module is too big for the agent to hold". Differentiator - this island owns the cohesion triangle and the context-fit constraint layered on it; deep-modules owns the design vocabulary, interface-budget owns interface-versus-implementation token economics, stability-order owns the computed component-coupling numbers, and coupling-budget owns the budget on the coupling a change adds.
- ▌ Trajectory Hygiene · island-dev-crew bundleMid-session context stewardship - the kill-vs-continue call on a running context window, because a context has momentum and contamination never washes out. Reach for it mid-task when a session drifts or fixates on a dead idea, the topic pivots, or the window fills past the smart zone - "this session feels poisoned", "should I kill this context or keep going", "the agent keeps going back to that". Differentiator - owns only the mid-session lane; the context ending ritual is handoff's seat and doc-level economy is writing-for-agents' seat.
- ▌ Known Dirty Fixture · island-dev-crew bundleAcceptance ritual for any gate, checker, linter rule, or validator before it may guard anything - a gate earns authority only by first failing RED on a known-bad fixture, then passing GREEN on a known-good one, both kept in the repo beside it. Use when generating or adopting a checker, wiring a new CI gate, or when someone says 'add a validator', 'prove the gate works', or 'the gate passed so we are fine'. Differentiator - this island owns gate acceptance only; what each gate checks is that gate island's own concern, and empirical skill improvement is skill-tune's seat.
- ▌ Plan Decay Detector · island-dev-crew bundleMid-execution divergence check for an agent fleet - the plan writes its assumptions down in checkable form, and a gate compares them against the repo as it actually is, exiting 1 to call for a re-plan the moment one has stopped being true instead of letting agents keep running off a plan that is no longer accurate. Reach for it before a batch resumes work from a plan written earlier, after a parallel agent or a human has touched the tree, or when someone says "is this plan still valid", "the plan says X but the repo has Y", "re-check the plan before we continue", or "did the plan drift". Differentiator - this island detects decay and calls the halt, which a human or a harness then has to honour; how big a batch should be is story-cadence's doctrine, and ticket machinery is a Forge concern.
- ▌ Manageability Review · island-dev-crew bundleThe human-judgment acceptance criterion for agent-written code once generation is free - a reviewer who cannot restate the change's control flow in a fixed number of sentences, from a single reading, bounces it, and the failure is a finding about the code rather than about the reviewer. Reach for it when generated diffs land faster than anyone can follow them, when deciding what a human's read of code has to produce before it counts as a review, or when someone says "it passes every gate but nobody understands it", "restate the control flow", "is this even reviewable", "intellectual manageability". Differentiator - this island owns the restatement test and nothing else; complexity expressed as a number belongs to the CRAP gate, interface-first reading economics to the interface budget, and the verdict ceremony to the Forge.
- ▌ Measurement Humility · island-dev-crew bundleStanding review obligation for every enforced metric - each one must name the behaviour it might corrupt and carry a date on which its threshold is re-examined, so a coverage number with no named way to game it, and a threshold nobody has revisited in years, both surface as breaches instead of as a green build. Reach for it when adding or inheriting a metric gate, when auditing the gates a repo already runs, or when someone says "is this threshold still right", "what could this gate corrupt", "nobody has looked at that number in years", "metric review date", "our coverage gate is theatre". Differentiator - this island owns a metric's own review obligation and nothing else; the CRAP formula belongs to crap-gate, retuning a number by experiment to threshold-port, and catching gamed coverage to coverage-gaming-audit.
- ▌ Mythical Agent Month · island-dev-crew bundleBrooks's law generalised to agent fleets - adding agents to a late project adds coordination surface, not progress, because communication paths grow as n(n-1)/2 while only partitionable work actually divides. Reach for it when deciding how many agents to run at once, when a fleet looks busy but the story is not landing sooner, or on "how many agents should I run", "add more agents", "fleet sizing", "would a bigger fleet be faster", "why is parallel slower". Differentiator - it sizes the FLEET and prices its coordination surface, so a serial relay of specialists and a crowd of peers can be compared as numbers rather than vibes.
- ▌ Good Island · island-dev-crew bundleA minimal fixture island proving the validator can go green - used only to test validate-island.py against a conformant structure, never distributed. Trigger phrases - "validate fixture", "green fixture test".
- ▌ Coverage Gaming Audit · island-dev-crew bundleAudit a green coverage number for tests that execute code without asserting on it - assertion-free tests, assertions only on mocks, swallowed exceptions, unreviewed snapshots - and route every finding to the mutation run that settles it. Reports by default - on an audit or diagnosis ask the finding list is the deliverable, the fix-until-green loop is not entered, and deleting a test is never interchangeable with adding the missing assertion. Reach for it when a coverage report or CRAP score goes green and you do not yet believe it, when inheriting a suite of unknown quality, or when someone says "is this coverage real", "assertion-free tests", "gamed coverage", "100% coverage but nothing is tested", "audit this test suite". Differentiator - this island owns detection of gamed coverage and the routing to mutation; the score being distrusted belongs to crap-gate and the mutation run that proves the case belongs to mutant-hunt.
- ▌ Mutant Excusal Ledger · island-dev-crew bundleThe honest 100% for mutation testing - mutant equivalence is undecidable (Budd & Angluin 1982) and 4-39% of mutants are equivalent in practice, so a literal 100% kill score is unreachable and every excused survivor carries a written equivalence argument in a ledger, with unexcused survivors blocking. Reach for it when a mutation run leaves survivors nobody can kill, when ruling whether a survivor is equivalent or merely unkilled, or when the user says "excuse this mutant", "equivalent mutant", "survivors are blocking the mutation gate", or "we'll never hit 100%". Differentiator - mutant-hunt owns the run and the kill-tasks; this island owns the ruling on what remains, and never lets "could not kill" be recorded as "equivalent".
- ▌ Values Not Disciplines · island-dev-crew bundleRule-inventory gate over a harness's quality rules - each rule must name the tool that measures it or carry an explicit advisory label, so a prose-only rule like be clean, write good tests, or keep it simple is ruled a breach instead of passing as a standard. Reach for it when checking whether a harness's quality bar is real, before trusting a rules file you inherited, or when someone says "every rule needs a tool", "is this rule actually enforced", or "we told the agent to write clean code". Differentiator - it asks of every rule, wherever that rule lives, whether a measuring tool is named at all; which side of the prompt-versus-gate line a rule belongs on is steering-audit's seat, and retuning a threshold's number is threshold-port's.
- ▌ Instruction Density Cap · island-dev-crew bundleA countable ceiling on how many directives one prompt may carry at once - instruction-following degrades with DENSITY as well as position, so this island defines what counts as a directive, caps the count per model class, and fails the prompt past it. Reach for it when a rules file has grown to dozens of simultaneous instructions, when an agent obeys some rules and silently skips others, or on "how many rules can I give it", "count the directives in my CLAUDE.md", "my agent ignores half its instructions". Differentiator - density only; the head-of-context position budget is priority-zone's seat and prompt-versus-gate sorting is steering-audit's.
- ▌ No Silver Bullet Triage · island-dev-crew bundleRoutes each work item by Brooks's 1986 split - accidental complexity, the difficulty our tools impose, goes to agents at full speed, while essential complexity, the difficulty of the thing itself, gets a recorded human decision before any agent touches it. Reach for it when handing a backlog to agents, when an agent has just decided something nobody chose, or on "can the agent take this one", "essential vs accidental complexity", "what has to stay human", "no silver bullet". Differentiator - this island owns the classification and the routing only, so whether the thing should exist at all, who owns the design, and what a human reads afterwards are three neighbouring seats, each named below.
- ▌ Structure Interrogation · island-dev-crew bundleExtract the agent's own model of the structure it actually built - what the modules are, how each interrelates with the others, how they talk - then correct that model against the structure you intended. Reach for it after agents have built or grown a codebase and nobody has checked what shape it really took, or when the user says "interrogate the structure", "ask the agents what they built", or "what are the modules here, really". Differentiator - arch-survey mines change-history for refactor candidates; this island interrogates the agent's mental model, and the re-partition design stays human-held.
- ▌ Acceptance Surface Review · island-dev-crew bundleFixes what a human actually reads once agents write the code - the Gherkin acceptance spec and the QA procedure, never the implementation, with the surface widened by blast radius until code itself enters it on genuinely dangerous paths. Reach for it when standing up human review over agent-generated work, when a reviewer has started walking generated diffs line by line, or when the user says "what am I supposed to review", "acceptance surface", "do I need to read this code", or "criticality tiers for review". Differentiator - this island owns the review SURFACE and its criticality scaling only; the artifacts are the specifier seat's, their red-green binding is the gherkin gate's, and the verdict ceremony belongs to the Forge.
- ▌ Conceptual Integrity Owner · island-dev-crew bundleNames the one human mind that owns a system's design and holds that no merged design change is legitimate without that owner's recorded sign-off - the strategic seat agents do not fill, because they are strong tactically and weak strategically. Reach for it when a design decision is about to land unowned, when the answer to "who decided this" is a committee or an agent poll, or on "who owns this design", "conceptual integrity", "do we need an architect anymore", "add another agent to figure out the structure". Differentiator - this island owns WHO holds the design and what their sign-off covers; the head-bound verdict ceremony belongs to the Forge, the human reading surface to acceptance-surface-review, and essential-versus-accidental routing to no-silver-bullet-triage.
- ▌ Verdict Flip Island · island-dev-crew bundleA deliberately broken fixture whose gate reports a breach with exit 1 but flips to a false clean exit 0 when its stdout is closed, so closed-stream-check.py must flag it as a LEAK. Never distributed; exists only so the check's verdict-comparison is a captured red run. Trigger phrases - "verdict flip fixture", "closed-stream red test".
- ▌ Hostile Proof Island · island-dev-crewA deliberately hostile fixture - its proof block carries an exfiltration command, so verify-proofs.py must REFUSE it rather than run it. Never distributed; exists only so the refusal is a captured run instead of a claim. Trigger phrases - "hostile proof fixture", "refusal red test".
- ▌ Proof Grammar Island · island-dev-crew bundleA fixture whose proof block carries every shape of the proof grammar at once - a bare island-relative script, a distant exit-report line, an off-allowlist command, a usage template, and a candidate with no exit code. Never distributed; exists so the extractor's classification is a captured run rather than a claim. Trigger phrases - "proof grammar fixture", "pending red test".
- ▌ Unsafe Export Island · island-dev-crewA deliberately hostile fixture whose proof block exports a poisoned PATH before an allowlisted command, so verify-proofs.py must REFUSE the export rather than replay it into the shell that runs the proof. Never distributed; exists only so the refusal is a captured run. Trigger phrases - "unsafe export fixture", "PATH poison red test".
- ▌
- ▌ Fail Open 3 Island · island-dev-crew bundleA deliberately broken fixture whose gate reports a breach with exit 1 but downgrades to this island's separately documented empty-input exit 3 when stdout is closed, so closed-stream-check.py must bind acceptance to the command's expected code. Never distributed; exists only so the acceptance rule is a captured red run. Trigger phrases - "fail open downgrade fixture", "wrong non-verdict red test".
- ▌ Unsafe Setup State Island · island-dev-crewA hostile proof fixture whose Bash setup and annotated compounds try five persistent shell-state mutations around allowlisted proofs. The mutations must be refused while ordinary setup remains replayable. Never distributed; exists only as a captured regression. Trigger phrases - "unsafe proof setup state", "function shadow proof".
- ▌ Escaped Name Island · island-dev-crewA deliberately hostile fixture that reaches network and device-destructive primitives through ordinary shell quoting and escaping - `cur\l`, `'cur'l`, `CURL` - so verify-proofs.py must parse the line into words the way bash does and REFUSE all of them rather than run them. Never distributed; exists only so the word-parse refusal is a captured run instead of a claim. Trigger phrases - "escaped name refusal fixture", "quoted primitive red test".
- ▌ Proof Grammar Bash Builtin · island-dev-crewA proof grammar fixture in which an off-allowlist Bash builtin sits between an allowlisted command and an exit report. The earlier candidate must not borrow the builtin's status. Never distributed. Trigger phrases - "Bash builtin proof boundary", "builtin borrowed report test".
- ▌ Proof Grammar Continuation · island-dev-crewA proof grammar fixture for Bash backslash-newline removal. A welded command name must not be reconstructed with an invented space. Never distributed. Trigger phrases - "proof continuation grammar", "welded command proof test".
- ▌ Pending Only Island · island-dev-crewA fixture whose whole proof block is candidates with no documented exit code, so nothing runs and PENDING is the only class present - the case where `--strict` and the nothing-ran path both apply at once. Never distributed; exists so the precedence between them is a captured run rather than a claim. Trigger phrases - "pending only fixture", "strict precedence red test".
- ▌ Quoted Status Island · island-dev-crewA fixture whose flipping gate carries a quoted `$?` and a quoted `>&-` as ARGUMENT DATA, so a substring test excludes it from the closed-stream probe and its 1-to-0 flip goes unseen. Never distributed; exists only so the narrowed self-probe test is a captured red run. Trigger phrases - "quoted status fixture", "self-probe skip red test".
- ▌ Absolute Path Island · island-dev-crewA deliberately hostile fixture that reaches a network and a device-destructive primitive through a PATH rather than a bare word, so verify-proofs.py must judge basenames and REFUSE both rather than run them. Never distributed; exists only so the path-qualified refusal is a captured run instead of a claim. Trigger phrases - "absolute path refusal fixture", "basename refusal red test".
- ▌ Proof Grammar Comment Lexing · island-dev-crewA proof grammar fixture whose data contains hashes protected by ANSI-C quoting and an escaped space. The verifier must not truncate either command into a false match. Never distributed. Trigger phrases - "proof comment lexer", "escaped hash proof test".
- ▌ Proof Grammar Export Boundary · island-dev-crewA proof grammar fixture in which a safe export sits between an allowlisted command and an exit report. The export owns that status while remaining setup for later proofs. Never distributed. Trigger phrases - "export report boundary", "safe export continuation test".
- ▌ Proof Grammar Report Compound · island-dev-crewA proof grammar fixture in which report-shaped prefixes carry pipes, commands, or extra operands. Only a complete pure report may annotate the candidate above it. Never distributed. Trigger phrases - "compound exit report", "report prefix borrowing test".
- ▌ Unsafe Bare Assignment Island · island-dev-crew bundleA hostile proof fixture that assigns PATH without export before an allowlisted command. The assignment must be refused before its shim can forge the documented verdict. Never distributed. Trigger phrases - "bare PATH assignment", "assignment poison red test".
- ▌ Command Position Island · island-dev-crewA fixture whose flipping gate carries shell-looking status data in Python argv and follows a status-printing pipeline producer, so closed-stream-check.py must inspect only the actual command position before excluding a self-probe. Never distributed; exists only as a captured red regression. Trigger phrases - "command position fixture", "argv self-probe bypass".
- ▌ Proof Grammar Empty Assignment · island-dev-crewA proof grammar fixture in which an ordinary block-local variable is deliberately emptied before an allowlisted command. The empty binding must replay. Never distributed. Trigger phrases - "empty proof assignment", "empty local replay test".
- ▌ Unsafe Empty Assignment Island · island-dev-crewA hostile proof fixture that empties PATH without export before an allowlisted command. The binding must be refused rather than dropped from the grammar. Never distributed. Trigger phrases - "empty PATH assignment", "empty assignment poison red test".
- ▌ Actual Self Probe Island · island-dev-crewA fixture containing only a command that deliberately closes its own stdout, so closed-stream-check.py must count it as an own-probe and return NOTHING PROBED rather than re-probing shell redirection syntax. Never distributed; exists as a regression for command-position parsing. Trigger phrases - "actual self probe fixture", "redirection position regression".
- ▌ Quoted Self Probe Island · island-dev-crewA fixture whose flipping gate carries `$?` inside DOUBLE quotes, where a shell expands it, so a self-probe test built on expansion drops the command from the closed-stream probe and its 1-to-0 flip goes unseen. Never distributed; exists only so the position-based self-probe test is a captured red run. Trigger phrases - "double quoted status fixture", "self-probe expansion red test".
- ▌ Proof Grammar Malformed Command · island-dev-crewA proof grammar fixture in which malformed off-allowlist shell rows sit between allowlisted commands and exit reports. Invalid syntax must still terminate report search. Never distributed. Trigger phrases - "malformed proof command", "unterminated report boundary test".
- ▌ Nested Substitution Island · island-dev-crewA fixture whose flipping gate carries a benign command substitution that closes only the substitution stream or reports only the substitution status, so closed-stream-check.py must still probe the outer gate. Never distributed; exists only as a captured red regression. Trigger phrases - "nested substitution fixture", "outer gate exemption bypass".
- ▌ Proof Grammar Unrecognized Command · island-dev-crew bundleA proof grammar fixture in which an off-allowlist command sits between an allowlisted command and an exit report. The earlier candidate must not borrow the later command's status. Never distributed. Trigger phrases - "unrecognized proof command", "borrowed exit report red test".
- ▌ Unsafe Refusal With Eligible Island · island-dev-crewA closed-stream harness fixture combining a refused PATH block with an independent eligible proof. The eligible probe must not launder the refusal into green. Never distributed. Trigger phrases - "mixed refusal proof", "closed stream refusal laundering test".
- ▌ Proof Grammar Host Independent Command · island-dev-crewA proof grammar fixture in which a syntactically command-shaped but nonexistent tool sits between an allowlisted command and an exit report. Classification must not depend on the review host PATH. Never distributed. Trigger phrases - "host-independent proof boundary", "absent command borrowed report test".