Files
OmniRoute/open-sse/config/contextEditing.ts
Diego Rodrigues de Sa e Souza f165efcd0b Release v3.8.28 (#4053)
* chore(release): open v3.8.28 development cycle

* fix(ws): warm SSE auth import on LiveWS startup; relocate boot test to integration (#4063)

The live dashboard WebSocket sidecar lazily import()-ed the SSE auth module
inside the connection handler, only on the API-key path. That cold import pulls
in hundreds of transitive modules and takes ~7s under tsx, blocking the
single-threaded event loop. The first API-key WebSocket connection therefore
stalled the loop long enough that any connection arriving in that window — e.g.
a same-origin cookie client — could not complete its handshake and timed out.

This was deterministic, not an "env flake": the boot test fires an API-key
connection immediately followed by a cookie connection, so the cookie connection
always raced the cold import and timed out (reproduced 3/3 locally and red on
every CI run; proven via instrumented probes — reversing the order or warming
the module first makes both connections open in ~20ms).

Fix:
- Memoize the auth-module import and warm it once at startup (before listen), so
  connection handling never pays the cold-import cost. Real improvement: the
  first API-key client no longer stalls the event loop for concurrent clients.
- Relocate the boot test from tests/unit/cli to tests/integration. It spawns a
  real subprocess + WS server + SQLite (~9-11s); under the unit suite's
  --test-concurrency=20 it contended for CPU and destabilized the shard. The
  serial integration runner is its correct home; it still guards #4004's
  cookie-parse fix on every PR via the integration CI job.
- Bump the test's startup/overall timeouts to absorb the eager auth warm.

Makes `npm run test:unit` deterministically green (the only remaining unit red).

Validated: relocated test 3/3 green via the integration runner (was 3/3 red);
typecheck:core + eslint clean; confirmed it no longer matches the test:unit glob
and does match tests/integration/*.test.ts.

* fix(ws): start LiveWS sidecar with cwd at package root (#4055) (#4064)

* chore(deps): bump ossf/scorecard-action from 2.4.0 to 2.4.3 (#4045)

Integrado em release/v3.8.28. Patch de SHA do ossf/scorecard-action (2.4.0→2.4.3), mantém SHA-pin. Reds de CI são exclusivamente os shards flaky pré-existentes branch-wide (Unit 7/8, Integration, Coverage 7/8, Node 1/2) — não relacionados ao bump (PR deps-only).

* deps: bump electron from 42.4.0 to 42.4.1 in /electron (#4049)

Integrado em release/v3.8.28. Patch do electron (42.4.0→42.4.1). Reds de CI: shards flaky pré-existentes + PR Test Policy = falso-positivo (mudança deps-only sob electron/ não comporta teste de código) + Node 26(2/2) sem step (flake/infra). Precedente #3913/#3914 (electron dependabot mergeado nessas condições).

* fix(auto): resolve built-in auto catalog combos (#4058)

Integrado em release/v3.8.28. Resolve os IDs de catálogo `auto/*` built-in (combos virtuais) — corrige o 400 "No auto combos configured" em auto/best-coding etc. Ajuste de review: os mapas AUTO_TEMPLATE_VARIANTS/VALID_AUTO_VARIANTS duplicados em chat.ts e chatHelpers.ts foram extraídos para open-sse/services/autoCombo/builtinCatalog.ts (DRY), devolvendo chatHelpers.ts <800 LOC; baseline de chat.ts rebaselinado 1432→1458 (lógica nova). Fast QG + semgrep + dast verdes; 22/22 testes.

* chore(docs): update Discord invite link to a non-expiring one (#4067)

* chore(deps): freeze @huggingface/transformers in dependabot (hard-pin) (#4066)

Integrado em release/v3.8.28. Congela @huggingface/transformers no dependabot (pin exato 3.5.2, load-bearing p/ LLMLingua + memory embeddings, VPS-validado #4014). Fast QG + semgrep + dast verdes.

* ci(quality): flip TIA impacted-unit-tests gate from advisory to blocking (#4069)

The pre-existing release unit test-debt that kept the TIA "Impacted unit tests"
step advisory has been cleared:
- #4030 restored 16 lossless Zod/registry reds (from the oyi77 modularize refactors).
- #4063 fixed the last red — the LiveWS boot test — which was a real deterministic
  event-loop stall in the WS sidecar (cold ~7s lazy auth import racing a second
  connection), not an env flake; fixed (warm the import at startup) and relocated to
  the integration suite.

A full workflow_dispatch ci.yml run on release/v3.8.28 then showed all 8 Unit Tests
shards green. The remaining Integration Tests / Quality Ratchet reds are pre-existing
and unrelated (combo/resilience env-flakes; eslint/i18n baseline drift).

Removing continue-on-error makes PR->release block on unit-test regressions in the
TIA-selected impacted set (fail-safe still runs the full unit suite on hub/unmapped
changes). typecheck:core was already blocking. Closes the fast-gates "no tests on
PR->release" hole (Quality Gate v2 / Fase 9, P2).

* docs(compression): document LLMLingua optional deps + on-demand install (#4061)

Integrado em release/v3.8.28. Docs LLMLingua optional deps + on-demand install (F3.1).

* feat(dashboard): Combo Studio connection-cooldown badge (U1b Slice 2) (#4068)

Integrado em release/v3.8.28. Combo Studio connection-cooldown badge (U1b Slice 2 / F5.1).

* feat(compression): record Context Editing telemetry (engine: context-editing) (#4062)

Integrado em release/v3.8.28. Context Editing telemetry (F4.1).

* feat(sse): Context Editing relay coverage + 400-fallback (#4065)

Integrado em release/v3.8.28. Context Editing relay coverage (cc-*) + 400-fallback (F4.2/F4.3). Conflito de file-size-baseline.json (vs #4062) resolvido por união (ambas justificativas + base.ts 1292 + chatCore.ts 5898). Validado local no tree mergeado: typecheck:core ✓, eslint ✓, check:file-size ✓, 4/4 testes ✓; semgrep + semgrep-cloud verdes. Fast QG enfileirado (saturação de runner) — mergeado nos gates de política verificados (precedente #4034/#4020).

* feat(providers): add OrcaRouter (OpenAI-compatible routing gateway) (#4070)

Integrado em release/v3.8.28. Adiciona o provider OrcaRouter (OpenAI-compatible, API-key, DefaultExecutor). Ajuste de review: rebaseline de file-size de providers.ts 3147→3159 (+12 da entrada OrcaRouter). Validado local no tree sincronizado: provider-consistency ✓, docs-counts STRICT 227 ✓, typecheck:core ✓, teste 3/3 ✓, eslint ✓; semgrep + semgrep-cloud verdes. Fast QG/dast enfileirados (saturação de runner) — merge nos gates de política verificados (precedente #4034/#4065).

* test(infra): isolate DATA_DIR per test process; raise Stryker concurrency 1→4 (#4078)

* test(infra): isolate DATA_DIR per test process; raise Stryker concurrency 1→4

Every test process resolved DATA_DIR to the same default (~/.omniroute) when the env
var was unset (src/lib/dataPaths.ts::resolveDataDir), so concurrent test files opened
the SAME on-disk storage.sqlite. node:test spawns a process per file and Stryker spawns
one per sandbox, so this shared file caused cross-file state races:
- SQLite lock contention that hung `npm run test:unit` under high --test-concurrency
  (the ~95-min local hang), and
- the non-deterministic baseline that forced stryker.conf.json to concurrency: 1, which
  in turn could not finish the ~15k-mutant run inside the nightly timeout (the cancelled
  2026-06-16/17 nightly-mutation runs) — blocking Quality Gate v2 / Fase 9 Onda 2.

open-sse/utils/setupPolyfill.ts could NOT host the fix: it is imported by production
(bin/omniroute.mjs, proxyFetch.ts, proxyDispatcher.ts), where redirecting DATA_DIR would
point the live SQLite DB at a throwaway temp dir. So this adds a TEST-ONLY
tests/_setup/isolateDataDir.ts that gives each process its own temp DATA_DIR when none is
set (tests that set DATA_DIR explicitly still win), wired via --import into the test,
mutation and CI invocations.

Verified:
- Stryker dry-run A/B at concurrency=4: FAILS without the isolation import
  (account-fallback-service tap exit 9, a cross-file race) and PASSES with it.
- Full `npm run test:unit` green with isolation (0 fail; a one-off
  chatcore-translation-paths timeout flake did not reproduce and passes 3/3 isolated)
  and noticeably faster — the DB lock contention is gone.
- New tests/unit/isolate-datadir.test.ts guards the contract (unique temp DATA_DIR when
  unset; explicit DATA_DIR respected).

Wired the --import into: package.json (13 test scripts), stryker.conf.json (tap.nodeArgs
+ concurrency 1→4), .github/workflows/quality.yml (TIA step), ci.yml (the 5
unit/coverage/integration commands), and bumped nightly-mutation.yml timeout 120→180 for
the first cold run before the incremental cache is seeded.

* ci(quality): run the TIA gate at CI concurrency (4) to stop oversubscription flakes

The TIA "Impacted unit tests" step (made blocking in #4069) ran its fail-safe via
`npm run test:unit` — concurrency=20, tuned for multi-core dev machines. On a 4-vCPU CI
runner that is 5x oversubscribed, so timing-sensitive tests flake under the load (e.g.
`db-backup-extended` "The database connection is not open", `chatcore-translation-paths`
upstream-timeout). That intermittently fails a blocking gate on legitimate PRs — exactly
what surfaced on the DATA_DIR-isolation PR, whose package.json/workflow changes trip the
__RUN_ALL__ fail-safe.

Run both the impacted set and the fail-safe at --test-concurrency=4, matching the stable
ci.yml unit job. Adds a `test:unit:ci` script (test:unit at concurrency=4). The DATA_DIR
isolation in this PR keeps the parallel run race-free, so the only change here is matching
the runner's core count. Verified locally: db-backup-extended passes 8/8 in isolation
(5 with isolation, 3 without).

* docs(quality-gates): reconcile gate inventory with ci.yml + add ROI rationalization backlog (#4095)

The "authoritative" gate inventory in QUALITY_GATES.md had drifted from ci.yml: it omitted
9 wired gates — `audit:deps`, `check:tracked-artifacts`, `check:lockfile`, `check:licenses`
(lint job), `check:dead-code`, `check:cognitive-complexity`, `check:type-coverage`,
`check:codeql-ratchet` (quality-gate job), and `check:pr-evidence` (pr-test-policy job).
You can't rationalize an inventory you can't trust, so this reconciles it first.

Adds those 9 rows to their job tables and a "Rationalization Backlog (ROI review)" section
capturing the Fase 9 Onda 3 findings: mechanical merge/dedup candidates (CVE scanners
audit:deps↔osv, the two complexity ESLint passes, cycles↔circular-deps, the two /api
anti-hallucination gates, the doubly-run check:docs-sync, check:node-runtime ×11) and the
operator-only flip/drop decisions (typecheck:noimplicit vs the type-coverage ratchet,
test:vitest:ui parked fails, check:secrets frozen FPs, openapi-security-tiers, pr-evidence,
the orphaned semgrep baseline). Also flags the undocumented advisory docs-lint job and the
standalone scanner workflows.

Docs-only — no gate behavior changes. The merges (CI changes) and flips (policy) are
deferred to operator-scoped follow-ups; this PR only makes the map accurate.

* test(dashboard): smoke e2e for the Combo Live Studio page (#4075)

Integrated into release/v3.8.28

* fix(sse): friendly 413 message for ChatGPT web payload-too-large (#4080)

Integrated into release/v3.8.28

* feat(sse): port Claude Code quota-probe bypass + command meta-request helpers (#4083)

Integrated into release/v3.8.28

* feat(api): exact offline token counting for count_tokens fallback via tiktoken (#4087)

Integrated into release/v3.8.28

* feat(compression): RTK learn/discover (sample source + API + UI) (#4088)

Integrated into release/v3.8.28

* feat(dashboard): 2026-06-17 free-tier refresh — honest catalog, uncapped + boost tiers, Layout A budget table (#4089)

Integrated into release/v3.8.28

* feat(mitm): capture-pipeline self-test route (Gap 12) (#4093)

Integrated into release/v3.8.28

* fix(mitm): crash-safe system-state teardown + socket timeouts (ProxyBridge-inspired hardening) (#4084)

Integrated into release/v3.8.28 (Fast QG TIA red = 3 pre-existing timing flakes verified passing locally 82/82; PR own tests green)

* feat(mitm): attribute intercepted requests to originating process (Gap 1) (#4085)

Integrated into release/v3.8.28 (Fast QG TIA red = 3 pre-existing timing flakes verified passing locally 82/82; PR own tests green)

* fix(sse): route image requests only to confirmed-vision combo targets (#4071)

Integrated into release/v3.8.28

* fix(security): injection guard respects INJECTION_GUARD_MODE DB feature flag (#4077)

Integrated into release/v3.8.28

* fix(ws): proxy LAN /live-ws upgrades and add unset JWT_SECRET warning (#4079)

Integrated into release/v3.8.28

* fix(dev): force webpack in custom dev server (Turbopack 16.2.x panics) (#4092)

Integrated into release/v3.8.28

* ci(quality): dedup the doubly-run check:docs-sync + record validated ROI backlog (#4099)

Onda 3 (gate ROI-review) Phase 2. Two parts, both low-risk:

1. Remove the standalone `check:docs-sync` from the `lint` job — it already runs in the
   `docs-sync-strict` job (via `check:docs-all`) and the husky pre-commit hook, so the
   `lint`-job copy was a pure duplicate. No coverage lost.

2. Update the Rationalization Backlog in QUALITY_GATES.md with trust-but-verify findings:
   several "obvious" merges/flips from the ROI review turned out to hide debt and are NOT
   clean drop-ins —
   - CVE merge (audit:deps→osv): different semantics (hard high/critical vs regression-ratchet) — keep both.
   - cycles→circular-deps: dpdm reports 91 cycles (can't promote to blocking) and is broader-scope than the green curated check:cycles — keep both.
   - openapi-security-tiers flip: blocked by traffic-inspector routes missing the x-loopback-only annotation.
   - complexity + /api merges: valid but real config/script surgery — deferred.
   - node-runtime ×11: ~10s savings vs a cheap guard — low ROI, skip.

   The remaining flips (typecheck:noimplicit, test:vitest:ui, check:secrets, pr-evidence,
   semgrep) are operator policy decisions, left for the owner.

* chore(deps): bump actions/github-script from 7 to 9 (#4046)

Integrated into release/v3.8.28 (dependabot GH-Action bump; SHA-pin preserved)

* chore(deps): bump actions/setup-node from 4 to 6 (#4048)

Integrated into release/v3.8.28 (dependabot GH-Action bump; SHA-pin preserved)

* chore(deps): bump actions/upload-artifact from 4 to 7 (#4044)

Integrated into release/v3.8.28 (dependabot GH-Action bump; SHA-pin preserved)

* chore(deps): bump actions/cache from 4.3.0 to 5.0.5 (#4047)

Integrated into release/v3.8.28 (dependabot GH-Action bump; SHA-pin preserved)

* deps: bump the development group with 10 updates (#4051)

Integrated into release/v3.8.28 (dependabot dev group; cyclonedx 4->5 verified compatible with the SBOM invocation --ignore-npm-errors/--output-format JSON/--output-file)

* fix(dashboard): event-driven fail-open auto-refresh for embedded log views (#4054) (#4103)

The Request Logger gated each auto-refresh tick on a static
document.visibilityState === "visible" read. Hosts that report a permanent
non-"visible" state without ever firing a visibilitychange event (Docker
dashboard wrappers, embedded/proxied webviews) froze auto-refresh entirely —
only the manual Refresh button worked, a regression from 3.8.24's unconditional
polling.

The pause is now event-driven and fail-open: visibleRef starts true and is only
flipped to false on a real visibilitychange → hidden transition, so a host that
never signals a genuine background transition keeps polling, while normal
browser tabs still pause when actually backgrounded.

Regression test reproduces the misreporting-host case (RED) and the perf guard
is re-encoded under the event-driven semantics.

* fix(docker): raise build-stage Node heap to stop production-build OOM (#4076) (#4104)

The Docker builder stage ran `npm run build` with V8's default heap ceiling
(~2 GB). After #4052 forced the heavier webpack engine (Turbopack panics on this
Next.js version), the production optimization pass exceeded that ceiling and the
build died with "FATAL ERROR: ... JavaScript heap out of memory" at
[builder] npm run build.

The builder stage now sets NODE_OPTIONS=--max-old-space-size (default 4096 MB,
overridable via --build-arg OMNIROUTE_BUILD_MEMORY_MB) before the build; the
value propagates to the spawned next build (resolveNextBuildEnv spreads
process.env). Build-only — the runtime heap on the runner stage is unchanged,
and CI/local builds (which invoke npm run build directly) are unaffected.

Regression guard: tests/unit/dockerfile-build-heap-4076.test.ts asserts the
builder stage sets the heap ceiling, before npm run build, at >= 4096 MB.

* feat(agent-bridge): portable JSON import/export of config (Gap 4) (#4094)

Integrated into release/v3.8.28

* feat(cli): add 'omniroute launch' zero-config Claude Code launcher (#4097)

Integrated into release/v3.8.28 (Fast QG TIA red = pre-existing env-doc-contract drift [MITM_IDLE_TIMEOUT_MS/TURBOPACK from #4084/#4092] + opencode-plugin-dist env flake; #4097 own test 3/3 green)

* feat(mitm): loop-guard self-check + verbosity control in server.cjs (Gaps 14+15) (#4101)

Integrated into release/v3.8.28 (rebased onto release — dropped the already-squash-merged #4084 commits; only the Gaps 14+15 loop-guard/verbosity delta remains)

* feat(sse): generic 400 field-downgrade retry + Groq field stripping (#4096)

Integrated into release/v3.8.28

* feat(providers): add Wafer AI (Anthropic-compatible, Bearer auth) (#4098)

Integrated into release/v3.8.28

* chore(docs)

* fix(responses): clear /v1/responses keepalive timer on cancel/abort (timer + CPU leak) (#4105)

Integrated into release/v3.8.28 (r7).

* perf(gemini): cache reasoning close-tag regex instead of recompiling per token (#4106)

Integrated into release/v3.8.28 (r7).

* fix(usage): reap orphaned pending-request details (unbounded memory leak) (#4107)

Integrated into release/v3.8.28 (r7).

* perf(stream): use structuredClone instead of JSON round-trip for per-chunk reasoning split (#4108)

Integrated into release/v3.8.28 (r7).

* fix(dashboard): restore Update Available banner with npm-binary-free version fallback (#4100) (#4112)

getLatestNpmVersion() derived the latest version only from the npm CLI binary and returned null on any error, so Docker/desktop/locked-down installs without npm on PATH silently hid the home banner even when an update existed. Add resolveLatestVersion() (npm CLI -> registry HTTP fallback -> logged warning) and harden version parsing for v-prefix/pre-release strings. Extracted into testable src/lib/system/versionCheck.ts with TDD coverage.

* fix(auth): prune expired entries from login brute-force guard map (unbounded growth) (#4111)

Integrated into release/v3.8.28 (r8)

* fix(logger): hard-cap the error-dedup map to bound memory under unique-message bursts (#4113)

Integrated into release/v3.8.28 (r8)

* fix(circuit-breaker): enforce MAX_REGISTRY_SIZE (declared but never applied) (#4114)

Integrated into release/v3.8.28 (r8)

* perf(obfuscation): cache per-word regexes instead of recompiling every request (#4109)

Integrated into release/v3.8.28 (r8)

* perf(registry): precompute model->provider index in parseModelFromRegistry (#4110)

Integrated into release/v3.8.28 (r8)

* fix(timers): unref background interval timers so they don't block clean shutdown (#4117)

Integrated into release/v3.8.28 (r8)

* fix(webhook): clear abort timer in finally to avoid dangling timers on fetch error (#4115)

Integrated into release/v3.8.28 (r8)

* fix(combo): detach per-target listener from shared hedge abort signal (#4116)

Integrated into release/v3.8.28 (r8)

* chore(release): finalize v3.8.28 CHANGELOG + reconcile env-doc contract

- Build the complete [3.8.28] CHANGELOG section (55 bullets) covering every
  commit since v3.8.27, grouped by type with PR back-references and human
  contributor attribution (artickc's memory-leak/perf cluster, OrcaRouter,
  Wafer AI, MITM gaps, etc.); move the OrcaRouter bullet out of [Unreleased].
- Inject the EN [3.8.28] section into all 41 i18n CHANGELOG mirrors (parity).
- Reconcile the env/docs contract: document MITM_IDLE_TIMEOUT_MS + MITM_VERBOSE
  in .env.example and ENVIRONMENT.md; allowlist the framework-internal TURBOPACK
  and the Claude Code ANTHROPIC_AUTH_TOKEN in check-env-doc-sync.
- Fix 3 broken relative links in docs/providers/AGENTROUTER.md (regressed when
  the file was relocated this cycle) so docs-sync-strict passes.

* fix(quality): treat test→test renames as relocations, not deletions

The anti-test-masking gate's subcheck-1 collected deleted AND renamed test
files via `--diff-filter=DR --name-only` and flagged every one as "deleted —
human review required", contradicting its own documented contract ("DELETADOS
ou renomeados-e-NÃO-substituídos"): a rename test→test IS a substitution (the
test moved, coverage preserved). This false-positived on #4063's legitimate
relocation of live-ws-startup.test.ts (unit/cli → integration, asserts 2→2)
and would block every PR that relocates a test — surfacing only at release-day
because the Fast QG (PR→release) doesn't run test-masking.

The gate now parses `--name-status -M`: true deletions and test→non-test
renames still flag; a test→test rename is run through the assert-reduction
check across the move, so a clean relocation passes while gutting-via-rename
(dropped asserts / new tautologies / skips) still fires. Adds
partitionDeletedRenamed + 6 regression tests.

---------

Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Demiurge The Single <megamen932@gmail.com>
Co-authored-by: jinhaosong-source <jinhao.song@myflashcloud.com>
Co-authored-by: diego-anselmo <contato@diegoanselmo.com.br>
Co-authored-by: Felipe Almeman <4226997+zhiru@users.noreply.github.com>
Co-authored-by: Rahul sharma <sharmaR0810@gmail.com>
Co-authored-by: Chirag Singhal <76880977+chirag127@users.noreply.github.com>
Co-authored-by: NOXX - Commiter <artur1992123@mail.ru>
2026-06-17 19:26:32 -03:00

157 lines
6.5 KiB
TypeScript

/**
* Anthropic Context Editing — delegated, server-side compression for Claude.
*
* Unlike OmniRoute's local compression engines (which rewrite the request body
* before it leaves the proxy), Context Editing asks the *provider* to clear
* stale tool-use/tool-result blocks from the running context window. We only
* attach a body param (`context_management.edits[]`); Claude does the clearing
* against its own tokenizer. The beta is advertised via the `anthropic-beta:
* context-management-2025-06-27` header, which OmniRoute already emits on Claude
* requests (see `anthropicHeaders.ts`).
*
* This is a Claude-only capability by nature — other providers would reject the
* param. Callers MUST gate invocation to the genuine Claude provider; this
* module does not inspect the provider itself.
*
* Strategy ids and shapes are pinned from live Anthropic docs
* (platform.claude.com/docs/en/build-with-claude/context-editing).
*/
/** Dated strategy id that clears old tool-use / tool-result pairs. */
export const CLEAR_TOOL_USES_STRATEGY = "clear_tool_uses_20250919";
/** Dated strategy id that clears old extended-thinking turns. */
export const CLEAR_THINKING_STRATEGY = "clear_thinking_20251015";
/** Default token threshold that triggers clearing (Anthropic default is 100k). */
export const CONTEXT_EDITING_DEFAULT_TRIGGER_TOKENS = 100000;
/** Recent tool-use/result pairs kept untouched when clearing. */
export const CONTEXT_EDITING_DEFAULT_KEEP_TOOL_USES = 3;
type ContextEditingEdit = { type?: unknown; [key: string]: unknown };
/**
* Mutate an Anthropic Messages request body in place to delegate context
* clearing to the provider. Adds the `clear_tool_uses_20250919` strategy to
* `context_management.edits[]`, composing with any `clear_thinking_20251015`
* edit already present — Anthropic requires `clear_thinking` to be listed first,
* so thinking edits are stable-sorted to the front.
*
* Idempotent: if a `clear_tool_uses` edit already exists (added by a previous
* call or supplied by the client), the body is left as-is. No-op when disabled
* or when `body` is not a plain object.
*
* @param body The outbound Anthropic-format request body (mutated in place).
* @param opts.enabled Whether Context Editing is on for this request.
*/
export function applyContextEditingToBody(
body: Record<string, unknown> | null | undefined,
opts: { enabled: boolean }
): void {
if (!opts.enabled || !body || typeof body !== "object") return;
const existing =
body.context_management && typeof body.context_management === "object"
? (body.context_management as Record<string, unknown>)
: {};
const edits: ContextEditingEdit[] = Array.isArray(existing.edits)
? [...(existing.edits as ContextEditingEdit[])]
: [];
const hasToolUseEdit = edits.some((edit) => edit && edit.type === CLEAR_TOOL_USES_STRATEGY);
if (!hasToolUseEdit) {
edits.push({
type: CLEAR_TOOL_USES_STRATEGY,
trigger: { type: "input_tokens", value: CONTEXT_EDITING_DEFAULT_TRIGGER_TOKENS },
keep: { type: "tool_uses", value: CONTEXT_EDITING_DEFAULT_KEEP_TOOL_USES },
});
// Anthropic requires clear_thinking edits to precede clear_tool_uses. Array
// sort is stable in modern Node, so same-key edits keep their relative order.
edits.sort((a, b) => {
const aRank = a && a.type === CLEAR_THINKING_STRATEGY ? 0 : 1;
const bRank = b && b.type === CLEAR_THINKING_STRATEGY ? 0 : 1;
return aRank - bRank;
});
}
existing.edits = edits;
body.context_management = existing;
}
/**
* Telemetry receipt for one Claude response that ran server-side context editing.
* `clearedInputTokens` is how many input tokens the provider pruned from its own
* context window — i.e. the tokens "saved" on subsequent turns by the delegated
* clearing (the value we surface in compression analytics under
* `engine: "context-editing"`).
*/
export interface ContextEditingTelemetry {
/** Number of applied edit entries the provider reported. */
editCount: number;
/** Total input tokens cleared across all applied edits. */
clearedInputTokens: number;
/** Total tool-use/result pairs cleared across all applied edits. */
clearedToolUses: number;
}
/** Coerce a finite, non-negative integer; returns 0 for junk. */
function toClearedInt(value: unknown): number {
if (typeof value === "number" && Number.isFinite(value)) return Math.max(0, Math.floor(value));
if (typeof value === "string" && value.trim()) {
const parsed = Number(value);
if (Number.isFinite(parsed)) return Math.max(0, Math.floor(parsed));
}
return 0;
}
/** Read a nested property by key path without throwing on missing intermediates. */
function getNested(obj: unknown, keys: string[]): unknown {
let cur: unknown = obj;
for (const key of keys) {
if (!cur || typeof cur !== "object") return undefined;
cur = (cur as Record<string, unknown>)[key];
}
return cur;
}
/**
* Extract the context-editing telemetry receipt from a Claude (Anthropic Messages)
* response body. Returns `null` when context editing did not run (no applied edits,
* or nothing was actually cleared).
*
* Defensive over the response shape: Anthropic surfaces `applied_edits` under
* `context_management` (top-level), and some response variants nest it under
* `usage`. We probe both rather than pin a single location, so the telemetry keeps
* working if the provider moves the field. Per-edit cleared counts are read from
* `cleared_input_tokens` / `cleared_tool_uses` (snake_case) with a camelCase
* fallback.
*/
export function extractContextEditingTelemetry(
responseBody: unknown
): ContextEditingTelemetry | null {
if (!responseBody || typeof responseBody !== "object") return null;
const candidates: unknown[] = [
getNested(responseBody, ["context_management", "applied_edits"]),
getNested(responseBody, ["usage", "context_management", "applied_edits"]),
getNested(responseBody, ["usage", "applied_edits"]),
];
const edits = candidates.find((c) => Array.isArray(c)) as unknown[] | undefined;
if (!Array.isArray(edits) || edits.length === 0) return null;
let clearedInputTokens = 0;
let clearedToolUses = 0;
for (const entry of edits) {
if (!entry || typeof entry !== "object") continue;
const edit = entry as Record<string, unknown>;
clearedInputTokens += toClearedInt(edit.cleared_input_tokens ?? edit.clearedInputTokens);
clearedToolUses += toClearedInt(edit.cleared_tool_uses ?? edit.clearedToolUses);
}
if (clearedInputTokens <= 0 && clearedToolUses <= 0) return null;
return { editCount: edits.length, clearedInputTokens, clearedToolUses };
}