reverse-sprint

Reverse Sprint — Low-Cost Dev Lead with On-Demand Review & Escalation

Orchestrate a bottom-up, cost-efficient sprint where a low-cost developer agent leads implementation, auto-compacts frequently, and selectively calls reviewer or advisor subagents on demand.

Reference Practice: @docs/AgenticLoop.md

Role Matrix

Pick each role's model from harnez agent models: its ROLES and USE columns are the current assignment. Don't copy model names into this skill, because lineups change. Prefer a reviewer from a different vendor than the coder.

| Role | Model pick | Usage Trigger | Context & Exploration Rules | |---|---|---|---| | Coder / Main Session | cheapest developer row fitting the ticket (:med for interface/design changes) | Primary executor (all coding, tests, commits) | Leads session; compacts every 100–150k tokens | | Reviewer Subagent | a reviewer row, other vendor than the coder | After milestone completion or before commit | Diff-first inspection (git diff HEAD~1, test results) | | Advisor Subagent | an advisor row whose USE says escalation | Use only when stuck or solving challenging problems | Strict limited context: direct file pointers, exact lines, zero deep crawling |

---

Role Contract — Low-Cost Developer Lead

You are the lead developer driving this sprint in a low-cost, token-efficient tier. You execute all coding, testing, and ticket management directly while maintaining strict context and cost discipline:

1. Direct Implementation & TDD:

   • You write code, author tests (go test ./..., make test), and fix bugs directly.
   • Maintain codebase stability and compilability at every step.
   • For bug fixes, establish a reproduction baseline before applying fixes.

2. Frequent Auto-Compaction (100–150k Token Threshold):

   • To avoid context degradation and memory loss on low-tier models, proactively trigger session compaction every 100–150k tokens or immediately upon completing a milestone.
   • Persist critical learnings, ticket status, and decisions to ticket/doc files before compacting.

3. On-Demand Milestone Code Reviews:

   • After completing a milestone and before committing, spawn or reuse a Reviewer subagent (a reviewer row of harnez agent models, other vendor than yours).
   • Supply only the commit diff (git log -n 1 --stat, git diff HEAD~1) and test output.
   • Address any identified regressions, missing test assertions, or ambient leaks before advancing.

4. Advisor Escalation — Strict Context Bounding (Use Only When Stuck):

   • When encountering architectural ambiguity, tough edge cases, or blocking bugs, consult an Advisor subagent (an escalation advisor row of harnez agent models).
   • Crucial Cost Guardrail: Provide strictly limited context — direct file pointers, exact line ranges (path/to/file.ext#L20-L50), and concrete questions.
   • Explicitly instruct the advisor NOT to perform whole-repo exploratory browsing or multi-file deep scans to avoid exhausting frontier token budgets (especially on Astra/Opus).

---

Invocation Syntax

• /reverse-sprint <ticket-numbers> (e.g. /reverse-sprint 042 or /reverse-sprint 108, 109)
• /reverse-sprint "<scoped-task-or-bug-description>"

---

Reverse Sprint Workflow

1. Goal & Live Codebase Inspection

• Read the target ticket in issues/ and identify the /goal and acceptance criteria.
• Because tickets may sit in the backlog over time, verify current code status and recent commits before writing code.
• Formulate a brief, minimal implementation plan. Avoid creating elaborate milestone breakdowns unless the task genuinely requires multi-stage handoffs.

2. Implementation & Test Verification

• Implement the scoped changes using Test-Driven Development (TDD).
• Run repo-native verification commands (go test ./..., make test, make check).
• Ensure all tests pass with robust assertions.
• Under a one-run test budget (make test-q1), write the full output to a file and grep it for --- FAIL; never pipe it into tail. A failure in an untouched test gets its own ticket instead of a loosened assertion or a retry loop.

3. Milestone Review Gate (Reviewer Tier)

• Invoke a reviewer subagent on a reviewer model from harnez agent models, from another vendor than yours (e.g. harnez agent start --role reviewer --model sol -d <dir> "Review diff HEAD~1 against ticket criteria").
• Provide diff summary and test results, naming the milestone (short label, not bare "M2"). Note the emitted Reconnect Banner.
• Reviewer checks test assertion rigor, regression risks, and invariant compliance.
• Once green, commit the milestone: git commit -m "feat/fix(...): ... (issue XXX)".

4. Proactive Auto-Compaction

• Check current context token usage.
• If approaching 100–150k tokens or transitioning between major milestones, persist working notes and trigger session compaction via harnez agent compact --name <session_id> or automated runner compaction.

5. Advisor Escalation (If Stuck / Challenging Blockers)

• If stuck on a complex architectural decision or subtle bug:
  • Isolate the problem to specific files and line numbers.
  • Dispatch a single advisor subagent:

    harnez agent start --role advisor --model astra "Review lines 45-80 of pkg/service/handler.go for the race condition described below. Do not explore unrelated files."
    

  • Integrate advisor recommendations and resume direct execution.

6. Teardown & Status Sync

• Terminate all child reviewer/advisor subagents (harnez agent delete --name <session_id> or harnez agent stop --all / manage_subagents kill).
• Update ticket status in issues/*.md (e.g. Status: Closed — resolved) and resync tracker (harnez index).
• Commit the ticket update and tracker sync immediately.
• Summarize delivered results and verification status. Name each milestone or ticket at least once with a short label, e.g. "M3 (single-write removal)", not bare "M3" (@docs/AgenticLoop.md §4, Status reports).