weekly-2026-03-26

99 commits, 8 bot PRs, 7 CI routing patches, 0 finished features that actually ship.

Orchestrator: Inflation to Self-Cover

This week the orchestrator got structural evolution:

Two-level dispatch + employee queues
Deterministic rebase of worktree onto main
Manager-led LLM triage for task assignment
Single-worktree-per-employee enforcement
PR convergence and orphan cleanup automation
Review gates + PR/merge feedback dedup

Biggest feature set of the week. But here's the uncomfortable question: how many of these features are fixing last week's design flaws?

single-worktree-per-employee means multi-worktree was previously allowed
deterministic rebase means rebase was previously non-deterministic
review gates means review wasn't previously enforced

Verdict: The orchestrator is inflating through self-iterating — each new feature patches the previous design mistake. That's not evolution, that's snowballing.

CI Runner Routing: Config as Debt

7 CI commits this week trying to route jobs to the right runner:

route all Gemini jobs to macOS self-hosted runner
route label jobs to self-hosted runner  
route debugger job to self-hosted runner
use pwsh shell for Gemini workflows on Windows self-hosted runner
revert dispatch/fallthrough to ubuntu-latest, keep self-hosted for Gemini only
...

Seven commits to answer: "which job should run on which runner?"

The problem isn't GitHub Actions knowledge. It's that there's no stable mental model of the runner topology, so every routing decision is trial-and-error.

Questions that don't have documented answers:

Which jobs must run on macOS?
Which must run on Linux?
Which can fall through to GitHub's default runners?

No docs, no principles. Just commit messages reading "oops, wrong runner again."

Issue #69: The Automation That Kept Hitting the Wall

Blog auto-deploy failed. 8 PRs opened. All by outbird-autodev[bot].

This isn't self-healing automation. This is a machine ramming into the same wall, backing up, and ramming again. Each PR: "retry with longer timeout" → fails → "retry with health check" → fails → "retry with retry logic" → finally human介入.

Lesson: Adding retry to a failing script doesn't fix the problem. It masks the symptom while the real issue compounds.

Qdrant Vector Dimension Mismatch: The Data Contract That Never Existed

Issue #80: vector dimension inconsistency between orchestrator write and qdrant read.

This isn't a bug. This is evidence that the data contract was never formally defined.

Orchestrator writes vectors
Qdrant stores vectors
Neither side ever sat down and said "this is the dimension, this is the schema"

Data layer always gets deprioritized until it becomes a blocking issue. Classic.

OpenClawBridge: Another Bridge Layer

orchestrator-api → OpenClawBridge → persona bot. Three layers of bridging. Each layer is technical debt.

If you need a bridge to make two systems "interact," it means the boundary definitions of those systems were wrong from the start.

Brutal Self-Critique

Pattern #1: Commit Count = Progress (Subconscious)

99 commits this week. 71 from the same human.

Not because product needed it. Because commit count = perceived velocity is still running in the background.

Real progress this week: the blog deploy finally works. Qdrant issue surfaced and documented. Those are two things. Two.

Everything else was either patching a previous patch or adding a new layer of indirection.

Pattern #2: Automation Theater

20 bot commits + 71 human commits, side by side, is darkly ironic.

Bot runs CI → fails → retry → fails → retry... Human commits feature → design flaw surfaces → commits new feature to patch it...

Neither system stopped to ask: are we solving the right problem?

Pattern #3: Reactive CI Topology

Seven CI routing commits aren't "iterative development." They're 现场学习 (on-site learning) without top-level design.

Every failed commit was preventable. The fix is not another routing rule. It's a one-page doc: "Runner Topology — What Goes Where."

Pattern #4: Data Layer Always Last

Qdrant vector inconsistency wasn't caught during development. It was caught during a retrospective issue. This means the data contract was never part of the development checklist.

Feature ships, data schema gets wing-and-a-prayer. This is how silent data corruption happens — not from malicious actors, but from "we'll define it later" engineering culture.

Course Correction

No new orchestrator features next week. Run the existing dispatch/rebase/worktree logic against real tasks. Validate with outcomes, not commit logs.
Write the Runner Topology doc. One page. End of CI routing mystery. No more trial-and-error commits.
Audit Qdrant data contracts. Dimension, metadata schema, query contract. Define them before the next feature touches them.
Deflate the commit快感 threshold. Progress isn't "record-high commit count." Progress is "X problems that used to exist no longer exist."

Quote of the Week

Every layer of abstraction you add is a silent vote of no confidence in the previous one.

Generated by TestUser bot from git log. The brutal view is for debugging purposes only.