mirror of https://github.com/mksglu/context-mode.git synced 2026-10-02 04:14:38 +08:00

Files

T

Mert Koseoglu d05a37584e feat(prose): retire prose-style enforcement entirely (#482 )

Issue #482 (makoMakoGo) reported that context-mode's caveman/terse
injection pressures the model toward brevity on its FINAL ANSWER, not
just on tool-output reporting. Cited evidence: Moonshot AI on
kimi-k2.5 (anomalyco/opencode#20258, PR #20259) — aggressive brevity
prompts measurably degrade coding/reasoning benchmarks because the
model drops assumptions, caveats, verification evidence, failure
modes, and security warnings the user actually needs.

Considered:
  A — config switch ("injectCommunicationStyle: false"). Rejected:
      switches default-on become dead code.
  B — close FR with rationale. Rejected: ignores valid evidence.
  C — refine wording ("compress when reporting raw tool output, be
      complete for technical answers"). Rejected: still text
      injection, model-dependent, half-measure.
  D — full strip everywhere. Adopted.

The decision after grilling: context-mode's value is data routing
(sandbox, FTS5, session continuity), not prose styling. The brevity
injection conflated three goals — keeping raw data out of context
(real, hard-enforced), summarizing tool output compactly (LLMs auto-
calibrate), and final-answer prose style (the wrong target). Strip
all 22 sites where prose-style language landed.

Sites stripped (A-Z):

  hooks/routing-block.mjs
    - <communication_style> block  (Terse like caveman, fragments OK,
      auto-expand for security warnings)
    - <response_format> block       (Concise summary, 2-3 bullets)

  src/server.ts (5 MCP tool descriptions + 2 cosmetic comments)
    - ctx_execute            "When reporting results — terse..."
    - ctx_execute_file       same
    - ctx_search             same
    - ctx_fetch_and_index    same (URL + commands shapes, both)
    - cosmetic comment       "Caveman style — terse status line"
    - rewrote concurrency note: "Indexing is serial regardless of
      concurrency" → "Fetches parallelize up to your concurrency
      setting; FTS5 indexing serializes the writes after (SQLite
      single-writer rule)." — same fact, less jargon.

  configs/ (15 adapter MD files — every shipped system prompt)
    antigravity/GEMINI.md, claude-code/CLAUDE.md, codex/AGENTS.md,
    cursor/context-mode.mdc, gemini-cli/GEMINI.md, jetbrains-copilot/
    copilot-instructions.md, kilo/AGENTS.md, kiro/KIRO.md, omp/SYSTEM.md,
    openclaw/AGENTS.md, opencode/AGENTS.md, pi/AGENTS.md, qwen-code/
    QWEN.md, vscode-copilot/copilot-instructions.md, zed/AGENTS.md
    All had identical "## Output" block: 3 caveman lines stripped,
    workflow lines ("Write artifacts to FILES", "Descriptive source
    labels") kept.

  CLAUDE.md (repo root — internal dev instructions)
    Same caveman block stripped. We don't ship this file but we do
    eat our own dog food.

  README.md
    Pillar 4 ("Output Compression — Terse like caveman...") rewritten
    to "No prose-style enforcement" — explicitly cites the kimi-k2.5
    benchmark evidence as the rationale.

  web/index.html
    Removed Ch 4b entirely (the "Output compression" chapter with
    before/after example pushing terse style on the model).

Tests:

  - tests/session/continuity.test.ts: SessionStart routing-block
    assertion flipped from "must include 'Terse like caveman'" to
    "must NOT include caveman/terse-style directive".
  - tests/core/server.test.ts: Task hook injection assertion same
    flip. Two cosmetic comment renames ("Caveman style — terse status
    line" → "Status line: counts + sections + size"), test name
    rename ("caveman style" → "compact format"). Added new
    "prose-style policy (#482)" describe block at end of file with 3
    negative-pin tests covering server.ts MCP descriptions,
    routing-block, and README.

  Full suite: 82/82 files passed, 2645 passed, 20 skipped, 0 failed.
  Net +3 new tests (the policy describe block).

CONTRIBUTING.md
  New "Prose-style policy (#482)" section documents the decision so
  future contributors don't re-add the injection. Cites the Moonshot
  benchmark evidence + the regression test that pins the deletion.

This addresses #482 in full. Closing the issue with a comment that
walks the requester through the decision and links the policy section.

2026-05-10 14:45:41 +03:00

3.8 KiB

Raw Blame History

context-mode — MANDATORY routing rules

context-mode MCP tools available. Rules protect context window from flooding. One unrouted command dumps 56 KB into context. Zed has NO hooks — these instructions are ONLY enforcement. Follow strictly.

Think in Code — MANDATORY

Analyze/count/filter/compare/search/parse/transform data: write code via mcp:context-mode:ctx_execute(language, code), console.log() only the answer. Do NOT read raw data into context. PROGRAM the analysis, not COMPUTE it. Pure JavaScript — Node.js built-ins only (fs, path, child_process). try/catch, handle null/undefined. One script replaces ten tool calls.

BLOCKED — do NOT use

curl / wget — FORBIDDEN

Do NOT use curl/wget in shell. Dumps raw HTTP into context. Use: mcp:context-mode:ctx_fetch_and_index(url, source) or mcp:context-mode:ctx_execute(language: "javascript", code: "const r = await fetch(...)")

Inline HTTP — FORBIDDEN

No node -e "fetch(...", python -c "requests.get(...". Bypasses sandbox. Use: mcp:context-mode:ctx_execute(language, code) — only stdout enters context

Direct web fetching — FORBIDDEN

Raw HTML can exceed 100 KB. Use: mcp:context-mode:ctx_fetch_and_index(url, source) then mcp:context-mode:ctx_search(queries)

REDIRECTED — use sandbox

Shell (>20 lines output)

Shell ONLY for: git, mkdir, rm, mv, cd, ls, npm install, pip install. Otherwise: mcp:context-mode:ctx_batch_execute(commands, queries) or mcp:context-mode:ctx_execute(language: "shell", code: "...")

File reading (for analysis)

Reading to edit → reading correct. Reading to analyze/explore/summarize → mcp:context-mode:ctx_execute_file(path, language, code).

grep / search (large results)

Use mcp:context-mode:ctx_execute(language: "shell", code: "grep ...") in sandbox.

Tool selection

GATHER: mcp:context-mode:ctx_batch_execute(commands, queries) — runs all commands, auto-indexes, returns search. ONE call replaces 30+. Each command: {label: "header", command: "..."}.
FOLLOW-UP: mcp:context-mode:ctx_search(queries: ["q1", "q2", ...]) — all questions as array, ONE call.
PROCESSING: mcp:context-mode:ctx_execute(language, code) | mcp:context-mode:ctx_execute_file(path, language, code) — sandbox, only stdout enters context.
WEB: mcp:context-mode:ctx_fetch_and_index(url, source) then mcp:context-mode:ctx_search(queries) — raw HTML never enters context.
INDEX: mcp:context-mode:ctx_index(content, source) — store in FTS5 for later search.

Parallel I/O batches

For multi-URL fetches or multi-API calls, always include concurrency: N (1-8):

mcp:context-mode:ctx_batch_execute(commands: [3+ network commands], concurrency: 5) — gh, curl, dig, docker inspect, multi-region cloud queries
mcp:context-mode:ctx_fetch_and_index(requests: [{url, source}, ...], concurrency: 5) — multi-URL batch fetch

Use concurrency 4-8 for I/O-bound work (network calls, API queries). Keep concurrency 1 for CPU-bound (npm test, build, lint) or commands sharing state (ports, lock files, same-repo writes).

GitHub API rate-limit: cap at 4 for gh calls.

Output

Write artifacts to FILES — never inline. Return: file path + 1-line description. Descriptive source labels for search(source: "label").

ctx commands

Command	Action
`ctx stats`	Call `stats` MCP tool, display full output verbatim
`ctx doctor`	Call `doctor` MCP tool, run returned shell command, display as checklist
`ctx upgrade`	Call `upgrade` MCP tool, run returned shell command, display as checklist
`ctx purge`	Call `mcp:context-mode:purge` MCP tool with confirm: true. Warns before wiping knowledge base.

After /clear or /compact: knowledge base and session stats preserved. Use ctx purge to start fresh.

3.8 KiB Raw Blame History