feat(pipelines): unify platform-specific implement pipelines into a single configurable pipeline with skill-based flavoring

[nextlevelshit]

Problem

Wave currently maintains separate pipelines for each Git hosting platform (GitHub, GitLab, Gitea, Bitbucket) across 7 pipeline families — 25 files total. This leads to:

• Duplicated pipeline definitions with only minor platform-specific differences
• Platform-specific API instructions embedded directly in persona prompts (e.g., raw curl commands for Bitbucket REST API)
• Maintenance burden: every pipeline change must be replicated across all platform variants
• Copy-paste bugs and inconsistencies accumulating across variants

As a developer, I typically use only one hosting platform, so having all variants is unnecessary.

Affected Pipeline Families

Family	Duplicated Files	Platform-Agnostic Steps	Platform-Specific Steps
implement-epic	gh-, gl-, gt-, bb- (4)	Most of the DAG	CLI commands in fetch/create steps
implement	gh-, gl-, gt-, bb- (4)	Steps 2-3 (plan, implement)	Steps 1, 4 (fetch-assess, create-pr)
scope	gh-, gl-, gt-, bb- (4)	Step 2 (scope-and-create logic)	Steps 1, 3 (fetch-epic, verify-report)
research	gh-, gl-, gt-, bb- (4)	Steps 2-4 (100% identical)	Steps 1, 5 (fetch-issue, post-comment)
rewrite	gh-, gl-, gt-, bb- (4)	Scoring/enhancement logic	CLI commands in both steps
refresh	gh-, gl-, gt-, bb- (4)	Step 2 (100% identical)	Steps 1, 3 (gather-context, apply-update)
pr-review	gh- only (1)	Steps 1-4 (diff, security, quality, summary)	Step 5 only (publish comment)

Total: 25 pipeline files → 7 unified pipelines

Existing Bugs from Duplication

The copy-paste approach has already introduced inconsistencies:

Bug	Family	Details
Hardcoded test command	implement	gl/gt hardcode go test -race ./... instead of {{ project.test_command }}
Missing release scanning	refresh	gt-refresh skips release/tag scanning entirely (sub-step absent)
Missing --repo flag	refresh	gl-refresh omits --repo in apply-update commands
Contradictory instructions	scope	bb-scope verify-report says "post a summary comment" then says "Do NOT post comments"
Missing prompt sections	implement	bb-implement is missing Agent Usage, Tool Usage, and Error Handling sections
GitHub-named schemas for all platforms	rewrite	All 4 variants reference github-enhancement-plan.schema.json
Different plan structure	implement	bb-implement plan step has completely different structure (4 steps vs 5, no spec/plan/tasks)

Proposed Solution

Consolidate all platform-specific pipelines into single configurable pipelines that:

1. Platform configuration in wave.yaml: Add an option to specify the target Git hosting platform (e.g., platform: github)
2. Skill-based flavoring: Replace inline platform-specific prompts with proper skills (/bitbucket, /github, /gitlab, /gitea) that encapsulate all platform knowledge
3. Programmatic API interactions: Replace raw curl + jq instructions in prompts with actual Go code that handles API calls — LLMs should not be executing raw REST API commands

Example of what gets eliminated

Platform-specific prompt blocks like:

Step 2: Fetch issues via the Bitbucket REST API.
- Single issue:
  curl -s -H "Authorization: Bearer $BB_TOKEN" \
    "https://api.bitbucket.org/2.0/repositories/WORKSPACE/REPO/issues/NUM" \
    | jq '{id, title, content: .content.raw, kind, url: .links.html.href}'

…would be replaced by a /bitbucket skill that provides the persona with the necessary context and tooling.

Bitbucket as the worst offender

Every bb-* pipeline embeds raw curl + jq REST API calls with $BB_TOKEN referenced directly in prompt text. The other three platforms use dedicated CLIs (gh, glab, tea) which handle authentication internally. This means:

• Bearer tokens appear in prompt text sent to the LLM
• Hardcoded API URLs and JSON payload construction in prompts
• Significantly more verbose and fragile prompts

Acceptance Criteria

• A single implement-epic pipeline replaces gh-implement-epic, gl-implement-epic, gt-implement-epic, and bb-implement-epic
• A single implement pipeline replaces gh-implement, gl-implement, gt-implement, and bb-implement
• A single scope pipeline replaces gh-scope, gl-scope, gt-scope, and bb-scope
• A single research pipeline replaces gh-research, gl-research, gt-research, and bb-research
• A single rewrite pipeline replaces gh-rewrite, gl-rewrite, gt-rewrite, and bb-rewrite
• A single refresh pipeline replaces gh-refresh, gl-refresh, gt-refresh, and bb-refresh
• pr-review pipeline supports all four platforms (currently GitHub-only)
• Platform selection is configurable via wave.yaml or pipeline parameter
• Platform-specific knowledge is encapsulated in skills (e.g., /github, /bitbucket)
• Raw API call instructions are removed from persona prompts
• Existing functionality for all four platforms is preserved
• Tests cover each unified pipeline with each platform flavor
• Existing duplication bugs (see table above) are fixed during consolidation

Out of Scope

• Adding support for new hosting platforms (that can follow as separate issues)

[nextlevelshit]

Cross-Reference: #296

The new #296 (AI-steered project optimization) is a key enabler for this issue. Once pipelines are unified (one implement pipeline instead of gh-implement/bb-implement/gl-implement/gt-implement), the forge detection from internal/forge/detect.go + the project profile from #296 determine which platform skills and forge CLI to inject. This makes pipeline unification practical — the project: config block and forge type drive the per-platform behavior instead of separate pipeline files.

Active implementation note

Forge detection (internal/forge/detect.go) is currently being implemented and already provides ForgeType, CLITool, and PipelinePrefix — the foundation for platform-aware unified pipelines.

[nextlevelshit]

Pipeline Unification Audit (March 2026)

Full audit of all 25 platform-specific pipeline YAML files + 32 prompt files + 16 personas.

Inventory

Family	gh	gl	gt	bb	Total
implement	✓	✓	✓	✓	4
implement-epic	✓	✓	✓	✓	4
scope	✓	✓	✓	✓	4
research	✓	✓	✓	✓	4
rewrite	✓	✓	✓	✓	4
refresh	✓	✓	✓	✓	4
pr-review	✓	—	—	—	1
Total	7	6	6	6	25

Plus: 32 prompt files in 8 forge-specific directories, 16 forge-specific personas in wave.yaml.

Bugs Found During Audit

#	Bug	Family	Severity
1	gl/gt plan.md references nonexistent create-new-feature.sh	implement	Critical — step fails
2	gl/gt implement.md hardcodes go test -race ./... instead of {{ project.test_command }}	implement	High — breaks non-Go
3	bb-implement plan.md + implement.md are completely different, degraded prompts	implement	High — missing TDD, agent usage
4	gt-refresh missing release checking step (step 3 absent)	refresh	Medium
5	gl-refresh apply-update missing --repo flag	refresh	Medium
6	bb-implement-epic uses $BITBUCKET_TOKEN while all other bb-* use $BB_TOKEN	implement-epic	Low
7	All 4 rewrite variants reference github-enhancement-plan.schema.json	rewrite	Low — works but confusing
8	bb-scope verify-report intro says "post a summary comment" then says "Do NOT post comments"	scope	Medium — contradictory
9	gt-research uses gh-style --json flags that tea CLI doesn't support	research	High — step may fail
10	gh-rewrite has URL parsing, gl/gt/bb lack it	rewrite	Low — feature gap

Sharing Analysis Per Family

Family	Steps	Shared	Platform-specific	Key difference
implement	4	Steps 1-3 structure identical	Step 4 (create-pr) CLI commands, step 1 CLI	bb prompts are completely different
implement-epic	3	Topology + artifacts identical	Personas, CLI commands, child_pipeline ref	bb env var inconsistency
scope	3	Structure identical	Personas + inline CLI commands	All reference github-named schemas
research	5	Steps 2-4 word-for-word identical	Steps 1,5 CLI commands	Cleanest candidate for unification
rewrite	2	Structure identical	Personas + CLI in inline prompts	Schema naming bug
refresh	3	Step 2 word-for-word identical	Steps 1,3 CLI commands	gt missing release step
pr-review	5	Steps 1-4 platform-agnostic	Step 5 only (publish)	No gl/gt/bb variants exist

Prerequisites for Unification

1. Add {{ forge.* }} template variables to PipelineContext:

   • forge.type, forge.cli_tool, forge.prefix, forge.host
   • forge.pr_term ("Pull Request" vs "Merge Request")
   • forge.pr_command ("pr" vs "mr" vs "pulls")

2. Persona resolution: Solve how unified pipelines reference forge-specific personas:

   • Option A: persona: "{{ forge.prefix }}-commenter" (template in persona field)
   • Option B: persona_map: {github: github-commenter, ...}
   • Option C: 4 generic personas with dynamic tool permissions (separate issue)

3. Conditional requires.tools: "{{ forge.cli_tool }}" or a requires.forge_tools: true flag

Implementation Plan

Phase	What	Risk	Size
1	Add ForgeInfo to PipelineContext, expose {{ forge.* }} vars	Low	S
2	Fix 10 bugs above in existing pipelines	Low	S
3	Unify research family (cleanest, steps 2-4 already identical)	Low	S
4	Unify refresh family (step 2 already identical)	Low	S
5	Unify implement family (most important, most battle-tested)	Medium	M
6	Unify remaining families (scope, rewrite, implement-epic)	Medium	M
7	Extend pr-review to all forges	Low	S
8	Cleanup old files (24 YAMLs, 32 prompts)	Low	S

Total: 24 files removed, 7 unified files created, 2 unified prompt directories.

Persona consolidation (16 → 4) should be a separate issue.

Critical Dependency: Bitbucket

The bb-* pipelines are fundamentally different — they use curl/jq against the REST API instead of a dedicated CLI. This means:

• Bearer tokens appear in prompt text
• More verbose prompts with full API URLs
• Conditional prompt sections or forge-specific snippets may be needed rather than simple variable substitution

[nextlevelshit]

Fixed 10 pipeline bugs in 9d0c69e and dadc55f:

1. Removed nonexistent script references in gl/gt/gh plan.md
2. Replaced hardcoded go test -race ./... with {{ project.test_command }} across all implement.md files
3. Rewrote degraded bb-implement prompts to match gh/gl/gt pattern
4. Added release checking to gt-refresh
5. Added --repo flag to gl-refresh glab commands
6. Fixed $BITBUCKET_TOKEN → $BB_TOKEN in bb-implement-epic
7. Created forge-neutral schema files, removed stale github-prefixed schemas
8. Fixed bb-scope contradictory verify-report instructions
9. Fixed gt-* pipelines --json → --output json
10. Added URL parsing to gl/gt/bb rewrite scan-and-score steps