@link-assistant/hive-mind 1.58.0 → 1.59.0

package/CHANGELOG.md

@@ -1,5 +1,150 @@
 # @link-assistant/hive-mind
 
+## 1.59.0
+
+### Minor Changes
+
+- 903b10e: Add `--auto-input-until-mergeable` (issue #1708): a new experimental
+  mode that extends a single Claude session for as long as possible by
+  streaming PR/issue comments, CI/CD failures, uncommitted-changes
+  status, and PR/issue title/body updates as NDJSON `user` frames into
+  the live `claude --input-format stream-json` process — instead of
+  killing the process and restarting with the feedback prepended to a
+  fresh prompt.
+
+  What it ships:
+  - Three new flags in `src/solve.config.lib.mjs`, all defaulting to
+    `false` and marked `[EXPERIMENTAL]`:
+    - `--auto-input-until-mergeable` — top-level opt-in for the new
+      behavior. Implies `--accept-incomming-comments-as-input` and
+      defaults to `--queue-comments-to-input` so the AI can finish its
+      current step before being interrupted.
+    - `--stream-comments-to-input` — forward each comment immediately
+      as it arrives. Default for `--accept-incomming-comments-as-input`
+      on its own (preserves the existing #817 behavior).
+    - `--queue-comments-to-input` — buffer comments while the AI is
+      busy and flush them only on `result` events. Default delivery
+      mode for `--auto-input-until-mergeable`. Mutually exclusive with
+      `--stream-comments-to-input`; queue mode wins if both are set.
+  - Queue-vs-stream delivery wired into
+    `src/bidirectional-interactive.lib.mjs#createBidirectionalHandler`:
+    - New `deliveryMode` option (`'stream'` / `'queue'`) plus
+      `markAiBusy()` / `markAiIdle()` lifecycle methods exposed on the
+      handler.
+    - In queue mode, comment frames and status frames are buffered in
+      `pendingFrames` while busy and FIFO-flushed to stdin on the next
+      `result` event. In stream mode, frames go to stdin immediately as
+      today.
+  - Status streaming (only when `--auto-input-until-mergeable` is on)
+    in `src/bidirectional-interactive.lib.mjs#checkForStatusChanges`:
+    - New parallel poller emits one-shot NDJSON frames for: PR
+      title/body changes, issue title/body changes (Issue #1708 G1),
+      uncommitted local changes (`git status --porcelain`), and CI
+      blockers (via `getMergeBlockers`).
+    - Each change is keyed by a stable signature so the same failing
+      check doesn't re-emit on every poll; failures in any sub-check
+      are swallowed and logged so the poller never breaks the live
+      Claude session.
+  - Stream parser in `src/claude.lib.mjs#executeClaudeCommand` now
+    signals `markAiBusy()` on `assistant` / `tool_use` / `tool_result`
+    events and `markAiIdle()` on `result` events, so queue-mode
+    buffering tracks the actual AI lifecycle.
+  - `src/solve.auto-merge.lib.mjs#watchUntilMergeable` logs a
+    "streaming-first" banner when `--auto-input-until-mergeable` was
+    active, so it is clear the auto-restart loop is the fallback rather
+    than the primary handler.
+  - For non-Claude tools, the validator continues to warn and disable
+    all four flags — the existing #817 fallback path. The default
+    behavior of every existing flag
+    (`--auto-restart-until-mergeable`, `--auto-merge`, etc.) is
+    preserved (R4: "must not break any existing features").
+  - Tests:
+    `tests/test-auto-input-until-mergeable-1708.mjs` (59 assertions)
+    and 11 new assertions in
+    `tests/test-bidirectional-interactive.mjs` cover flag composition,
+    queue-vs-stream routing, FIFO flushing on idle, busy-flag
+    preservation across stream-mode writes, default-deliveryMode is
+    stream, status-frame stamping with the right header per kind
+    (`comment` / `ci` / `uncommitted` / `metadata`), and metadata
+    diff/snapshot helpers.
+
+  The case study at `docs/case-studies/issue-1708/` is updated to
+  reflect that R1, R2 (Claude path), R3 (PR/issue title+body, CI,
+  uncommitted, comments), R4, R5, R6, plus G1, G5, G7 are addressed
+  here. Codex/Agent/OpenCode still degrade gracefully (no mid-session
+  NDJSON channel upstream) and use the existing `watchUntilMergeable`
+  loop as documented in G4.
+
+- 6efcab4: Fix cost / token calculation correctness, unify Total / sub-session format,
+  add verbose budget trace, and case study for issue #1710
+
+  Resolves the four "strange things" the issue reported by changing both the
+  public-pricing math and the rendered output:
+  - **R1 — `$0.040000` residual eliminated.** `calculateModelCost`
+    ([`src/claude.lib.mjs`](./src/claude.lib.mjs)) now bills Anthropic
+    server-side tools. `web_search` is charged at the documented
+    $10 / 1 000 requests rate (= $0.01 / req) via the new constants module
+    [`src/anthropic-server-tool-pricing.lib.mjs`](./src/anthropic-server-tool-pricing.lib.mjs).
+    For the issue's PR #1707 run that comes out to exactly the previously-shown
+    $0.040000 / +0.16% delta, so the public-pricing total now reconciles with
+    Anthropic's reported `total_cost_usd`. `accumulateModelUsage`
+    ([`src/claude.budget-stats.lib.mjs`](./src/claude.budget-stats.lib.mjs))
+    also picks up `usage.server_tool_use.web_search_requests` from JSONL.
+  - **R2 — Haiku sub-session line includes input information.** Sub-agent
+    models never appear as the responding model in the parent JSONL, so
+    `peakContextUsage` stays at `0`. The fallback in `buildBudgetStatsString`
+    now emits the cumulative `(X new + Y cache writes [+ Z cache reads])`
+    phrase instead of dropping the input information entirely.
+  - **R3/R5 — Sub-session and Total reconcile.** The bullet line is now
+    labelled `peak request: …` so it cannot be confused with the cumulative
+    Total line. `requestContext` (the source of `peakContextByModel`) excludes
+    cache reads, so the bullet figure is `input + cache_creation` and is
+    reconcilable with the cumulative non-cached total. Cache reads remain
+    visible — and visible separately — on the Total line.
+  - **R4 — Total always splits cache reads / cache writes when present.**
+    The conditional that previously keyed on `cacheReadTokens` only is replaced
+    with a `buildCumulativeInputPhrase` helper that emits
+    `(X new + W cache writes + Y cache reads) input tokens` when both kinds of
+    cache activity exist, `(X new + W cache writes)` when only writes exist
+    (the Haiku case that triggered the issue), and the back-compat
+    `(X + Y cached)` form when only reads exist (so common Opus-only output
+    is unchanged). Cache writes are billed at 1.25× / 2× of input — fusing
+    them silently into the input figure was a real semantic bug, not a
+    cosmetic one.
+
+  Both `displayBudgetStats` (solver-log renderer) and `buildBudgetStatsString`
+  (PR-comment renderer) share the helper, so the two paths render identically.
+
+  Also adds **`dumpBudgetTrace`**
+  ([`src/claude.budget-stats.lib.mjs`](./src/claude.budget-stats.lib.mjs)),
+  a verbose-only structured per-model trace (peak request, cumulative
+  input/cache_write 5m+1h split/cache_read/output, server-tool counts with
+  implied dollar cost, public and Anthropic-reported costs, and the data
+  source) that fires from `displayBudgetStats` only when `{verbose: true}` is
+  set, so the default solver output is unchanged. The trace captures all the
+  inputs that drive the renderer in one place, so the next "calculation
+  correctness" report can be triaged from a saved log alone.
+
+  Tests:
+  - `tests/test-issue-1710-budget-trace.mjs` — 10 cases for the verbose trace.
+  - `tests/test-issue-1710-format-fixes.mjs` — 8 cases locking each requirement
+    to numbers from `docs/case-studies/issue-1710/facts.md` (the actual
+    PR #1707 result event the issue quotes).
+
+  Documentation: `docs/case-studies/issue-1710/` contains the root-cause
+  analysis (per symptom, with file:line citations), the captured facts, and
+  the (now-implemented) solution plans.
+
+  Also fixes the hosted-CI flake that surfaced while validating this PR:
+  `use-m` occasionally hands back a truncated/corrupt global package after
+  `npm install -g`, surfacing as either
+  `Failed to import module from '...': SyntaxError: Unexpected end of input`
+  or `Failed to resolve the path to '<pkg>'` when use-m loads `getenv` /
+  `links-notation` from `src/config.lib.mjs` and `src/lino.lib.mjs`. Adds
+  `src/use-with-retry.lib.mjs`, a small wrapper around `use(...)` that
+  recognises both flake modes, removes the broken alias directory, and
+  re-fetches once. Covered by `tests/test-use-with-retry.mjs` (13 cases).
+
 ## 1.58.0
 
 ### Minor Changes

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@link-assistant/hive-mind",
-  "version": "1.58.0",
+  "version": "1.59.0",
   "description": "AI-powered issue solver and hive mind for collaborative problem solving",
   "main": "src/hive.mjs",
   "type": "module",

package/src/anthropic-server-tool-pricing.lib.mjs

@@ -0,0 +1,34 @@
+#!/usr/bin/env node
+
+/**
+ * Issue #1710: Anthropic server-side tool pricing.
+ *
+ * `calculateModelCost` historically only billed token-based usage (input,
+ * cache_creation, cache_read, output). When a sub-agent uses Anthropic's
+ * server-side web_search tool, the result event reports `webSearchRequests`,
+ * which Anthropic bills at $10 / 1 000 searches ($0.01 / request) per
+ * <https://platform.claude.com/docs/en/about-claude/pricing#web-search-tool>.
+ *
+ * Without billing it locally, the public-pricing estimate disagreed with
+ * Anthropic's reported `total_cost_usd` by exactly that amount — the
+ * "Difference: $0.040000 (+0.16%)" line that issue #1710 quotes.
+ *
+ * Centralising the constants in this module keeps the source-of-truth in one
+ * file: bumping a price is a one-line edit, and `calculateModelCost` /
+ * `dumpBudgetTrace` both read from the same map.
+ */
+export const SERVER_TOOL_PRICING_USD = Object.freeze({
+  // $10 per 1 000 searches = $0.01 per request.
+  // https://platform.claude.com/docs/en/about-claude/pricing#web-search-tool
+  web_search: { costPerRequest: 0.01, source: 'https://platform.claude.com/docs/en/about-claude/pricing#web-search-tool' },
+  // web_fetch is currently free for paying customers; kept here for
+  // completeness and so a future price change is a one-line edit.
+  web_fetch: { costPerRequest: 0, source: 'https://platform.claude.com/docs/en/about-claude/pricing#web-fetch-tool' },
+});
+
+/**
+ * Returns the per-request USD price for a server-side tool, or 0 if unknown.
+ * @param {string} tool - canonical tool name (e.g. "web_search")
+ * @returns {number} per-request price in USD
+ */
+export const getServerToolPrice = tool => SERVER_TOOL_PRICING_USD[tool]?.costPerRequest || 0;