dsh-cot-profile v0.2.0

v0.2.0 — verified judgment, working records, controlled probes

Judgment (verified against real data)

• Transition-band detection — low confidence or we/letMe mixing reports ambiguous instead of a possibly-wrong hard label
• Strong-signal override — runaway indicators (e.g. letMe100 ≈ 1700 from a tool-heavy session) now judge as confident react/spec instead of being swallowed into "uncertain" (found by real probe data, golden regression passes)
• Judgment threshold lowered to 3 blocks (golden data: single-block direction is reliable; confidence + transition guard cover short tasks)

Record mode (actually works now)

• Records were written only at session disposal — but DSH sessions rarely dispose, so nothing ever landed. Now writes a per-turn cumulative snapshot (turn/final fields) plus a final record at disposal. Verified on live sessions.

Calibration

• Settings panel: per-group judgment distribution, nearest-baseline diff, one-click apply as profile family
• baselineDiff strong directional override (aggregation no longer mis-matches to gray-like on runaway indicators)
• Panel Live / Records tabs

Controlled probes & experiments

• scripts/probe-verify.mjs — replay session logs under an expected trajectory side, fail on direction errors
• Probe workspace system (isolated a/b workspaces, git hard-reset, two-bug task)
• Experiment archive: probe-1 (feasibility) and probe-2 (n=3 assembly contrast: standard→react, minimal→spec, confirmed post-fix)
• Methodology note: session-header agentPreset is unreliable; actual assembly = agent-preset/selected event + tool-call surface

UI

• Two-layer panel: community classification label up front (minimal-like / standard-like / ambiguous), professional details (fault-line axis, indicators, baseline distances) behind a fold
• i18n fix: correct DSH locale id zh

Honest limits (unchanged)

• Wording is a fault-line fingerprint, not model identity; built-in baselines are estimates — calibrate with record data
• Web settings needs the temporary apiproxy patch on DSH 0.1.0-rc.6; track panel mode is experimental

dsh-cot-profile v0.1.0

Real-time chain-of-thought trajectory profiling for DeepSeek Harness

v0.1.0 — first release candidate. Monitors the reasoning stream live, tallies signature wording indicators (let me / we / let's / I, first-line patterns, block length, interim replies), and judges the session's trajectory side (spec-like / react-like / gray-like) against built-in baselines — with an explicit transition-band state instead of forcing a possibly-wrong label.

Features

• Live UI: session-header badge + floating panel (default) or experimental layout-track right column, driven by session-projection push frames — no polling, no custom RPC
• Judgment: weighted-distance match + confidence; transition-band detection (low confidence or We/The/Let mixing, per router-standard's measured fault line)
• Settings UI: thresholds, weights, profile families, panel mode, record mode, and one-click semi-automatic calibration
• Record mode: per-session aggregate JSON records (event + optional JSONL) — aggregates only, never raw reasoning text
• Bilingual UI: zh-CN + en, follows the DSH locale
• Persistence for free: stats survive restarts via the projection cache subsystem

Verification (four layers)

• 34 unit tests (zero-dependency)
• Golden verification against 119 real V4 Pro / V4 Flash probe runs: spec-side 83/83 (100%), zero direction errors, react-side never misjudged spec-side, ambiguous → transition band ≥55%
• End-to-end replay verification against real session logs: all blocks, counts, first-line patterns, and block-length medians match an independent ground truth exactly
• Record-mode calibration for long-term baseline measurement

Honest limits

• Wording is a fault-line fingerprint, not model identity — the plugin reports trajectory side, never "this is model X"
• Built-in baselines are estimates; calibrate with record-mode data
• Web settings section needs the optional temporary apiproxy patch on DSH 0.1.0-rc.6 (see README)
• track panel mode is experimental (direct DOM grid manipulation)

Install

dsh plugin --profile web add github:Chloride233/dsh-cot-profile