/cf-help — Coding Friend Help
Catalog = what exists. Flags / config / hooks → lookup rule. Never guess.
Workflow
Step 0: Custom Guide
bash "<plugin-root>/lib/load-custom-guide.sh" cf-help
If the block above printed anything, apply only the ## Before, ## Rules, and ## After sections; if it shows the raw command instead of output, re-run that exact load-custom-guide.sh fence now.
Step 1: Understand the question
Classify (more than one is fine): general overview · specific skill · specific agent · hooks / fewer prompts · CLI · setup / config · memory / MCP · workflow · troubleshooting.
Lookup rule (do not skip)
Never invent a flag, config key, hook name, CLI command, or host difference.
- "What exists / which skill?" → catalog below.
- Anything more specific → Read
<plugin-root>/skills/cf-help/topics.md (or plugin/skills/cf-help/topics.md in this repo) — topic → source map, flags, hooks, CLI, config, native prompt reduction.
- Then Read the source that index points to (
skills/<name>/SKILL.md, agents/<name>.md, hooks/<file>).
- If still unclear: Glob / Grep under
<plugin-root>/skills/, agents/, hooks/.
Step 2: Provide overview (if general question)
Coding Friend is a lean toolkit for disciplined engineering workflows in Devin CLI. Check skills first; TDD optional (--add-tests or tdd: true); /cf-plan autopilot optional (--auto or planAuto: true, --no-auto forces off); verify before claiming done.
Hosts: Claude Code (default), Codex CLI, omp, Google Antigravity (--agent agy / --agy), Devin CLI (beta; --agent devin / --devin). Plugin works alone; coding-friend-cli is optional. Skills never call cf. Flags / config / native prompt-reduction: topics.md.
Slash Commands (user triggers with /)
/cf-advise [decision] — ⚡⚡ — Advisory interview. --quick, --save
/cf-ask [question] — ⚡⚡ — Codebase Q&A → docs/memory/
/cf-plan [task] — ⚡⚡ — Phased plans. --fast/--quick, --hard, --auto, --no-auto, --inline/--no-file, --gui/--human, --model <alias> pin the model for cf-planner at the brainstorm step. Autopilot also from config planAuto: true (local overrides global) unless --no-auto. Autopilot review-fix cap: review.maxRounds (default 5; 0 = unlimited until no progress). --commit-per-task / --no-commit-per-task (or config planCommitPerTask: true, autopilot only): commit each task as phase N/M task i/K <title>, review once per phase over its commits, [parallel] phases run sequentially. Normal/hard mode also writes brief.md (request, Q&A, assumptions) for /cf-plan-review.
/cf-plan-resume <plan> — ⚡⚡ — Resume a saved plan. Honors that plan's auto: true + ## AUTOPILOT section. Does not read config planAuto / planCommitPerTask; honors frontmatter commitPerTask: true (--commit-per-task / --no-commit-per-task rewrite it).
/cf-plan-review [plan] — ⚡⚡ — Review a saved plan with a fresh in-session reviewer before implementing. --codex, --gemini, --claude, --cursor, --grok add external reviewers in parallel (the flag matching the current host is skipped). --fix (--inline) applies 🚨/⚠️ immediately and re-reviews until clean (≤ review.maxRounds, 0 = unlimited); config planReviewFix: true (local overrides global) makes that the default, --no-fix forces off; otherwise offers to apply. Writes review.md.
/cf-later <text> — ⚡ — Defer a side-task to docs/later/ from a forked subagent (one-line result)
/cf-later-do [item] — ⚡⚡ — Resolve docs/later/ via /cf-fix or /cf-plan
/cf-review [target] — ⚡⚡ — Dispatch review. Flags: --with-codex/--codex, --claude, --gemini, --cursor, --grok run headless external reviewers in parallel and merge into one report; --out exports a /cf-review-out prompt with Claude's findings embedded. Set review.withCodex: true in config to enable Codex by default (likewise review.withClaude, review.withGrok, review.withCursor, review.withGemini); review.agentTimeout (default 300s) bounds each external agent; review.nativeTimeout (default 600s) bounds each in-session reviewer job; review.maxRounds (default 5) caps the autopilot fix loop. Unavailable agents are skipped with a warning. --fix applies 🚨/⚠️ findings and re-reviews (capped by review.maxRounds; 0 = until clean, stops when a re-review makes no progress); --commit implies --fix and commits once the final review is COMPLETE with no 🚨/⚠️ — only the reviewed uncommitted paths, after a secret scan; a hook failure stops it. Both are refused with a commit-range target or --out. Config review.fix / review.commit (local overrides global) turn them on by default — also for the auto-review after /cf-fix and /cf-optimize; unless you typed /cf-review (e.g. after /cf-fix / /cf-optimize, or a "review this" request) a config-driven commit is skipped when the tree has paths this session did not touch (a /cf-review you type commits like --commit); --no-fix / --no-commit force off, and /cf-plan / /cf-tdd autopilot and /cf-commit pass --no-fix (they own their loop or change set). With a commit range or --out, a config default is skipped with a notice (not refused). The two timeouts are different mechanisms: nativeTimeout is a cooperative budget carried in the in-session reviewer's prompt and bounding the skill's wait — at the deadline the skill cancels the job only if the host offers a cancel, otherwise it marks it timed out and merges what arrived — while agentTimeout is really enforced on the external subprocess (TERM, 2s grace, then KILL). External reviewers are opt-in extras — they never count toward the native 1 / 1 / 2. Every report ends with Review status: COMPLETE | PARTIAL | FAILED: COMPLETE means the required coverage finished, not that there were no findings, and PARTIAL / FAILED blocks the autopilot commit. Scope, depth table, and the trade-off: topics.md.
/cf-review-out [label] — ⚡⚡ — Prompt + diff → docs/reviews/
/cf-review-in <label> [service] — ⚡⚡ — Read external review, offer to fix
/cf-commit [hint] — ⚡ — Conventional commit
/cf-design [mode] — ⚡⚡ — Scan / design / modify UI
/cf-showcase [--video|--poster] [--default] — ⚡⚡ — Intro video (mp4) / poster (png/pdf/html) from the project's real content → docs/showcase/; --default skips the interview and uses saved showcase.* defaults
/cf-ship [hint] — ⚡ — Verify, commit, push, PR (--dry-run)
/cf-fix [bug] — ⚡⚡ — Quick bug-fix
/cf-optimize [target] — ⚡⚡ — Baseline → optimize → measure
/cf-scan [description] — ⚡⚡⚡ — Bootstrap memory
/cf-remember [topic] — ⚡⚡ — Project knowledge → docs/memory/. Also auto-invoked.
/cf-learn [topic] — ⚡⚡ — Educational notes. Also auto-invoked.
/cf-teach [topic] — ⚡⚡ — Story of what happened
/cf-research [topic] — ⚡⚡ — Web research → docs/research/
/cf-session [label] — ⚡⚡ — Save the Claude session verbatim to session.dir via cf session save (needs the CLI)
/cf-warm [user] — ⚡⚡ — Git history catch-up
/cf-checkpoint [additional-prompt] — ⚡⚡ — Checkpoint → docs/context/checkpoints/
/cf-checkpoint-from <slug> [message] — ⚡⚡ — Restore checkpoint, then act. --recap
/cf-help [question] — ⚡⚡⚡ — This skill. Also auto-invoked.
Auto-Invoked Skills (activate automatically when relevant)
- cf-tdd — ⚡⚡ — Code-writing gate. Default direct; TDD with
--add-tests or tdd: true. --auto = review + fix + commit (review.maxRounds, default 5; 0 = unlimited).
- cf-sys-debug — ⚡⚡ — 4-phase debug
- cf-verification — ⚡ — Evidence gate before claiming done
- cf-learn / cf-remember / cf-help — also auto-invoked (see above)
Agents (run in forked sessions — separate context window)
- cf-reviewer — ⚡ — Reviews the diff itself, 5 layers.
/cf-review dispatches it directly: 0 in TRIVIAL, 1 in QUICK, 1 in STANDARD, 2 in DEEP — the second is cf-reviewer-security (sonnet). Flat graph — no sub-dispatch, no merge agent.
- cf-reviewer-plan (sonnet), cf-reviewer-quality (haiku), cf-reviewer-tests (haiku), cf-reviewer-rules (haiku), cf-reviewer-reducer (haiku) — ⚡ — retained, directly callable; off
/cf-review's default path
- cf-implementer — ⚡ — Writes code; TDD with
--add-tests. [CF-RESULT: success|failure]. No autopilot.
- cf-explorer — ⚡ — Repo map + context files
- cf-planner — ⚡ — Approaches + phased tasks
- cf-writer — ⚡ — Straightforward markdown
- cf-writer-deep — ⚡ — Nuanced technical docs
Hooks (automatic — not slash commands)
hooks/hooks.json (+ *.agy.* / *.devin.*; Codex transformed). Keys + native modes: topics.md.
session-init.sh · rules-reminder.sh · privacy-block.sh · scout-block.cjs · auto-approve.cjs (autoApprove; Codex / agy / Devin only, deterministic; Claude / omp use native modes) · session-log.sh · task-tracker.sh · agent-tracker.sh · memory-capture.sh · statusline.sh (Claude; cf statusline)
Mods plugin (coding-friend-mods) — optional, Claude Code only
Second plugin in the same marketplace (needs coding-friend): quota toast + band at 80/90/95% (also writes the quota snapshot), toast when /cf-review or a plan phase finishes, plan progress pane (mods slash command cf-plan-status [slug]), band when a CF MCP server fails. Install: cf install --mods or /plugin install coding-friend-mods@coding-friend-marketplace. Details: topics.md (Mods).
CLI (coding-friend-cli, binary cf) — optional (except /cf-session)
Lifecycle install|uninstall|enable|disable|update (--user|--project|--local, --agent claude|codex|omp|agy|devin, install --mods Claude only); setup init config permission statusline; memory / learn / mcp; status version doctor project task ask reply done clean session decide test|eval|log|review-fp sync guide dev; --json on read commands. Only /cf-session requires the CLI. Full flags: topics.md.
Tiers ⚡ / ⚡⚡ / ⚡⚡⚡: https://cf.dinhanhthi.com/.
Step 3: Read specific files (if detailed question)
Follow the lookup rule. Start at skills/cf-help/topics.md, then skills/<name>/SKILL.md, agents/<name>.md, hooks/. Config keys and custom guides: topics.md. Repo-only schema: docs/config-schema.md.
Step 4: Common Workflows (if workflow question)
Present the workflows in topics.md. Distinction: /cf-remember = project knowledge for AI; /cf-learn = notes for the human.
Step 5: Troubleshooting (if troubleshooting question)
- Skill not triggering? Invoke
/cf-<skill-name> manually.
- Config / custom guides?
topics.md.
- Permission prompts? Native modes: Claude
auto / acceptEdits / sandbox.autoAllowBashIfSandboxed; Codex approval_policy / Smart Approvals / --approve-for-me; AGY remembers per-conversation; omp yolo; Devin --permission-mode dangerous (CF PermissionRequest hook approve/block). CF hook autoApprove: Codex / AGY / Devin only. Details: topics.md.
- More? Troubleshooting.
Step 6: Answer concisely
Answer from sources you read. Link files for deeper dives. If you skipped the lookup rule on a flag / hook / CLI / config question, go back.
CLI Requirements (quick reference)
Plugin works without CLI, except /cf-session. NONE / OPTIONAL (MCP or grep) / REQUIRED (/cf-session). Matrix: docs/cli-requirements.md.
Plugin root: the PLUGIN_ROOT: path in the session bootstrap context (HOST: devin), or the parent of the skills/ folder that contains this SKILL.md. Replace <plugin-root> with it when running bundled scripts.
1---2name: cf-help-43description: >4 Answer questions about Coding Friend itself — skills, agents, hooks, CLI, config,5 memory, hosts, parameters. TRIGGER — "what skills are available?", "how does6 coding friend work?", "list all skills", "how do I use cf-plan?", "what is7 cf-tdd?", "is the CLI required?", or any question about a cf-* name, flag, hook,8 or config key. SKIP — general coding questions unrelated to Coding Friend, and9 requests to run a skill (invoke that skill instead).10user-invocable: true11model: haiku12allowed-tools: [Read, Glob, Grep]13created: 2026-02-1714updated: 2026-11-0515metadata:16 source: https://skillmd.com/skills/dinhanhthi/cf-help-417---1819# /cf-help — Coding Friend Help2021Catalog = what exists. Flags / config / hooks → lookup rule. Never guess.2223## Workflow2425### Step 0: Custom Guide2627```!28bash "<plugin-root>/lib/load-custom-guide.sh" cf-help29```3031If the block above printed anything, apply only the `## Before`, `## Rules`, and `## After` sections; if it shows the raw command instead of output, re-run that exact `load-custom-guide.sh` fence now.3233### Step 1: Understand the question3435Classify (more than one is fine): general overview · specific skill · specific agent · hooks / fewer prompts · CLI · setup / config · memory / MCP · workflow · troubleshooting.3637### Lookup rule (do not skip)3839Never invent a flag, config key, hook name, CLI command, or host difference.40411. "What exists / which skill?" → catalog below.422. Anything more specific → Read `<plugin-root>/skills/cf-help/topics.md` (or `plugin/skills/cf-help/topics.md` in this repo) — topic → source map, flags, hooks, CLI, config, native prompt reduction.433. Then Read the source that index points to (`skills/<name>/SKILL.md`, `agents/<name>.md`, `hooks/<file>`).444. If still unclear: Glob / Grep under `<plugin-root>/skills/`, `agents/`, `hooks/`.4546### Step 2: Provide overview (if general question)4748Coding Friend is a lean toolkit for disciplined engineering workflows in Devin CLI. Check skills first; TDD optional (`--add-tests` or `tdd: true`); `/cf-plan` autopilot optional (`--auto` or `planAuto: true`, `--no-auto` forces off); verify before claiming done.4950Hosts: Claude Code (default), Codex CLI, omp, **Google Antigravity** (`--agent agy` / `--agy`), **Devin CLI** (beta; `--agent devin` / `--devin`). Plugin works alone; `coding-friend-cli` is optional. Skills never call `cf`. Flags / config / native prompt-reduction: `topics.md`.5152### Slash Commands (user triggers with /)5354- `/cf-advise [decision]` — ⚡⚡ — Advisory interview. `--quick`, `--save`55- `/cf-ask [question]` — ⚡⚡ — Codebase Q&A → docs/memory/56- `/cf-plan [task]` — ⚡⚡ — Phased plans. `--fast`/`--quick`, `--hard`, `--auto`, `--no-auto`, `--inline`/`--no-file`, `--gui`/`--human`, `--model <alias>` pin the model for cf-planner at the brainstorm step. Autopilot also from config `planAuto: true` (local overrides global) unless `--no-auto`. Autopilot review-fix cap: `review.maxRounds` (default 5; `0` = unlimited until no progress). `--commit-per-task` / `--no-commit-per-task` (or config `planCommitPerTask: true`, autopilot only): commit each task as `phase N/M task i/K <title>`, review once per phase over its commits, `[parallel]` phases run sequentially. Normal/hard mode also writes `brief.md` (request, Q&A, assumptions) for `/cf-plan-review`.57- `/cf-plan-resume <plan>` — ⚡⚡ — Resume a saved plan. Honors that plan's `auto: true` + `## AUTOPILOT` section. Does **not** read config `planAuto` / `planCommitPerTask`; honors frontmatter `commitPerTask: true` (`--commit-per-task` / `--no-commit-per-task` rewrite it).58- `/cf-plan-review [plan]` — ⚡⚡ — Review a saved plan with a fresh in-session reviewer before implementing. `--codex`, `--gemini`, `--claude`, `--cursor`, `--grok` add external reviewers in parallel (the flag matching the current host is skipped). `--fix` (`--inline`) applies 🚨/⚠️ immediately and re-reviews until clean (≤ `review.maxRounds`, `0` = unlimited); config `planReviewFix: true` (local overrides global) makes that the default, `--no-fix` forces off; otherwise offers to apply. Writes `review.md`.59- `/cf-later <text>` — ⚡ — Defer a side-task to `docs/later/` from a forked subagent (one-line result)60- `/cf-later-do [item]` — ⚡⚡ — Resolve `docs/later/` via `/cf-fix` or `/cf-plan`61- `/cf-review [target]` — ⚡⚡ — Dispatch review. Flags: `--with-codex`/`--codex`, `--claude`, `--gemini`, `--cursor`, `--grok` run headless external reviewers in parallel and merge into one report; `--out` exports a `/cf-review-out` prompt with Claude's findings embedded. Set `review.withCodex: true` in config to enable Codex by default (likewise `review.withClaude`, `review.withGrok`, `review.withCursor`, `review.withGemini`); `review.agentTimeout` (default 300s) bounds each external agent; `review.nativeTimeout` (default 600s) bounds each in-session reviewer job; `review.maxRounds` (default 5) caps the autopilot fix loop. Unavailable agents are skipped with a warning. `--fix` applies 🚨/⚠️ findings and re-reviews (capped by `review.maxRounds`; `0` = until clean, stops when a re-review makes no progress); `--commit` implies `--fix` and commits once the final review is COMPLETE with no 🚨/⚠️ — only the reviewed uncommitted paths, after a secret scan; a hook failure stops it. Both are refused with a commit-range target or `--out`. Config `review.fix` / `review.commit` (local overrides global) turn them on by default — also for the auto-review after `/cf-fix` and `/cf-optimize`; unless you typed `/cf-review` (e.g. after `/cf-fix` / `/cf-optimize`, or a "review this" request) a config-driven commit is skipped when the tree has paths this session did not touch (a `/cf-review` you type commits like `--commit`); `--no-fix` / `--no-commit` force off, and `/cf-plan` / `/cf-tdd` autopilot and `/cf-commit` pass `--no-fix` (they own their loop or change set). With a commit range or `--out`, a config default is skipped with a notice (not refused). The two timeouts are different mechanisms: `nativeTimeout` is a cooperative budget carried in the in-session reviewer's prompt and bounding the skill's wait — at the deadline the skill cancels the job only if the host offers a cancel, otherwise it marks it timed out and merges what arrived — while `agentTimeout` is really enforced on the external subprocess (TERM, 2s grace, then KILL). External reviewers are opt-in extras — they never count toward the native 1 / 1 / 2. Every report ends with `Review status: COMPLETE | PARTIAL | FAILED`: COMPLETE means the required coverage finished, **not** that there were no findings, and PARTIAL / FAILED blocks the autopilot commit. Scope, depth table, and the trade-off: `topics.md`.62- `/cf-review-out [label]` — ⚡⚡ — Prompt + diff → `docs/reviews/`63- `/cf-review-in <label> [service]` — ⚡⚡ — Read external review, offer to fix64- `/cf-commit [hint]` — ⚡ — Conventional commit65- `/cf-design [mode]` — ⚡⚡ — Scan / design / modify UI66- `/cf-showcase [--video|--poster] [--default]` — ⚡⚡ — Intro video (mp4) / poster (png/pdf/html) from the project's real content → docs/showcase/; `--default` skips the interview and uses saved `showcase.*` defaults67- `/cf-ship [hint]` — ⚡ — Verify, commit, push, PR (`--dry-run`)68- `/cf-fix [bug]` — ⚡⚡ — Quick bug-fix69- `/cf-optimize [target]` — ⚡⚡ — Baseline → optimize → measure70- `/cf-scan [description]` — ⚡⚡⚡ — Bootstrap memory71- `/cf-remember [topic]` — ⚡⚡ — Project knowledge → docs/memory/. Also auto-invoked.72- `/cf-learn [topic]` — ⚡⚡ — Educational notes. Also auto-invoked.73- `/cf-teach [topic]` — ⚡⚡ — Story of what happened74- `/cf-research [topic]` — ⚡⚡ — Web research → docs/research/75- `/cf-session [label]` — ⚡⚡ — Save the Claude session verbatim to `session.dir` via `cf session save` (needs the CLI)76- `/cf-warm [user]` — ⚡⚡ — Git history catch-up77- `/cf-checkpoint [additional-prompt]` — ⚡⚡ — Checkpoint → docs/context/checkpoints/78- `/cf-checkpoint-from <slug> [message]` — ⚡⚡ — Restore checkpoint, then act. `--recap`79- `/cf-help [question]` — ⚡⚡⚡ — This skill. Also auto-invoked.8081### Auto-Invoked Skills (activate automatically when relevant)8283- **cf-tdd** — ⚡⚡ — Code-writing gate. Default direct; TDD with `--add-tests` or `tdd: true`. `--auto` = review + fix + commit (`review.maxRounds`, default 5; `0` = unlimited).84- **cf-sys-debug** — ⚡⚡ — 4-phase debug85- **cf-verification** — ⚡ — Evidence gate before claiming done86- **cf-learn** / **cf-remember** / **cf-help** — also auto-invoked (see above)8788### Agents (run in forked sessions — separate context window)8990- **cf-reviewer** — ⚡ — Reviews the diff itself, 5 layers. `/cf-review` dispatches it directly: 0 in TRIVIAL, 1 in QUICK, 1 in STANDARD, 2 in DEEP — the second is **cf-reviewer-security** (sonnet). Flat graph — no sub-dispatch, no merge agent.91- **cf-reviewer-plan** (sonnet), **cf-reviewer-quality** (haiku), **cf-reviewer-tests** (haiku), **cf-reviewer-rules** (haiku), **cf-reviewer-reducer** (haiku) — ⚡ — retained, directly callable; off `/cf-review`'s default path92- **cf-implementer** — ⚡ — Writes code; TDD with `--add-tests`. `[CF-RESULT: success|failure]`. No autopilot.93- **cf-explorer** — ⚡ — Repo map + context files94- **cf-planner** — ⚡ — Approaches + phased tasks95- **cf-writer** — ⚡ — Straightforward markdown96- **cf-writer-deep** — ⚡ — Nuanced technical docs9798### Hooks (automatic — not slash commands)99100`hooks/hooks.json` (+ `*.agy.*` / `*.devin.*`; Codex transformed). Keys + native modes: `topics.md`.101102**session-init.sh** · **rules-reminder.sh** · **privacy-block.sh** · **scout-block.cjs** · **auto-approve.cjs** (`autoApprove`; Codex / agy / Devin only, deterministic; Claude / omp use native modes) · **session-log.sh** · **task-tracker.sh** · **agent-tracker.sh** · **memory-capture.sh** · **statusline.sh** (Claude; `cf statusline`)103104### Mods plugin (`coding-friend-mods`) — optional, Claude Code only105106Second plugin in the same marketplace (needs `coding-friend`): quota toast + band at 80/90/95% (also writes the quota snapshot), toast when `/cf-review` or a plan phase finishes, plan progress pane (mods slash command `cf-plan-status [slug]`), band when a CF MCP server fails. Install: `cf install --mods` or `/plugin install coding-friend-mods@coding-friend-marketplace`. Details: `topics.md` (Mods).107108### CLI (`coding-friend-cli`, binary `cf`) — optional (except `/cf-session`)109110Lifecycle `install|uninstall|enable|disable|update` (`--user|--project|--local`, `--agent claude|codex|omp|agy|devin`, `install --mods` Claude only); setup `init` `config` `permission` `statusline`; memory / learn / mcp; `status` `version` `doctor` `project` `task` `ask` `reply` `done` `clean` `session` `decide test|eval|log|review-fp` `sync` `guide` `dev`; `--json` on read commands. Only `/cf-session` requires the CLI. Full flags: `topics.md`.111112Tiers `⚡` / `⚡⚡` / `⚡⚡⚡`: https://cf.dinhanhthi.com/.113114### Step 3: Read specific files (if detailed question)115116Follow the lookup rule. Start at `skills/cf-help/topics.md`, then `skills/<name>/SKILL.md`, `agents/<name>.md`, `hooks/`. Config keys and custom guides: `topics.md`. Repo-only schema: `docs/config-schema.md`.117118### Step 4: Common Workflows (if workflow question)119120Present the workflows in `topics.md`. Distinction: `/cf-remember` = project knowledge for AI; `/cf-learn` = notes for the human.121122### Step 5: Troubleshooting (if troubleshooting question)123124- **Skill not triggering?** Invoke `/cf-<skill-name>` manually.125- **Config / custom guides?** `topics.md`.126- **Permission prompts?** Native modes: Claude `auto` / `acceptEdits` / `sandbox.autoAllowBashIfSandboxed`; Codex `approval_policy` / Smart Approvals / `--approve-for-me`; AGY remembers per-conversation; omp `yolo`; Devin `--permission-mode dangerous` (CF `PermissionRequest` hook approve/block). CF hook `autoApprove`: Codex / AGY / Devin only. Details: `topics.md`.127- **More?** [Troubleshooting](https://cf.dinhanhthi.com/#help).128129### Step 6: Answer concisely130131Answer from sources you read. Link files for deeper dives. If you skipped the lookup rule on a flag / hook / CLI / config question, go back.132133## CLI Requirements (quick reference)134135Plugin works without CLI, except `/cf-session`. **NONE** / **OPTIONAL** (MCP or grep) / **REQUIRED** (`/cf-session`). Matrix: [docs/cli-requirements.md](../../../docs/cli-requirements.md).136137> Plugin root: the `PLUGIN_ROOT:` path in the session bootstrap context (HOST: devin), or the parent of the `skills/` folder that contains this SKILL.md. Replace `<plugin-root>` with it when running bundled scripts.