Skip to content

2.3.0: prompt mode speed + quality refresh (dense prompt v2) - #34

Merged
agr77one merged 4 commits into
mainfrom
feat/2.3.0-prompt-v2
Jul 13, 2026
Merged

2.3.0: prompt mode speed + quality refresh (dense prompt v2)#34
agr77one merged 4 commits into
mainfrom
feat/2.3.0-prompt-v2

Conversation

@agr77one

Copy link
Copy Markdown
Owner

2.3.0 — prompt mode speed + quality refresh

Rewrites the default prompt: output to a dense, high-signal prompt v2 and lowers the token budget, cutting latency well under the 15 s goal while improving quality.

Measured (committed A/B evidence, FLM qwen3.5:4b)

  • Speed gate PASS: v2 p50 ≈ 18.6% of v1 (~5× faster), ≤15 s ✓, p90 ≤20 s ✓, no >25 s outliers ✓.
  • Quality gate PASS: judge-scored median ≥ v1 ✓, zero invented requirements ✓, clean-section rate ≥ v1 ✓.

Changes

  • Dense prompt v2 is the new default (CLAUDE_PROMPT_SYSTEM_PROMPT_V2); old verbose prompt kept as ..._V1 for instant rollback.
  • Prompt-mode token caps 700/900/1200 → 240/320/420.
  • Background keep-FLM-warm to avoid the cold-load tax.
  • New committed speed+quality A/B evaluator (tools/prompt_speed_quality_eval.py) + artifacts; +33 tests.

Gates: ruff clean, 381 pytest, version 2.3.0 consistent across the 4 files.

🤖 Generated with Claude Code

agcycle added 4 commits July 10, 2026 09:46
Cites V24, V27, V28, V30, and V32.
Closes T26. Enforces V18, V24, V29, V30, V33, V34, V35, and V36.
@agr77one
agr77one merged commit 04b88c9 into main Jul 13, 2026
5 checks passed
@agr77one
agr77one deleted the feat/2.3.0-prompt-v2 branch July 13, 2026 11:11

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6c3e04722d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

intent,
legacy_system_prompt=ffp_config.CLAUDE_PROMPT_SYSTEM_PROMPT,
prompt_v1_system_prompt=ffp_config.CLAUDE_PROMPT_SYSTEM_PROMPT_V1,
prompt_v2_system_prompt=ffp_config.CLAUDE_PROMPT_SYSTEM_PROMPT_V2,

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Apply v2 grounding in live prompt-builder eval

When tools/prompt_builder_eval.py --live runs the default Claude cases, this now sends the v2 draft prompt, which intentionally asks the model for a one-sentence draft; production then calls ground_prompt_v2_output() before validating. This evaluator still validates the raw draft against the four-section XML contract, so live eval artifacts for the default cases will report false failures instead of measuring shipped behavior.

Useful? React with 👍 / 👎.

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.

2 participants