Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add commands/sjarmak/coding-agent-workflows/researchgit clone --depth 1 https://github.com/sjarmak/coding-agent-workflowsWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00023 | $0.00436 |
| Opus 5 | $0.00012 | $0.00218 |
| Sonnet 5 | $0.00005 | $0.00087 |
| Haiku 4.5 | $0.00002 | $0.00044 |
Grade A, and why
research scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
What it actually says
Workflow: Research
For open questions and design decisions where the obvious answer may be wrong. Three phases, each a distinct cognitive mode, kept separate so one doesn't contaminate the next.
Inputs
| Input | Source | Description |
|---|---|---|
question |
caller | The topic or decision to research. |
context |
caller | What's known, constraints, prior art. |
Steps
1. diverge
Run diverge: spawn independent investigation angles with uncorrelated
context, so findings aren't anchored to a single framing. Breadth first; the
goal is to surface options and evidence, not to decide. Where the runtime has
subagents, fan the angles out as independent agents; otherwise work them one at
a time in separate passes, resetting framing between each so they stay
uncorrelated.
Exit: multiple independent findings, each with its own evidence.
2. converge (needs: diverge)
Run converge: reconcile the divergent findings into a single synthesis with a
recommended direction. Name the trade-offs explicitly; don't hide the ones the
recommendation loses on. Where subagents are available this runs as a structured
multi-agent debate; otherwise reconcile the findings directly in one pass.
Exit: one recommended direction + its trade-offs, traceable to the findings.
3. premortem (needs: converge)
Run premortem on the recommendation: assume it shipped and failed, then
enumerate the most likely failure modes and what would have to be true for each. Feed the
serious ones back as constraints on the direction.
Exit: a direction that has survived its own failure analysis, with the known risks written down.
Note
Research output is decision support, not artifacts. Capture the decision and
its rationale (see the architecture-decision-records skill); the intermediate
exploration is ephemeral.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 54 lines · 23 tokens per session scan A 5f4efa545804
research is a command published in the GitHub repository sjarmak/coding-agent-workflows (2 stars, last pushed 1mo ago), licensed MIT. It adds 23 tokens to every session and 436 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other commands, from other repositories
plugin-update
Upgrade DeepInit to the latest version — pulls the newest from the marketplace (on one confirm), then guides the host-correct reload. Ends the "is my plugin stale?" dance.
check
Is the context layer still true? 0-token staleness + broken-citation audit (no LLM, CI-friendly). Add --status for the fast hash-only subset.
customize
Tune a DeepInit run with buttons — depth, issue detection, outputs, scope, cost, and the freshness/notification settings (disable the nudge, change its cadence/time-window) — no flags to type. Opens a native multiple-choice picker, then runs.
version
Which DeepInit version is actually running right now? Prints the LOADED version, checks it against the on-disk version, and tells you if you need to reload. No analysis.
doctor
DeepInit preflight — tools, scope, resolved config (and whether it's valid), enabled issue families, estimated cost. 0 tokens, no LLM. Offers to install the freshness hooks.
help
Show all DeepInit commands + key options, grouped and ordered by how often you'll use them. Instant, no analysis.