changes.md

changes — senpi-monorepo root

Unify the shared and evals Vitest runners (2026-09-21)

What changed

• package.json pins the root development runner and its V8 coverage provider to 5.0.1 so the hoisted runner can load coverage.
• .gitignore excludes the .vitest/ artifact directory.
• package-lock.json and bun.lock resolve Vitest and V8 coverage 5.0.1 across every workspace, including evals.
• package.json overrides vitest-evals 0.17.0's Vitest peer edge to 5.0.1.
• bun.lock retains configuration version 0 and the existing hoisted install layout.

Why

• PTY and codemode invoke the hoisted runner without declaring it. A root pin makes their shared runner version explicit under both npm and Bun.
• vitest-evals 0.17.0 declares Vitest >=4 <5. Its npm peer override makes the single-major installation explicit; runtime tests and TypeScript checks verify compatibility instead of preserving a split runner graph.
• Bun hoists the harness beside the root runner even when a lock entry requests workspace nesting. Keeping every runner on 5.0.1 avoids mixed TaskMeta types without changing the native workflows' root dependency paths.

Why an extension could not handle it

• Package managers select test runners and resolve peer dependencies before extensions load.

Expected merge conflict zones

• The root development dependencies and generated dependency locks.

Run the two dev lanes through run-workspaces --parallel and drop concurrently (2026-09-21)

What changed

• package.json: the root dev script is node scripts/run-workspaces.mjs --parallel --workspace packages/ai --workspace packages/coding-agent dev; the concurrently devDependency is removed and shell-quote 1.10.0 is declared as a root devDependency — three repository scripts import it directly but it only reached node_modules as concurrently's transitive dependency (its version was already pinned by the root override). package-lock.json / bun.lock are regenerated the repository way (bun.lock stays configVersion: 0).

Why

• concurrently was the last root script that bypassed the package-manager-agnostic runner from #1447; npm run dev, bun run dev and pnpm run dev now all start both lanes through the same driver, with prefixed output and one Ctrl-C reaching every lane (senpi#1895).

Why an extension could not handle it

• Root scripts and the dependency closure are resolved by the package manager before any extension loads.

Expected merge conflict zones

• The root scripts.dev line and the root devDependency block, on every upstream tooling bump.

Refresh the dependency pins and pin past the reachable advisories (2026-09-21)

What changed

• package.json: the root overrides block moves fast-uri to 3.1.8, brace-expansion to 5.0.12 and @anthropic-ai/sdk to 0.127.0, and gains express-rate-limit 8.7.0, hono 4.13.8, ip-address 10.7.2, qs 6.16.0 and a nested @earendil-works/gondolin > undici 6.28.1. @types/node moves to 26.6.2, @biomejs/biome to 2.5.14 and tsx to 4.23.13.
• biome.json: the $schema URL follows the Biome pin to 2.5.14.
• packages/telemetry/package.json: @types/node moves to 26.6.2.

Why

• Every advisory npm audit and bun audit could reach came in through a transitive edge the fork does not declare: fast-uri and ajv, and the @modelcontextprotocol/sdk subtree that carries hono, qs and express-rate-limit > ip-address. scripts/regenerate-bun-lock-isolated.mjs seeds its island with the committed bun.lock, so re-resolving only the npm lock left Bun on the vulnerable copies; declaring the versions as overrides moves both lockfiles together without drifting the 56 unrelated transitives a from-scratch Bun resolution touched.
• tsx stops at 4.23.13 because 4.23.14 and 4.23.15 were both published 2026-09-20, inside the .npmrc min-release-age=2 window npm enforces.

Why an extension could not handle it

• Dependency resolution and formatter configuration are read by the package manager and the toolchain before any extension is loaded.

Expected merge conflict zones

• LOW: the overrides block and the devDependency versions, on every upstream manifest bump.

Make B.AI credentials available to development environments (2026-09-18)

What changed

• .devcontainer/devcontainer.json exposes an optional B.AI secret alongside the other provider keys.
• pi-test.sh, pi-test.ps1, test.sh, and packages/coding-agent/scripts/qa-app-server/lib/env.mjs scrub BAI_API_KEY from hermetic test processes.

Why

• The native B.AI provider should work consistently in local checkouts and dev containers without storing credentials in tracked files.

Why an extension could not handle it

• Development environment bootstrapping and container secret declarations run before Senpi or its extensions.

Expected merge conflict zones

• LOW: the provider-key arrays in the setup script and devcontainer secret block.

Root check verifies formatting instead of rewriting it (2026-09-17)

What changed

• package.json: the root check script now runs biome check --error-on-warnings . (read-only) instead of biome check --write --error-on-warnings ., so format drift fails the script rather than being silently repaired. A new check:fix script keeps the autofix form (biome check --write --error-on-warnings . && npm run check) for local use.
• .husky/pre-commit: runs npm run check:fix, preserving the hook's existing autofix-then-verify behavior now that check no longer writes.
• .github/workflows/releasability.yml: drops the hand-inlined read-only biome step plus its verbatim copy of the remaining check sub-scripts and calls npm run check directly; that workaround existed only because check autofixed, and its copy had already drifted from the real chain (missing check:entry-graphs and check:claude-sdk-platform-lock).

Why

• #1443: the CI Static checks job runs npm run check, whose leading biome check --write reformats offending files inside the runner and exits 0. The rewrite is discarded when the runner exits, so a formatting regression could never fail CI while drift accumulated on main. Root AGENTS.md also requires the local check and CI to stay in sync; with autofix in the shared script they disagreed by construction.

Why an extension could not handle it

• The root package.json script chain, the Husky hook, and the workflow step are build-time and repository-policy gates that execute before any Senpi runtime loads; no runtime extension participates in them.

Expected merge conflict zones

• LOW: the check script string in root package.json and the adjacent check:fix entry.
• LOW: the check invocation line in .husky/pre-commit.

claude-sdk-oauth re-login refreshes the slot; stored pool blocks bind to credential revisions (2026-09-17)

What changed

• packages/coding-agent/src/core/extensions/builtin/anthropic-subscription/accounts.ts: new upsertAccount() — a same-name slot is replaced in place (fresh token material, block stamps cleared, displayName preserved); unknown names still append.
• packages/coding-agent/src/core/extensions/builtin/anthropic-subscription/oauth-login.ts: a re-login now targets an existing slot instead of minting account-N+1 — a lone slot or the pool's one auth_error-blocked slot is refreshed in place; anything else stays append-only unless the user types an existing name, so a blank or headless re-login never overwrites the newest working slot in a multi-account pool. The Anthropic import is now a move: accepting it removes the grant from the anthropic provider so two stores never refresh one single-use token.
• packages/coding-agent/src/core/extensions/builtin/anthropic-subscription/index.ts: wires removeAnthropicCredential through the locked auth.json backend.
• packages/coding-agent/src/core/extensions/builtin/anthropic-subscription/{affinity,guidance,stream-guidance}.ts: AllAccountsBlockedError carries the dominant block reason and the all-blocked guidance names an authentication failure explicitly, so the outer credential-pool classifier maps it to auth_error instead of laundering it into a rate-limit cooldown via the generic "(rate limit or auth errors)" wording.
• packages/coding-agent/src/core/credential-pool/{state-store,rotation-stream}.ts and packages/coding-agent/src/core/credential-accounts.ts: stored-lane sidecar health is bound to a credential revision (HMAC over the installation key and slot material, never raw material) — the stored-lane twin of the env revision rule — so a re-login or token refresh retires the block the old material earned. Legacy rows without a revision are retired on first read.
• Tests: new test/claude-sdk-oauth-login-refresh.test.ts (refresh matrix, import move, dominant reason) and test/credential-pool-stored-revision.test.ts (legacy/foreign/current revision, revision stamping); test/credential-error-taxonomy.test.ts gains the guidance→classifier composition cases; test/credential-accounts.test.ts and test/model-runtime-credential-rotation.test.ts fixtures now stamp the matching revision for blocks that must apply.

Why

• omo#7084 (two fresh field reports on 2026.9.16-3): /login claude-sdk-oauth never refreshed the existing slot — it appended account-N+1 or threw on a duplicate name — while auth_error blocks were permanent by design ("until login refreshes the slot"), so the documented recovery could never fire and the pool dead-ended at "blocked until re-login". The import path also copied the Anthropic grant into a second store, guaranteeing a later invalid_grant.
• omo#8383: the lane's generic all-blocked wording let the outer classifier string-match "rate limit" and stamp a cooldown for what was a 401 revocation, and the stored lane's sidecar had no credential-replacement signal at all, so blocks outlived the credential that earned them.

Why an extension could not handle it

• The slot store, the OAuth login flow, the failover block policy, and the credential-pool sidecar are engine internals; the recovery contract spans the lane's auth.json stamps and the generic pool's sidecar, which no extension surface reaches.

Expected merge conflict zones

• packages/coding-agent/src/core/extensions/builtin/anthropic-subscription/oauth-login.ts (login naming and the import branch).
• packages/coding-agent/src/core/credential-pool/rotation-stream.ts (stored-lane listing and persistBlock).

Type-check the qa scripts (2026-09-16)

What changed

• tsconfig.json includes scripts/qa/**/*.ts so root tsc --noEmit type-checks the qa runners.
• package.json: npm run check also runs tsc --noEmit -p scripts/tsconfig.json after the root program.

Why

• scripts/qa/*.ts sat outside every tsconfig, so implicit-any import errors there never failed CI.

Why an extension could not handle it

• Root tsconfig.json include globs and the package.json check script are compile-time gates; no runtime extension can add files to tsc.

Expected merge conflict zones

• LOW: the include array in root tsconfig.json and the check script string in root package.json.

Re-wire check:entry-graphs into the root check chain (2026-09-13)

What changed

• package.json: npm run check runs check:entry-graphs after check:ts-imports, matching the original 5507d76ee gate. scripts/check-entry-graphs.mjs prints each entry's file count on success so a green run still reports the harness/session size.
• packages/agent/src/harness/messages.ts: session no longer value-imports the AI barrel; see packages/agent/src/changes.md.

Why

• The session subpath is a cost contract (budget 25, no packages/ai/src/index.ts). The script existed but was not in check, so the barrel regression stayed red until someone ran it by hand.

Why an extension could not handle it

• Root package.json scripts and the source import graph are build-time inputs; no runtime extension can restore either.

Expected merge conflict zones

• LOW: the check script string in root package.json.

Pin the bundled chord workspace to upstream's published version (2026-09-12)

What changed

• packages/chord/package.json returns to upstream's own 0.85.1 version instead of the fork CalVer stamp, so the bundled workspace keeps @earendil-works/chord's published release identity.

Why

• chord is bundled into the senpi tarball but the fork does not publish it. CalVer-stamping it made the packaged manifests declare @earendil-works/chord@^<CalVer>, which no registry version answers, so bun add @code-yeongyu/senpi failed (issue #1632). Keeping chord on upstream's 0.85.1 — which exists on the registry and is byte-for-byte our bundled copy apart from packaging metadata — makes every declared edge resolvable while the bundled copy shadows it at runtime.

Why an extension could not handle it

• packages/chord/package.json is static manifest data consumed by the package manager and the release/publish pipeline, never reachable from the runtime extension system.

Expected merge conflict zones

• The version field in packages/chord/package.json.

Scrub VENICE_API_KEY in the hermetic test environments (2026-09-10)

What changed

• pi-test.sh and pi-test.ps1 add VENICE_API_KEY to the provider credentials cleared before the suite runs, alongside the existing CEREBRAS_API_KEY/XAI_API_KEY entries. test.sh and packages/coding-agent/scripts/qa-app-server/lib/env.mjs gained the same entry.

Why

• Venice is now a built-in provider, and several suites key opt-in live behavior off the mere presence of a provider API key. Leaving VENICE_API_KEY in the inherited environment would let a developer's real credential change test behavior or reach the network.

Why an extension could not handle it

• These are the shell entry points that build the test environment before any senpi process starts.

Expected merge conflict zones

• LOW: the unset/credential-name lists in pi-test.sh and pi-test.ps1 when upstream adds providers.

Root scripts reach workspaces only through scripts/run-workspaces.mjs (2026-09-07)

What changed

• package.json: test, clean, eval, dev, dev:tsc, generate:models, generate:model-catalog, hydrate:model-data, and check:model-data delegate into workspaces through node scripts/run-workspaces.mjs [--if-present] [--workspace <path>] <script> instead of npm run --workspaces --if-present <script>, npm --workspace=<name> run, npm --prefix <dir> run, or cd <dir> && npm run lanes inside concurrently. dev keeps only the packages/ai and packages/coding-agent lanes, the two workspaces that define a dev script. version:* keep npm version --workspaces (npm's version bookkeeping, not a script delegation); refresh-lock, publish*, and release* are untouched.

Why

• Under bun the old shapes worked only where bun happened to rewrite npm run to bun run, and bun's --workspaces fans out in parallel while npm runs sequentially; --prefix, --workspace=, and cd <dir> && npm run never reach bun or pnpm and always execute real npm, against the scripts/AGENTS.md rule of not hardcoding the child package manager. The runner executes every workspace script with the manager that launched the root script, sequentially and in path order, with one PASS / SKIP / FAIL summary, so bun run test, npm run test, and pnpm run test behave identically. The packages/agent and packages/tui dev lanes pointed at scripts that do not exist.

Why an extension could not handle it

• Root manifest scripts run before any Senpi runtime starts; the package manager is the only surface above them.

Expected merge conflict zones

• LOW: the nine script lines in the root package.json scripts block. Upstream still spells these in npm's dialect; keep the runner form on sync.

bun.lock refreshed wherever package-lock.json is refreshed (2026-09-04)

What changed

• package.json: version:patch, version:minor and version:major append bun install --lockfile-only after npm install --package-lock-only --ignore-scripts.
• package.json: refresh-lock does the same, so the manual lockfile-refresh path and the release path agree.

Why

• bun install --frozen-lockfile broke on main twice in one day: release v2026.9.4-2 (a79aa2080) rewrote package-lock.json while bun.lock kept 2026.9.3 workspace versions, and d052dbb6b added the @anthropic-ai/sdk override without regenerating it (fixed as a one-off in #1364).
• Both incidents share one cause: every script that refreshes a lockfile refreshes only the npm one, and CI installs with npm ci, so the drift recurs at each bump with nothing observing it. #1364 cleared the symptom; this closes the source.

Why an extension could not handle it

• These are the repo's own release and lockfile-maintenance scripts; nothing outside package.json decides which lockfiles a version bump rewrites.

Expected merge conflict zones

• package.json scripts — the version:* and refresh-lock lines.

@anthropic-ai/sdk 0.123.0 pin for the v0.84.4 upstream sync (2026-09-04)

What changed

• package.json pins @anthropic-ai/sdk at 0.123.0 (was 0.120.0), the version the 2026-09-03 upstream sync of badlogic/pi-mono v0.84.4 resolved against.
• .npmrc adds min-release-age-exclude[]=@anthropic-ai/sdk next to the existing excludes, with an in-file note to remove it after 2026-09-05.

Why

• The fork's min-release-age=2 supply-chain gate refuses packages younger than two days, so the freshly published 0.123.0 pin would fail installs until 2026-09-05. The exclude is time-boxed by its removal note instead of weakening the policy for every package.

Why an extension could not handle it

• Root manifest pins and npm install policy execute before any package code, let alone an extension, runs.

Expected merge conflict zones

• LOW: the package.json root dependency overrides block and the .npmrc exclude list on future upstream syncs and release bumps.

Root workspace fan-out scripts no longer recurse under bun (2026-09-02)

What changed

• package.json: the three root scripts that fan out to workspaces stop passing workspace flags after the script name. test and clean move the flags before the script name (npm run --workspaces --if-present <script>), and eval moves its flag before run (npm --workspace=@code-yeongyu/senpi-evals run eval --). npm behavior is unchanged in all three cases.
• scripts/root-workspace-scripts.test.mjs (new): parses the root manifest and fails a root script for either recursion-prone shape — a workspace flag after the script name, or a singular --workspace on an npm run call (which bun ignores, re-entering the root script). Both shipped shapes are covered; a mutation check confirms reverting eval to npm run --workspace=<name> eval fails the guard.

Why

• Bun rewrites npm run <name> to bun run <name> inside script text, and bun appends flags placed after the script name to the script itself instead of parsing them. npm run test --workspaces --if-present therefore re-invoked the ROOT script with an ever-growing flag suffix (bun run test --workspaces --if-present --workspaces --if-present ...) and spun forever instead of running the workspace suites — it never failed, so it read as a slow suite. clean and eval had the same defect. Verified in a throwaway fixture: the flag-before form fans out under both npm and bun, while the singular --workspace=<name> form still recurses under bun (bun does not recognize it), which is why eval needs the flag before run so no npm run substring remains to rewrite.

Why an extension could not handle it

• These are root package manifest scripts consumed by the release gate (scripts/release.mjs, scripts/local-release.mjs run CI=1 npm test) and by contributors directly; no extension surface exists above the package manager.

Expected merge conflict zones

• LOW: the test, clean, and eval lines in the root package.json scripts block.

Shared-host rendering isolation (2026-08-30)

Shared socket clients now register rendered_components through additive set_client_info capabilities. Factory-rendered component records are filtered per connection, including capability-aware snapshot replay. Capabilities remain connection-wide across sessions and are cleared only on socket release; explicit close removes only the closing width. Shared bindings retain factories while disposing live renderers and footer providers when no capable connection remains, recreating them for later capable joiners.

Root tracker for repository-level divergence from upstream badlogic/pi-mono. Owns every audited production path whose nearest tracker is the repository root.

CodeGraph reference cleanup (2026-09-02)

What changed

• biome.json drops the !!**/.codegraph ignore entry.

Why

• The omo product removed its CodeGraph integration, so nothing writes a .codegraph directory anymore. An ignore entry for a directory that is never created is dead configuration that implies the integration still exists.
• The matching EXCLUDED_ROOT_PATHS change in packages/coding-agent/src/beta/omo-local-update-fingerprint.ts is recorded in packages/coding-agent/src/changes.md, that path's nearest ancestor tracker.

Why this lives in the fork

• The biome ignore list is fork-owned: it is an omo-specific surface that upstream badlogic/pi-mono does not carry.

Expected merge conflict zones

• LOW: biome.json ignore list ordering during upstream syncs.

@anthropic-ai/sdk peer alignment (2026-08-26)

What changed

• package.json bumps the pinned @anthropic-ai/sdk from 0.91.1 to 0.120.0 so the pin satisfies @anthropic-ai/claude-agent-sdk@0.3.241's >=0.93.0 peer range.

Why

• Every bun install printed warn: incorrect peer dependency "@anthropic-ai/sdk@0.91.1"; the SDK floor moved to 0.93.0 when the agent SDK gained its credentials subsystem.

Why this lives in the fork

• The root pin set is fork-owned dependency policy; upstream does not pin these packages together.

Expected merge conflict zones

• LOW: package.json root dependency pins during upstream syncs.

Root config and package identities re-diverge from upstream dcd4619 (2026-08-25)

What changed

• biome.json keeps the fork lint surface: schema 2.5.10, preset: "recommended" syntax, and the **/api/cursor-agent/gen and **/.codegraph exclusions.
• packages/agent/package.json keeps the senpi calver (2026.8.24), tsc build (upstream uses tsgo), and the fork dependency set (diff 9, typebox 1.3.18, calver workspace ranges).
• packages/session-backends/sqlite-node/package.json keeps the fork package name @earendil-works/pi-storage-sqlite-node, tsc build, and vitest 4.1.11.
• packages/telemetry/package.json keeps calver, @types/node 26, vitest 4.1.11, and private: true.
• packages/tui/package.json keeps calver, tsc build, the --import tsx + multiplexer-env test loader, node >=24, marked 18.0.10, and the bench:frame-cost script.

Why

These are fork-owned product surfaces (senpi branding, provider wire behavior, fork runtime features) that upstream does not carry; the sync must re-assert them on top of upstream's tree.

Why this lives in the fork

The divergence lives in core wiring, package identity, or build plumbing that executes before any extension loads, so no extension hook can express it.

Expected merge conflict zones

• Version/name/scripts blocks of every listed package.json on each upstream release bump; biome.json whenever upstream migrates Biome versions.

Vitest source alias for ai auth subpaths (2026-08-25)

What changed

• vitest.base.ts: added a resolve alias mapping @earendil-works/pi-ai/auth/* to packages/ai/src/auth/*.ts so vitest resolves the new auth/pool/slots subpath to source during tests.

Why

• Workspace tests import @earendil-works/pi-ai/auth/pool/slots; without a source alias vitest resolves to the built dist, which does not exist for the new module, breaking test runs.

Why an extension could not handle it

• Test runner aliasing is repository-level tooling configuration.

Expected merge conflict zones

• LOW: single additive alias line in vitest.base.ts.

Release dependency refresh (2026-08-24)

What changed

• package.json: @biomejs/biome 2.5.9 -> 2.5.10.
• packages/agent/package.json: typebox 1.3.16 -> 1.3.18.
• packages/ai/package.json: typebox 1.3.16 -> 1.3.18.
• packages/coding-agent/package.json: typebox 1.3.16 -> 1.3.18.
• packages/protocol/package.json: typebox 1.3.16 -> 1.3.18.
• packages/senpi-codemode/package.json: typebox 1.3.16 -> 1.3.18.
• packages/{ai,coding-agent}/package.json: @aws-sdk/client-bedrock-runtime 3.1115.0 -> 3.1116.0.
• packages/coding-agent/package.json: @anthropic-ai/claude-agent-sdk 0.3.238 -> 0.3.241.
• Root and generated release locks were regenerated from those exact pins.

Why

• These are the repository-audited patch-level or same-line upgrades available for the 2026.8.24 release. TypeBox must remain single-instanced across the shared protocol/runtime packages, and the Bedrock pin must remain identical in ai and coding-agent. The Claude Agent SDK update also requires regenerating its platform lock and the published/install dependency closures.
• @anthropic-ai/sdk remains at 0.91.1 because the minimum peer-compatible 0.93.0 still introduces browser-breaking credential-chain imports, while 0.120.0 is likewise unsafe. Deliberate breaking holds remain unchanged for openai 6.26.0 and signal-exit 3.0.7.

Why an extension could not handle it

• Dependency resolution, exact pins, generated release locks, and platform-package selection happen before the runtime and extension system load.

Expected merge conflict zones

• HIGH: root and coding-agent dependency blocks and generated lock artifacts.
• MEDIUM: the shared TypeBox pins across five package manifests.

Dependency pin refresh, unused-dependency removal, and lock regeneration (2026-08-20)

What changed

• package.json: root devDependencies bumped esbuild 0.28.1 -> 0.28.2 and tsx 4.23.1 -> 4.23.12; declared concurrently 10.0.5 (the root dev script invoked it while it was undeclared and absent from the lock); dropped the unused @anthropic-ai/sandbox-runtime and jiti devDependencies and the unused get-east-asian-width dependency. Overrides bumped @hono/node-server 2.0.10 -> 2.1.1, postcss 8.5.18 -> 8.5.26, brace-expansion 5.0.8 -> 5.0.9, esbuild 0.28.1 -> 0.28.2, rimraf 6.1.2 -> 6.1.3 (including the nested gaxios.rimraf pin), shell-quote 1.9.0 -> 1.10.0, vite 8.0.16 -> 8.2.2, and ws 8.21.1 -> 8.21.3, while fast-uri stays on 3.x and protobufjs on 7.x and @anthropic-ai/sdk stays pinned at 0.91.1.
• .npmrc: rewrote the min-release-age exemption list as package-name patterns (@hono/node-server, @anthropic-ai/claude-agent-sdk, @aws-sdk/*, @google/genai, @smithy/*, typebox, vite) so the freshly published target versions resolve under the repository's two-day supply-chain window.
• packages/agent/package.json, packages/protocol/package.json: typebox moved to 1.3.16 (from 1.3.8 and from the inconsistent 1.3.7).
• packages/telemetry/package.json: @types/node 24.12.4 -> 26.2.0, matching the rest of the repository.
• packages/tui/package.json: marked 18.0.7 -> 18.0.10.
• crates/senpi-pty/Cargo.toml, crates/senpi-pty/package.json, and the workspace Cargo.toml pins: libc =0.2.174 -> =0.2.189, napi =3.10.3 -> =3.12.1, napi-derive =3.5.9 -> =3.6.3, napi-build =2.3.2 -> =2.4.1, @napi-rs/cli 3.7.2 -> 3.8.6.
• scripts/rolldown-platform-lock.test.mjs: the asserted Rolldown binding version tracks 1.0.3 -> 1.2.4, which is what vite 8.2.2 resolves.

Why

• These pins had drifted behind their current releases while the repository enforces exact pins through npm run check:pinned-deps, so refreshing them in one pass keeps every workspace on one resolved version and keeps the shared typebox identity single-instanced. The removals delete manifest entries with zero source references, and declaring concurrently makes the root manifest truthful about what npm run dev actually needs. @anthropic-ai/sdk is deliberately held at 0.91.1 because 0.120.0 adds credential-chain modules whose node:fs and node:path imports break the browser-bundle invariant enforced by scripts/check-browser-smoke.mjs. The .npmrc rewrite fixes an exemption list that could never match: npm compares these patterns against the package name only, so the previous name@version string was inert.

Why an extension could not handle it

• Dependency resolution, override pinning, the supply-chain age gate, and Cargo pin selection are all performed by the package managers before any runtime exists, so no extension can influence which versions get installed or locked.

Expected merge conflict zones

• HIGH: the overrides and devDependencies blocks in package.json, which upstream edits on nearly every release.
• MEDIUM: the per-package typebox/@types/node pins and the workspace Cargo.toml dependency table.
• LOW: .npmrc and the Rolldown binding version constant.

Repository-wide upstream divergence audit (2026-08-17)

What changed

Canonical backfill seeded from the pre-backfill audit report under local-ignore/qa-evidence/20260817-changes-md-audit/pre-backfill-audit.json (upstream pin badlogic/pi-mono v0.84.2, 914cf1472e715297caa30db4b9535d534a9eb718). Every remaining audited production path with no nearer tracker than the root:

• .npmrc: adds min-release-age-exclude=@hono/node-server@2.0.10 on top of the upstream min-release-age supply-chain policy.
• biome.json: biome schema 2.3.5 -> 2.5.5, recommended: true migrated to preset: "recommended", and extended ignore sets for generated and tool-owned trees (!**/api/cursor-agent/gen, !!**/.codegraph).
• package.json: monorepo renamed pi-monorepo -> senpi-monorepo, packages/pty joined the workspace, chained-cd build scripts replaced by scripts/build-all.mjs with build:npm/build:bun/build:pnpm entry points, root check swapped tsgo --noEmit for tsc --noEmit and added check:claude-sdk-platform-lock plus script-based browser smoke, and fork-only verify:pms orchestration was added.
• pnpm-workspace.yaml: mirrors the root npm workspace's nested packages/session-backends/* glob so the pnpm parity build installs and links the sqlite session backend's workspace dependencies before scripts/build-all.mjs builds it.
• tsconfig.base.json: target/lib raised from ES2022 to ES2024.
• tsconfig.json: reformatted to the fork's biome multi-line layout; workspace path mappings are semantically unchanged.
• vitest.base.ts: added the workspace source alias mapping @earendil-works/pi-ai/utils/* to packages/ai/src/utils/* so shared test configs resolve utils from source.
• packages/agent/package.json: private CalVer 2026.8.16, tsgo -> tsc build/typecheck, fork dependency pins (@earendil-works/pi-ai/pi-telemetry ^2026.8.16, diff 9.0.0, ignore 7.0.6).
• packages/client/package.json: CalVer 2026.8.16, tsgo -> tsc, @earendil-works/pi-protocol pinned exactly to 2026.8.16.
• packages/client/src/unix.ts: typed the socket data callback chunk as Buffer.
• packages/protocol/package.json: CalVer 2026.8.16, tsgo -> tsc.
• packages/session-backends/sqlite-node/package.json: renamed @earendil-works/pi-session-backend-sqlite-node -> @earendil-works/pi-storage-sqlite-node, made private and independently versioned at 0.83.0, tsgo -> tsc, and keeps its runtime pi-agent-core / pi-ai dependencies on lockstep semver ranges so npm, Bun, and pnpm all link the live workspace packages.
• packages/session-backends/sqlite-node/src/sqlite/repo.ts: optional-chaining refactor of the message-target guard.
• packages/telemetry/package.json: private CalVer 2026.8.16.
• packages/telemetry/src/index.ts: type-layout reformat under the fork's biome/TypeScript settings; no contract change.
• packages/tui/package.json: private CalVer 2026.8.16, tsgo -> tsc, tests run under tsx with test/setup-multiplexer-env.mjs, added bench:frame-cost, Node engine >=24.0.0, pinned bumps (marked 18.0.7).
• .pi/extensions/prompt-url-widget.ts: deleted; relocated into global builtins (see the focused section below).
• .pi/extensions/tps.ts: deleted; relocated into global builtins (see the focused section below).

Why

• Senpi is a fork with its own identity, CalVer release trains, and an npm/bun/pnpm install matrix; root manifests, compiler settings, and lint configuration carry that policy, so they intentionally diverge from the upstream npm-only 0.x layout.
• Non-published support packages (agent, telemetry, tui, sqlite storage backend) are private and lockstep-versioned or independently pinned per AGENTS dependency policy, which shows up as manifest-level divergence with no deeper tracker of its own.
• The two deleted .pi/extensions/* files were repository-local development extensions that the fork promoted into shipped product behavior; the deletion itself is the audited divergence and is recorded here because .pi/ has no tracker of its own.

Why an extension could not handle it

• Every path in this section is repository, build, toolchain, or non-coding-agent package metadata that executes before any Senpi session, extension loader, or runtime exists. Extensions load inside a coding-agent session and cannot rename a monorepo, retarget compilers, reshape git hooks, reversion packages, or alter dependency policy.

Expected merge conflict zones

• HIGH: root package.json scripts/workspaces and packages/*/package.json version blocks on every upstream sync; upstream 0.x bumps must be reconciled into CalVer deliberately.
• MEDIUM: biome.json, tsconfig.base.json, tsconfig.json, and vitest.base.ts whenever upstream bumps toolchain majors or adds workspaces.
• MEDIUM: .pi/extensions/prompt-url-widget.ts and .pi/extensions/tps.ts — upstream still owns these files, so syncs will propose edits to deleted paths; resolve to the deletion and re-port any upstream improvement into the builtin copies.

Deleted repo-local .pi extensions, relocated into global builtins (2026-04-27)

What changed

• Deleted .pi/extensions/prompt-url-widget.ts and .pi/extensions/tps.ts, which the upstream pin still ships as repository-local dev extensions.
• Relocated their functionality into always-on global builtins at packages/coding-agent/src/core/extensions/builtin/prompt-url-widget.ts and packages/coding-agent/src/core/extensions/builtin/tps.ts, registered with the other fork builtins and covered by packages/coding-agent/src/core/extensions/builtin/changes.md.
• Subsequent fork releases hardened the TPS builtin (monotonic timing in 7f6097bf3, cache-hit notice in c7874fda3) with regression coverage in packages/coding-agent/test/suite/tps-extension.test.ts.
• Context: sibling .pi/extensions/import-repro.ts and .pi/extensions/redraws.ts moved the same way and are rename-tracked under the builtin tracker, so they do not appear in the canonical audit list above.

Why

• Repository-local .pi/extensions only load for sessions started inside this clone and require per-repo wiring. Senpi ships the URL prompt widget and tokens-per-second notice as product affordances for every user and session, versioned, registered, and tested together with the coding agent instead of living in an unaudied dot-directory.

Why an extension could not handle it

• Remaining a repo-local extension is exactly what this change removed: an extension cannot distribute itself to other clones or sessions. Promoting the behavior into the builtin set is the mechanism; there is no extension-side equivalent of "ship enabled-by-default for all users".

Expected merge conflict zones

• Upstream-side edits to the deleted .pi/extensions/prompt-url-widget.ts and .pi/extensions/tps.ts on every sync (resolve to deletion, re-port improvements).
• Builtin registration and widget internals under packages/coding-agent/src/core/extensions/builtin/ if upstream reworks extension loading or adds overlapping notices.

Pnpm parity for the nested SQLite session backend (2026-08-19)

What changed

• pnpm-workspace.yaml now includes packages/session-backends/*, matching the root npm workspace and the package set explicitly built by scripts/build-all.mjs.
• packages/session-backends/sqlite-node/package.json declares its shipped pi-agent-core / pi-ai imports as lockstep runtime dependencies instead of packed file: dev dependencies.
• scripts/sync-versions.js keeps the backend's own 0.83.0 version independent while synchronizing those lockstep dependency ranges during Senpi releases.

Why

• The release pre-commit gate verifies npm, Bun, and pnpm. Pnpm previously excluded the nested backend from its workspace and then, once included, packed its file: dependencies before their declarations were built. The ordered build therefore reached the backend with unresolved pi-agent-core / pi-ai types even though npm and Bun passed.

Why an extension could not handle it

• This is package-manager workspace topology and release-version synchronization. Runtime extensions load only after packages install and build, so they cannot repair missing workspace membership, dependency links, or manifest pins.

Expected merge conflict zones

• Upstream changes to the SQLite backend's dependency placement or independent-version policy.
• Future workspace additions under nested packages/*/* paths, which must remain aligned across root npm workspaces, pnpm-workspace.yaml, and scripts/build-all.mjs.

Upstream sync (upstream/main@71dca871) integration repairs (2026-09-12)

What changed

• package.json: the root manifest stays the fork's senpi-monorepo (Node >= 24, packages/pty workspace, build/clean/test routed through scripts/build-all.mjs and scripts/run-workspaces.mjs, the fork check chain with check:claude-sdk-platform-lock and tsc --noEmit, refresh-lock, preinstall bin stubs, Bun/pnpm trustedDependencies/onlyBuiltDependencies, and the held overrides such as protobufjs 7.6.5, @anthropic-ai/sdk 0.123.0, esbuild 0.28.2); upstream's check:runtime-deps/check:entry-graphs/check:package-install scripts exist but are not wired into check.
• packages/chord/package.json: differs from the pin only by version fields: the fork CalVer 2026.9.12 instead of 0.85.1, vitest 4.1.11 instead of 4.1.9 so the held pin stays single-instanced, and private: true because chord is bundled into the senpi tarball rather than published.
• packages/chord/src/types.ts: the same declarations as upstream; the only difference is biome 2.5.10 formatting of the nested conditional types (JsonRepresentation, InvalidJsonPart, InvalidRemoteMember), which the fork's --error-on-warnings check rewrites.
• packages/telemetry/package.json: version fields only: CalVer 2026.9.12, @types/node 26.2.0, vitest 4.1.11, private: true.
• tsconfig.json: keeps the fork path map (@code-yeongyu/senpi, @code-yeongyu/senpi/hooks, @code-yeongyu/senpi-server, @earendil-works/pi-pty, @earendil-works/pi-agent-core/session/testing) unioned with upstream's Chord root and subpath entries; the file is expanded one-entry-per-line by the fork formatter.
• vitest.base.ts: unions upstream's aiUtils alias with the fork aiAuthPool alias (@earendil-works/pi-ai/auth/* -> packages/ai/src/auth/*) beside the Chord aliases upstream added.

Why

• The fork ships under its own package names, CalVer lockstep, Node 24 floor, mixed npm/Bun/pnpm build orchestration and held dependency pins; the root manifest, TypeScript path map and vitest aliases are where those choices are declared, so the sync cannot take upstream's versions of them verbatim.

Why an extension could not handle it

• Workspace manifests, compiler path maps and test-runner aliases are build-time inputs read before any runtime code loads; no extension hook can rename packages, change the engine floor or register a module alias.

Expected merge conflict zones

• HIGH: root package.json scripts, devDependencies, overrides and engines whenever upstream bumps tooling or adds a check:* step.
• MEDIUM: tsconfig.json paths when upstream adds a workspace or subpath export; vitest.base.ts alias list for the same reason.
• LOW: packages/chord/package.json and packages/telemetry/package.json version lines on every upstream release; packages/chord/src/types.ts re-wraps whenever upstream edits those conditional types.