Sprint 152 — Blog [Technical Challenge] — Sliding Window for Agent Context Optimization

Sprint 152 — Blog [Technical Challenge] — Sliding Window for Agent Context Optimization

  • Status: Completed ✅
  • Period: 2026-05-13 (single day)
  • Trigger: User request (blog post authoring)
  • start_commit: 4313561 (Sprint 151 hotfix end)
  • end_commit: 3873f6d (Sprint 152 English version end)
  • Merged PRs: 3 (squash merge)

1. Background / User Request

Operational experience accumulated in AlgoSu's agent collaboration workflow — particularly the need for maintenance-perspective documentation + context optimization via sliding window system introduction — was compiled into a [Technical Challenge] category blog post. Sprint 145~151 accumulated zero-regression operational results were used as empirical evidence.

2. Decisions

2.1 Post Concept

  • Category: challenge (same as the 3 existing posts: toward-model-agnostic-harness / baekjoon-gone / session-policy-sync, order: 10 — previous max=9)
  • 3-step core logic (presented directly by user):
    1. Memory is volatile → documentation is essential (especially in agent collaboration)
    2. Referencing all documents every session = "pouring water into an already-full bucket" (paradox)
    3. Drawing on the sliding window algorithm → introduced as a solution
  • Title: "Memory Fades, Documents Pile Up — Optimizing Agent Context with a Sliding Window"

2.2 i18n Policy

  • All existing 9 posts have both KR (blog/content/posts/) + EN (blog/content/posts-en/) versions
  • New posts are also obligated to have both versions — missed in this sprint and immediately corrected after user prompt (PR #235)
  • English version tone reference: toward-model-agnostic-harness.mdx (1st person, dash emphasis, bold pattern)

2.3 Body Format

  • mdx components: MetricGrid/MetricCard × 2, Callout × 1
  • mermaid × 2 (state machine + 4-layer storage structure)
  • ASCII diagrams × 2 (context accumulation + sliding window principle)
  • 7 GitHub URLs embedded directly in body (no footnotes) — ADR directory / sprint-145·151 ADR / CLAUDE.md / .claude/commands/agents/

2.4 Immediate Re-routing on Broken Plan Assumption (Sprint 147 Lesson Directly Applied)

  • User's plan specified "librarian dispatch" but persona cross-check revealed librarian = DB persistence/migration specialist (outside document curation domain)
  • Immediately re-routed to Scribe solo handling — Scribe's exact domain: "exclusively responsible for accuracy and currency of all project memory, documents, and prompts"
  • Blog work has a fast user review cycle → Oracle proceeded directly in Scribe persona instead of tmux dispatch

3. Implementation Flow

3.1 PR #233 — Korean Version New (9034a66)

1 commit, 1 file, +297

  • blog/content/posts/sliding-window-agent-context.mdx newly created
  • frontmatter: title / date(2026-05-13) / category(challenge) / order(10) / tags 5 types (agent, documentation, sliding-window, context-optimization, claude-code)
  • Body 10 sections:
    1. The 152nd Sprint, and Accumulated ADRs (introduction)
    2. Memory Is Volatile — Both Humans and Agents Forget
    3. Documentation Really Worked — Empirical Evidence: 8 Regression Blocking Dimensions Accumulated from Sprint 145~151
    4. But — The Documents Themselves Started Becoming the Enemy
    5. The Full Bucket Analogy — lost-in-the-middle, attention dispersion
    6. The Sliding Window Algorithm — Returning to Algorithm Time
    7. sprint-window.md size 2 + idle/active status state machine
    8. 4-Layer Storage Structure — Separating "Where We Forget" from "Where We Remember"
    9. Operational Results — 7 Sprints Zero Regressions + Auto-Critic One-Set Philosophy
    10. Learning to Forget Was Learning to Remember (closing)

3.2 PR #234 — Fact Correction Hotfix (9c8caa3)

1 commit, 1 file, +4 -4

  • Immediately after merge, user fact-check pointed out: sprint-by-sprint ADRs were introduced starting Sprint 62 and have accumulated to nearly 90, but the draft had 4 inaccurate expressions
  • 4 corrections:
    • H2: "The 152nd Sprint, and 152 ADRs" → "Accumulated ADRs"
    • Introduction: "from sprint-001 to sprint-151" → Sprint 62 introduction background + "from sprint-62 to sprint-151, nearly 90"
    • Body mid-section: "When the ADR count exceeded 100" → "When sprint-by-sprint ADRs had accumulated to nearly 40"
    • Closing: "ADRs will grow from 152 to 200, 300" → "Sprint-by-sprint ADRs will grow from 90 to 100, 200"
  • Factual basis: git ls-tree -r main --name-only | grep "docs/adr/sprints/" → sprint-62.md ~ sprint-148.md, 85 tracked (sprint-149/150/151 untracked separately)

3.3 PR #235 — English Version Added (3873f6d)

1 commit, 1 file, +297

  • Immediately after merge, user pointed out i18n omission
  • blog/content/posts-en/sliding-window-agent-context.mdx newly created
  • English title: "Memory Fades, Documents Pile Up — A Sliding Window for Agent Context"
  • All PR #234 fact corrections simultaneously applied (Sprint 62 introduction / nearly 90 / 40 accumulated point / 90200 outlook)
  • Tone match: toward-model-agnostic-harness.mdx reference — 1st person + dash emphasis + bold pattern
  • mdx components / mermaid / ASCII / 7 GitHub URLs all 1:1 identical to Korean version

4. Verification

PRCImergeStateStatusNotes
#23328 SUCCESS / 0 FAIL / 11 SKIPPEDBEHIND → CLEANupdate-branch then merge
#23428 SUCCESS / 0 FAIL / 11 SKIPPEDCLEANhotfix passed immediately
#23528 SUCCESS / 0 FAIL / 11 SKIPPEDCLEANEnglish version passed immediately
  • i18n verification: KR (blog/content/posts/) 10 posts = EN (blog/content/posts-en/) 10 posts ✅
  • mdx component existence verification: MetricGrid/MetricCard (blog/src/components/blog/metric-grid.tsx), Callout (blog/src/components/blog/callout.tsx)
  • Branch discipline ✅: All 3 PRs use new branches + Squash merge — 18 consecutive sprints compliant (since Sprint 134 violation), 0 direct commits to main

5. New Patterns

  • Immediate re-routing on broken plan assumption (Sprint 147 lesson directly re-applied): User's plan specified "librarian dispatch" but persona cross-check revealed domain mismatch → immediately re-routed to Scribe solo handling, 0 cycle impact. Same pattern recurred after 5 sprints
  • Blog work dispatch bypass pattern: 0 code changes + fast user review cycle → Oracle directly adopts persona instead of tmux dispatch. Suitable for single-post unit (plan→supplement→write) cycle (feedback-blog-workflow)
  • User direct fact-check as the last safety net for posts: Auto-Critic and CI cannot catch factual accuracy. Domain facts (ADR start sprint / accumulation count, etc.) require grep cross-check procedure before merge — automation candidate
  • i18n omission immediate hotfix pattern: New blog posts are obligated to have both KR + EN versions, but omitted in this sprint → immediately corrected PR after user prompt. Plan stage i18n obligation specification procedure needed — automation candidate
  • Single-day 3 PR bundled response pattern (Sprint 150/151 hotfix pattern directly inherited): User prompt → new PR → CI green → Squash merge average ~5 minute cycle. Fact correction / i18n omission both responded to rapidly with separate hotfix PRs (no waiting/deferral)

6. Lessons Learned

  • Plan agent mappings require domain cross-check (Sprint 147 lesson reconfirmed): librarian = DB persistence, scribe = documentation — must verify ## Role & Core Responsibilities + ## Prohibited Actions in persona files (.claude/commands/agents/{name}.md). Skipping domain verification at plan stage impacts cycle
  • User direct fact-check is the last safety net that cannot be automated: CI/Critic are code consistency verification tools, not domain fact verification tools. Blog post domain facts (sprint numbers / ADR introduction point / accumulation count, etc.) need grep cross-check explicitly specified in plan stage checklist
  • i18n omission is the most frequent debt in new content creation: Dual-version policy is documented but easily missed at plan stage. Plan template needs to include "EN English version" work item as default
  • Post-merge prompt → immediate hotfix is the standard (Sprint 150/151 verified → Sprint 152 reconfirmed): Fact corrections / i18n omissions / Trivy CVEs all responded to rapidly with separate hotfix PRs. Waiting/deferral risks debt accumulation. Single-day 3 PRs also have low cycle impact
  • Single-post unit (plan→supplement→write) cycle for blog work validated: feedback-blog-workflow policy applied correctly. One post = single sprint, plan stage user agreement → scribe writes → user review → merge flow works well. Inline handling sufficient without tmux dispatch

7. Sprint 153 Carryover Seeds

UAT User Direct (Outside Oracle's Scope)

  • Seed #5: Programmers resubmission scoring pass confirmation (9 sprints accumulated — Sprint 145~152)
  • Seed #9: English environment + production Grafana CB dashboard ai-analysis visual consistency (9 sprints accumulated)

Automation Candidates (2 New Seeds from Sprint 152)

  • Seed #18: Blog post pre-merge domain fact cross-check automation (e.g., ADR accumulation count / sprint numbers / GitHub URL validity). Specification in plan template checklist + lint rule if possible
  • Seed #19: KR/EN dual simultaneous authoring plan obligation for new blog posts — plan template update + CI rule ("when posts/X.mdx added, posts-en/X.mdx mandatory" matching)

Follow-up (Sprint 151 unchanged + 0 new from this sprint)

  • create/edit page.tsx category UI addition (palette assessment: low user scenario frequency)
  • Programmers URL automatic category inference (sourceUrl pattern matching)
  • Existing SQL problem data backfill (manual ADMIN or import script)
  • Sprint 150 unresolved 3 candidates (.claude-tools/ cleanup / CI paths filter bypass debt check automation / prom-client default metric stale check)