test: add unit coverage for cost-tracker, compact, knowledge, tool-utils

[stealthwhizz]

Summary

Several core modules shipped with no unit tests. This PR adds deterministic, offline node:test suites for the pure logic in four of them — +26 test cases, all placed under test/ so npm test actually runs them.

Coverage before / after

Existing tests covered sdk (tool helper, schema builder, programmatic hooks, query error paths) and telemetry. These modules had none:

Module	New cases	What's covered
cost-tracker.ts	6	Session + per-model token/cost aggregation, derived totalTokens fallback, defensive get() copy, reset()
compact.ts	10	estimateTokens heuristic, per-message-type accounting (incl. tool_use overhead), 75% compaction threshold, tool-result head/tail truncation, transcript rendering (drops deltas/system noise), compact-prompt building
knowledge.ts	6	Index loading: always_load preload + trim, available listing, missing-file skip, malformed index → empty; prompt formatting (inline docs + read hint)
tool-utils.ts	4	GCToolDefinition → AgentTool mapping, string vs {text, details} handler results, abort-signal forwarding

Notes

• All suites are pure/offline — no API keys, no network, temp dirs cleaned up.
• Full suite: 53 passing (27 existing + 26 new).

Follow-up worth flagging (not in this PR)

The npm test script globs only test/*.test.ts, so the existing suites under src/**/__tests__/ (memory, sandbox-memory, skill-learner, task-tracker) and src/__tests__/telemetry.test.ts are never run by CI. Wiring those into the test script (or moving them under test/) would be a good separate cleanup.

[shreyas-lyzr]

26 new test cases across four previously-uncovered modules. All look correct.

A few things I checked explicitly:

• estimateTokens arithmetic (ceil(5/4)=2), estimateMessageTokens summation (1+2+51+1=55), and the 75% compaction threshold (800/1000=0.8) all hold.
• totalTokens fallback (input+output when field absent) is tested correctly.
• The defensive-copy test for CostTracker.get() properly verifies internal state survives external mutation of the snapshot.
• knowledge tests use mkdtemp + rm recursive cleanup — no leaked temp dirs.
• before() blocks are correctly placed at module scope, not inside describe(), which is required for node:test.
• All suites are offline — no network, no API keys, no side effects.

The follow-up note in the PR description about test/tests not being wired into the npm test glob is a real gap worth tracking separately.

[Copilot]

Pull request overview

Adds deterministic node:test unit suites for several core “pure logic” modules so their behavior is exercised offline and without external dependencies.

Changes:

• Added new unit tests for CostTracker session + per-model aggregation behavior.
• Added new unit tests for compaction/token estimation utilities, truncation logic, and transcript/prompt building.
• Added new unit tests for knowledge index loading/formatting and tool-definition conversion utilities.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
test/tool-utils.test.ts	Tests toAgentTool mapping behavior, handler result normalization, and abort-signal forwarding.
test/knowledge.test.ts	Tests knowledge index loading (including malformed/missing cases) and prompt formatting output.
test/cost-tracker.test.ts	Tests CostTracker accumulation, per-model aggregation, defensive snapshotting, and reset behavior.
test/compact.test.ts	Tests token estimation, compaction thresholding, tool-result truncation, transcript rendering, and compact-prompt building.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[shreyas-lyzr]

Good addition. Reviewed the four new test suites against their source implementations and ran the full suite locally after building.

Verified:

• cost-tracker: all six cases map cleanly to the implementation. The defensive-copy test correctly validates that key deletion on a snapshot doesn't corrupt internal state (shallow copy of modelUsage object is sufficient for that guarantee).
• compact: token arithmetic checks out — ceil(5/4)=2, tool_use overhead of 50 added, needsCompaction threshold at 0.75 all match the source. truncateToolResults head/tail splitting and messagesToText filtering (drops delta/system) are correctly exercised.
• knowledge: temp-dir setup/teardown is clean, covers all four index loading paths (missing, always_load, missing file skip, malformed YAML), and the formatKnowledgeForPrompt assertion matches the exact tag/attribute ordering the source produces.
• tool-utils: covers the two handler result shapes (string vs object with details) and abort signal forwarding, which are the only real branches in toAgentTool.
• mcp.test.ts fixes: the four changed assertions update from the old string return shape to the current content[0].text shape — correct.

Full suite: 65 pass, 0 fail.

Security pass: no new dependencies added, no secrets or credentials in the diff, no injection surface (pure offline unit tests with temp dirs).

One thing worth a follow-up (noted in the PR description): the existing suites under src/**/tests/ and src/tests/telemetry.test.ts are not picked up by the test glob and never run in CI. Worth wiring those in separately.