coding

name	coding
description	Use when performing coding tasks (implement features, fix bugs, refactor code). Orchestrates language detection, mode switching (normal/autopilot/full-auto), quality enforcement, and verification.
allowed-tools	["shell","apply_patch","read_file","write_file","update_plan"]
metadata	{"short-description":"General coding orchestrator with modes"}

Coding Orchestrator

This skill serves as the primary entry point for general coding tasks. It provides:

• Mode switching: normal (confirm all) / autopilot (confirm risky) / full-auto (minimal confirmation)
• Language detection: Auto-detect project language and apply appropriate guidance
• Quality enforcement: Integrate task-discipline, code-quality, verification patterns
• Risk-based confirmation: Consistent behavior based on operation risk tier

Read these references:

• references/workflow.md - Universal coding workflow
• references/quality.md - Quality priorities (correctness > safety > maintainability)
• references/safety.md - Runtime + security safety
• references/risk-taxonomy.md - 4-tier operation risk classification
• references/guardrails.md - Per-mode guardrails and limits
• ~/.config/agent/core/methodology.md - 5-phase problem-solving
• ~/.config/agent/core/task-management.md - Task discipline patterns

---

Mode Selection

Active Mode Determination

Mode is determined by (priority order):

1. Explicit argument: mode:autopilot or mode:full-auto prefix in arguments
2. Session context: Previously set mode (sticky within session)
3. Default: normal

Mode Definitions

Mode	Philosophy	Confirmation Behavior
normal	Maximum safety	Confirm for all modifications (Tier 1/2/3)
autopilot	Balanced autonomy	Confirm for risky operations (Tier 1/2), auto for safe (Tier 3/4)
full-auto	Maximum autonomy	Confirm for critical only (Tier 1), auto for all else

See references/risk-taxonomy.md for full tier definitions.

Mode Switching

Explicit mode in arguments:

mode:autopilot fix the bug in handler.go
mode:full-auto add tests to all packages

Conversational switching: User says any of:

• "switch to autopilot mode" / "use autopilot"
• "switch to full-auto mode" / "go full auto"
• "back to normal mode" / "use normal mode"

Mode persistence: Once set, mode persists for the entire session until explicitly changed. Each new session starts in normal mode.

Escalation safeguards:

• normal → autopilot: Acknowledge, proceed
• normal → full-auto: Warn about risks, confirm intent, proceed
• autopilot → full-auto: Warn about risks, confirm intent, proceed
• De-escalation (to more restrictive): Acknowledge, proceed immediately

---

Language Detection

Detect the project's primary language to apply language-specific guidance.

Detection Priority

1. Explicit in request: "fix the Go handler", "update the Python script"
2. File extension: .go, .py, .rs, .ts, .js, .java, etc.
3. Project markers (check working directory):
   • go.mod → Go
   • Cargo.toml → Rust
   • package.json → JavaScript/TypeScript
   • pyproject.toml or requirements.txt → Python
   • pom.xml or build.gradle → Java
4. Ask if ambiguous: Multiple languages detected or unclear

Language-Specific Guidance

When a language is detected, read and apply ALL guidance from that language's domain:

Go (detected):

Read and apply:
- `~/.config/agent/domain/coding/go/style.md`
- `~/.config/agent/domain/coding/go/idioms.md`
- `~/.config/agent/domain/coding/go/concurrency.md`
- `~/.config/agent/domain/coding/go/testing.md`
- `~/.config/agent/domain/coding/go/tooling.md`
- `~/.config/agent/domain/coding/go/organization.md`
- `~/.config/agent/domain/coding/go/performance.md`

Python (future):

Read and apply:
- `~/.config/agent/domain/coding/python/style.md`
- `~/.config/agent/domain/coding/python/idioms.md`
- (etc.)

Other languages: Apply generic workflow from references/workflow.md and references/quality.md.

---

Workflow

The coding skill follows a 5-phase workflow with mode-specific confirmation behavior.

Phase 1: Context Gathering

Objectives:

• Determine active mode
• Detect project language
• Read relevant files before modifying
• Load language-specific guidance

Steps:

1. Parse arguments for mode prefix (mode:autopilot, mode:full-auto)
2. If no explicit mode, check session context for previously set mode
3. If no session mode, use normal as default
4. Detect language using priority algorithm
5. Read target files (never modify before reading)
6. Load language-specific guidance modules

Mode behavior: All modes identical (read-only phase).

Phase 2: Planning (if non-trivial)

When to plan:

• Tasks requiring 3+ distinct steps
• Multi-file modifications
• Complex logic changes
• Non-obvious implementation approach

Apply task-discipline patterns:

1. Create TODO list with atomic, testable tasks
2. Use imperative form for content: "Fix bug in parser"
3. Use present continuous for activeForm: "Fixing bug in parser"
4. Ensure exactly ONE task is in_progress at any time
5. Mark tasks completed immediately after finishing

Mode behavior: All modes identical (planning phase).

Phase 3: Implementation

Apply universal coding workflow from references/workflow.md:

1. Understand before modifying: Read code, understand context
2. Make minimal, targeted changes: Don't refactor unrelated code
3. Test incrementally: Don't save all testing for the end
4. Document non-obvious decisions: Add comments where logic isn't self-evident

Mode-specific confirmation behavior:

Normal mode:

• Propose complete plan upfront
• Confirm before EACH file modification
• Show exact changes (diff if useful)
• Offer options: [P]roceed / [D]iff / [S]kip / [A]bort

Autopilot mode:

• Proceed automatically for Tier 3/4 operations
• Report progress: "Created X, Modified Y"
• Pause only for Tier 1/2 operations (see references/risk-taxonomy.md)
• Create checkpoint (git stash) before destructive operations
• Track operation count and time limits (see references/guardrails.md)
• Commits are never auto-run; require explicit confirmation/manual action

Full-auto mode:

• Proceed automatically for Tier 2/3/4 operations (except never-auto overrides)
• Audit log all operations with timestamps (see references/guardrails.md)
• Create checkpoints before major operations (>5 files or git push)
• Pause for Tier 1 CRITICAL operations and never-auto overrides (e.g., commits)
• Report progress every 10 operations or 5 minutes
• Stop after 3 consecutive failures, offer recovery options

Phase 4: Quality Gates

Before claiming completion, enforce quality checks:

Apply code-quality patterns:

1. Correctness: Does the code do what it's supposed to?
2. Safety: Any runtime safety issues (nil dereferences, races, leaks)?
3. Maintainability: Is the code clear and simple?
4. Performance: Any obvious inefficiencies?
5. Style: Does it follow language conventions?

Language-specific quality checks:

Go:

# Run tests (if applicable)
go test -race -count 1 ./...

# Run linter (if applicable)
golangci-lint run --new=false ./...

# Format code
golines --base-formatter=gofumpt --ignore-generated --max-len=79 --tab-len=4 --no-reformat-tags --chain-split-dots -w .
gofumpt -w .

Mode behavior for quality checks:

• Normal: Confirm before running each tool
• Autopilot: Auto-run tests/lint/format (Tier 3), report results
• Full-auto: Auto-run all quality checks, only stop if failures block progress

Phase 5: Verification

Apply verification-before-completion patterns:

NEVER claim completion without evidence:

1. Run the verification command (tests, lint, build)
2. Confirm output supports the claim
3. ONLY THEN declare completion

Examples of evidence:

• Tests passing: ok github.com/user/repo/package 0.123s
• Lint clean: golangci-lint run with no output or exit 0
• Build successful: go build -o /dev/null ./... with exit 0
• Manual verification: Show output of commands that confirm fix

Mode behavior:

• All modes must provide evidence before completion
• Cannot skip verification even in full-auto
• If verification fails, apply recovery procedures (see references/guardrails.md)

---

Cross-Skill Coordination

Absorb (Apply Inline)

These skills' patterns are absorbed and applied directly within the coding workflow:

Skill	When	What to Apply
task-discipline	3+ step tasks	TODO list with exactly ONE in_progress at any time
code-quality	All coding	Priority order: correctness > safety > maintainability > performance > style
verification-before-completion	Before completion claims	Evidence requirement: run verification, show output, then claim success

Delegate (Recommend Invocation)

Recommend explicit skill invocation when appropriate:

Condition	Recommended Skill	When to Recommend
Bugs/test failures encountered	systematic-debugging	After 2 failed fix attempts or unclear root cause
Security-sensitive code	safety-check	When modifying auth, crypto, input validation, or data access
Pre-commit checks needed	go-precommit (or language equivalent)	Before creating commits
Code review requested	review	User explicitly asks for review
Strict TDD requested	tdd-enforce	User explicitly requests test-first approach

Recommendation format:

I encountered test failures. I recommend using the `systematic-debugging` skill
for a methodical root-cause analysis.

Would you like me to:
1. Continue debugging manually
2. Invoke the systematic-debugging skill
3. Ask you for guidance

---

Operation Risk Classification

Before performing any operation, classify it using the risk taxonomy:

Tier 1 (CRITICAL): Force push main, credentials, prod DB, external API mutations Tier 2 (HIGH): File deletion, git state changes (commit/push/merge/rebase), non-deterministic line edits, package management Tier 3 (MEDIUM): New files, tests, deterministic formatting/lint fixes, builds Tier 4 (LOW): Read-only operations

See references/risk-taxonomy.md for complete classification algorithm and examples.

Apply confirmation behavior based on mode and tier:

Tier	Normal	Autopilot	Full-Auto
1	Confirm	Confirm	Confirm
2	Confirm	Confirm	Auto
3	Confirm	Auto	Auto
4	Auto	Auto	Auto

Never-auto overrides (e.g., git commit) supersede this matrix.

---

Guardrails and Limits

Apply limits from references/guardrails.md based on active mode:

Autopilot limits:

• Max 10 files modified per operation
• Max 100 lines modified per file
• Auto-stash before destructive operations

Full-auto limits:

• Report progress every 10 operations or 5 minutes
• Pause after 30 minutes
• Stop after 3 consecutive failures

Absolute never-auto (all modes including full-auto):

• Git commits (git commit, git commit --amend)
• Force push to main/master
• Credential/secret operations
• Production deployments
• Database migrations on prod

---

Recovery and Safety

Checkpoints

Autopilot mode:

# Before Tier 2 destructive operations
git stash push -m "autopilot-checkpoint-$(date +%s)"

Full-auto mode:

# Before operations affecting >5 files or git push
git stash push -m "full-auto-checkpoint-$(date +%s)"

Rollback on Failure

If an operation fails and a checkpoint exists:

Operation failed: tests not passing after modification

Recovery options:
1. [R]evert to checkpoint
2. [D]ebug the issue
3. [C]ontinue anyway (not recommended)
4. [S]witch to normal mode

Choice: _

Audit Logging

Full-auto mode logs all operations. Autopilot logs Tier 2/3 operations. See references/guardrails.md for file location and redaction requirements.

Example format:

[12:00:01] [FULL-AUTO] [TIER-3] Created: api/handler.go [SUCCESS]
[12:00:03] [FULL-AUTO] [TIER-3] Created: api/handler_test.go [SUCCESS]
[12:00:05] [FULL-AUTO] [TIER-2] Modified: api/routes.go (+15 lines) [SUCCESS]

---

Integration with Existing Skills

This skill complements (does not replace) existing skills:

Existing Skill	Relationship
go-development	coding applies Go guidance when Go detected; users can still invoke go-development directly
feature	coding handles tactical implementation; feature handles strategic 7-phase workflow
code-quality	coding absorbs quality patterns inline
task-discipline	coding absorbs task patterns for 3+ step work
verification-before-completion	coding absorbs verification gate requirement
systematic-debugging	coding recommends invocation when bugs encountered
safety-check	coding recommends invocation for security-sensitive code
review	coding recommends invocation when review requested

---

Activation Triggers

This skill activates for:

• General coding requests: "implement X", "fix bug in Y", "refactor Z"
• Mode-prefixed requests: "mode:autopilot add error handling"
• Generic file modifications (when not caught by more specific skills)

This skill does NOT activate for:

• Planning/design requests (use planner)
• Review requests (use review)
• Explicit debugging requests (use systematic-debugging)
• Language-specific tool invocations (use go-test, go-lint, etc.)

---

Arguments

Mode: ${MODE} Target: ${ARGUMENTS}

Argument patterns:

• mode:normal implement user authentication
• mode:autopilot fix all lint errors
• mode:full-auto add tests for parser package
• fix bug in handler.go:42 (defaults to normal mode)