pxpipe/CHANGELOG.md

Changelog

All notable changes to pxpipe are documented here. This project adheres to Semantic Versioning (pre-1.0: minor = features / behavioral changes, patch = fixes).

0.6.2 — 2026-06-22

Fixed

• GPT history image cap. A hard cap (16) on GPT history images. Long OpenCode sessions could otherwise render 80+ images — token-cheap but slow enough that gpt-5.5 times out before first token. The cap images the oldest sections up to the limit and leaves the rest as text; no context dropped, no live tool state orphaned. (openai-history.ts)
• GPT "live request" guard. A developer-role note after the history image tells the model the image is prior context, not the current request — reinforcing the turn-index so a stale opening turn doesn't read as live. (openai.ts)
• Honest cache math. Savings are priced warm whenever a cache read was actually observed (cache_read > 0), even when pxpipe has no in-memory warmth prior (after a restart/eviction or on the first tracked turn). Pricing those turns cold billed the text counterfactual a 1.25× create on a prefix we know was cached — fabricating inflated "saved" rows. Applied across the live dashboard, its replay path, and the Sessions panel. (dashboard.ts, sessions.ts)

0.6.1 — 2026-06-21

Fixed

• GPT opening prompt no longer reads as the live request. The static slab is now inserted as its own dedicated image item instead of being bundled onto the first real user message. That message stayed protected from collapse, so the opening prompt floated at the front as un-collapsed, live-looking text (right next to the slab image) — and gpt-5.x would answer it instead of the actual latest turn. The opening prompt is now collapsible, turn-indexed history like every other old turn; only the slab item is protected. Both Chat Completions and Responses paths. (openai.ts)

0.6.0 — 2026-06-21

Structure-through role attribution and a turn-index recency anchor for collapsed history images, plus per-model GPT rendering profiles. The model now reads who said each turn and how recent it is, instead of reconstructing it from a flattened, role-ambiguous blob (which led it to resurface the opening turn as if it were the live request).

Added

• Turn-index recency anchor. Each collapsed turn serializes as <user t="N"> / <assistant t="N"> with an absolute turn index (larger N = more recent), so the model can distinguish turn 1 from turn 60 and stops treating the opening turn as the live request. Absolute (never relative) so frozen chunks stay byte-identical and keep hitting the prompt cache. Applied on both the Anthropic (history.ts) and GPT (openai-history.ts) paths.
• colorByRole role tinting. <user>/<assistant> tags in the history image are colored via a parallel "slot string" carried from serialize time, replacing the parse-back that miscolored a body quoting a literal tag. (render.ts)
• Per-model GPT profiles (gpt-model-profiles.ts). Vision-cost regime, strip width, and max image height per model id, retunable via PXPIPE_GPT_PROFILES (a JSON model-id-prefix map) without a code change. Built-ins are behavior-identical to the prior hardcoded values.

Changed

• ↵-packing for sentinel-bearing content. A pre-existing ↵ (U+21B5) in content is swapped to ⏎ (U+23CE) in render-prep so reflow packs newlines instead of bailing to a raw, unpacked render — common when the content is about pxpipe itself (rendered dumps, OCR). Render-only; originals are preserved.

Fixed

• Banner single source of truth. The GPT intro/outro now alias the Anthropic constants in history.ts instead of being byte-copies, so the two paths cannot silently drift on turn-attribution wording.
• Slot/text alignment. slotCopyBody neutralizes literal slot-marker control chars to a width-equivalent control char (U+0003) instead of a space, which minifyForRender would strip as trailing whitespace and desync the slot from the text (smearing role colors).

0.5.0 — 2026-06-20

Cache-stable history-collapse imaging for both providers, with GPT-5.6 promoted to the default imaged scope. Old conversation history collapses into rendered PNG sections so the model reads a compact image instead of re-billed text, while prompt caching and tool-call behavior are preserved.

Added

• GPT history collapse (openai-history.ts). Append-only, o200k token-length sectioning. Sections seal only at a tool-closed boundary (the open call-id set is empty), so a function_call and its function_call_output never split across the collapse cut.
• GPT + Anthropic dashboard rendering. Per-family model toggles, persisted metrics, thumbnail-expired session UI, and reflow/newline handling.

Changed

• Default imaged scope is now claude-fable-5 + gpt-5.6. GPT-5.6 is promoted from opt-in (0.4.0) to on by default. gpt-5.5 and claude-opus-4-8 stay opt-in: they degrade reading dense imaged history (gist drift), so silently imaging them by default is wrong. Promotion is gated on an OCR/recall threshold.
• Anthropic cache contract (history.ts). Append-only per-chunk rendering; cache_control markers are preserved/moved, never added; chunk boundaries align with caller marker seams for byte-stable prefix caching.
• GPT image budget (openai.ts). detail:'original' for gpt-5.x, flagship vision-multiplier fix, and a patch cap; schema-strip preserves real descriptions.
• Savings accounting (openai-savings.ts). Now computed on a cached_tokens + vision-token basis.

Fixed

• OpenAI 400 on long Responses-API sessions. "No tool call found for function call output with call_id ..." no longer occurs — the tool-closed sectioning boundary keeps each function_call/function_call_output pair intact across the collapse cut.

0.4.0 — 2026-06-19

New library surface for harness authors, opt-in GPT-5.x / Responses API support, and a round of dashboard-honesty and cache-correctness fixes.

Added

• Library API (pxpipe/transform): transformAnthropicMessages now accepts keepSharp (pin specific blocks as text so the caller controls what stays legible) and emitRecoverable (a provenance-recovery channel surfaced on info.recoverable). New exported types KeepSharpBlock, RecoverableBlock.
• Edge / Workers-safe packaging. process.env access is typeof-guarded; @napi-rs/canvas moved to devDependencies (the atlas is baked at build time), so the runtime is pure-JS and runs on Node and Cloudflare Workers unchanged.
• GPT-5.x family + Responses API (opt-in, off by default). isPxpipeSupportedGptModel gates the gpt-5/5.5/5.6/-mini/-nano family; a 768 px portrait-strip render profile avoids OpenAI's mandatory shortest-side-768 downscale; an OpenAI vision-token cost model replaces the Anthropic 750 px/token math for the GPT path; transformOpenAIResponses compresses /v1/responses (Codex). Still gated off until the day-one OCR-fidelity eval on a released GPT-5.x model.

Fixed

• Cache anchor relocation. The single cache breakpoint moved from the static slab onto the last history image (which sits after the slab in prefix order), so slab + history cache as one stable prefix — created once, then read at 0.1×. Previously the ~141k-token history image re-created at the 1.25× rate on warm turns, turning a real compression win into a net loss. Marker count is unchanged: pxpipe still never adds a breakpoint, only relocates the caller's.
• Dashboard honesty. The Details headline and the session hero now use cache-weighted tokens (matching the Saved column) instead of dividing the raw count_tokens baseline by sent tokens — which over-claimed "fewer tokens" even on requests that were a net loss after caching.
• Restart restore. Replayed rows now reconstruct the Saved delta and the Details breakdown from the persisted JSONL; image thumbnails are honestly marked expired rather than left blank.
• CI: regenerated pnpm-lock.yaml after the canvas dependency move (--frozen-lockfile had been rejecting every push).
• Node: clean exit on Ctrl+C during in-flight streams and idle keep-alive.

Changed

• Dashboard redesign: light flame theme, plain-language hero, and per-request image-vs-text transparency (which context became an image, which stayed text).
• Docs: lead with token reduction rather than dollar savings (pricing is a side-effect of tokens saved); added HISTORY_CACHE_MODEL.md; large comment cleanup across the core (~1,640 lines trimmed) with deep math moved to docs.
• Deps: patched 6 advisories (vite, undici, esbuild).

0.3.1 — 2026-06-17

Changed

• Demo: the side-by-side A/B clip moved to Google Drive (the committed copy was too low-res to read). The README keeps the preview thumbnail and links out to the video; the 8.9 MB video binary is no longer in the package/repo tree.

0.3.0 — 2026-06-17

Render-sizing overhaul, dashboard transparency, honest savings accounting, and a multi-agent code review with five confirmed fixes. Reviewed at extra-high recall (10 finder angles → verify → sweep).

Changed

• Render page ceiling raised to ~1932×1932. Fable 5 / Opus 4.8 accept images up to 2576 px long edge / 4784 visual tokens, but a request with >20 images (pxpipe always sends many) is held to the stricter ≤2000 px/side rule — so the real ceiling is ~1932×1932 (1928×1928 = 69×69 = 4761 tokens). MAX_HEIGHT_PX 1568→1932; dense tool/history pages now DENSE_CONTENT_COLS=384 / DENSE_CONTENT_CHARS_PER_IMAGE=92160 (1928×1928 full page) — fewer image blocks at the same OCR-validated 5×8 cell. The static slab is unchanged (313 cols / 1573×1280). Pages never trip a server-side downscale. Note: the larger per-page density uses the validated 5×8 cell and stays within Anthropic's pixel/token limits, but OCR legibility at this page size has not been independently re-eval'd (revert = the four render constants).
• Opus is OFF by default. Production scope defaults to Fable-5 only; Opus 4.8/4.7 are opt-in (they read imaged content at a measurable tax — see FINDINGS.md). Opt in via PXPIPE_MODELS or the dashboard chips.
• Honest savings accounting. Per-turn/session savings are the real baseline_eff − actual_eff with no ≥0 floor — a net-losing turn (e.g. a cache_create-heavy image rewrite) now reports the real loss instead of a fabricated 0. (Dashboard renders negatives explicitly.)

Added

• Dashboard "how your context works" panel — per-request token flow (as-text → real) + the exact-char breakdown of what became images + a gallery of every rendered page, reached via a "view" link on each recent-requests row.
• Flexible "compress models" chips — the toggle set is the union of a model catalog (Fable 5, Opus 4.8/4.7, Sonnet 4.6, Haiku 4.5), the PXPIPE_MODELS env scope, and the currently-active scope, so any env-enabled model stays toggleable (off ↔ on). Runtime-only override of the compress scope.
• Demos — demo/cost-ab/ (cost A/B on a real coding task) and demo/effective-context/ (recall-at-scale needle test), each with a model arg: defaults to Fable, a.sh opus to override. Plus eval/ab/ token-savings scripts.

Fixed (from the code review)

• Tool_result over-truncation (regression): the paging/break-even gate and truncateForBudget predicted against the slab geometry (313 cols / 159 rows) while the dense renderer emits 384 cols / 240 rows — so large tool_results were truncated far earlier than the 10-image cap required, silently dropping output that would have rendered. The gate, paging budget, and image-count estimate now price the same page the renderer produces.
• Garbled session headline: a net-losing session showed "-7% fewer tokens"; now phrased honestly as "N% more tokens".
• Context-map "view" mis-resolution: contextHistory capped at 30 while the recent table showed 50 rows, so older rows' "view" links silently showed the latest request's breakdown. Caps aligned; an evicted/unrecorded request now shows an explicit "no longer available" message instead of wrong data.
• Multi-col token cap: the multi-col width ceiling now respects the 4784 visual-token limit at full page height (was bounded only by the 2000 px side limit, which could produce a 4968-token page that the API rejects).
• Doc/code contradictions: baseline.ts and caller comments no longer claim a ≥0 clamp the code intentionally doesn't apply.

Docs

• Rewrote docs/RENDER_SIZING.md and updated docs/TRANSFORM_INFO.md, README.md, and in-code comments for the new ceiling and limits.