Skip to content

docs(readme): v5 — model-driven focus, unit consistency, move cache section - #21

Merged
ranxianglei merged 2 commits into
masterfrom
2026-06-27_readme-v5
Jun 27, 2026
Merged

ranxianglei merged 2 commits into
masterfrom
2026-06-27_readme-v5

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

Summary

README refinement, 5 rounds of user feedback consolidated. Docs-only (no lib/ changes). EN + ZH kept consistent.

  • Why ACP: "ACP hands all context-management authority to the model itself — best implementation to date." Two effects: saves ~2/3 tokens (a 1M-window model runs in the 200,000–300,000 range); ultra-long sessions without losing key content (500M-level context, 100K messages). Units unified; 37-bug-fixes footnote dropped (explained later).
  • Proven at scale: one-line summary lead (500M-level context, p95 ~30%, avg cache hit >85% — average not per-session, explained in the cache section). Dropped Span / Model turns / Tokens-reclaimed rows. Sessions anonymized.
  • How It Works: model 100% responsible; tools = compress/decompress/delete. Lifecycle as a 3-object mermaid state diagram (Raw ⇄ Compressed → Deleted). Compression strategy with the source priority list (from lib/prompts/system.ts: agent results, verbose command output, dead-end exploration, ...). Decompression + Deletion strategies (deletion irrecoverable, replaces forced GC).
  • Impact on Prompt Caching: ~87% hit rate; beats traditional compression (which acts at 80–90% + forces full re-hit); ACP keeps context ~30% vs traditional 50–80%. Moved to right after How It Works.

Verification

  • Docs-only — no lib/ touched, no build/typecheck/test impact.
  • Language cross-links present in both READMEs ([English](./README.md) | [中文](./README.zh-CN.md)).
  • No real project/session names leak (anonymized to Session 1 / Session 2).
  • Devlog: devlog/2026-06-27_readme-v3/.

Non-goals

  • No version bump (1.4.0 is already on npm; a follow-up release will pick this up if desired).
  • LICENSE unchanged (AGPL-3.0-or-later) — MIT relicense is not legally possible on an AGPL fork.

Branch pushed via the GitHub git database API because direct git push to github.com:443 times out in this environment (GnuTLS); api.github.com is reachable. Reviewed against ranxianglei/opencode-acp master e981bb51.

Per user feedback (5 rounds of README refinement):

- Why ACP: "ACP hands all context-management authority to the model itself" —
  best implementation to date. Two effects: saves ~2/3 tokens (a 1M-window
  model runs in the 200,000-300,000 range); ultra-long sessions without losing
  key content (500M-level context, 100K messages). Unify units; drop the
  37-bug-fixes footnote (explained later).
- Proven at scale: lead with a one-line summary (500M-level context, p95 ~30%,
  avg cache hit >85% — average not per-session, explained in the cache section).
  Drop Span / Model turns / Tokens-reclaimed rows.
- How It Works: model is 100% responsible; tools = compress/decompress/delete.
  Lifecycle state-machine (mermaid) simplified to 3 objects. Compression
  strategy with the source priority list (agent results, verbose command
  output, dead-end exploration, ...). Decompression + Deletion strategies.
- Impact on Prompt Caching: ~87% hit rate; beats traditional compression;
  context ~30% vs traditional 50-80%; conclusion = higher hit rate AND no key
  info lost. Moved to right after How It Works.

EN + ZH consistent. Docs-only. Includes anonymized real-world session stats.

Reviewed via dual-agent code review (AGENTS.md §5.3) before this PR.
Reviewer B (dual-agent review, AGENTS.md §5.3) flagged that the devlog still
described the earlier v3 revision (wrong numbers like 2.6M/1.8M reclaimed,
H3 subsections, 'demoted to footnote', stale 'targets Gitea fork' note) which
do not match the v5 diff. Rewrote REQ.md + WORKLOG.md to accurately reflect
PR #21's shipped content.

mark_block/'delete' framing retained per maintainer decision (Option B): it is
conceptually deletion (mark -> eventually truly deleted); the merge-vs-true-
delete mechanism is intentionally not elaborated in the README.
@ranxianglei
ranxianglei merged commit c1dc4af into master Jun 27, 2026
2 checks passed
@ranxianglei ranxianglei mentioned this pull request Jun 27, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant