Claude 3t

claude-3t (plugin)

Structured session management for Claude Code: Sonnet + Advisor + file-based memory. No installer, no server, no container.

You (Sonnet — implements, decides, reviews)
  → /advisor  (Opus — hard/irreversible reasoning; produces no artifacts)

Memory: .claude/context/
  → Hot:  MEMORY.md, CONTEXT.md, EXECUTOR_MEMORY.md  — loaded each session
  → Cold: cold/INDEX.md + cold/*.md  — indexed, loaded on demand

---

Install

claude plugin marketplace add JimCline/claude-3t
claude plugin install claude-3t@claude-3t

Local development (no push needed):

claude plugin marketplace add /path/to/claude-3t

---

Per-project, even when installed globally

The plugin gates on a committed marker (.claude/context/cold/INDEX.md):

• No marker → hook emits nothing; session is untouched.
• Marker present, session start/resume → the executor is asked "Start the workflow for this session? (yes/no)". Yes runs /claude-3t:3t-start; no proceeds normally.
• Marker present, compact/clear → if the session was active (.claude/.3t-active flag set by /3t-start), the protocol is reloaded automatically.

---

Use

/claude-3t:3t-init        # once per project — scaffolds memory + settings.json + marker, then restart
/claude-3t:3t-start       # each session — loads protocol + hot memory
/claude-3t:3t-status      # memory state, cold index, model status
/claude-3t:3t-tokens      # measured session token usage by category / exchange
/claude-3t:3t-checkpoint  # write a session snapshot now
/claude-3t:3t-debrief     # post-work debrief — route lessons to memory / upstream
/claude-3t:3t-feedback    # capture a baseline-protocol gap for upstream
/claude-3t:3t-leaving     # departure protocol for autonomous work
/claude-3t:3t-remove      # deactivate or fully remove from this project
/claude-3t:3t-reset       # wipe to zero so /3t-init can re-run (destructive — confirms first)

---

Advisor

The advisor is the native advisorModel escalation: when you need stronger judgement on a hard or irreversible decision, type /advisor at the prompt. Opus reasons through it and returns; the executor continues. It produces no artifacts — only sharpens your decision.

/claude-3t:3t-init writes advisorModel: "opus" to the project .claude/settings.json. Change it to any model alias or pinned ID; it takes effect next session. It is a native Claude Code setting — survives plugin updates.

---

Memory model

Layer	Files	Committed?	Loaded
Hot	MEMORY.md, CONTEXT.md, EXECUTOR_MEMORY.md, OVERRIDE_LOG.md	CONTEXT.md only	in full each session
Cold	cold/INDEX.md + cold/*.md	yes	index always scanned; topic files on demand

Store only what is durable (true a year from now), cold (not needed every task), and expensive to re-derive. Never store volatile facts (signatures, config values, versions) — re-derive those from code.

---

Token tooling

bin/session-tokens.mjs reads exact API-reported counts from any session transcript — current session auto-detected, or pass a UUID / .jsonl path:

node bin/session-tokens.mjs                   # current session
node bin/session-tokens.mjs <uuid>            # any past session by id
node bin/session-tokens.mjs /path/to/file.jsonl

Reports output, cache-creation, uncached input, and cache-read separately (do not sum them — the cache-read prefix is re-counted every turn).

bin/token-baseline.mjs statically estimates what every context artifact costs, so optimizations are judged by a measured delta:

node bin/token-baseline.mjs [project-path]

---

Auto-verify (.claude/.3t-verify, opt-in)

Drop one shell command into .claude/.3t-verify (e.g. npm test -s or dotnet build). The post-agent.mjs hook runs it after tool-call-heavy work and injects a PASS/FAIL summary — one fewer tool call for the executor to spend on a build check. 120s timeout; absent file → no-op.

---

Removing / deactivating

Run /claude-3t:3t-remove:

• Deactivate — drops .claude/.3t-disabled; 3t goes dormant while files stay. Reverse: rm .claude/.3t-disabled.
• Full remove — deletes scaffolded memory + cold files, strips the 3t .gitignore block and advisorModel key. Committed files recoverable with git restore.

To remove the plugin globally: claude plugin uninstall claude-3t.

---

Contributing feedback

1. Capture — run /claude-3t:3t-debrief after a rough session (routes lessons to project memory and writes protocol gaps to .claude/context/3t-plugin-feedback.md), or /claude-3t:3t-feedback for a standalone protocol-gap capture.
2. File — open a GitHub issue on JimCline/claude-3t with label feedback, pasting the entry block as the body.
3. Apply — in this repo, run /3t-apply-feedback to cluster, propose, and apply approved changes.
4. Update — claude plugin marketplace update claude-3t && claude plugin update claude-3t.

---

Releasing changes

Bump version in .claude-plugin/plugin.json on every change you want installs to pick up. Without a bump, claude plugin update reports "already current."

claude plugin marketplace update claude-3t   # refresh cached listing
claude plugin update claude-3t               # install newer version
# restart Claude

---

Layout

claude-3t-plugin/
├── .claude-plugin/plugin.json
├── hooks/hooks.json
├── hooks/session-start.mjs          # gated SessionStart anchor
├── hooks/pre-agent.mjs              # injects gate checklist (see Legacy)
├── hooks/post-agent.mjs             # auto-verify + post-delegation audit (see Legacy)
├── bin/session-probe.mjs            # one-shot state scan for /3t-start
├── bin/session-tokens.mjs           # measured token report from session transcript
├── bin/token-baseline.mjs           # static token-cost estimate of context artifacts
├── context/3t-core.md               # session protocol, loaded once/session
├── context/3t-reference.md          # advisor flow, fork mode, occasional protocols
├── context/3t-gate.md               # PRE-AGENT checklist (Legacy — implementor use)
├── context/3t-workflow-mode.md      # workflow delegation protocol (Legacy)
├── docs/measurement-protocol.md     # token economy experiment results
├── skills/{3t-start,3t-init,3t-status,3t-tokens,3t-checkpoint,
│          3t-debrief,3t-feedback,3t-leaving,3t-remove,3t-reset}/SKILL.md
├── agents/implementor.md            # Haiku subagent (Legacy — see below)
└── templates/                       # blank project files /3t-init copies in

---

Legacy: Haiku implementor tier

The plugin ships a claude-3t:implementor Haiku subagent and delegation protocol (3t-core.md, 3t-gate.md, 3t-workflow-mode.md). These are not removed — they function and may be useful for narrow mechanical batch work (apply the same transform to many files, transcribe a settled document to disk). But they are no longer the recommended path for general feature implementation.

The short version of why: on unfamiliar-API work, Haiku and Sonnet cost the same in tokens while Haiku burns 1.8× more build iterations and runs near the turn budget on tasks a real project would consider simple. The failure tax (~3× a clean run when a delegation fails) wipes out any tier-pricing advantage. See docs/measurement-protocol.md for the full controlled experiment and data.

If you want to use the implementor for an explicitly mechanical batch, it is still there: invoke claude-3t:implementor directly and follow the PRE-AGENT CHECKLIST in 3t-gate.md. The protocol has not been removed — it just is not the default recommendation anymore.