Deep analysis

Skill stevenfackley/opencode-amplifier/.opencode/skills/deep-analysis

Contract-governed OpenCode config that amplifies constrained LLMs (Sonnet 4.5, GPT-5.1, cheap corp models) into near-frontier coding agents: multi-agent pipeline with per-agent models, independent test-gen + locked tests, golden-pattern corpus, cross-model review, and an eval harness.

Install

npx -y skills add stevenfackley/opencode-amplifier --skill deep-analysis

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when orchestrating /deep-design, /deep-debug, /deep-review, /deep-read, or /council — packet assembly and artifact discipline for fresh-context worker fan-out

SKILL.md

2.6 KB, as published. Nobody here has run it

Deep analysis — disposable workers, durable artifacts

Calls are cheap; context is the scarce resource. Heavy reading and reasoning happen in fresh-context workers that die after returning a capped artifact. The orchestrating session holds packets and artifacts only — rot cannot accumulate in a context that is thrown away.

Packet assembly (step 1 of every deep command)

A packet is everything a worker sees. Build it from:

The question/goal, verbatim.
TASK.json invariants + forbidden files (if a ledger exists).
File REFERENCES (paths + line ranges + a sentence on relevance). Inline full contents only when small; the mapper is the only worker built for raw dumps.
The matching section of memory/briefs/CODEBASE-BRIEF.md, if one exists.

Packet cap: ~200 lines. Needing more is the signal to run /deep-read first and reference the brief instead.

Staleness check (before using any brief)

Compare the brief's frontmatter git_sha to git rev-parse HEAD. If drifted, warn the user and offer an incremental /deep-read refresh (changed subsystems only) before proceeding.

Artifact discipline

Workers return text; only the synthesizer writes files. One artifact per command run, following the matching template in .opencode/templates/.
Ephemeral analyses → .analysis/ (gitignored). Durable briefs → memory/briefs/ (committed).
Caps are hard: 150 lines per artifact (40 per brief section), frontmatter always filled (command, date, git_sha, models, inputs, status).
The orchestrator NEVER pastes raw worker output into the main thread. Report the artifact path + a compact verdict + the dissent register. That's it.

Blindness

Fan-out workers answer the SAME packet independently. Never include one worker's output in another worker's input — correlation destroys the ensemble's value. Only the judge and synthesizer see multiple outputs, and they see them labeled by model.

Degradation ladder

Worker output non-conforming (oversized/off-format) → retry that worker ONCE with a stricter format reminder → still bad? proceed with survivors. A council of 2 is a council.
All workers fail → answer single-model inline, but the artifact says status: degraded and the report says so honestly.
A model missing from the proxy → use the fallback named in the agent file's frontmatter comment; /setup verifies all council models against opencode models.