# Changelog

All notable changes to pi-web are recorded here.

## [0.4.1] - 2026-07-22

### Fixed

- Removed the artificial pi `0.80.x` upper compatibility ceiling. pi-web keeps
  pi `>=0.80.10` as the reproducible minimum development and compatibility
  baseline while allowing newer tested pi releases.

## [0.4.0] - 2026-07-21

pi-web is now a strict native subscription-OAuth extension for both OpenAI
Codex and Anthropic. This release is a clean break: the old configurable
Anthropic-first product surface is rejected rather than silently translated.

### Added

- Native OpenAI Codex search and native HTML/PDF open through fixed ChatGPT
  first-party endpoints.
- Verified two-step Codex prompted fetch: native open must prove retrieval
  before ordinary same-model grounded processing.
- Native Anthropic search plus local-first prompted fetch using the accepted
  subscription compatibility envelope.
- Per-invocation active-model observation and automatic active-first routing.
- Pre-semantic fallback only. Provider switching stops permanently at native
  search, native open, or grounded-processing dispatch.
- Strict canonical-model trust checks and request-time pi-managed OAuth
  resolution. API keys, environment keys, header-owned credentials, custom
  bases, and generic providers are rejected.
- Direct-IP local retrieval with all-address DNS validation, pinned peer proof,
  redirect-hop revalidation, Host/SNI preservation, public-address enforcement,
  fixed deadlines, and bounded extraction/output.
- Snapshot-only `/web-status` diagnostics for search, prompted fetch, and
  no-prompt fetch.
- Compact renderers with explicit truncation state and bounded expanded
  previews; preview clipping does not alter the bounded result returned to the
  agent.
- Runtime host gate for pi `>=0.80.10 <0.81.0` and an isolated packed-package
  compatibility matrix for Node and TypeBox boundaries.

### Clean-break migration

| Surface | Removed behavior | Replacement in 0.4.0 |
|---|---|---|
| `web_search` fields | `max_results`, `tier`, `orchestrator_model`, `include_synthesis`, `bypass_cache` | `query`, optional strict `provider`, and one optional native domain filter |
| `web_fetch` fields | `backend`, `mode`, `summary_tier`, `summary_model`, `thinking_budget`, `answer_max_tokens`, `raw_max_bytes`, `bypass_cache`, `max_content_tokens`, `require_fetch` | `url`, optional normalized `prompt`, optional strict `provider` |
| Provider selection | Anthropic-only search and arbitrary-provider summarizers | Exact `openai-codex` or `anthropic` subscription routes only |
| Model selection | Tiers, model pins, family/version resolution, agent-selected model IDs | Active trusted model, fixed inactive Anthropic fallback, or deterministic Codex catalog match |
| Global config | Multi-key models/search/fetch/cache/security/display objects | One key: `provider: auto | openai-codex | anthropic` |
| Project config | `.pi/pi-web.json` | Removed; only the global config file is read |
| Environment variables | `PI_WEB_SEARCH_TIER`, `PI_WEB_FETCH_TIER`, `PI_WEB_SUMMARY_MODEL`, `PI_WEB_THINKING_BUDGET`, `PI_WEB_CACHE_TTL`, `PI_WEB_USER_AGENT_CONTACT`, `PI_WEB_DISABLE_CACHE` | Removed |
| CLI flags | `--web-no-cache`, `--web-summary-model`, `--web-debug` | Removed |
| Slash commands | `/web-config`, `/web-models`, `/web-cache`, `/web-debug` | `/web-status` only |
| Fetch strategy | Client/server backend switch and Anthropic server fetch | Operation-specific Codex native open or secure local retrieval |
| Citation state | Encrypted search-result retention and conversation replay | Removed; opaque provider state is discarded immediately |
| Caching | Memory cache, disk cache, citation cache, cache details and commands | Removed; existing cache files are left untouched |
| Generic processing | pi-ai routing to OpenAI API, Google, Bedrock, and other configured providers | Ordinary processing is limited to the selected trusted Codex or Anthropic OAuth route |
| Diagnostics | Cost, token, thinking, cache, model, citation, and transport fields in stable results | Stable normalized source/fetch metadata plus explicit truncation only |
| Security escapes | Configurable private-IP/file access, caller-controlled limits and headers | Fixed public HTTP(S) boundary with no escape switch |

Malformed or legacy multi-key configuration now produces an observable
migration error. Removed tool fields and unknown properties are rejected by the
final closed TypeBox schemas.

Historical entries below describe the behavior of those releases and are not
current documentation.

## [0.3.2]

- `web_search.max_results` schema bound lowered from 1-20 to 1-10.
  Anthropic's server-side `web_search_20250305` tool exposes no
  `max_results` parameter and returns at most 10 results per
  invocation (confirmed empirically; the API docs do not state the
  number but consistently cap there). The prior 1-20 bound was a
  lie: pi-web told the orchestrator "return up to N", Anthropic
  returned 10, pi-web sliced `[0, N]` on a 10-element array and the
  caller silently got 10. Schema, runtime clamp
  (`web-search.ts`), and config-loader clamp
  (`default_max_results` in `loader.ts`) all now top out at 10.
  Description rewritten to spell out the cap. Slicing behavior for
  1-10 is unchanged.

## [0.3.1]

- Renderer no longer prints the bold `WebSearch ` / `WebFetch ` header
  twice on settled tool rows. The tool-execution row composes
  `renderCall` and `renderResult` into the same container; both slots
  were emitting the tool name, so a settled row showed two stacked
  bold headers. The result-line head now leads with the outcome
  (`5 results` / `summary` / `raw` / `summary [server·cite]`) and lets
  the call slot above carry the tool name. Matches the precedent set
  by built-in `bash` (call: `$ <cmd>`, result: just output and timing).
- Unified head structure between `web_search` and `web_fetch` settled
  rows: `<outcome>  [middle]  <model> (cached) (elapsed) [cost]`,
  built by joining a filtered segments array. `(cached)` and
  `(elapsed)` sit in the same column in both tools.
- `web_fetch` raw mode no longer renders a misleading `raw` token in
  the model slot; the mode is already shown to the left as the
  outcome.
- `web_fetch` call line uses a parts-array pattern so an empty flag
  list no longer leaves trailing whitespace.
- `web_fetch` expanded preview shows the page body, not the
  `Fetched: ...` / `Title: ...` / `Summarized via ...` lines that
  already appear above as structured fields. New `extractFetchBody`
  helper splits on the `\n\n---\n\n` divider (raw and server modes) or
  the first blank line (summary mode); server-mode pre-header
  skip-note is correctly skipped along with the structured header.
- `web_fetch` expanded view no longer prints a redundant `Cache: hit`
  line. The collapsed head already shows `(cached)`.
- `web_search` in-band `web_search_tool_result_error` codes
  (`too_many_requests`, `invalid_input`, `max_uses_exceeded`,
  `query_too_long`, `unavailable`) are mapped to friendlier labels.
  Unknown codes fall through to the raw code.
- `web_search` expanded view now shows the `Usage: in=N out=M` line,
  matching what `web_fetch` already displayed.

## [0.3.0]

- Server-side `web_fetch` skip detection. The orchestrator may answer
  from prior knowledge instead of invoking the declared
  `web_fetch_20250910` tool. `WebFetchDetails` gains
  `serverFetchInvoked: boolean`, derived from the presence of a
  `web_fetch_tool_result` block in the response. New schema param
  `require_fetch?: boolean` (default `true`, server-only): on skip,
  pi-web throws a recoverable error pointing the caller at three
  remediations (more specific prompt, switch to `backend: "client"`,
  or `require_fetch: false`). Soft mode (`require_fetch: false`)
  returns the prior-knowledge answer with a skip-note prepended.
  Neither skips nor server-side errors are cached.
  `buildServerFetchMessages` uses imperative wording ("Use your
  web_fetch tool... Do not answer from prior knowledge") in both the
  citation and no-citation paths. Fetch renderer surfaces a
  `[server·skipped]` warning tag (collapsed) and a `Fetch invoked: no`
  row (expanded).
- Server-side `web_fetch` backend (`backend: "server"`). Routes the fetch
  through Anthropic's server-side `web_fetch_20250910` tool instead of
  the client-side GET + Readability + Turndown pipeline. The motivation
  is citation continuity with `web_search`: pi-web maintains an
  in-memory `CitationContextCache` keyed by URL that records the prior
  `web_search` assistant turn (server_tool_use + web_search_tool_result
  with their opaque encrypted_content blobs); on server-mode fetch the
  cached turn is replayed into the `/v1/messages` body so Anthropic's
  server-side URL-provenance check links the fetch to the originating
  search. Surfaces `details.backend`, `details.citationLinked`,
  `details.citationQuery`, `details.retrievedAt`,
  `details.serverFetchErrorCode`, and `details.maxContentTokens`.
  Server backend is anthropic-only, has no `raw` mode (server-side
  always pays for an orchestrator turn), and does not support
  `thinking_budget`.
- `max_results` on `web_search` is now actually enforced. Anthropic's
  server-side tool always returns 10; we now slice the parsed results
  to the caller's `max_results` cap before they reach the LLM or
  `details.results`.
- Citation-context cache is in-memory only; encrypted_content blobs from
  `web_search` results are never persisted to disk.
- Fetch cache key includes `backend` and `maxContentTokens` so client and
  server results never alias on the same URL+prompt.
- Fetch renderer surfaces `[server]` / `[server·cite]` tag in the
  collapsed header and Backend / Citation / Retrieved / max_content_tokens
  / Server error rows in the expanded view.
- PDF extraction in `web_fetch`. PDFs at any reachable URL flow through
  Mozilla pdf.js (via `pdfjs-dist`) instead of being rejected with a
  deferred-feature note. Output is page-segmented as `## Page N` headings
  so the agent can cite by page. PDF metadata (`Title`, `Author`) is
  promoted to `pageTitle` / `byline`.
- `WebFetchDetails` gains `contentKind` and `pageCount`; the expanded
  fetch renderer surfaces "Kind: PDF (12 pages)".
- `extractContent` is now async to accommodate PDF parsing.

## [0.2.0]

- CLI flags: `--web-no-cache`, `--web-summary-model <provider/id>`,
  `--web-debug`. Layered above env vars, below per-call args.
- Cross-provider summarizer routing for `web_fetch summary` mode.
  Anthropic stays on the hand-rolled `/v1/messages` path; OpenAI /
  Google / Bedrock / etc. route through pi-ai's `completeSimple`,
  picking up the same API-key resolution the rest of pi uses.
- Disk-persistent cache (opt-in via `cache.persist_to_disk`). Entries
  survive `/reload` and process restart; corrupt files are silently
  skipped on load.
- New slash commands: `/web-cache` (stats, list, clear, clear-expired)
  and `/web-debug` (toggle).
- `MemoryCache` gained `setPersistence(hook)` and `restore(entries)`
  for write-through to external stores.
- README documents the cross-provider section and the persistent
  cache section.

## [0.1.0]

- `web_search` tool: orchestrator-based search via Anthropic's server-side
  `web_search` tool. Supports `tier` (`fast` / `balanced` / `strong`),
  `orchestrator_model` escape hatch, domain filters, configurable result
  count, optional synthesis text.
- `web_fetch` tool: `raw` / `summary` / `auto` modes. HTML→markdown via
  Readability + Turndown. Optional side-channel summarization with
  per-call `summary_tier` / `summary_model` / `thinking_budget`.
- Tier auto-resolution against `ctx.modelRegistry` (newest version wins,
  alias preferred over dated id).
- Three-layer config: defaults -> `~/.pi/agent/extensions/pi-web/config.json`
  / `.pi/pi-web.json` -> env vars -> per-call tool args.
- In-memory LRU+TTL cache.
- URL safety: no `file://`, no private IP space, redirect re-check,
  response size cap, request timeout.
- Custom renderers with optional cost display.
- Slash commands `/web-config`, `/web-models`.
- System-prompt integration via `promptSnippet` + `promptGuidelines`.