27 KiB
Development Workflow
Core Principles
- Plan before code — figure out what to do before you start
- Specs injected, not remembered — guidelines are injected via hook/skill, not recalled from memory
- Persist everything — research, decisions, and lessons all go to files; conversations get compacted, files don't
- Incremental development — one task at a time
- Capture learnings — after each task, review and write new knowledge back to spec
Trellis System
Developer Identity
On first use, initialize your identity:
python3 ./.trellis/scripts/init_developer.py <your-name>
Creates .trellis/.developer (gitignored) + .trellis/workspace/<your-name>/.
Spec System
.trellis/spec/ holds coding guidelines organized by package and layer.
.trellis/spec/<package>/<layer>/index.md— entry point with Pre-Development Checklist + Quality Check. Actual guidelines live in the.mdfiles it points to..trellis/spec/guides/index.md— cross-package thinking guides.
python3 ./.trellis/scripts/get_context.py --mode packages # list packages / layers
When to update spec: new pattern/convention found · bug-fix prevention to codify · new technical decision.
Task System
Every task has its own directory under .trellis/tasks/{MM-DD-name}/ holding prd.md, implement.jsonl, check.jsonl, task.json, optional research/, info.md.
# Task lifecycle
python3 ./.trellis/scripts/task.py create "<title>" [--slug <name>] [--parent <dir>]
python3 ./.trellis/scripts/task.py start <name> # set active task (session-scoped when available)
python3 ./.trellis/scripts/task.py current --source # show active task and source
python3 ./.trellis/scripts/task.py finish # clear active task (triggers after_finish hooks)
python3 ./.trellis/scripts/task.py archive <name> # move to archive/{year-month}/
python3 ./.trellis/scripts/task.py list [--mine] [--status <s>]
python3 ./.trellis/scripts/task.py list-archive
# Code-spec context (injected into implement/check agents via JSONL).
# `implement.jsonl` / `check.jsonl` are seeded on `task create` for sub-agent-capable
# platforms; the AI curates real spec + research entries during Phase 1.3.
python3 ./.trellis/scripts/task.py add-context <name> <action> <file> <reason>
python3 ./.trellis/scripts/task.py list-context <name> [action]
python3 ./.trellis/scripts/task.py validate <name>
# Task metadata
python3 ./.trellis/scripts/task.py set-branch <name> <branch>
python3 ./.trellis/scripts/task.py set-base-branch <name> <branch> # PR target
python3 ./.trellis/scripts/task.py set-scope <name> <scope>
# Hierarchy (parent/child)
python3 ./.trellis/scripts/task.py add-subtask <parent> <child>
python3 ./.trellis/scripts/task.py remove-subtask <parent> <child>
# PR creation
python3 ./.trellis/scripts/task.py create-pr [name] [--dry-run]
Run
python3 ./.trellis/scripts/task.py --helpto see the authoritative, up-to-date list.
Current-task mechanism: task.py start writes the task path through the active-task resolver into session/window-scoped runtime state under .trellis/.runtime/sessions/. If no context key is available from hook input, TRELLIS_CONTEXT_ID, or a platform-native session environment variable, there is no active task and task.py start fails with a session identity hint. task.py finish deletes the current session file. task.py archive <task> also deletes any runtime session files that still point at the archived task.
Workspace System
Records every AI session for cross-session tracking under .trellis/workspace/<developer>/.
journal-N.md— session log. Max 2000 lines per file; a newjournal-(N+1).mdis auto-created when exceeded.index.md— personal index (total sessions, last active).
python3 ./.trellis/scripts/add_session.py --title "Title" --commit "hash" --summary "Summary"
Context Script
python3 ./.trellis/scripts/get_context.py # full session runtime
python3 ./.trellis/scripts/get_context.py --mode packages # available packages + spec layers
python3 ./.trellis/scripts/get_context.py --mode phase --step <X.Y> # detailed guide for a workflow step
Phase Index
Phase 1: Plan → figure out what to do (brainstorm + research → prd.md)
Phase 2: Execute → write code and pass quality checks
Phase 3: Finish → distill lessons + wrap-up
Phase 1: Plan
- 1.0 Create task
[required · once] - 1.1 Requirement exploration
[required · repeatable] - 1.2 Research
[optional · repeatable] - 1.3 Configure context
[required · once]— Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi - 1.4 Completion criteria
Phase 2: Execute
- 2.1 Implement
[required · repeatable] - 2.2 Quality check
[required · repeatable] - 2.3 Rollback
[on demand]
Phase 3: Finish
- 3.1 Quality verification
[required · repeatable] - 3.2 Debug retrospective
[on demand] - 3.3 Spec update
[required · once] - 3.4 Commit changes
[required · once] - 3.5 Wrap-up reminder
Rules
- Identify which Phase you're in, then continue from the next step there
- Run steps in order inside each Phase;
[required]steps can't be skipped - Phases can roll back (e.g., Execute reveals a prd defect → return to Plan to fix, then re-enter Execute)
- Steps tagged
[once]are skipped if already done; don't re-run
Skill Routing
When a user request matches one of these intents, load the corresponding skill (or dispatch the corresponding sub-agent) first — do not skip skills.
[Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
| User intent | Route |
|---|---|
| Wants a new feature / requirement unclear | trellis-brainstorm |
| About to write code / start implementing | Dispatch the trellis-implement sub-agent per Phase 2.1 |
| Finished writing / want to verify | Dispatch the trellis-check sub-agent per Phase 2.2 |
| Stuck / fixed same bug several times | trellis-break-loop |
| Spec needs update | trellis-update-spec |
Why trellis-before-dev is NOT in this table: you are not the one writing code — the trellis-implement sub-agent is. Sub-agent platforms get spec context via implement.jsonl injection / prelude, not via the main thread loading trellis-before-dev.
[/Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
[Kilo, Antigravity, Windsurf]
| User intent | Skill |
|---|---|
| Wants a new feature / requirement unclear | trellis-brainstorm |
| About to write code / start implementing | trellis-before-dev (then implement directly in the main session) |
| Finished writing / want to verify | trellis-check |
| Stuck / fixed same bug several times | trellis-break-loop |
| Spec needs update | trellis-update-spec |
[/Kilo, Antigravity, Windsurf]
DO NOT skip skills
[Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
| What you're thinking | Why it's wrong |
|---|---|
| "This is simple, I'll just code it in the main thread" | Dispatching trellis-implement is the cheap path; skipping it tempts you to write code in the main thread and lose spec context — sub-agents get implement.jsonl injected, you don't |
| "I already thought it through in plan mode" | Plan-mode output lives in memory — sub-agents can't see it; must be persisted to prd.md |
| "I already know the spec" | The spec may have been updated since you last read it; the sub-agent gets the fresh copy, you may not |
| "Code first, check later" | trellis-check surfaces issues you won't notice yourself; earlier is cheaper |
[/Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
[Kilo, Antigravity, Windsurf]
| What you're thinking | Why it's wrong |
|---|---|
| "This is simple, just code it" | Simple tasks often grow complex; trellis-before-dev takes under a minute and loads the spec context you'll need |
| "I already thought it through in plan mode" | Plan-mode output lives in memory — must be persisted to prd.md before code |
| "I already know the spec" | The spec may have been updated since you last read it; read again |
| "Code first, check later" | trellis-check surfaces issues you won't notice yourself; earlier is cheaper |
[/Kilo, Antigravity, Windsurf]
Loading Step Detail
At each step, run this to fetch detailed guidance:
python3 ./.trellis/scripts/get_context.py --mode phase --step <step>
# e.g. python3 ./.trellis/scripts/get_context.py --mode phase --step 1.1
Phase 1: Plan
Goal: figure out what to build, produce a clear requirements doc and the context needed to implement it.
1.0 Create task [required · once]
Create the task directory and set it as current:
python3 ./.trellis/scripts/task.py create "<task title>" --slug <name>
python3 ./.trellis/scripts/task.py start <task-dir>
--slug is the human-readable name only. Do not include the MM-DD- date
prefix; task.py create adds that prefix automatically.
Skip when python3 ./.trellis/scripts/task.py current --source already points to a task.
1.1 Requirement exploration [required · repeatable]
Load the trellis-brainstorm skill and explore requirements interactively with the user per the skill's guidance.
The brainstorm skill will guide you to:
- Ask one question at a time
- Prefer researching over asking the user
- Prefer offering options over open-ended questions
- Update
prd.mdimmediately after each user answer
Return to this step whenever requirements change and revise prd.md.
1.2 Research [optional · repeatable]
Research can happen at any time during requirement exploration. It isn't limited to local code — you can use any available tool (MCP servers, skills, web search, etc.) to look up external information, including third-party library docs, industry practices, API references, etc.
[Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
Spawn the research sub-agent:
- Agent type:
trellis-research - Task description: Research
- Key requirement: Research output MUST be persisted to
{TASK_DIR}/research/
[/Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
[Kilo, Antigravity, Windsurf]
Do the research in the main session directly and write findings into {TASK_DIR}/research/.
[/Kilo, Antigravity, Windsurf]
Research artifact conventions:
- One file per research topic (e.g.
research/auth-library-comparison.md) - Record third-party library usage examples, API references, version constraints in files
- Note relevant spec file paths you discovered for later reference
Brainstorm and research can interleave freely — pause to research a technical question, then return to talk with the user.
Key principle: Research output must be written to files, not left only in the chat. Conversations get compacted; files don't.
1.3 Configure context [required · once]
[Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
Curate implement.jsonl and check.jsonl so the Phase 2 sub-agents get the right spec context. These files were seeded on task create with a single self-describing _example line; your job here is to fill in real entries.
Location: {TASK_DIR}/implement.jsonl and {TASK_DIR}/check.jsonl (already exist).
Format: one JSON object per line — {"file": "<path>", "reason": "<why>"}. Paths are repo-root relative.
What to put in:
- Spec files —
.trellis/spec/<package>/<layer>/index.mdand any specific guideline files (error-handling.md,conventions.md, etc.) relevant to this task - Research files —
{TASK_DIR}/research/*.mdthat the sub-agent will need to consult
What NOT to put in:
- Code files (
src/**,packages/**/*.ts, etc.) — those are read by the sub-agent during implementation, not pre-registered here - Files you're about to modify — same reason
Split between the two files:
implement.jsonl→ specs + research the implement sub-agent needs to write code correctlycheck.jsonl→ specs for the check sub-agent (quality guidelines, check conventions, same research if needed)
How to discover relevant specs:
python3 ./.trellis/scripts/get_context.py --mode packages
Lists every package + its spec layers with paths. Pick the entries that match this task's domain.
How to append entries:
Either edit the jsonl file directly in your editor, or use:
python3 ./.trellis/scripts/task.py add-context "$TASK_DIR" implement "<path>" "<reason>"
python3 ./.trellis/scripts/task.py add-context "$TASK_DIR" check "<path>" "<reason>"
Delete the seed _example line once real entries exist (optional — it's skipped automatically by consumers).
Skip when: implement.jsonl has agent-curated entries (the seed row alone doesn't count).
[/Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
[Kilo, Antigravity, Windsurf]
Skip this step. Context is loaded directly by the trellis-before-dev skill in Phase 2.
[/Kilo, Antigravity, Windsurf]
1.4 Completion criteria
| Condition | Required |
|---|---|
prd.md exists |
✅ |
| User confirms requirements | ✅ |
research/ has artifacts (complex tasks) |
recommended |
info.md technical design (complex tasks) |
optional |
[Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
| implement.jsonl has agent-curated entries (not just the seed row) | ✅ |
[/Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
Phase 2: Execute
Goal: turn the prd into code that passes quality checks.
2.1 Implement [required · repeatable]
[Claude Code, Cursor, OpenCode, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
Spawn the implement sub-agent:
- Agent type:
trellis-implement - Task description: Implement the requirements per prd.md, consulting materials under
{TASK_DIR}/research/; finish by running project lint and type-check
The platform hook/plugin auto-handles:
- Reads
implement.jsonland injects the referenced spec files into the agent prompt - Injects prd.md content
[/Claude Code, Cursor, OpenCode, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
[Codex]
Spawn the implement sub-agent:
- Agent type:
trellis-implement - Task description: Implement the requirements per prd.md, consulting materials under
{TASK_DIR}/research/; finish by running project lint and type-check
The Codex sub-agent definition auto-handles the context load requirement:
- Resolves the active task with
task.py current --source, then readsprd.mdandinfo.mdif present - Reads
implement.jsonland requires the agent to load each referenced spec file before coding
[/Codex]
[Kiro]
Spawn the implement sub-agent:
- Agent type:
trellis-implement - Task description: Implement the requirements per prd.md, consulting materials under
{TASK_DIR}/research/; finish by running project lint and type-check
The platform prelude auto-handles the context load requirement:
- Reads
implement.jsonland injects the referenced spec files into the agent prompt - Injects prd.md content
[/Kiro]
[Kilo, Antigravity, Windsurf]
- Load the
trellis-before-devskill to read project guidelines - Read
{TASK_DIR}/prd.mdfor requirements - Consult materials under
{TASK_DIR}/research/ - Implement the code per requirements
- Run project lint and type-check
[/Kilo, Antigravity, Windsurf]
2.2 Quality check [required · repeatable]
[Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
Spawn the check sub-agent:
- Agent type:
trellis-check - Task description: Review all code changes against spec and prd; fix any findings directly; ensure lint and type-check pass
The check agent's job:
- Review code changes against specs
- Auto-fix issues it finds
- Run lint and typecheck to verify
[/Claude Code, Cursor, OpenCode, Codex, Kiro, Gemini, Qoder, CodeBuddy, Copilot, Droid, Pi]
[Kilo, Antigravity, Windsurf]
Load the trellis-check skill and verify the code per its guidance:
- Spec compliance
- lint / type-check / tests
- Cross-layer consistency (when changes span layers)
If issues are found → fix → re-check, until green.
[/Kilo, Antigravity, Windsurf]
2.3 Rollback [on demand]
checkreveals a prd defect → return to Phase 1, fixprd.md, then redo 2.1- Implementation went wrong → revert code, redo 2.1
- Need more research → research (same as Phase 1.2), write findings into
research/
Phase 3: Finish
Goal: ensure code quality, capture lessons, record the work.
3.1 Quality verification [required · repeatable]
Load the trellis-check skill and do a final verification:
- Spec compliance
- lint / type-check / tests
- Cross-layer consistency (when changes span layers)
If issues are found → fix → re-check, until green.
3.2 Debug retrospective [on demand]
If this task involved repeated debugging (the same issue was fixed multiple times), load the trellis-break-loop skill to:
- Classify the root cause
- Explain why earlier fixes failed
- Propose prevention
The goal is to capture debugging lessons so the same class of issue doesn't recur.
3.3 Spec update [required · once]
Load the trellis-update-spec skill and review whether this task produced new knowledge worth recording:
- Newly discovered patterns or conventions
- Pitfalls you hit
- New technical decisions
Update the docs under .trellis/spec/ accordingly. Even if the conclusion is "nothing to update", walk through the judgment.
3.4 Commit changes [required · once]
The AI drives a batched commit of this task's code changes so /finish-work can run cleanly afterwards. Goal: produce work commits FIRST, then bookkeeping (archive + journal) commits land after — never interleaved.
Step-by-step:
-
Inspect dirty state:
git status --porcelainSnapshot every dirty path. If the working tree is clean, skip to 3.5.
-
Learn commit style from recent history (so drafted messages blend in):
git log --oneline -5Note the prefix convention (
feat:/fix:/chore:/docs:...), language (中文/English), and length style. -
Classify dirty files into two groups:
- AI-edited this session — files you wrote/edited via Edit/Write/Bash tool calls in this session. You know what changed and why.
- Unrecognized — dirty files you did NOT touch this session (could be the user's manual edits, leftover WIP from a previous session, or unrelated work). Do NOT silently include these.
-
Draft a commit plan. Group AI-edited files into logical commits (1 commit per coherent change unit, not 1 commit per file). Each entry:
<commit message>+ file list. List unrecognized files separately at the bottom. -
Present the plan once, ask for one-shot confirmation. Format:
Proposed commits (in order): 1. <message> - <file> - <file> 2. <message> - <file> Unrecognized dirty files (NOT in any commit — confirm include/exclude): - <file> - <file> Reply 'ok' / '行' to execute. Reply with edits, or '我自己来' / 'manual' to abort. -
On confirmation: run
git add <files>+git commit -m "<msg>"for each batch in order. Do not amend. Do not push. -
On rejection (user replies "不行" / "我自己来" / "manual" / any pushback on the plan): stop. Do not attempt a second plan. The user will commit by hand; you skip ahead to 3.5 once they confirm.
Rules:
- No
git commit --amendanywhere — three-stage three-commit flow (work commits → archive commit → journal commit). - Never push to remote in this step.
- If the user wants different message wording but accepts the file grouping, edit the message and re-confirm once — but if they reject the grouping, exit to manual mode.
- The batched plan is one prompt; do not prompt per commit.
3.5 Wrap-up reminder
After the above, remind the user they can run /finish-work to wrap up (archive the task, record the session).
Workflow State Breadcrumbs
[workflow-state:no_task]
No active task. A Direct answer — pure Q&A / explanation / lookup / chat; no file writes + one-line answer + repo reads ≤ 2 files → AI judges, no override needed.
B Create a task — any implementation / code change / build / refactor work. Entry sequence: (1) python3 ./.trellis/scripts/task.py create "<title>" to create the task (status=planning, breadcrumb switches to [workflow-state:planning] for brainstorm + jsonl phase guidance) → (2) load trellis-brainstorm skill to discuss requirements with the user and iterate on prd.md → (3) once prd is done and jsonl is curated, run task.py start <task-dir> to enter [workflow-state:in_progress] for the implementation skeleton. For research-heavy work, dispatch trellis-research sub-agents — main agent must NOT do 3+ inline WebFetch / WebSearch / gh api calls. "It looks small" is NOT grounds for downgrading B to A or C.
C Inline change (per-turn only, escape hatch for B) — the user's CURRENT message MUST contain one of: "skip trellis" / "no task" / "just do it" / "don't create a task" / "跳过 trellis" / "别走流程" / "小修一下" / "直接改" / "先别建任务" → briefly acknowledge ("ok, skipping trellis flow this turn"), then inline. Without seeing one of these phrases you must NOT inline on your own; do not invent an override the user never said.
[/workflow-state:no_task]
[workflow-state:planning]
Load the trellis-brainstorm skill and iterate on prd.md with the user.
Phase 1.3 (required, once): before task.py start, you MUST curate implement.jsonl and check.jsonl — list the spec / research files sub-agents need so they get the right context injected. You may skip only if the jsonl already has agent-curated entries (the seed _example row alone doesn't count).
Then run task.py start <task-dir> to flip status to in_progress.
Research output must land in {task_dir}/research/*.md, written by trellis-research sub-agents. The main agent should not inline WebFetch / WebSearch — the PRD only links to research files.
[/workflow-state:planning]
[workflow-state:in_progress]
Flow: trellis-implement → trellis-check → trellis-update-spec → commit (Phase 3.4) → /trellis:finish-work.
Main-session default (no override): dispatch the trellis-implement / trellis-check sub-agents — the main agent does NOT edit code by default. Phase 3.4 commit (required, once): after trellis-update-spec, or whenever implementation is verifiably complete, the main agent drives the commit — state the commit plan in user-facing text, then run git commit — BEFORE suggesting /trellis:finish-work. /finish-work refuses to run on a dirty working tree (paths outside .trellis/workspace/ and .trellis/tasks/).
Sub-agent self-exemption: if you are already running as trellis-implement, implement directly from the loaded task context and do NOT spawn another trellis-implement; if you are already running as trellis-check, review/fix directly and do NOT spawn another trellis-check. The default dispatch rule applies to the main session only.
Sub-agent dispatch protocol (all platforms, all sub-agents EXCEPT trellis-research): When you spawn trellis-implement / trellis-check, your dispatch prompt MUST start with one line: Active task: <task path from \task.py current`>. No exceptions. On class-2 platforms (codex / copilot / gemini / qoder) the sub-agent depends on this line because there is no hook to inject task context. On class-1 platforms (claude / cursor / opencode / kiro / codebuddy / droid) the line is normally redundant — the hook injects context directly — but it serves as a critical fallback when the hook fails (Windows + Claude Code PreToolUse silent skip, --continueresume, fork distribution, hooks disabled, etc.).trellis-research` does not need this line because it operates without a task binding.
Inline override (per-turn only, escape hatch for sub-agent dispatch): the user's CURRENT message MUST explicitly contain one of: "do it inline" / "no sub-agent" / "你直接改" / "别派 sub-agent" / "main session 写就行" / "不用 sub-agent". Without seeing one of these phrases you must NOT inline on your own; do not invent an override the user never said.
[/workflow-state:in_progress]
[workflow-state:completed]
Code committed via Phase 3.4; run /trellis:finish-work to wrap up (archive the task + record session).
If you reach this state with uncommitted code, return to Phase 3.4 first — /finish-work refuses to run on a dirty working tree.
task.py archive deletes any runtime session files that still point at the archived task.
[/workflow-state:completed]
[workflow-state:planning-inline]
Load the trellis-brainstorm skill and iterate on prd.md with the user.
Phase 1.3 jsonl curation is skipped in inline dispatch mode — the main session loads trellis-before-dev directly in Phase 2 and reads spec context itself, so there is no sub-agent to inject jsonl into.
Then run task.py start <task-dir> to flip status to in_progress.
Research output must land in {task_dir}/research/*.md. In inline mode the main session may do research itself or dispatch trellis-research sub-agents.
[/workflow-state:planning-inline]
[workflow-state:in_progress-inline]
Flow (inline mode): main session loads trellis-before-dev → main session edits code → main session loads trellis-check → run lint / type-check / tests → fix → trellis-update-spec → commit (Phase 3.4) → /trellis:finish-work.
Main-session default (inline dispatch_mode): the main agent edits code directly. Do NOT dispatch trellis-implement / trellis-check sub-agents. Load the trellis-before-dev skill before writing code; load the trellis-check skill before reporting completion.
Phase 3.4 commit (required, once): after trellis-update-spec, or whenever implementation is verifiably complete, the main agent drives the commit — state the commit plan in user-facing text, then run git commit — BEFORE suggesting /trellis:finish-work. /finish-work refuses to run on a dirty working tree (paths outside .trellis/workspace/ and .trellis/tasks/).
[/workflow-state:in_progress-inline]
[workflow-state:my-status] your per-turn prompt text [/workflow-state:my-status]