AGENT FAILURE LAB ARMED
PLAYABLE PORTFOLIO / 001
BREAK
THE AGENT.
KEEP THE
EVIDENCE.
Inject a known failure into an agent run. Follow the diagnosis. Leave with the receipt.
Deterministic UI simulation based on real open-source tool contracts. No model call. No staged success.
RUN / 0xC0DE0003INVARIANT Every tool symbol must resolve through one runtime registry.
00INSPECTcontext-lens
01BUILDplugin-starter
02DIAGNOSEdsh-doctor
03TRACEtrajectory
04RECEIPTreport-studio
FAULT LOCATEDdsh-doctor
ARMED
00 / 05 CONTRACTSLIVE / STDOUTWAITING
+
Choose a fault.
Run the system.
Interrogate the result.
SELECTED OPEN-SOURCE SYSTEMS / 2026
EVIDENCE
ARCHIVE
Five small tools. One opinion: an agent system should be inspectable before, during, and after the run.
00CROSS-LAYERAgent Context LensShows the instructions, skills, rules, and MCP configuration an agent will actually consume before it acts.LOCAL-FIRSTDETERMINISTIC90-TEST MATRIX01BUILDdsh-plugin-starterA zero-dependency scaffold for a host plugin, tool, runtime skill, tests, CI, and manifest.ONE COMMANDNO BUILD STEPSMOKE TEST02DIAGNOSEdsh-doctorOne-command health checks for versions, profiles, session logs, ports, and duplicate packages.ZERO DEPJSON ENVELOPEACTIONABLE CHECKS03TRACEdsh-trajectoryTurns durable session logs into one portable execution trace with tools, errors, and token ledger in context.OFFLINE HTMLMULTI-FRAME ZSTDSHA-25604DELIVERdsh-report-studioConverts sessions into reports and handoffs without letting the narrative outrun the evidence.4 TEMPLATESPATH CONFINEMENTRECEIPTS
FIELD NOTES / PUBLIC DEBUGGING
I don’t just use the harness.
I inspect its failure modes.
21
OUTSIDE THE TERMINAL
Business student.
Stage director.
Agent-tool builder.
I study how technical systems become legible products. Before building agent tooling, I coordinated a 50-person theatre team and shipped work for an audience of 5,000. Different medium; same obsession with state, timing, and what the audience can actually verify.