Reading Other People's Bug Trackers: The Bets Behind fakoli Are Getting Called

The views expressed here are my own and do not represent those of any current or former employer.

This week I read four hundred GitHub issues about software I didn’t write.

More precisely: I pointed research agents at the issue trackers of the two most successful tools adjacent to mine — superpowers, the dominant skills framework for disciplined agent coding, and claude-task-master, the most adopted PRD-to-tasks system — and asked one question. Not “what do these tools do?” Marketing answers that. The question was: where do they hurt the people who love them?

A bug tracker is the one place a product can’t perform. It’s thousands of users, under no NDA, telling you exactly which design decisions failed them, ranked by emoji reactions. It is free strategy research, and almost nobody reads it.

Here’s what it said — and what it confirmed about the bets I’ve been making with the fakoli stack.

What the trackers actually said

Prose rules fail in both directions. Superpowers’ most instructive issues are the ones where its mandatory process simply… didn’t happen. An agent decides a task is “straightforward” and exempts itself from review. Auto mode declares “I’ll make reasonable calls and proceed” and walks past every gate. And the same prose that fails to stop a determined model also over-fires: users report a five-line change generating six or seven agent invocations of pure ceremony. The maintainer tried prose-based calibration and abandoned it — the model used calibration language as permission to skip steps. Rules written in English are suggestions to a language model, in both directions.

Sessions die and take the truth with them. The second superpowers cluster: plans whose checkboxes never get checked, executions that complete with the tracking document still showing nothing done, twenty-minute planning runs that time out and lose everything. State that lives in a context window dies with the context window.

A single mutable file is a corruption machine. Task-master’s deepest wounds are structural. Its entire task state lives in one JSON file, rewritten wholesale on every operation. The tracker reads like a pathology report: status updates that silently mutate ID types on every write, a multi-tag system that destroyed sibling workstreams on expand, and — my favorite war story in the whole corpus — a power user running sixty-five concurrent agent processes whose state file corrupted so reliably that he wrote his own replacement server rather than keep using the official one. There is no claim concept, so two terminals happily work the same task. File locking arrived twenty months after launch.

People resent paying twice. The third-most-reacted task-master issue ever is a demand to stop requiring separate API keys when users already pay for Claude. And the most-wanted integration — GitHub Issues sync — has been open since March 2025, deflected to a commercial waitlist.

I want to be honest about what reading this felt like. It was not schadenfreude. Both tools are good; one of them is vastly more adopted than mine and earned it. It felt like being handed the failure analysis for my own roadmap, written by someone else’s users, eighteen months before my users would have written it.

The bets, and who’s calling them

Every architectural bet in the fakoli stack now has external evidence on the table.

The wave bet. Two years of running multi-agent waves — parallel specialists with file ownership, dependency-ordered dispatch, review gates between phases — was a bet that this is how agent development would settle. Anthropic has now shipped it natively: Agent Teams with shared task lists, claiming, file locking, and dependency blocking are a platform feature. And superpowers’ single hottest issue — over a hundred reactions, more than double anything else — is users begging for exactly that integration. When the platform vendor and your competitor’s user base independently converge on your execution model, the bet is called. It’s also a warning: the dispatch mechanics themselves are becoming commodity. The value moves to what survives the session.

The enforcement bet. fakoli moved its critic gate from prose to hooks this month — a PreToolUse hook that denies dispatch until review happens, with a state machine that mechanically forces re-review after every fix. Superpowers’ tracker is the control group: the same gate, written as prose, self-exempting in the wild. I keep an “enforcement ladder” in the architecture docs now — every rule graded Hard or Convention, and nothing gets marketed as enforced unless it sits in the top half. That table exists because I caught my own docs claiming “cannot be skipped” about a rule nothing was enforcing.

The state bet. fakoli-state’s design — an append-only event log as the source of truth, SQLite as a disposable projection, lease-based claims, evidence required before anything counts as done — looked like over-engineering for a personal stack. Task-master’s corruption pathology is what the alternative looks like at scale: normalize-on-write schema drift cannot happen to an append-only log, lost updates cannot happen when the log is the truth, and “two terminals on one task” cannot happen when claims are leases that expire. The sixty-five-process user who abandoned ship? That’s the exact persona a state engine with real concurrency semantics exists for.

The verification bet. Nobody in either tracker even asks for evidence-gated completion — and both trackers are full of its absence: tasks marked done on say-so, parsers reporting success while upstream calls returned auth errors. The 2026 consensus in agent-verification writing is “inspect artifacts, never transcripts.” fakoli-state shipped that as a database transition, not a convention. Being ahead of the complaint curve is a strange place to be; the complaints arrive after people scale, and the field is scaling now.

Growing on evidence, not enthusiasm

The ecosystem grew this week, and the way it grew matters more than the growth.

The crew gained its ninth agent: warden, a read-only security auditor — injection surfaces, secret leakage, supply chain, and the permission surfaces of Claude Code plugins themselves. A second standing gate beside the critic: one reviews for correctness, the other for exploitability. But the reason warden exists isn’t that security sounded important. A field scan across every serious subagent roster found the security auditor was the only consensus archetype fakoli-crew lacked — it appears in every major collection and three separate first-party Anthropic surfaces. The same scan recommended holding a test-author agent until the pipeline has a TDD wave to put it in, and adding nothing else. Nine agents, not sixty. The sprawling-roster experiment has been run by others, and the market already graded it.

The same discipline now runs in reverse: the positioning docs keep a retired ideas table. This month it gained its first entry — a unification refactor that looked obviously right until reconnaissance showed both surfaces were already thin wrappers over the same managers, and the refactor would have added two hundred lines of indirection to save fifty. The idea is dead, the evidence is written next to the grave, and it can’t be re-proposed by a future me who forgot.

And the state engine closed its scoring loop: complexity scores now drive decomposition — tasks above threshold queue for expansion automatically instead of decorating a report. That idea came straight from task-master’s tracker too, where users have spent a year asking for scores that recurse instead of going stale.

The lesson I keep relearning: my own conviction is the weakest evidence I have access to. Four hundred strangers arguing in someone else’s issue tracker is better data than any architecture review I could run on myself.

Where this is heading

The direction hasn’t changed since last post — state that travels with the repository, hash-chained events, git as the sync layer — and the field research only sharpened it. Task-master users describe task files in git as “conflict armageddon.” An append-only log with collision-free IDs merges as a union. That contrast is the whole pitch, written by someone else’s users.

What did change is my confidence in the division of labor. The platform will own dispatch — Anthropic is better at shipping orchestration than any plugin author will ever be, and that’s fine. The opinion layer stays cheap and portable. The moat was, and remains, the boring part: durable, evidence-bearing, project-scoped truth.

Read your competitors’ bug trackers. Not their landing pages, not their launch threads — their trackers, sorted by reactions, oldest grievances first. Your roadmap is in there, pre-validated, written by the only people who never lie about software: the ones stuck using it.